WO2022175767A1 - Equipment fault detection system using artificial intelligence(icms) - Google Patents

Equipment fault detection system using artificial intelligence(icms) Download PDF

Info

Publication number
WO2022175767A1
WO2022175767A1 PCT/IB2022/050672 IB2022050672W WO2022175767A1 WO 2022175767 A1 WO2022175767 A1 WO 2022175767A1 IB 2022050672 W IB2022050672 W IB 2022050672W WO 2022175767 A1 WO2022175767 A1 WO 2022175767A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
block
expert
algorithm
ads
Prior art date
Application number
PCT/IB2022/050672
Other languages
French (fr)
Inventor
Sareh BAHMANPOUR
Original Assignee
Bahmanpour Sareh
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Bahmanpour Sareh filed Critical Bahmanpour Sareh
Publication of WO2022175767A1 publication Critical patent/WO2022175767A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition

Definitions

  • an intelligent artificial intelligence-based status monitoring system is presented in the form of web-based software that can be easily extended to all systems.
  • the only prerequisite for the use of the present invention is the availability of historical data of the system in question for monitoring the status.
  • this invention uses the historical data of the system and asking the user (expert) to determine the range of healthy performance of the system, this invention attempts to learn the normal behavior of the system in question and then notifies the user (expert) of any change in the behavior of the system and
  • the invention next identifies the cause of the change through two-way information interaction with the user (expert). Then, over time, the invention's knowledge of the monitored system expands and enables the user (expert) to predict future events.
  • the purpose of present invention is to introduce methodology to know the status of the grid interconnected/power transformer commissioned in transmission or distribution substation.
  • the algorithm developed and tested is based on application of Artificial Neural Network (ANN).
  • ICMS Intelligent Condition Monitoring System
  • the invention discloses an intelligent vehicle condition monitoring method based on deep learning.
  • An information collecting module obtains output information of each sensor of an automobile, then an information processing module performs preliminary processing on the information, and the road traffic environment is perceived through a convolutional neural network; the well processed information is transmitted to a cloud platform through a communication module; and after the information is further processed through a BP neutral network of a remote cloud platform, a man-machine interaction module feeds vehicle condition state information of the automobile back to a user through vision and sound signals.
  • the intelligent vehicle condition monitoring method the intelligent monitoring of the vehicle interior and exterior conditions is achieved, and timely feeds prompt or warning information such as automobile faults and danger warning back to the user so that the user can knowing real time whether the working state of the automobile is good, and measures are taken timely to avoid dangerous accidents. Therefore, the method not only effectively improves the efficiency of the use of the automobile, but also improves the driving safety of the automobile.
  • a system for condition monitoring and fault diagnosis includes a data collection function that acquires time histories of selected variables for one or more of the components, a pre-processing function that calculates specified characteristics of the time histories, an analysis function for evaluating the characteristics to produce one or more hypotheses of a condition of the one or more components, and a reasoning function for determining the condition of the one or more components from the one or more hypotheses.
  • the invention discloses a network online monitoring system for a transformer substation.
  • the network online monitoring system comprises a network monitoring device, a monitoring plug-in component for a switchboard of the transformer substation and a monitoring plug-in component for network messages of the transformer substation;
  • the network monitoring device is applicable to the intelligent transformer substation, and the monitoring plug-in component for the switchboard of the transformer substation and the monitoring plug-in component for the network messages of the transformer substation realize a network monitoring function;
  • the network monitoring device comprises a special interface card for multiple network communication interfaces, a main CPU (central processing unit) control panel and relevant control and communication interfaces;
  • the monitoring plug-in component for the switchboard of the transformer substation acquires a state of the network management switchboard and displays data warning and abnormity information in real time;
  • the monitoring plug-in component for the network messages of the transformer substation acquires the network transmission messages of the transformer substation and displays message abnormity information and variation of data states of the messages in real time.
  • the network online monitoring system has the advantages that the network online monitoring system can be used for monitoring and warning display for a communication state of a network in the intelligent transformer substation in real time, and important processed state information can be directly transmitted into the monitoring system for the transformer substation, so that the transformer substation can be monitored in a unified manner, and an effective solution is provided for problems of incapability of visually monitoring network information in an existing intelligent transformer substation in real time, difficulty in tracking procedure abnormality and difficulty in analyzing shortcomings of an intelligent device.
  • An intelligent monitoring system includes an intelligent monitoring device, a mobile terminal, and a garage door device.
  • the intelligent monitoring device includes a first communicating module and an image capturing module for collecting a multi-media information and transmitting it to the first communicating module.
  • the mobile terminal includes a displaying module, a second communicating module receiving and sending the multi-media information to the displaying module, and a transmitting module.
  • a user determines whether the garage door is in the open state or the closing state, and sends an instruction information via the transmitting module.
  • the garage door device includes a third communicating module, a controlling module, and a driving module, the third communicating module receiving and sending the instruction information to the controlling module, the controlling module sending a controlling signal to the driving module, the driving module opens the garage door or to closes the garage door.
  • the intelligent transformer monitoring system includes a GSM-GPRS, a measurement and instrumentation module, a control relay module, a Trivector energy measurement, and a GPS module.
  • the GSM-GPRS includes a microcontroller along with a GSM_GPRS modem to execute remote communication on GSM-GPRS.
  • the Measurement and Instrumentation module includes eleven temperature measurement channels with 8-digital temperature sensors and 3-RTD.
  • the control relay module includes 4 SPDT relays to execute output controls such as load trip and cooling motor etc.
  • the GPS module acquires the latitude, longitude, and time data from the satellite for location sharing.
  • the Power supply module is an AC/DC SMPS power supply to convert 240V/415V AC to 12VDC for the intelligent transformer monitoring system.
  • the present invention does not need to be labeled with defects to teach and diagnose anomalies and identify defects, and will only be able to detect anomalies by having normal (no-fault) performance data of the trained system and will enable him to identify defects by exchanging information with an expert.
  • clusters in normal data are identified and then a single-class SVM classifier is taught for each cluster in the test phase, the trained classifiers are used to determine whether the data is normal/abnormal.
  • clusters in normal data are first identified by an innovative algorithm, and then a single-class SVM classifier, a single-class LOF classifier, and a multi-class SVM classifier are taught for each cluster in the evaluation phase, by voting, the total opinions of these categories can be used to determine the anomalies of the samples in the data.
  • ADS Asset Performance Management
  • the system receives and analyzes temperature, vibration, acceleration, and rotational speed data in real-time for early signs of electrical or mechanical problems or declining efficiencies.
  • the results are provided to GE engineers and power plant operators through an augmented reality communication gateway. With the help of the results of these analyzes, it is possible to predict the occurrence of defects and identify the defective part.
  • Bently-Nevada has developed a product called AnomAlert to detect abnormalities in motors, generators, loads, and drivers that can send mechanical and electrical fault detection results to System 1 software, a software platform for delivering results to the user. There is. AnomAlert also can detect a short circuit event, detect a bearing malfunction, and disconnect the coil. This product uses a mathematical model for diagnosis.
  • This project presents a system for detecting equipment failures using artificial intelligence.
  • This design is an artificial intelligence-based condition monitoring system capable of diagnosing, detecting, and predicting the occurrence of faults in various systems, provided that the historical data of this system is available. If the information requirements are met in the ADS system design process, this system can be implemented on any type of industrial and non-industrial system.
  • This architecture includes two basic blocks, which are a collection of sub-blocks.
  • the system ADS detects anomalies using one-class classifiers. To detect anomalies, the data distribution of the monitored system is compared to the known data distribution, revealing the occurrence of unknown behavior. After detection, the system receives information and completes its knowledge by informing the expert about the details of the anomaly. Afterward, it repairs the faults.
  • NBM normal performance mode
  • KPI performance indices
  • system health indicators system health indicators
  • artificial intelligence-based systems are presented in the form of web-based software that can be easily extended to all systems. The only prerequisite for the use of the present invention is the availability of historical data of the system in question for monitoring the status.
  • this invention uses the historical data of the system and asking the user (expert) to determine the range of healthy performance of the system, this invention attempts to learn the normal behavior of the system in question and then notifies the user (expert) of any change in the behavior of the system and the cause of that change. The invention next identifies the cause of the change through two-way information interaction with the user (expert). Then, over time, the invention's knowledge of the monitored system expands and enables the user (expert) to predict future events.
  • the present invention is an artificial intelligence-based condition monitoring system that is capable of inspecting, detecting, and predicting faults in various systems only when historical data of that system is available. Therefore, with the availability of historical system data, this invention can be used for any type of system, whether industrial or non-industrial.
  • Such a property has not existed in any other invention, i.e., all previous inventions have been developed only for a specific system and cannot be extended to other systems in practice. This is true even for the industrial products under study. It can be argued that no single product has ever been presented with such high generalizability; a single system that can monitor the status of various types of systems. Concerning some products that provide status monitoring services on cloud platforms, it can be claimed that since the present invention is designed and implemented as a web application, it can be used on any cloud platform that provides data to it.
  • ADS is the abbreviation for "Anomaly Detection System"
  • the goal of this project is to intelligently diagnose, detect and predict faults in a system. Due to its generalizability in the ADS system design process, in case of the availability of information requirements, this system can be implemented on any type of industrial and non-industrial system.
  • This architecture consists of two main blocks. The main blocks consist of several sub-blocks, which are fully described below. To better understand the function of the blocks and sub-blocks, we should first look at the algorithm of the ADS system. For this purpose, we would first introduce the ADS system algorithm and then explain in detail how blocks and sub-blocks work.
  • the algorithm of the ADS system is based on anomaly detection using one-class classifications. Anomaly detection is the detection of an unusual or unknown condition in the monitored system.
  • the system ADS informs the expert about the details of the diagnosed anomaly and tries to get information from him to complete his knowledge.
  • the data distribution of the monitored system is compared with the data distribution known to it to detect the occurrence of unknown behavior. If the known behaviors of the system include the normal (error-free) modes of operation of the monitored system in all its modes of operation, then detecting the anomaly is the same as detecting the error or change in data distribution in the normal mode.
  • the one-class classification algorithm is a semi-regular classification algorithm used in anomaly detection. This type of classification with only normal data attempts to detect the generated data at the time of the occurrence of the error (abnormal data). These types of classifications attempt to detect abnormal data by learning the distribution of normal data. The occurrence of abnormal data is detected when it does not follow the known distribution of normal data. If we talk about the operation of the system ADS, we can generally claim that through the normal status data of the monitored system and the knowledge of an expert, over time, by detecting the occurrence of abnormal status, it tries to learn how to distribute abnormal data to reach the required maturity to detect the error in the monitored system.
  • the ADS system in the first days of installation, the ADS system is only able to detect the deviation of the monitored system from the normal state, with time and using the knowledge of the expert, it becomes more mature.
  • One ADS is matured, it would not only be able to detect anomalies, but also classify anomalies and label them accordingly. The more mature versions would also be able to predict future occurrences. It is expected that the ADS system would generate three categories of information as it matures. This information includes:
  • FIG. 1 shows a diagram of the input/output structure of the ADS system. As can be seen from the figure, three communication ports are dedicated to the system. These three ports include:
  • FIG. 1 shows the blocks and sub-blocks of the ADS system. Based on the diagram block structure of the ADS system, it can be classified into two general parts.
  • Evaluator block This block can be seen at the bottom of . This block assesses the status of the monitored system, using the knowledge embedded in it. This block consists of a sub-block set that is discussed in the following sections.
  • Improver block This block can be seen at the top of .
  • This block uses the knowledge of the expert along with the results of the analysis performed by the evaluator block to improve the knowledge embedded in the ADS system.
  • This block consists of a sub-block set that is discussed in the following sections.
  • This block evaluates and determines the status of the system from the perspective of being normal, abnormal, and fault occurrence.
  • the operating mechanism of this block is derived from the operating mechanism of the algorithm.
  • the data after preprocessing and normalization, use an executor to apply the preprocessed data to the appropriate classifiers and receive their opinion on the belonging of the input data to those classes and provide it to the decision-making block.
  • the decision-making block announces the final opinion, taking into account the opinions of various categories.
  • the process of retrieving data from the data system is usually affected by the three components of Nan, invalid data and ambient noise.
  • the input data must be purified from the presence of any error components.
  • Preprocessor sub-block purifies the data of Nan, Outlier, and noise.
  • the raw data input After being purified by the data preprocessing sub-block, the raw data input, must be applied to the machine learning algorithms. Since the domain of the data obtained from the process is not uniform, the output of the algorithm would be affected by the data with a bigger domain and the effect of the information contained in other data would be diminished. To avoid the diminishing of the effect of the information contained in the smaller domain data, the data should be unified in terms of domain, the process of data domain unification is called data normalization.
  • Training phase and evaluation phase in the training phase, the data attributed to each flag are normalized separately to train the classification of that flag.
  • the data obtained after passing the normalizer sub-block turns into zero average data in the interval [1, 1-]. This way, all data would have the same value in terms of machine learning algorithms.
  • the evaluation phase for a precise evaluation, the received data attributed to each flag should be normalized according to the data normalization factors of that flag in the training phase. Otherwise, according to the ADS algorithm, the results of the evaluation phase are incorrect. shows the modules that make up the Normalizer sub-block in the training phase.
  • this sub-block is activated. Subsequently, through the flag Identifier module, it is determined to which flag the received data is attributed, then the normalizer module, after normalizing the data, stores the data normalization parameters in the train data set’s normalization factors module. shows the modules that make up the normalizer sub-block in the evaluation phase. This sub-block is activated after activating the pre-process complete input.
  • the train data set’s normalization factors module provides the Scaler module with the parameters needed to normalize the data obtained during the corresponding classification training process.
  • the Scaler module receives the parameters needed to normalize the data and unify the input data. As it is known, each module is informed of the completion of the previous module process through its input and starts operating.
  • classifier manipulator sub-block and classifiers bank sub-block these two sub-blocks are examined in this section. shows the modules that make up these two sub-blocks.
  • the classifiers bank sub-block consists of two classes of modules: the classifiers repository and repositories manager. In the ADS system, there is a classifiers repository for each defined flag, in which the trained classifiers for that flag are stored.
  • the repositories manager module is responsible for managing repositories in the process of reading/writing information from/to them and responding to requests received from the classifiers manipulator sub-block.
  • the classifiers manipulator sub-block is responsible for retrieving classifiers and evaluating the input data through them.
  • the classifier manager module is responsible for retrieving the categories corresponding to the input data flag from the classifiers bank sub-block and introducing them to the suitable module.
  • the sub-block structure of the classifiers manipulator is designed based on a group classification algorithm. In this structure, various classification methods are used simultaneously to improve the decision result. After Scale Complete input is activated, the method x normal determiner module receives the normal classification classifier appropriate to the input data flag from the classifier manager and determines through the normal output whether in its view there is input data in the normal class.
  • the normal output of the method x normal determiner modules is applied to the decision-maker block and the final decision regarding the normalcy of the system would be made in this sub-block. Normal complete binary output is activated after all classifiers have commented as a sign of the completion of the comment process.
  • the implementation of the group classification mechanism in fault detection is that the method x fault determiner modules are activated in case the decision-maker sub-block detects that the system is not normal (zero normal input) and each of them from a various point of view, comment on the attribution of input data to various classes of faults and submit their comments to the decision-maker sub-block.
  • the function of the classifiers manipulator sub-block is such that if the data is not normal, the method x fault determiner modules are asked for comments, and if the data input is normal, no comment is received from them.
  • the normal binary signal is generated by the decision-maker sub-block, which is discussed in the following section.
  • the fault complete binary output is activated after all classifiers have commented as a sign of completion of the comment process.
  • This sub-block announces the final comment on the status of the system in terms of normal, abnormal, and defective labels.
  • the faults concluder module summarizes the comments of the fault label detection categories after the fault complete input is activated.
  • the normal concluder module summarizes the comments of the normal classification classifiers when the normal complete input is activated.
  • the interpreter module is responsible for interpreting the normal and fault signals and translating them to the system status in terms of normal, abnormal, or fault occurrence. Normal output is a binary signal, we talked about this signal previously, and it can be used in other sub-blocks as an activator input.
  • This block obtains information about the occurrence and increases the knowledge of the ADS system in occurrence detection by monitoring the performance of the evaluator block and using the knowledge of the expert. After the reveal of the anomaly occurrence, this block comes into action and informs the expert about its details, and obtains additional information from him. After announcing the opinion of the expert, this block takes appropriate measures such as training new categories or re-training existing categories. In the following, the sub-blocks of the improvement block are examined.
  • This sub-block as the input of the Improver block is responsible for detecting the occurrence of the anomaly by monitoring the output of the evaluator block. If the result of the evaluator block operation shows an anomaly, the anomaly occurrence detector sub-block becomes actives and generates a suitable signal through its output to launch the next sub-blocks of the improver block. shows the modules that make up the anomaly occurrence detector sub-block.
  • This sub-block consists of a single interpreter module whose task is to interpret the output of the evaluator block and detect anomalies. The output of this module is a binary signal called an anomaly.
  • the anomaly signal would be zero when the output of the evaluator block is normal or a fault occurrence and 1 when the output is an anomaly.
  • 2-2- Stimulus parameter identifier sub-block The stimulus parameter identifier sub-block provides the expert with a set of data that are the prime cause of the system going out of normal state to facilitate the process of detecting the causes of the anomaly. The expert can simply comment on the occurred anomaly by using the stimulus data of the occurrence of the anomaly and analyzing other factors. shows the modules that make up the stimulus parameter identifier sub-block. According to the figure, this sub-block consists of a module called a deviation detector.
  • the deviation detector module is informed of the time of the anomaly occurrence, and by analyzing the preprocessed data; it detects the anomaly occurrence stimulus signals and provides it to other sub-blocks as deviated data. 2-3 under the expert messenger block.
  • This sub-block generates the content of the message based on the deviated data, which is the stimulus parameter identifier sub-block output.
  • the message generated by this sub-block contains the information of the anomaly stimulus data. shows the modules of this sub-block.
  • This sub-block is a communication port that establishes communication between the ADS system and the expert.
  • This sub-block should display the messages generated by the expert Messenger sub-block to the expert and receive the result of the expert's analysis. shows the modules that generate this sub-block.
  • the interpreter sub-block receives expert analysis and translates it into new and modified signals.
  • expert analysis has two interpretations.
  • the first case is the result of the analysis of the expert analysis of the production of a new class in the system data.
  • the new output under the interpreter sub-block must be activated to continue the related processes.
  • the second case is the result of the expert analysis of the correction of a pre-existing class.
  • the modified output of the interpreter sub-block must be active so that related processes can be done.
  • New and modified signals are binary signals and only one of them is active at a time.
  • the data repository module temporarily stores the output data of the normalizer sub-block after the normalized complete input is enabled so that it provides the required data to the retrain classifier module when necessary.
  • the retrain classifier module retrieves the corresponding classifier through the classifier manager module.
  • the classifier manager module based on the data flag input, retrieves the desired classifier from the classifiers bank sub-block and provides it to the retrain classifier module. After retraining the desired classification by the retrain classifier module, the trained classification is transferred to the classifiers bank through the classifier manager module.
  • the New class classifier trainer sub-block is activated in case the new output of the Interpreter sub-block is enabled, and its job is to train the new classification based on new information. shows the modules of this sub-block.
  • the function of this sub-block is quite similar to the function of the already classifiers modifier sub-block and the only difference is in the train classifier module, which does not require the classification information available in the classifiers bank to train the new classification
  • the function of the ADS system is based on anomaly detection and incremental learning.
  • ADS acquires from him the knowledge of the cause of the abnormality and institutionalizes it within itself so that in similar upcoming occurrences it can be used by the expert as an auxiliary element.
  • the incremental learning desired by the ADS system refers to how knowledge is communicated and extracted from the expert. That is to say, the ADS system increases its knowledge on analyzing the status of the monitored system by exchanging information with an expert in the form of incremental learning.
  • Anomaly detection refers to finding the data with patterns that do not match the known and expected patterns of the system.
  • anomalies Data with unknown patterns in various fields are also referred to as anomalies, out-of-class, inconsistent observations, exceptions, deviations, surprises, unusual properties, or pollutants among which, anomaly and out-of-class expressions are much more common.
  • anomaly detection algorithm designed for the ADS system. This unit uses machine learning techniques in the data space of the N dimension to analyze the status of the system. shows the flowchart of the desired anomaly detection algorithm. Below, each of the blocks is examined in detail. Based on , the desired anomaly detection algorithm consists of two general parts, namely training and evaluation.
  • the training part of this algorithm is equivalent to the improvement block and the evaluation part is equivalent to the ADS system evaluator block.
  • ADS increases its knowledge by learning from an expert.
  • ADS evaluation phase with the use of the learned knowledge, it analyzes the status of the monitored system.
  • This phase is activated when there is a need to update or increase the knowledge of the ADS system.
  • the normalized samples must first be retrieved.
  • the normalization process of the samples is performed in the normalizer sub-block.
  • the clusters in the received samples for each flag should be identified, this is performed by the clustering algorithm developed by the ADS system design team, the design details of this algorithm are reviewed below.
  • a one-class classification is trained for each cluster.
  • the trained one-class classifications are stored in a repository called OCC’s Bank. This is done for later use in the evaluation phase. The details of the performance of each of the blocks discussed in this section are examined below.
  • the algorithm in the first step after determining the required parameters in the configure algorithm step by an expert, retrieves the desired data to extract the clusters in it.
  • the required parameters include the desired settings that the expert uses to determine how to implement the algorithm; we would discuss how to use them below.
  • the drop duplicates parameter is set to True value by an expert to reduce the computational volume and speed up the clustering algorithm, duplicate samples are removed and only one sample remains to represent the other samples. If the expert does not want to do this, he can set the drop duplicates parameter to False value. It is recommended that the drop duplicates parameter be set to True only if it does not change the result of the clustering algorithm.
  • the display raw data step is executed. Display raw data flowchart is shown in . As presented in the figure, in case the dimensions of the data to be displayed are more than three dimensions, Then, to display the data while preserving the information contained in it, its dimensions are reduced to two by t-SNE and MDS algorithms (depending on expert opinion). Otherwise, (or after reducing the data dimension) three-dimensional or two-dimensional display of data is done depending on the dimensions of the input data.
  • the t-SNE and MDS dimension reduction algorithms used in this stage are derived from the Manikold package of the Scikit-Learn machine learning library.
  • the t-SNE algorithm is a nonlinear machine learning algorithm for dimension reduction suitable for high-dimensional data for illustration. This algorithm maps the samples in the high-dimensional space to two or three-dimensional space so that similar samples locate next to each other and dissimilar samples space apart. Thus this algorithm can visualize high-dimensional data.
  • the MDS algorithm is also a nonlinear visualizing algorithm: First, the data distance matrix is calculated in the space of high dimensions, then by placing the samples in a two or three-dimensional space which is based on the information in the distance matrix, attempts to display the samples in the high dimensions’ space. Next, if the expert sets the estimated clusters parameter to True value, estimating the number of clusters comes into action, otherwise it would not be executed and the algorithm would continue.
  • Cluster number estimation algorithm is a network search algorithm based on the K-means clustering algorithm in which the criterion for selecting the best result is the Silhouette criterion, which is used as a descriptive measure of the accuracy of sample clustering.
  • the Silhouette criterion produces a number between [1, -1] for each sample in the data based on the clustering result. The number 1 indicates that the sample belongs to the correct cluster; the number -1 indicates that the sample belongs to the wrong cluster, and zero indicates that the sample is located between two clusters with equal distance.
  • the Silhouette criterion is calculated for every single sample, so to have a single quantity, the average Silhouette criterion is usually used to evaluate the accuracy of clustering results.
  • the closer the average Silhouette criterion for clustering the samples is to the number 1, the more accurate the clustering is, and vice versa. shows the flowchart of the cluster number estimation algorithm. As shown in the figure, the cluster number estimation algorithm starts working to get the minimum and the maximum number of possible clusters in the data. The expert must specify the range of number of clusters in the data for the algorithm as the minimum number of clusters, maximum number of clusters.
  • the estimation algorithm then performs clustering operations for all elements within the specified range using the K-means algorithm and reports the element whose average Silhouette criterion is higher than the other elements as the estimation of the number of clusters.
  • the clustering algorithm is shown in in the "get number of clusters from expert" needs to receive the precise number of clusters in the data.
  • the expert can determine the number of clusters in the data using the display tools and estimate the number of clusters installed during the algorithm in the "get the number of clusters from expert” stage.
  • the clustering algorithm decides how to continue the algorithm according to the number of definite clusters entered by the expert. In case the number of specified clusters is 1, all the samples in the data are identified as one cluster, otherwise, the fine-tune clustering result algorithm would be performed.
  • the flowchart of the fine-tune clustering result algorithm is shown in .
  • This algorithm is an innovative clustering algorithm using the DBSCAN algorithm.
  • the DBSCAN algorithm is a sample density-based clustering algorithm that uses the epsilon and minimum sample parameters to determine the minimum density required to detect clusters. Density refers to the accumulation of samples in one part of space.
  • the DBSCAN algorithm requires the minimum acceptable density to identify a cluster. This algorithm examines the samples in the data and considers the set of samples that have the desired minimum density as a cluster.
  • the main property of the DBSCAN clustering algorithm which is also used as the core of the ADS clustering algorithm, is the criterion for defining the cluster included in its design.
  • the algorithm uses the DBSCAN algorithm to explore data space with various densities to extract the desired number of clusters. Obviously, in such a process it is possible to extract the desired number of clusters for various parameters. It is also possible to extract the desired number of clusters from the data space for a set of epsilon and minimum sample parameters.
  • the set of parameters that bring about the extraction of the desired clusters is stored as the final selected candidate to be decided upon after the search process.
  • the preferred condition over the number of clusters is the assessment of the degree of contamination of the clustering results to out-of-category data. The reason for considering this condition is that the data in the preprocessing process is cleared of out-of-class data, so the result of clusters that have reasonable out-of-class data is valid. The above was a description of the steps that are performed before the "select one of the nominees as a suggestion" step in the fine-tune clustering result algorithm.
  • the algorithm In the "select one of nominees as suggestion” step, the algorithm must select an offer from the nominees obtained in the previous steps and provide it to the expert.
  • the criterion for selecting a suggestion among nominees in the "select one of nominees as a suggestion” step is the maximum average of the Silhouette criterion. In fact, among the obtained nominees, the nominee whose average Silhouette criterion is the highest would be presented to the expert as a suggestion. Sometimes the expert may not agree with the result proposed by the algorithm and may want to choose another nominee. In such a case, the expert must set the show nominee parameter to True to examine the other nominees to display the algorithm of the other nominees to select the final result.
  • the expert can determine the result of the clustering based on his wishes by choosing his acceptable nominee. If the show nominee parameter is set to False, the clustering result is the same as the algorithm proposed.
  • evaluate clustering result the accuracy of the clustering result performed in the previous steps is evaluated as follows. The idea for this evaluation is derived from the definition of a cluster.
  • each cluster is a set of samples that are similar to each other in a specific pattern and at the same time various from other clusters.
  • the criterion of similarity is the Euclidean distance. So if the problem changes to a classification problem after the clustering process, the clustering would be correct and a linear or nonlinear multi-class classifier can also perform the classification operation with high accuracy. In this step, the problem is changed from unsupervised to supervised state, and the accuracy of linear or nonlinear multi-class classification is evaluated with labels obtained from the clustering algorithm. With these interpretations, a decrease in the accuracy of linear and nonlinear classifiers is considered as inappropriate clustering results and vice versa. In this project, SVM multi-class classifier is used as a nonlinear classification and the logistic regression classifier is used as a linear multi-class classification.
  • the criterion of the accuracy of the result of classification algorithms here is the criterion F1, which is the weighted average of the criteria of accuracy and retrieval.
  • the accuracy criterion expresses the percentage of correct samples available as a result of classification in all classified samples.
  • the retrieve criterion expresses the percentage of correct samples in the output of the classifier in all samples belonging to that class.
  • the ADS performance mechanism after extracting the clusters in the desired data space, a single class classification should be trained for each cluster, so that at the end of the training process, ADS can evaluate the status of the monitored system with the help of the obtained classifiers.
  • the flowchart of the one-class classification training process is shown in .
  • the first step is to retrieve the settings considered by the execution of the algorithm.
  • the settings of the algorithm it proceeds to remove the clusters that contain a small number of samples; the criterion of smallness is announced by the expert as a percentage of the total samples.
  • an expert can use the cluster_sample_thd parameter to express the threshold number of acceptable cluster samples as a percentage of the sample set present in the result of clustering.
  • the expert can remove similar data to speed up the training process. After the data is ready, it is time to train the classification of a one-class per cluster. To count the clusters, as shown in the flowchart, X is used as a counter.
  • one-class classification training is performed.
  • the one-class classification used here is a combination of two one-class classifiers and one multi-class classifier.
  • the desired classification training process is a network search algorithm with a cross-validation algorithm with F1 selection criteria. This process attempts to produce a reliable classifier.
  • Cross-validation is a model evaluation method that determines how the resulting model is generalizable and independent of the training data. This method is specifically used in forecasting to determine how useful the model could be in practice.
  • the category is specifically entitled to be stored in OCC’s Bank for future use.
  • the ADS system is ready to assess the status of the monitored system.
  • the evaluation phase as in the training phase, it is necessary to retrieve normalized data.
  • the retrieved data should retrieve the classifiers belonging to the mentioned flag from OCC’s Bank to comment on the normality or abnormality of the samples in the evaluated data by a set of OCC’s steps.
  • comments are made regarding the normality or abnormality of the samples of that flag by aggregating the opinions of the one-class classifiers in each flag.
  • FIG. 19 and 20 show an overview of the ADS system interface environment in the training and evaluation phases, respectively. As can be seen from Figures 19 and 20, this interface includes two tabs, enhancement, and evaluation.
  • the enhancement section is appropriate to the improver block shown in and is dedicated to model training and improvement processes.
  • the evaluation section is appropriate to the evaluator block shown in and is dedicated to evaluation processes.
  • the interface shown in Figures 19 and 20 is designed to use the ADS system in power plants, therefore, the examples and parameters discussed below are specified for power plant applications.
  • the enhancement section consists of several drop-down menus and keys, which we would cover below. The user can use the drop-down menu (1) to select the desired power plant, the drop-down menu (2) to select the desired unit, and the drop-down menu (3) to select the domain.
  • the domain menu is available to the user to analyze part of the system.
  • the user can determine which system signals are intended for ADS training. With this option, the user can consider all or only a certain part of the signals such as vibration signals of a gas turbine for ADS training.
  • the keys available to the user are the Config/Reconfig key, the Train / Retrain key, the Mind key, and the Kill key, respectively, which we would discuss below.
  • the Config /Reconfig key is used to configure the user desired domain for ADS training. Clicking on it shows the menu shown in .
  • the user can set the specifications of his desired domain through the available options.
  • the user must determine the various operating modes of the system.
  • the user can determine the list of sensors that are intended to be located in the flagging space.
  • the user can determine the list of sensors that are intended to be located in the decision space.
  • the user can use option (4) to determine the normal (no-fault) operating interval of the system by setting a start date and an end date.
  • Option (5) enables the user to allow or disallow the pre-process operation of the data received from the system.
  • Option (6) allows the user to decide whether to remove similar samples from the data to speed up the training process.
  • Option (7) allows the user to decide on the estimated number of clusters by ADS.
  • the user can determine the size of the network search space of clustering and classifier training processes, respectively.
  • the user can specify the data interval required for each execution of the analysis process in the evaluation phase. For example, selecting 1 is considered as the execution of the evaluation phase for the data one day.
  • the ADS system uses a sequence filter to prevent the generation of faulty anomaly alarms and to ma sure of the occurrence of that anomalies.
  • the user can specify the number of consecutive samples required to activate the anomaly alarms in the evaluation phase.
  • the user can specify a name for the configured domain and use the Save key to save it.
  • the user can command the execution of ADS training operations, and then the modules shown in the panel below is sequentially performed, and include the ADS configuration according to the intended settings, data retrieval from the database, data preprocessing, data standardization, flag identification, anomaly detection model training, database update using the obtained information. After successful completion of the above steps, ADS is trained and ready to perform the analysis process.
  • the Mind key allows the user to view the data space in three dimensions for better analysis and viewing the training results. An example of the mind output is shown in .
  • the Kill key is also installed for times when the user for any reason does not want to maintain the trained models so that he can eliminate the model and its related information. shows a view of the evaluation section of the ADS system interface. The user can use the models trained in the enhancement section to analyze the status of the monitored system.
  • the user can use option (1) to select the power plant, using option (2) to select the unit, and using option (3) to select the domain.
  • the keys available in this section are the Analyze key, the Feedback key, the Forget key, the Histogram key, and the monitored system key for the specified intervals.
  • the results are displayed in sections (4) to (8).
  • the user can use the Feedback key to give feedback to the ADS of the summary results obtained after viewing the results of the ADS analysis. shows a view of this panel. For detected sequences, anomalous samples request feedback from the user.
  • the user can use the Save key to save them in the database of the ADS system to be used in the next execution of the training process.
  • the Forget key is included to enable the possibility of clearing ADS memory from expert feedback and resume analysis.
  • the user requests ADS to clear the database of stored feedback by clicking Forget.
  • the Histogram key the user can view the histogram of abnormal samples along with the histogram of normal samples and other labels for the stimulus parameters selected in panel (7). shows an example of the output of the Histogram key.
  • the Mind key has the same function as the Mind key described in the enhancement section.
  • Section (4) is a panel called System Status, which is responsible for displaying the status of the system in terms of the occurrence of faults and anomalies.
  • Section (5) is a panel called Anomaly Trend, which displays the volume of anomalous occurrence over time in terms of the percentage of analyzed samples and thus provides the user with information about the health status of the system over time.
  • Section (6) is a panel called Anomaly Sequence, which displays the anomalous event sequences detected in the analysis process, and the user can use it to view information such as the number of anomalous events and the volume of anomalous samples in each sequence and anomalous event control period. By clicking on each sequence, the user can view additional information such as the stimulus factors for anomalies in that sequence in the panel (7).
  • the time signal of the selected stimulus in panel (8) along with the location of the occurrence of that sequence is displayed for further investigation. In this way, the user with the above-mentioned information and his previous experiences can provide appropriate feedback.
  • the ADS system can be used to monitor the status of any type of system in any industry.
  • the benefits of using ADS in different industries are some of the benefits of using ADS in different industries.
  • This invention can be used to monitor the situation to detect, detect, and predict the occurrence of defects in various industrial and non-industrial systems.
  • this invention has very high generalizability that can be used to monitor the status of any system, whether industrial or non-industrial, only if data is available.
  • the ICMS system interface is designed as a Web application, it can be used both on PC and in cloud computing systems. It will also be possible to access it through a variety of local networks and the Internet. In this way, with the help of the ICMS system, it will be possible to easily monitor the status of different systems, provided that their data is available remotely.
  • Faults can occur in systems and one of the common challenges in the industry is the occurrence of faults. Fault occurrence is normal and the effect of this fault on the system must be eliminated so that the system can return to normal operation. After a fault occurrence, issues like defect occurrence detection and occurrence effects elimination are faced.
  • the ADS system detects anomalies using one-class classifiers (a semi-regulatory classification that detects anomalies).
  • one-class classifiers a semi-regulatory classification that detects anomalies.
  • the data distribution of the monitored system is compared with the known data distribution, which reveals the occurrence of unknown behavior.
  • the system obtains information and completes its knowledge by informing the expert about the details of the anomaly. Then, it eliminates the faults and in the shortest time, causes the desired efficiency and correct and safe operation of the systems, and prevents the occurrence of failures in them.
  • the ICMS system can be used to monitor the status of any type of system in any industry.
  • the benefits of using ICMS in different industries are some of the benefits of using ICMS in different industries.
  • This design can be used to troubleshoot mechanical devices, medicine, electronic equipment, automotive, etc.
  • This invention can be used to monitor the situation to detect, detect, and predict the occurrence of defects in various industrial and non-industrial systems.

Abstract

In the present invention, an intelligent artificial intelligence-based status monitoring system is presented in the form of web-based software that can be easily extended to all systems. The only prerequisite for the use of the present invention is the availability of historical data of the system in question for monitoring the status. Using the historical data of the system and asking the user (expert) to determine the range of healthy performance of the system, this invention attempts to learn the normal behavior of the system in question and then notifies the user (expert) of any change in the behavior of the system and the cause of that change. The invention next identifies the cause of the change through two-way information interaction with the user (expert). Then, over time, the invention's knowledge of the monitored system expands and enables the user (expert) to predict future events.

Description

Equipment fault detection system using artificial intelligence(ICMS)
In the present invention, an intelligent artificial intelligence-based status monitoring system is presented in the form of web-based software that can be easily extended to all systems. The only prerequisite for the use of the present invention is the availability of historical data of the system in question for monitoring the status.
Using the historical data of the system and asking the user (expert) to determine the range of healthy performance of the system, this invention attempts to learn the normal behavior of the system in question and then notifies the user (expert) of any change in the behavior of the system and
The invention next identifies the cause of the change through two-way information interaction with the user (expert). Then, over time, the invention's knowledge of the monitored system expands and enables the user (expert) to predict future events.
Machine learning (G06N 20/00
The issue of monitoring the health of equipment and the use of artificial intelligence in improving performance and its results is not an emerging issue in the world and much work has been done in this area some are patents and some are articles and dissertations, each with its unique characteristics but the pioneering invention has a better feature than all of them, which we will examine in the following. For this purpose, the history of similar works is discussed first.
INTELLIGENT CONDITION MONITORING SYSTEM FOR GRID INTERCONNECTED / POWER TRANSFORMERS
IN2660/MUM/2013
The purpose of present invention is to introduce methodology to know the status of the grid interconnected/power transformer commissioned in transmission or distribution substation. The algorithm developed and tested is based on application of Artificial Neural Network (ANN).
The existing measurable parameters like oil temperature, winding temperature, low oil level, Moisture Content and Hydrogen Content by hydran meters, and efficiency deviation data patters applied as input parameters for the design of Intelligent Condition Monitoring System (ICMS). This ICMS avoids the unnecessary outages, prevent breakdowns thereby increases the availability of the equipment and cause considerable savings in the economy of utility. The main advantage of the model is that it can be applied for online monitoring to grid interconnected /power transformer of any ratings.
Intelligent vehicle condition monitoring method based on deep learning
CN107878450
The invention discloses an intelligent vehicle condition monitoring method based on deep learning. An information collecting module obtains output information of each sensor of an automobile, then an information processing module performs preliminary processing on the information, and the road traffic environment is perceived through a convolutional neural network; the well processed information is transmitted to a cloud platform through a communication module; and after the information is further processed through a BP neutral network of a remote cloud platform, a man-machine interaction module feeds vehicle condition state information of the automobile back to a user through vision and sound signals.
According to the intelligent vehicle condition monitoring method, the intelligent monitoring of the vehicle interior and exterior conditions is achieved, and timely feeds prompt or warning information such as automobile faults and danger warning back to the user so that the user can knowing real time whether the working state of the automobile is good, and measures are taken timely to avoid dangerous accidents. Therefore, the method not only effectively improves the efficiency of the use of the automobile, but also improves the driving safety of the automobile.
Intelligent condition monitoring and fault diagnostic system for preventative maintenance
US20190179298
A system for condition monitoring and fault diagnosis includes a data collection function that acquires time histories of selected variables for one or more of the components, a pre-processing function that calculates specified characteristics of the time histories, an analysis function for evaluating the characteristics to produce one or more hypotheses of a condition of the one or more components, and a reasoning function for determining the condition of the one or more components from the one or more hypotheses.
Network online monitoring system for transformer substation
CN103296755B
The invention discloses a network online monitoring system for a transformer substation. The network online monitoring system comprises a network monitoring device, a monitoring plug-in component for a switchboard of the transformer substation and a monitoring plug-in component for network messages of the transformer substation; the network monitoring device is applicable to the intelligent transformer substation, and the monitoring plug-in component for the switchboard of the transformer substation and the monitoring plug-in component for the network messages of the transformer substation realize a network monitoring function; the network monitoring device comprises a special interface card for multiple network communication interfaces, a main CPU (central processing unit) control panel and relevant control and communication interfaces; the monitoring plug-in component for the switchboard of the transformer substation acquires a state of the network management switchboard and displays data warning and abnormity information in real time; the monitoring plug-in component for the network messages of the transformer substation acquires the network transmission messages of the transformer substation and displays message abnormity information and variation of data states of the messages in real time.
The network online monitoring system has the advantages that the network online monitoring system can be used for monitoring and warning display for a communication state of a network in the intelligent transformer substation in real time, and important processed state information can be directly transmitted into the monitoring system for the transformer substation, so that the transformer substation can be monitored in a unified manner, and an effective solution is provided for problems of incapability of visually monitoring network information in an existing intelligent transformer substation in real time, difficulty in tracking procedure abnormality and difficulty in analyzing shortcomings of an intelligent device.
INTELLIGENT MONITORING SYSTEM
US20180192008
An intelligent monitoring system includes an intelligent monitoring device, a mobile terminal, and a garage door device. The intelligent monitoring device includes a first communicating module and an image capturing module for collecting a multi-media information and transmitting it to the first communicating module. The mobile terminal includes a displaying module, a second communicating module receiving and sending the multi-media information to the displaying module, and a transmitting module. A user determines whether the garage door is in the open state or the closing state, and sends an instruction information via the transmitting module. The garage door device includes a third communicating module, a controlling module, and a driving module, the third communicating module receiving and sending the instruction information to the controlling module, the controlling module sending a controlling signal to the driving module, the driving module opens the garage door or to closes the garage door.
INTELLIGENT TRANSFORMER MONITORING SYSTEM
WO2020075181
An intelligent transformer monitoring system to detect and monitor random failures in distribution transformers due to improper usage and poor maintenance is provided. The intelligent transformer monitoring system includes a GSM-GPRS, a measurement and instrumentation module, a control relay module, a Trivector energy measurement, and a GPS module. The GSM-GPRS includes a microcontroller along with a GSM_GPRS modem to execute remote communication on GSM-GPRS. The Measurement and Instrumentation module includes eleven temperature measurement channels with 8-digital temperature sensors and 3-RTD.
The control relay module includes 4 SPDT relays to execute output controls such as load trip and cooling motor etc. The GPS module acquires the latitude, longitude, and time data from the satellite for location sharing. The Power supply module is an AC/DC SMPS power supply to convert 240V/415V AC to 12VDC for the intelligent transformer monitoring system.
As it turns out, all previous similar inventions were designed specifically for a particular system and did not provide the possibility of generalizability, while one of the features of the present invention is its ability to be generalizable to any system in exchange for the availability of historical data of that system.
Among the articles and dissertations, a lot of research has been done on smart status monitoring. The following are two cases that are very similar to the invention in question. In an article entitled A method of anomaly detection and fault diagnosis with online adaptive learning under small training samples, published in Pattern Recognition in Elsevier in 2017 by Mr. Lee et al. An anomaly detection algorithm is presented along with the fault diagnosis. This algorithm needs labeled data from different defects to detect the defect, which will limit its use due to the limited labeled data in different industries, also, if the labeled data is not available, then this algorithm will only be able to detect the anomaly.
The present invention does not need to be labeled with defects to teach and diagnose anomalies and identify defects, and will only be able to detect anomalies by having normal (no-fault) performance data of the trained system and will enable him to identify defects by exchanging information with an expert.
In an article entitled Toward a more practical unsupervised anomaly detection system, published in the 2013 journal Information Sciences by Elsevier by Mr. Song et al. An anomaly detection algorithm is proposed which is very similar to the anomaly detection algorithm developed in this invention except that it is developed in the algorithm. in this invention, an ensemble of several different classifications has been created to increase the reliability of anomaly alarms using the Majority Voting algorithm. So, as mentioned, this article is similar only in terms of anomaly detection and does not have many of the capabilities presented in this invention, such as fault diagnosis, fault prediction, and incremental learning.
As it is known, in this article, first the clusters in normal data are identified and then a single-class SVM classifier is taught for each cluster in the test phase, the trained classifiers are used to determine whether the data is normal/abnormal. In this invention, however, clusters in normal data are first identified by an innovative algorithm, and then a single-class SVM classifier, a single-class LOF classifier, and a multi-class SVM classifier are taught for each cluster in the evaluation phase, by voting, the total opinions of these categories can be used to determine the anomalies of the samples in the data.
In the following, we will examine industrial products in the field of work similar to the present invention. In 2016, GE introduced its first smartwatch monitoring product, ADS, designed specifically for hydropower plants at the Hydro Vision Trade Fair. ADS is part of GE's Asset Performance Management (APM) solution, which utilizes machine learning techniques to increase the efficiency of the monitoring and maintenance process in power plants. The system receives and analyzes temperature, vibration, acceleration, and rotational speed data in real-time for early signs of electrical or mechanical problems or declining efficiencies. The results are provided to GE engineers and power plant operators through an augmented reality communication gateway. With the help of the results of these analyzes, it is possible to predict the occurrence of defects and identify the defective part.
Bently-Nevada has developed a product called AnomAlert to detect abnormalities in motors, generators, loads, and drivers that can send mechanical and electrical fault detection results to System 1 software, a software platform for delivering results to the user. There is. AnomAlert also can detect a short circuit event, detect a bearing malfunction, and disconnect the coil. This product uses a mathematical model for diagnosis.
Many companies now offer their status monitoring products in the form of cloud-based platforms and remotely. In addition to the availability of status monitoring services everywhere, this measure also enables the aggregation of data and the sharing of users' experiences. In such platforms, various status monitoring services such as system efficiency monitoring, fault detection, detection, and prediction, as well as increasing productivity can be provided according to customer needs.
This project presents a system for detecting equipment failures using artificial intelligence. With the increasing complexity of industrial systems, their repair and maintenance using intelligent monitoring systems are inevitable. This design is an artificial intelligence-based condition monitoring system capable of diagnosing, detecting, and predicting the occurrence of faults in various systems, provided that the historical data of this system is available. If the information requirements are met in the ADS system design process, this system can be implemented on any type of industrial and non-industrial system. This architecture includes two basic blocks, which are a collection of sub-blocks. The system ADS detects anomalies using one-class classifiers. To detect anomalies, the data distribution of the monitored system is compared to the known data distribution, revealing the occurrence of unknown behavior. After detection, the system receives information and completes its knowledge by informing the expert about the details of the anomaly. Afterward, it repairs the faults.
On account of the increasing complexity in industrial systems and the importance of their repair and maintenance, the use of status monitoring systems to increase productivity and reduce costs has become inevitable in many industries. Until now, many methods have been used together with training methods and an observer to monitor the status of various systems, such as models describing normal performance mode (NBM), performance indices (KPI), system health indicators, and artificial intelligence-based systems. They all work only for the intended system and are not readily applicable to other complex industrial systems. In the present invention, an intelligent artificial intelligence-based status monitoring system is presented in the form of web-based software that can be easily extended to all systems. The only prerequisite for the use of the present invention is the availability of historical data of the system in question for monitoring the status.
Using the historical data of the system and asking the user (expert) to determine the range of healthy performance of the system, this invention attempts to learn the normal behavior of the system in question and then notifies the user (expert) of any change in the behavior of the system and the cause of that change. The invention next identifies the cause of the change through two-way information interaction with the user (expert). Then, over time, the invention's knowledge of the monitored system expands and enables the user (expert) to predict future events.
As mentioned above, the present invention is an artificial intelligence-based condition monitoring system that is capable of inspecting, detecting, and predicting faults in various systems only when historical data of that system is available. Therefore, with the availability of historical system data, this invention can be used for any type of system, whether industrial or non-industrial. Such a property has not existed in any other invention, i.e., all previous inventions have been developed only for a specific system and cannot be extended to other systems in practice. This is true even for the industrial products under study. It can be argued that no single product has ever been presented with such high generalizability; a single system that can monitor the status of various types of systems. Concerning some products that provide status monitoring services on cloud platforms, it can be claimed that since the present invention is designed and implemented as a web application, it can be used on any cloud platform that provides data to it.
Solution of problem
The ADS system, like any other system, has an architecture designed to achieve its goals. ADS is the abbreviation for "Anomaly Detection System "
The goal of this project is to intelligently diagnose, detect and predict faults in a system. Due to its generalizability in the ADS system design process, in case of the availability of information requirements, this system can be implemented on any type of industrial and non-industrial system. This architecture consists of two main blocks. The main blocks consist of several sub-blocks, which are fully described below. To better understand the function of the blocks and sub-blocks, we should first look at the algorithm of the ADS system. For this purpose, we would first introduce the ADS system algorithm and then explain in detail how blocks and sub-blocks work. The algorithm of the ADS system is based on anomaly detection using one-class classifications. Anomaly detection is the detection of an unusual or unknown condition in the monitored system. After diagnosing the anomaly, the system ADS informs the expert about the details of the diagnosed anomaly and tries to get information from him to complete his knowledge. In the ADS system, to detect anomalies, the data distribution of the monitored system is compared with the data distribution known to it to detect the occurrence of unknown behavior. If the known behaviors of the system include the normal (error-free) modes of operation of the monitored system in all its modes of operation, then detecting the anomaly is the same as detecting the error or change in data distribution in the normal mode.
The one-class classification algorithm is a semi-regular classification algorithm used in anomaly detection. This type of classification with only normal data attempts to detect the generated data at the time of the occurrence of the error (abnormal data). These types of classifications attempt to detect abnormal data by learning the distribution of normal data. The occurrence of abnormal data is detected when it does not follow the known distribution of normal data. If we talk about the operation of the system ADS, we can generally claim that through the normal status data of the monitored system and the knowledge of an expert, over time, by detecting the occurrence of abnormal status, it tries to learn how to distribute abnormal data to reach the required maturity to detect the error in the monitored system.
In other words, in the first days of installation, the ADS system is only able to detect the deviation of the monitored system from the normal state, with time and using the knowledge of the expert, it becomes more mature. One ADS is matured, it would not only be able to detect anomalies, but also classify anomalies and label them accordingly. The more mature versions would also be able to predict future occurrences. It is expected that the ADS system would generate three categories of information as it matures. This information includes:
  1. Preparation of unknown status labeling (normal, abnormal, and defective).
  2. Identification of the fault-sensitive variables (fault-sensitive variables are the measurable variables of the system that have the greatest impact when a fault occurs)
  3. Prediction of the future status of the system in terms of normal, abnormal, and defective.
shows a diagram of the input/output structure of the ADS system. As can be seen from the figure, three communication ports are dedicated to the system. These three ports include:
  1. Input port: Through this raw data port, the monitored system enters the ADS system.
  2. Output port: The results of intelligent analysis on the input data are provided to the user through this port as the status of the monitored system.
  3. Expert port: This port is designed to acquire expert knowledge when necessary. When the ADS system encounters an occurrence about which it has no information, it engages in exchanging information with the expert through the above port to increase its knowledge about the occurrence.
ADS system architecture
shows the blocks and sub-blocks of the ADS system. Based on the diagram block structure of the ADS system, it can be classified into two general parts.
Evaluator block: This block can be seen at the bottom of . This block assesses the status of the monitored system, using the knowledge embedded in it. This block consists of a sub-block set that is discussed in the following sections.
Improver block: This block can be seen at the top of . This block uses the knowledge of the expert along with the results of the analysis performed by the evaluator block to improve the knowledge embedded in the ADS system. This block consists of a sub-block set that is discussed in the following sections.
1- Evaluator block
This block evaluates and determines the status of the system from the perspective of being normal, abnormal, and fault occurrence. The operating mechanism of this block is derived from the operating mechanism of the algorithm. First, the data, after preprocessing and normalization, use an executor to apply the preprocessed data to the appropriate classifiers and receive their opinion on the belonging of the input data to those classes and provide it to the decision-making block. The decision-making block announces the final opinion, taking into account the opinions of various categories. We examine the various sub-blocks of this block in the following sections
1-1- Data Preprocessor sub-block
The process of retrieving data from the data system is usually affected by the three components of Nan, invalid data and ambient noise. To increase the output accuracy of the algorithm, the input data must be purified from the presence of any error components. Preprocessor sub-block purifies the data of Nan, Outlier, and noise.
1-2- Normalizer sub-block
After being purified by the data preprocessing sub-block, the raw data input, must be applied to the machine learning algorithms. Since the domain of the data obtained from the process is not uniform, the output of the algorithm would be affected by the data with a bigger domain and the effect of the information contained in other data would be diminished. To avoid the diminishing of the effect of the information contained in the smaller domain data, the data should be unified in terms of domain, the process of data domain unification is called data normalization.
According to the ADS algorithm, data unification is performed in two phases. Training phase and evaluation phase, in the training phase, the data attributed to each flag are normalized separately to train the classification of that flag. In this phase, the data obtained after passing the normalizer sub-block turns into zero average data in the interval [1, 1-]. This way, all data would have the same value in terms of machine learning algorithms. In the evaluation phase, for a precise evaluation, the received data attributed to each flag should be normalized according to the data normalization factors of that flag in the training phase. Otherwise, according to the ADS algorithm, the results of the evaluation phase are incorrect. shows the modules that make up the Normalizer sub-block in the training phase.
After the pre-process complete input is activated, this sub-block is activated. Subsequently, through the flag Identifier module, it is determined to which flag the received data is attributed, then the normalizer module, after normalizing the data, stores the data normalization parameters in the train data set’s normalization factors module. shows the modules that make up the normalizer sub-block in the evaluation phase. This sub-block is activated after activating the pre-process complete input.
Then, through the flag Identifier module, it is determined which flag the received data belongs to, then the train data set’s normalization factors module provides the Scaler module with the parameters needed to normalize the data obtained during the corresponding classification training process. The Scaler module receives the parameters needed to normalize the data and unify the input data. As it is known, each module is informed of the completion of the previous module process through its input and starts operating.
1-3- Classifiers manipulator sub-block and classifiers bank sub-block
To better express the operation of classifier manipulator sub-block and classifiers bank sub-block, these two sub-blocks are examined in this section. shows the modules that make up these two sub-blocks. The classifiers bank sub-block consists of two classes of modules: the classifiers repository and repositories manager. In the ADS system, there is a classifiers repository for each defined flag, in which the trained classifiers for that flag are stored.
The repositories manager module is responsible for managing repositories in the process of reading/writing information from/to them and responding to requests received from the classifiers manipulator sub-block. The classifiers manipulator sub-block is responsible for retrieving classifiers and evaluating the input data through them. The classifier manager module is responsible for retrieving the categories corresponding to the input data flag from the classifiers bank sub-block and introducing them to the suitable module. The sub-block structure of the classifiers manipulator is designed based on a group classification algorithm. In this structure, various classification methods are used simultaneously to improve the decision result. After Scale Complete input is activated, the method x normal determiner module receives the normal classification classifier appropriate to the input data flag from the classifier manager and determines through the normal output whether in its view there is input data in the normal class.
The normal output of the method x normal determiner modules is applied to the decision-maker block and the final decision regarding the normalcy of the system would be made in this sub-block. Normal complete binary output is activated after all classifiers have commented as a sign of the completion of the comment process. The implementation of the group classification mechanism in fault detection is that the method x fault determiner modules are activated in case the decision-maker sub-block detects that the system is not normal (zero normal input) and each of them from a various point of view, comment on the attribution of input data to various classes of faults and submit their comments to the decision-maker sub-block. The function of the classifiers manipulator sub-block is such that if the data is not normal, the method x fault determiner modules are asked for comments, and if the data input is normal, no comment is received from them. The normal binary signal is generated by the decision-maker sub-block, which is discussed in the following section. The fault complete binary output is activated after all classifiers have commented as a sign of completion of the comment process.
1-4- decision-maker sub-block
shows the modules that make up the decision-maker block. This sub-block announces the final comment on the status of the system in terms of normal, abnormal, and defective labels. The faults concluder module summarizes the comments of the fault label detection categories after the fault complete input is activated. The normal concluder module summarizes the comments of the normal classification classifiers when the normal complete input is activated. To build the output of this sub-block, the interpreter module is responsible for interpreting the normal and fault signals and translating them to the system status in terms of normal, abnormal, or fault occurrence. Normal output is a binary signal, we talked about this signal previously, and it can be used in other sub-blocks as an activator input.
Improver block
The task of this block, as its name implies, is to improve the performance and increase the accuracy of the ADS algorithm. This block obtains information about the occurrence and increases the knowledge of the ADS system in occurrence detection by monitoring the performance of the evaluator block and using the knowledge of the expert. After the reveal of the anomaly occurrence, this block comes into action and informs the expert about its details, and obtains additional information from him. After announcing the opinion of the expert, this block takes appropriate measures such as training new categories or re-training existing categories. In the following, the sub-blocks of the improvement block are examined.
1-2- Anomaly occurrence detector sub-block
This sub-block as the input of the Improver block is responsible for detecting the occurrence of the anomaly by monitoring the output of the evaluator block. If the result of the evaluator block operation shows an anomaly, the anomaly occurrence detector sub-block becomes actives and generates a suitable signal through its output to launch the next sub-blocks of the improver block. shows the modules that make up the anomaly occurrence detector sub-block. This sub-block consists of a single interpreter module whose task is to interpret the output of the evaluator block and detect anomalies. The output of this module is a binary signal called an anomaly.
The anomaly signal would be zero when the output of the evaluator block is normal or a fault occurrence and 1 when the output is an anomaly. 2-2- Stimulus parameter identifier sub-block. The stimulus parameter identifier sub-block provides the expert with a set of data that are the prime cause of the system going out of normal state to facilitate the process of detecting the causes of the anomaly. The expert can simply comment on the occurred anomaly by using the stimulus data of the occurrence of the anomaly and analyzing other factors. shows the modules that make up the stimulus parameter identifier sub-block. According to the figure, this sub-block consists of a module called a deviation detector. First, through the anomaly signal, the deviation detector module is informed of the time of the anomaly occurrence, and by analyzing the preprocessed data; it detects the anomaly occurrence stimulus signals and provides it to other sub-blocks as deviated data. 2-3 under the expert messenger block. This sub-block generates the content of the message based on the deviated data, which is the stimulus parameter identifier sub-block output. The message generated by this sub-block contains the information of the anomaly stimulus data. shows the modules of this sub-block.
2-4- Expert interface sub-block
This sub-block is a communication port that establishes communication between the ADS system and the expert. This sub-block should display the messages generated by the expert Messenger sub-block to the expert and receive the result of the expert's analysis. shows the modules that generate this sub-block.
2-5-Interpreter sub-block
The interpreter sub-block receives expert analysis and translates it into new and modified signals. In this way, expert analysis has two interpretations. The first case is the result of the analysis of the expert analysis of the production of a new class in the system data. In such a case, the new output under the interpreter sub-block must be activated to continue the related processes. The second case is the result of the expert analysis of the correction of a pre-existing class. In such a case, the modified output of the interpreter sub-block must be active so that related processes can be done. New and modified signals are binary signals and only one of them is active at a time.
2-6- Already classifiers modifier sub-block
Already Classifiers Modifier sub-block is activated in case the modified output of the interpreter sub-block is enabled and its job is to retrain existing classifications based on new information. shows the modules of this sub-block. The data repository module temporarily stores the output data of the normalizer sub-block after the normalized complete input is enabled so that it provides the required data to the retrain classifier module when necessary. When modify input is enabled, the retrain classifier module retrieves the corresponding classifier through the classifier manager module. The classifier manager module, based on the data flag input, retrieves the desired classifier from the classifiers bank sub-block and provides it to the retrain classifier module. After retraining the desired classification by the retrain classifier module, the trained classification is transferred to the classifiers bank through the classifier manager module.
2-7- New class classifier trainer
The New class classifier trainer sub-block is activated in case the new output of the Interpreter sub-block is enabled, and its job is to train the new classification based on new information. shows the modules of this sub-block. The function of this sub-block is quite similar to the function of the already classifiers modifier sub-block and the only difference is in the train classifier module, which does not require the classification information available in the classifiers bank to train the new classification
Diagnosis of anomalies
As mentioned above, the function of the ADS system is based on anomaly detection and incremental learning. based on what has been said so far about the function of the ADS system, we can say that to determine the abnormal state of the system for the expert, ADS acquires from him the knowledge of the cause of the abnormality and institutionalizes it within itself so that in similar upcoming occurrences it can be used by the expert as an auxiliary element. The incremental learning desired by the ADS system refers to how knowledge is communicated and extracted from the expert. That is to say, the ADS system increases its knowledge on analyzing the status of the monitored system by exchanging information with an expert in the form of incremental learning. Anomaly detection refers to finding the data with patterns that do not match the known and expected patterns of the system.
Data with unknown patterns in various fields are also referred to as anomalies, out-of-class, inconsistent observations, exceptions, deviations, surprises, unusual properties, or pollutants among which, anomaly and out-of-class expressions are much more common. In fault diagnosis, the term anomaly is much more common. The first studies in the field of abnormal data detection were commenced by statisticians in the late 18th century. In this section, we would review the anomaly detection algorithm designed for the ADS system. This unit uses machine learning techniques in the data space of the N dimension to analyze the status of the system. shows the flowchart of the desired anomaly detection algorithm. Below, each of the blocks is examined in detail. Based on , the desired anomaly detection algorithm consists of two general parts, namely training and evaluation. The training part of this algorithm is equivalent to the improvement block and the evaluation part is equivalent to the ADS system evaluator block. In the training phase, at specific times, ADS increases its knowledge by learning from an expert. In the ADS evaluation phase, with the use of the learned knowledge, it analyzes the status of the monitored system.
1- Training phase
This phase is activated when there is a need to update or increase the knowledge of the ADS system. According to the training phase shown in , the normalized samples must first be retrieved. The normalization process of the samples is performed in the normalizer sub-block. Then the clusters in the received samples for each flag should be identified, this is performed by the clustering algorithm developed by the ADS system design team, the design details of this algorithm are reviewed below.
After extracting the clusters in the samples of each flag, in the next step, a one-class classification is trained for each cluster. The trained one-class classifications are stored in a repository called OCC’s Bank. This is done for later use in the evaluation phase. The details of the performance of each of the blocks discussed in this section are examined below.
1- Extraction of clusters
As mentioned above, after retrieving the normalized data, in the training phase, it is time to extract the clusters of the samples belonging to each flag. This is done by the algorithm whose flowchart is shown in . As shown in the figure, the algorithm in the first step, after determining the required parameters in the configure algorithm step by an expert, retrieves the desired data to extract the clusters in it. The required parameters include the desired settings that the expert uses to determine how to implement the algorithm; we would discuss how to use them below. After retrieving the desired data, if the drop duplicates parameter is set to True value by an expert to reduce the computational volume and speed up the clustering algorithm, duplicate samples are removed and only one sample remains to represent the other samples. If the expert does not want to do this, he can set the drop duplicates parameter to False value. It is recommended that the drop duplicates parameter be set to True only if it does not change the result of the clustering algorithm.
If the expert wishes to see the unprocessed data before running the clustering algorithm, he can do so by setting the display raw data parameter to True. Otherwise, he should set this parameter to the False value. If the display raw data parameter is set to True by the expert, the display raw data step is executed. Display raw data flowchart is shown in . As presented in the figure, in case the dimensions of the data to be displayed are more than three dimensions, Then, to display the data while preserving the information contained in it, its dimensions are reduced to two by t-SNE and MDS algorithms (depending on expert opinion). Otherwise, (or after reducing the data dimension) three-dimensional or two-dimensional display of data is done depending on the dimensions of the input data.
The t-SNE and MDS dimension reduction algorithms used in this stage are derived from the Manikold package of the Scikit-Learn machine learning library. The t-SNE algorithm is a nonlinear machine learning algorithm for dimension reduction suitable for high-dimensional data for illustration. This algorithm maps the samples in the high-dimensional space to two or three-dimensional space so that similar samples locate next to each other and dissimilar samples space apart. Thus this algorithm can visualize high-dimensional data. The MDS algorithm is also a nonlinear visualizing algorithm: First, the data distance matrix is calculated in the space of high dimensions, then by placing the samples in a two or three-dimensional space which is based on the information in the distance matrix, attempts to display the samples in the high dimensions’ space. Next, if the expert sets the estimated clusters parameter to True value, estimating the number of clusters comes into action, otherwise it would not be executed and the algorithm would continue.
The purpose of this algorithm is to help the expert determine the number of clusters because the clustering algorithm in the next steps needs to receive a definite number of clusters. If necessary, the expert can determine the definite number of clusters with the help of a cluster number estimation algorithm. Cluster number estimation algorithm is a network search algorithm based on the K-means clustering algorithm in which the criterion for selecting the best result is the Silhouette criterion, which is used as a descriptive measure of the accuracy of sample clustering. The Silhouette criterion produces a number between [1, -1] for each sample in the data based on the clustering result. The number 1 indicates that the sample belongs to the correct cluster; the number -1 indicates that the sample belongs to the wrong cluster, and zero indicates that the sample is located between two clusters with equal distance.
As mentioned, the Silhouette criterion is calculated for every single sample, so to have a single quantity, the average Silhouette criterion is usually used to evaluate the accuracy of clustering results. The closer the average Silhouette criterion for clustering the samples is to the number 1, the more accurate the clustering is, and vice versa. shows the flowchart of the cluster number estimation algorithm. As shown in the figure, the cluster number estimation algorithm starts working to get the minimum and the maximum number of possible clusters in the data. The expert must specify the range of number of clusters in the data for the algorithm as the minimum number of clusters, maximum number of clusters.
The estimation algorithm then performs clustering operations for all elements within the specified range using the K-means algorithm and reports the element whose average Silhouette criterion is higher than the other elements as the estimation of the number of clusters. After performing the above steps, the clustering algorithm is shown in in the "get number of clusters from expert" needs to receive the precise number of clusters in the data. In this stage, the expert can determine the number of clusters in the data using the display tools and estimate the number of clusters installed during the algorithm in the "get the number of clusters from expert" stage. In the next step, the clustering algorithm decides how to continue the algorithm according to the number of definite clusters entered by the expert. In case the number of specified clusters is 1, all the samples in the data are identified as one cluster, otherwise, the fine-tune clustering result algorithm would be performed.
The flowchart of the fine-tune clustering result algorithm is shown in . This algorithm is an innovative clustering algorithm using the DBSCAN algorithm. As known the DBSCAN algorithm is a sample density-based clustering algorithm that uses the epsilon and minimum sample parameters to determine the minimum density required to detect clusters. Density refers to the accumulation of samples in one part of space. To begin, the DBSCAN algorithm requires the minimum acceptable density to identify a cluster. This algorithm examines the samples in the data and considers the set of samples that have the desired minimum density as a cluster. The main property of the DBSCAN clustering algorithm, which is also used as the core of the ADS clustering algorithm, is the criterion for defining the cluster included in its design.
Since the DBSCAN algorithm for identifying clusters focuses on the accumulation of samples its can identify various types of clusters with various shapes in space and would not be sensitive to the structure of the cluster. Its main problem is the output sensitivity in valuing the epsilon and minimum sample parameters.
To solve this problem in the ADS clustering algorithm, an innovative solution has been introduced. Here, utilizing the DBSCAN algorithm is slightly various from what is stated in scientific references, and the number of expected clusters in the study space is considered as the input of the innovative DBSCAN algorithm and during a network search process epsilon and minimum sample, parameters are determined as described below so that the desired number of clusters can be extracted from the data. Based on , the expert first determines the number of clusters in the desired data space for the algorithm. Then, the algorithm creates a network space of epsilon and minimum sample parameters. This is done to execute the search process and extract the desired number of clusters for various parameters. This way, the algorithm uses the DBSCAN algorithm to explore data space with various densities to extract the desired number of clusters. Obviously, in such a process it is possible to extract the desired number of clusters for various parameters. It is also possible to extract the desired number of clusters from the data space for a set of epsilon and minimum sample parameters.
For further processing, the set of parameters that bring about the extraction of the desired clusters is stored as the final selected candidate to be decided upon after the search process. Based on , when selecting candidates, the preferred condition over the number of clusters is the assessment of the degree of contamination of the clustering results to out-of-category data. The reason for considering this condition is that the data in the preprocessing process is cleared of out-of-class data, so the result of clusters that have reasonable out-of-class data is valid. The above was a description of the steps that are performed before the "select one of the nominees as a suggestion" step in the fine-tune clustering result algorithm. In the "select one of nominees as suggestion" step, the algorithm must select an offer from the nominees obtained in the previous steps and provide it to the expert. The criterion for selecting a suggestion among nominees in the "select one of nominees as a suggestion" step is the maximum average of the Silhouette criterion. In fact, among the obtained nominees, the nominee whose average Silhouette criterion is the highest would be presented to the expert as a suggestion. Sometimes the expert may not agree with the result proposed by the algorithm and may want to choose another nominee. In such a case, the expert must set the show nominee parameter to True to examine the other nominees to display the algorithm of the other nominees to select the final result. In this case, the expert can determine the result of the clustering based on his wishes by choosing his acceptable nominee. If the show nominee parameter is set to False, the clustering result is the same as the algorithm proposed. As shown in the flowchart in , after the clustering process, in a step called evaluate clustering result, the accuracy of the clustering result performed in the previous steps is evaluated as follows. The idea for this evaluation is derived from the definition of a cluster.
As you know, each cluster is a set of samples that are similar to each other in a specific pattern and at the same time various from other clusters. The criterion of similarity here is the Euclidean distance. So if the problem changes to a classification problem after the clustering process, the clustering would be correct and a linear or nonlinear multi-class classifier can also perform the classification operation with high accuracy. In this step, the problem is changed from unsupervised to supervised state, and the accuracy of linear or nonlinear multi-class classification is evaluated with labels obtained from the clustering algorithm. With these interpretations, a decrease in the accuracy of linear and nonlinear classifiers is considered as inappropriate clustering results and vice versa. In this project, SVM multi-class classifier is used as a nonlinear classification and the logistic regression classifier is used as a linear multi-class classification.
By observing the accuracy of the above classifications, the expert would gain a correct understanding of the accuracy of the clustering result. The criterion of the accuracy of the result of classification algorithms here is the criterion F1, which is the weighted average of the criteria of accuracy and retrieval. The accuracy criterion expresses the percentage of correct samples available as a result of classification in all classified samples. The retrieve criterion expresses the percentage of correct samples in the output of the classifier in all samples belonging to that class. After successful completion of all the above steps, the results of the clustering algorithm are attached to the samples that were previously removed to speed up the algorithm. Finally, the clustering result is displayed graphically in the form of a scatter plot.
1-2 One-class classification training
According to the ADS performance mechanism, after extracting the clusters in the desired data space, a single class classification should be trained for each cluster, so that at the end of the training process, ADS can evaluate the status of the monitored system with the help of the obtained classifiers. The flowchart of the one-class classification training process is shown in . Based on , the first step is to retrieve the settings considered by the execution of the algorithm. Then, according to the settings of the algorithm, it proceeds to remove the clusters that contain a small number of samples; the criterion of smallness is announced by the expert as a percentage of the total samples. For example, an expert can use the cluster_sample_thd parameter to express the threshold number of acceptable cluster samples as a percentage of the sample set present in the result of clustering. This removes small clusters that may not contain information or possibly out-of-category data. In the next step, as stated in the clustering process, the expert can remove similar data to speed up the training process. After the data is ready, it is time to train the classification of a one-class per cluster. To count the clusters, as shown in the flowchart, X is used as a counter.
In the next step, entitled "Train an OCC by samples of a cluster", one-class classification training is performed. The one-class classification used here is a combination of two one-class classifiers and one multi-class classifier. The desired classification training process is a network search algorithm with a cross-validation algorithm with F1 selection criteria. This process attempts to produce a reliable classifier. Cross-validation is a model evaluation method that determines how the resulting model is generalizable and independent of the training data. This method is specifically used in forecasting to determine how useful the model could be in practice. At the end of the training, according to , the category is specifically entitled to be stored in OCC’s Bank for future use.
Evaluation phase
When the training process is done, the ADS system is ready to assess the status of the monitored system. According to , in the evaluation phase as in the training phase, it is necessary to retrieve normalized data. After the retrieve, according to the flag, the retrieved data should retrieve the classifiers belonging to the mentioned flag from OCC’s Bank to comment on the normality or abnormality of the samples in the evaluated data by a set of OCC’s steps. In the "decide on evaluated data" step, as the final step, comments are made regarding the normality or abnormality of the samples of that flag by aggregating the opinions of the one-class classifiers in each flag. To aggregate the views of one-class classifiers for samples belonging to each flag, one of the following three methods is used:
  1. A sample is normal if it belongs to only one classifier.
  2. A sample is abnormal if it belongs to more than one classifier.
  3. A sample is abnormal if it does not belong to any of the classifiers.
The interface
In this section, the interface of the ADS system is examined and its function is explained. The ADS system interface is designed as a Web-Application. This allows using of ADS in cloud networks and personal computers without any restrictions. Figures 19 and 20 show an overview of the ADS system interface environment in the training and evaluation phases, respectively. As can be seen from Figures 19 and 20, this interface includes two tabs, enhancement, and evaluation.
The enhancement section is appropriate to the improver block shown in and is dedicated to model training and improvement processes. Also, the evaluation section is appropriate to the evaluator block shown in and is dedicated to evaluation processes. The interface shown in Figures 19 and 20 is designed to use the ADS system in power plants, therefore, the examples and parameters discussed below are specified for power plant applications. As shown in , the enhancement section consists of several drop-down menus and keys, which we would cover below. The user can use the drop-down menu (1) to select the desired power plant, the drop-down menu (2) to select the desired unit, and the drop-down menu (3) to select the domain. The domain menu is available to the user to analyze part of the system. As a matter of fact, through the settings mentioned below, in monitored systems, the user can determine which system signals are intended for ADS training. With this option, the user can consider all or only a certain part of the signals such as vibration signals of a gas turbine for ADS training. The keys available to the user are the Config/Reconfig key, the Train / Retrain key, the Mind key, and the Kill key, respectively, which we would discuss below. The Config /Reconfig key is used to configure the user desired domain for ADS training. Clicking on it shows the menu shown in .
In this menu, the user can set the specifications of his desired domain through the available options. According to , through option (1) the user must determine the various operating modes of the system. Using option (2) the user can determine the list of sensors that are intended to be located in the flagging space. Using option (3) the user can determine the list of sensors that are intended to be located in the decision space. The user can use option (4) to determine the normal (no-fault) operating interval of the system by setting a start date and an end date. Option (5) enables the user to allow or disallow the pre-process operation of the data received from the system. Option (6) allows the user to decide whether to remove similar samples from the data to speed up the training process. Option (7) allows the user to decide on the estimated number of clusters by ADS. In this option, if the user selects yes, then he should determine the minimum and the maximum number of clusters through the embedded slider. Through options (8) and (9), the user can determine the size of the network search space of clustering and classifier training processes, respectively. Through option (10), the user can specify the data interval required for each execution of the analysis process in the evaluation phase. For example, selecting 1 is considered as the execution of the evaluation phase for the data one day. The ADS system uses a sequence filter to prevent the generation of faulty anomaly alarms and to ma sure of the occurrence of that anomalies. Through option (11), the user can specify the number of consecutive samples required to activate the anomaly alarms in the evaluation phase. Finally, through option (12), the user can specify a name for the configured domain and use the Save key to save it. By clicking on the Train / Retrain button, the user can command the execution of ADS training operations, and then the modules shown in the panel below is sequentially performed, and include the ADS configuration according to the intended settings, data retrieval from the database, data preprocessing, data standardization, flag identification, anomaly detection model training, database update using the obtained information. After successful completion of the above steps, ADS is trained and ready to perform the analysis process.
The Mind key allows the user to view the data space in three dimensions for better analysis and viewing the training results. An example of the mind output is shown in . The Kill key is also installed for times when the user for any reason does not want to maintain the trained models so that he can eliminate the model and its related information. shows a view of the evaluation section of the ADS system interface. The user can use the models trained in the enhancement section to analyze the status of the monitored system.
The various sections of this section are explained below. Like the enhancement section, the user can use option (1) to select the power plant, using option (2) to select the unit, and using option (3) to select the domain. The keys available in this section are the Analyze key, the Feedback key, the Forget key, the Histogram key, and the monitored system key for the specified intervals. After completing the analysis process, the results are displayed in sections (4) to (8).The user can use the Feedback key to give feedback to the ADS of the summary results obtained after viewing the results of the ADS analysis. shows a view of this panel. For detected sequences, anomalous samples request feedback from the user. After entering his opinion about the cause of the anomaly, the user can use the Save key to save them in the database of the ADS system to be used in the next execution of the training process. The Forget key is included to enable the possibility of clearing ADS memory from expert feedback and resume analysis. The user requests ADS to clear the database of stored feedback by clicking Forget. Using the Histogram key, the user can view the histogram of abnormal samples along with the histogram of normal samples and other labels for the stimulus parameters selected in panel (7). shows an example of the output of the Histogram key. The Mind key has the same function as the Mind key described in the enhancement section. The only difference is that here the Mind key, in addition to displaying the distribution of data in three-dimensional space, also displays the position of the anomalous samples detected after the analysis process relative to the normal and labeled samples. Section (4) is a panel called System Status, which is responsible for displaying the status of the system in terms of the occurrence of faults and anomalies.
This panel displays the labels detected in the analyzed samples. Section (5) is a panel called Anomaly Trend, which displays the volume of anomalous occurrence over time in terms of the percentage of analyzed samples and thus provides the user with information about the health status of the system over time. Section (6) is a panel called Anomaly Sequence, which displays the anomalous event sequences detected in the analysis process, and the user can use it to view information such as the number of anomalous events and the volume of anomalous samples in each sequence and anomalous event control period. By clicking on each sequence, the user can view additional information such as the stimulus factors for anomalies in that sequence in the panel (7). Additionally, by selecting the stimulus factors for anomalies by the user in panel (7), the time signal of the selected stimulus in panel (8) along with the location of the occurrence of that sequence is displayed for further investigation. In this way, the user with the above-mentioned information and his previous experiences can provide appropriate feedback.
Advantage effects of invention
As mentioned earlier, if data is available, the ADS system can be used to monitor the status of any type of system in any industry. Here are some of the benefits of using ADS in different industries.
This invention can be used to monitor the situation to detect, detect, and predict the occurrence of defects in various industrial and non-industrial systems.
Reduce unforeseen system outages due to fault detection and prediction
Avoid progress and defect and become a failure
Informing users of the occurrence of an unknown and abnormal state in the systems
In addition to all the capabilities provided in the field of fault detection and detection and prediction, this invention has very high generalizability that can be used to monitor the status of any system, whether industrial or non-industrial, only if data is available.
: System input / output structure
: System blocks and sub-blocks
: Modules under the Normalizer block in the training phase.
: Modules under the Normalizer block in the evaluation phase.
: Modules under the classifier executable blocks and the classifier bank.
: shows the Decision Maker sub-block modules.
: Modules under the Anomaly Event Detector block.
: Stimulus Parameter Identifier sub-block modules.
: Modules under the Expert Messenger block.
: Modules under the Expert Interface block.
: Modules under the Already Classifiers Modifier block
: Modules under the New Class Classifier Trainer block
: Flowchart of an anomaly detection algorithm as Data-Driven.
: Flowchart of the cluster extraction algorithm in the data.
: Display raw data flowchart algorithm.
: Flowchart of the Estimate number of clusters algorithm.
: Flowchart of Fine tune clustering result algorithm.
: Flowchart of a single class classification teaching algorithm.
: Overview of the system user interface environment in the training phase
: Overview of the system interface environment in the evaluation phase.
: View of the domain configuration menu
: view of the system output
: View of the feedback panel in the Evaluation section
: view of the output of the Histogram key
: System input / output structure
: System blocks and sub-blocks
: Modules under the Normalizer block in the training phase.
: Modules under the Normalizer block in the evaluation phase.
: Modules under the classifier executable blocks and the classifier bank.
: shows the Decision Maker sub-block modules.
: Modules under the Anomaly Event Detector block.
: Stimulus Parameter Identifier sub-block modules.
: Modules under the Expert Messenger block.
: Modules under the Expert Interface block.
: Modules under the Already Classifiers Modifier block
: Modules under the New Class Classifier Trainer block
: Flowchart of an anomaly detection algorithm as Data-Driven.
: Flowchart of the cluster extraction algorithm in the data.
: Display raw data flowchart algorithm.
: Flowchart of the Estimate number of clusters algorithm.
: Flowchart of Fine tune clustering result algorithm.
: Flowchart of a single class classification teaching algorithm.
: Overview of the system user interface environment in the training phase
: Overview of the system interface environment in the evaluation phase.
: View of the domain configuration menu
: view of the system output
: View of the feedback panel in the Evaluation section
: view of the output of the Histogram key
Examples
Because the ICMS system interface is designed as a Web application, it can be used both on PC and in cloud computing systems. It will also be possible to access it through a variety of local networks and the Internet. In this way, with the help of the ICMS system, it will be possible to easily monitor the status of different systems, provided that their data is available remotely.
Faults can occur in systems and one of the common challenges in the industry is the occurrence of faults. Fault occurrence is normal and the effect of this fault on the system must be eliminated so that the system can return to normal operation. After a fault occurrence, issues like defect occurrence detection and occurrence effects elimination are faced.
The ADS system detects anomalies using one-class classifiers (a semi-regulatory classification that detects anomalies). In this system, to detect anomalies, the data distribution of the monitored system is compared with the known data distribution, which reveals the occurrence of unknown behavior. After detection, the system obtains information and completes its knowledge by informing the expert about the details of the anomaly. Then, it eliminates the faults and in the shortest time, causes the desired efficiency and correct and safe operation of the systems, and prevents the occurrence of failures in them.
As mentioned earlier, if data is available, the ICMS system can be used to monitor the status of any type of system in any industry. Here are some of the benefits of using ICMS in different industries.
This design can be used to troubleshoot mechanical devices, medicine, electronic equipment, automotive, etc.
This invention can be used to monitor the situation to detect, detect, and predict the occurrence of defects in various industrial and non-industrial systems.
Reduce unforeseen system outages due to fault detection and prediction
Avoid progress and defect and become a failure
Informing users of the occurrence of an unknown and abnormal state in the systems.

Claims (5)

  1. A software system called ICMS, which generally consists of a database, an artificial intelligence algorithm, and a graphical web-based user interface. And its user interface consists of improvement and evaluation sections, which use artificial intelligence to assess the functional health status of a variety of complex industrial and non-industrial systems only by using the previous performance data of the system. This software also uses the knowledge of an expert and by expanding the content of its database, it provides the possibility of predicting the health status of systems in the future.
  2. According to claim 1, the above scheme is to detect, detect and predict fault occurrence intelligently in a system. Due to the ability to be generalized in the design process of the ADS system, if the information requirements are met, this system can be implemented on any type of industrial and non-industrial system and this architecture consists of two basic blocks.
  3. According to claim 2, the diagram block structure of the ADS system can be divided into two general parts.
    Evaluator block: The task of this block is to assess the status of the monitored system, using the knowledge embedded in it.
    Improver block: The task of this block is to use the knowledge of the expert along with the results of the analysis performed by the evaluator block to improve the knowledge institutionalized in the ADS system.
  4. According to claim 3, the anomaly detection algorithm consists of two general parts: training and evaluation.
    The training part of this algorithm is equivalent to the improvement block and the evaluation part is equivalent to the ADS system evaluator block. In the training phase, at defined times, ADS increases its knowledge by learning from an expert. In the ADS evaluation phase, with the help of the learned knowledge, it analyzes the status of the monitored system.
  5. According to claim4, When the training process is done, the ADS system is ready to assess the status of the monitored system. In the evaluation phase as in the training phase, it is necessary to retrieve normalized data. After the retrieve, according to the flag, the retrieved data should retrieve the classifiers belonging to the mentioned flag from OCC’s Bank to comment on the normality or abnormality of the samples in the evaluated data by a set of OCC’s steps. In the "decide on evaluated data" step, as the final step, comments are made regarding the normality or abnormality of the samples of that flag by aggregating the opinions of the one-class classifiers in each flag.
PCT/IB2022/050672 2021-02-21 2022-01-26 Equipment fault detection system using artificial intelligence(icms) WO2022175767A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
IR139950140003010421 2021-02-21
IR13993010421 2021-02-21

Publications (1)

Publication Number Publication Date
WO2022175767A1 true WO2022175767A1 (en) 2022-08-25

Family

ID=83226368

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IB2022/050672 WO2022175767A1 (en) 2021-02-21 2022-01-26 Equipment fault detection system using artificial intelligence(icms)

Country Status (1)

Country Link
WO (1) WO2022175767A1 (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116930666A (en) * 2023-09-15 2023-10-24 深圳凯升联合科技有限公司 Intelligent diagnosis system and diagnosis method for low-voltage complete switch cabinet

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10410135B2 (en) * 2015-05-21 2019-09-10 Software Ag Usa, Inc. Systems and/or methods for dynamic anomaly detection in machine sensor data
US20200312056A1 (en) * 2019-03-28 2020-10-01 GM Global Technology Operations LLC Monitoring and diagnosing vehicle system problems using machine learning classifiers

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10410135B2 (en) * 2015-05-21 2019-09-10 Software Ag Usa, Inc. Systems and/or methods for dynamic anomaly detection in machine sensor data
US20200312056A1 (en) * 2019-03-28 2020-10-01 GM Global Technology Operations LLC Monitoring and diagnosing vehicle system problems using machine learning classifiers

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116930666A (en) * 2023-09-15 2023-10-24 深圳凯升联合科技有限公司 Intelligent diagnosis system and diagnosis method for low-voltage complete switch cabinet

Similar Documents

Publication Publication Date Title
Jiang et al. Performance supervised plant-wide process monitoring in industry 4.0: A roadmap
Li et al. Data alignments in machinery remaining useful life prediction using deep adversarial neural networks
Lee et al. Prognostics and health management design for rotary machinery systems—Reviews, methodology and applications
Galar et al. Context awareness for maintenance decision making: A diagnosis and prognosis approach
Manco et al. Fault detection and explanation through big data analysis on sensor streams
Xu et al. Health management based on fusion prognostics for avionics systems
KR101713985B1 (en) Method and apparatus for prediction maintenance
CN102520697B (en) Onsite information preprocessing method of remote cooperative diagnosis
Rögnvaldsson et al. Self-monitoring for maintenance of vehicle fleets
Xiang et al. A generalized predictive framework for data driven prognostics and diagnostics using machine logs
CN109871002B (en) Concurrent abnormal state identification and positioning system based on tensor label learning
Brahimi et al. Development of a prognostics and health management system for the railway infrastructure—Review and methodology
Kobbacy et al. New technologies for maintenance
De Simone et al. LSTM-based failure prediction for railway rolling stock equipment
WO2022175767A1 (en) Equipment fault detection system using artificial intelligence(icms)
CN116720324A (en) Traction substation key equipment fault early warning method and system based on prediction model
Borissova et al. An integrated framework of designing a decision support system for engineering predictive maintenance
Liu et al. A machine learning-based clustering approach to diagnose multi-component degradation of aircraft fuel systems
Létourneau et al. A domain independent data mining methodology for prognostics
Hou et al. Fault detection and diagnosis of air brake system: A systematic review
Merkt Predictive models for maintenance optimization: an analytical literature survey of industrial maintenance strategies
Davari et al. A fault detection framework based on lstm autoencoder: A case study for volvo bus data set
Najjar et al. Predictive Maintenance of Urban Metro Vehicles: Classification of Air Production Unit Failures Using Machine Learning
Xia A systematic graph-based methodology for cognitive predictive maintenance of complex engineering equipment
Liao An adaptive modeling for robust prognostics on a reconfigurable platform

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22755642

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22755642

Country of ref document: EP

Kind code of ref document: A1