CN112101431A - Electronic equipment fault diagnosis system - Google Patents

Electronic equipment fault diagnosis system Download PDF

Info

Publication number
CN112101431A
CN112101431A CN202010891078.2A CN202010891078A CN112101431A CN 112101431 A CN112101431 A CN 112101431A CN 202010891078 A CN202010891078 A CN 202010891078A CN 112101431 A CN112101431 A CN 112101431A
Authority
CN
China
Prior art keywords
fault
layer
time
event
neural network
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202010891078.2A
Other languages
Chinese (zh)
Inventor
陈文豪
黄明
李骁
雷志雄
乔文昇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
CETC 10 Research Institute
Southwest Electronic Technology Institute No 10 Institute of Cetc
Original Assignee
Southwest Electronic Technology Institute No 10 Institute of Cetc
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 Southwest Electronic Technology Institute No 10 Institute of Cetc filed Critical Southwest Electronic Technology Institute No 10 Institute of Cetc
Priority to CN202010891078.2A priority Critical patent/CN112101431A/en
Publication of CN112101431A publication Critical patent/CN112101431A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G06F18/24137—Distances to cluster centroïds
    • G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/254—Fusion techniques of classification results, e.g. of results related to same input data
    • G06F18/256—Fusion techniques of classification results, e.g. of results related to same input data of results relating to different input data, e.g. multimodal recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Test And Diagnosis Of Digital Computers (AREA)

Abstract

The invention discloses a fault diagnosis system of electronic equipment, and aims to provide a system for carrying out fault diagnosis on the electronic equipment by using a cyclic neural network point process condition intensity function. The invention is realized by the following technical scheme: the data preprocessing module extracts acquired data from the first database and converts the fault and normal time of the electronic equipment into a timestamp; the point process cyclic neural network module inputs data preprocessed by the data preprocessing module into a time sequence and a fault event sequence, and sends the data into a network embedded in a mapping layer, the time sequence and the fault event sequence of a test result are fused into an RNN (neural network) embedding layer, faults are predicted based on a cyclic neural network point process condition intensity function, a predicted fault type, a fault event subtype and a predicted fault timestamp related to the fault event are output through a prediction layer of the fault prediction module, and the predicted fault timestamp and the result of the fault type are stored in a second database.

Description

Electronic equipment fault diagnosis system
Technical Field
The invention relates to a system for diagnosing faults of electronic equipment in the field of time point processes of artificial Neural networks, and provides a system for modeling a time sequence point process by using two Recurrent Neural Networks (RNNs) and diagnosing the faults of the electronic equipment by using a Recurrent Neural Network point process condition intensity function.
Background
With the rapid development of modern industrial and scientific technology, production equipment is becoming large-scale, high-speed, automatic and intelligent, and once the system fails, catastrophic accidents can be caused, resulting in great loss of personnel and property. The modern industrial production process is gradually enlarged, integrated and refined, and the structure is more and more complex. In order to maintain various devices and equipment, people must find out the faults of circuits by computers, and the diagnosis of the faults of the analog circuits is one of the most important problems in large-scale integrated circuits. The problem of fault diagnosis and positioning of the analog circuit not only draws wide attention, but also is a big problem of designing and using an electronic system by experts at home and abroad, and for a large-scale analog circuit, because of more elements, the diagnosis is more complex, wherein the fault, namely soft fault diagnosis under the condition of large-scale nonlinear complex circuit tolerance is also a problem which puzzles a great number of scientists. To date, there is little literature on systematic and effective methods for the diagnosis of soft faults, i.e., tolerance circuits, particularly for the diagnosis of faults in large-scale analog circuits. With the improvement of automation and intelligence of mechanical equipment and electronic systems, the detection and fault diagnosis level is also continuously improved, and higher requirements are placed on the fault diagnosis technology of electronic equipment. Faults of industrial control systems are mainly sensor faults, controller faults, actuator faults and controlled process element faults. The faults can be classified into additive faults and multiplicative faults according to the form in which the faults occur. The fault diagnosis has two meanings, one means that some special instruments are used for detecting whether some machine equipment works normally or not: the other method is that the computer analyzes and judges whether the production is normal, when the fault is caused by any reason, and what degree of the fault, etc. by using the analytic redundancy of the system to complete the working condition analysis, so as to draw a conclusion. The latter is generally referred to as a fault diagnosis technique commonly used at present. The main contents of fault diagnosis are to separate out the fault location, judge the fault type, estimate the fault size and time, and make evaluation and decision. Generally, fault detection is easier and takes less time. In contrast, fault diagnosis is difficult, and it takes much time to correctly separate the fault site and accurately estimate the size of the fault. The task of fault diagnosis can be divided into four aspects from low level to high level: (1) fault modeling: and establishing a mathematical model of the system fault according to the prior information and the input-output relation, and using the mathematical model as a basis for fault detection and diagnosis. (2) And (3) fault detection: and judging whether the running system has faults or not from the detectable and undetectable estimated variables, and sending an alarm once the system has unexpected changes. (3) Separation and estimation of faults: if the system has faults, the position of a fault source is given, and whether the fault reason is actuator fault, sensor fault or controlled object fault or extra disturbance is distinguished. The failure estimation is to calculate parameters such as the degree and size of a failure and the time of occurrence of the failure while clarifying the nature of the failure. (4) Classification, evaluation and decision of faults: and judging the severity of the fault, the influence and development trend of the fault on the system, and taking different measures aiming at different working conditions, wherein the measures comprise starting of a protection system, system reconstruction, fault tolerance and the like. An important way to improve the reliability and safety of the system is to perform real-time fault detection and diagnosis on the operation process of the system, thereby realizing fault-tolerant control, early warning on catastrophic faults and taking corresponding fault control measures. The traditional fault diagnosis method of the industrial control system usually requires to establish an accurate mathematical model of a diagnosed object, and in fact, because a control system is a complex of electronic, mechanical, software and other factors, the fault of the control system has the problems of complexity, nonlinearity and the like, and the mathematical model is difficult to establish accurately; when a certain link of an electronic system breaks down or is abnormal, if the failure or the abnormality is not timely processed, the failure is expanded, and a major accident is caused. Therefore, an efficient and accurate real-time fault diagnosis method is established, the hidden trouble of the fault is eliminated, the fault is timely eliminated, and safe, stable and high-quality production and work are ensured, which becomes the key point of the whole production and electronic system working process. In engineering practice, there are a large number of multiple faults, multiple processes, sudden faults and the need to monitor and diagnose large machines or engineering systems. Because the neural network has the capability of self-learning and fitting any continuous nonlinear function, the research of fault diagnosis of the industrial control system by applying the neural network theory and method becomes a popular research subject. A great deal of research results show that the neural network fault diagnosis technology provides a new approach for the fault diagnosis of the industrial control system. The neural network diagnosis method is mainly used for generating residual errors by adopting the neural network, further diagnosing by adopting the neural network, compensating self-adaptive errors by adopting the neural network and diagnosing faults by adopting the fuzzy neural network.
For equipment fault diagnosis evaluation, the key point is to effectively model asynchronous event sequences over a continuous time domain. The time-series point process may model data generated in a number of real scenarios, such as equipment fault logs, location and magnitude of earthquakes, and so forth. These data are all multidimensional asynchronous event data that interact and exhibit complex dynamics in the continuous time domain. Unlike the discrete nature of a synchronous time series of equally spaced samples, the time stamps of asynchronous events are in the continuous time domain. The sequential point process has become an important solution to this problem. The time point process is often used for prediction of future events and discovery of causal associations, these two tasks in a sense corresponding to two core problems in machine learning: accuracy of prediction and model interpretability. For example, most parametric time sequence point process models achieve the purpose of model learning by optimizing a log-likelihood function or its lower bound. For the problem of modeling asynchronous event sequences in a continuous time domain, a time sequence point process has become an important solution. From the perspective of machine learning, the development of the time-series point process can be divided into two directions: a conventional count point process and a depth point process. There are many point process function forms of statistical point processes, and the success of modeling often depends on the correct model selection. Since the functional form and the parametric representation of the method generally have a little physical significance, the statistical point process has stronger interpretability and has smaller dependence on the number of samples. In contrast, the deep-point process (also referred to as the neural-point process) exploits the powerful capacity of neural networks, trying to learn a more fitting model through large-scale data, and reducing the dependence on a priori knowledge, while sacrificing model explanatory power to some extent.
The parameterization method in the point process has a definite intensity function form, and the corresponding intensity function form needs to be manually determined in advance before modeling. One major limitation of the parameterized form of the point process is its specialization and limited expressive power for arbitrarily distributed event data, which tends to be overly simplified or even impossible to capture the complexity of the problem in practical applications. Furthermore, it risks model misappropriation due to misjudgment of model selection. Under the conditions of poor prior knowledge and complex data distribution, the point process is low in performance of mining hidden event rules so as to predict future events. With the development of deep neural networks, deep learning techniques exhibit strong capabilities in multiple fields, and the fusion of point processes and deep learning techniques has become a trend.
The current methods based on signal processing usually directly analyze measurable signals by using signal models, such as correlation function, frequency spectrum, autoregressive moving average, wavelet transform, etc., and extract characteristic values such as variance, amplitude, frequency, etc., so as to detect faults. The method based on the analytical model is to process and diagnose the measured information according to a certain mathematical method on the basis of knowing a mathematical model of a diagnosis object, and can be divided into a state estimation method, an equivalent space method, a parameter estimation method, a knowledge-based fault diagnosis method, an expert system fault diagnosis method, a fuzzy fault diagnosis method, a fault tree fault diagnosis method, a neural network fault diagnosis method and a data fusion fault diagnosis method; for an online monitoring or diagnostic system, the contents of the database are real-time monitored working data; for off-line diagnosis, the detection data can be stored during fault, and some characteristic data can be artificially detected, namely, various information knowledge bases required and generated in the reasoning process are stored: the stored knowledge can be the working environment of the system, and the system knowledge (reflecting the working mechanism and the structure knowledge of the system); the rule base stores a group of rules, reflects the causal relationship of the system and is used for fault reasoning. The knowledge base is a collection of expert domain knowledge. And comprehensively applying various rules to carry out fault diagnosis according to the acquired information, and outputting a diagnosis result to form an organization control structure of the expert system. Current control systems become more complex and it is difficult to obtain accurate mathematical models of the systems in many cases, and knowledge-based methods do not require accurate mathematical models and therefore have good application prospects. The fault diagnosis method based on the expert system is the most active branch in the field of artificial intelligence, and is widely applied to a process monitoring system. The method is independent of a mathematical model of the system, and a set of intelligent computer program is designed according to long-term practical experience and a large amount of fault information knowledge of people, so that the problem of fault diagnosis of a complex system is solved. Expert system troubleshooting limitations rely on expert domain knowledge acquisition, which is recognized as a bottleneck in expert system research and development. In addition, there are also different degrees of limitations in adaptive learning, learning ability, and real-time performance. The fuzzy fault diagnosis method establishes a fuzzy relation matrix R between faults and symptoms, also called a membership matrix. The size of each element in the matrix indicates how closely they are related to each other. Constructing a membership function for fuzzy fault diagnosis is a premise for realizing fuzzy fault, but the membership function is artificially constructed and contains certain subjective factors; in addition, certain requirements are also provided for the selection of the characteristic elements, and if the selection is not reasonable, the diagnosis precision is reduced, and even the diagnosis fails. Logical reasoning diagnostics: the top-down test method includes starting from a fault top event, testing an initial intermediate event, judging and testing a next-level intermediate event according to a test result of the intermediate event until a test bottom event, and searching a fault reason and a fault part. The limitation of the fault tree approach is the reliance on building a correct and reasonable fault tree. This diagnostic method will fail if the fault tree is not fully built or is incorrect once. Because the neural network has the functions of processing complex multi-modes and performing association, speculation and memory, the fault diagnosis method based on the neural network is very suitable for being applied to a fault diagnosis system. The system has self-organizing and self-learning capabilities, and can overcome the defect that the traditional expert system cannot work when heuristic rules are not considered. The neural network fault diagnosis learning process of the neural network fault diagnosis method comprises the following steps: based on a certain standard pattern sample, a neural network classifier is designed according to a certain classification rule and is trained. The diagnosis process is to put the unknown pattern into a trained classifier to diagnose the fault class of the unknown pattern. Both preprocessing and feature extraction learning and diagnosis processes include preprocessing and feature extraction. Pretreatment: the other type of fault mode is obtained by deleting useless information in the original data, the sample space is mapped into the data space, and fault diagnosis is facilitated through certain transformation. Feature extraction: the subject to be diagnosed can be generally viewed as a set of time series from the data obtained. By sampling segments of the time series, the input data can be mapped to points of sample space. Such data may contain information on the type, extent and location of the fault. However, the distribution of these feature information varies from the sample space, and thus, they cannot be directly used for classification in general. Effective fault features need to be extracted with appropriate changes. Neural networks commonly used for fault diagnosis classification are: the combination of BP network, Bidirectional Associative Memory (BAM) network, adaptive resonance theory neural network and other fault diagnosis methods is combined with an expert system. For a large-scale complex network, if each specific element fault information is directly stored in a neural network, the information quantity is too large, the structure of the neural network for diagnosing the element fault information is more complex, and the diagnosis speed is influenced. The neural network fault diagnosis has the limitations that training samples are difficult to obtain, the experience knowledge of field experts is ignored, and the expression mode of the network weight form is difficult to understand. The data fusion of the data fusion fault diagnosis method is a data processing process which utilizes a computer to automatically analyze and synthesize information from a plurality of sensors according to certain criteria so as to complete needed decision and judgment.
Disclosure of Invention
The invention aims to: aiming at the existing problems, the system can reduce the dependence on the prior knowledge and the design workload of the complex learning algorithm, can avoid the risk caused by improper model selection of the traditional point process model, and carries out the fault diagnosis of the electronic equipment by using the strength function of the point process condition of the recurrent neural network.
To achieve the above object, an electronic device fault diagnosis system includes: the data preprocessing module, the point process circulation neural network module and the failure prediction module that connect in series in order between two databases, its characterized in that: the data preprocessing module extracts acquired data from the first database, filters, screens and reads the ID number of the electronic equipment, the fault or normal time of the electronic equipment and the fault type of the electronic equipment, and converts the fault and normal time of the electronic equipment into a timestamp; the point process cyclic neural network module interprets the condition intensity function of the point process as nonlinear mapping, data preprocessed by the data preprocessing module is input into a time sequence and a fault event sequence and is sent into a network of an embedded mapping layer fusing information from two long-short term memory neural networks LSTM network, the LSTM adopts the time sequence with the length of T at uniform intervals to simulate the background intensity of the occurrence of the fault event, and learns the long-term event dependency relationship by using the fault event sequence with the length of N to train and recognize to obtain a test result, the time sequence and the fault event sequence of the test result are fused into an RNN embedded layer of the cyclic neural network, the RNN is used as a composite neural network for collaborative modeling, the composite neural network prediction model predicts the fault based on the condition intensity function of the point process of the cyclic neural network, and the predicted fault type is output through a prediction layer of the fault prediction module, A sub-type of the failure event and a predicted failure timestamp associated with the failure event, and storing the result of the predicted failure timestamp and the failure type in a second database.
In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:
(1) has strong information comprehensive capability. The invention adopts a point process cyclic neural network module which comprises a time sequence and a fault event sequence RNN and is fused into an embedded layer and a fault prediction module which comprises a fault time prediction layer and a fault event prediction layer, simultaneously processes a large amount of input information of different types, and well solves the problems of complementarity and redundancy between the input information by using the strong capacity of a neural network RNN (RecurrentNeralNet) and trying to learn a model with stronger fitting capability through large-scale data.
(2) The method adopts a data preprocessing module to extract collected data from a database, reads the ID number of the electronic equipment, the fault or normal time of the equipment and the fault type of the equipment through filtering and screening, and converts the fault and normal time of the equipment into a timestamp; and predicting the whole model of an output layer through the fault event type and the fault timestamp, fusing the time sequence and the condition intensity function of the LSTM instantiation point process of the fault event sequence, and simulating the background and the processing historical event of the fault event sequence. End-to-end training can be realized, and features can be automatically extracted.
(3) The invention adopts a point process cyclic neural network module to interpret the conditional strength function of a point process as nonlinear mapping, two RNNs which adopt a Long-Short Term Memory neural network (LSTM) for training are used as a composite neural network for collaborative modeling, data are respectively input into a time sequence and a fault event sequence from a database through data preprocessing and feature extraction, and then the data are connected into a network embedded in a mapping layer together, and the Long-Short Term Memory network LSTM (Long Short-Term Memory) is used as a time Recurrent Neural Network (RNN). In order to model the non-linear intensity mapping without any a priori knowledge. LSTM can perform better in longer sequences than normal RNNs.
(4) The invention uses the evenly spaced time sequence with the length of T to simulate the background intensity of the occurrence of the fault event, uses the fault event sequence with the length of N to learn the long-term event dependency relationship, and utilizes two LSTMs to respectively model the time sequence and the fault type sequence, thereby reducing the dependency on the prior knowledge. The accuracy of fault diagnosis and prediction is obviously improved, and the obtained fault information is more diversified.
(5) The invention uses the prediction layer to output the predicted main type, the sub-type of the fault event and the timestamp associated with the fault event, and stores the result of the predicted fault timestamp and the fault type in the database, thereby reducing the design workload of a parameter or semi-parameter point process model and a complex learning algorithm thereof, simplifying the model and avoiding the risk caused by improper model selection of the traditional point process model.
Drawings
FIG. 1 is a flow chart of the cooperation of the modules of the fault diagnosis system based on the process condition intensity function of the recurrent neural network point;
FIG. 2 is a schematic diagram of the present invention for collaborative modeling of time series and fault event series;
FIG. 3 is a schematic of the LSTM of the time series and fault event series of the present invention;
FIG. 4 is a schematic diagram of a point process cycle neural network module and a fault prediction module, such as LSTM.
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following embodiments and accompanying drawings.
Detailed Description
See fig. 1. In a preferred embodiment described below, an electronic device fault diagnosis system includes: the data preprocessing module, the point process circulation neural network module and the failure prediction module that connect in series in order between two databases, its characterized in that: the data preprocessing module extracts acquired data from the first database, filters, screens and reads the ID number of the electronic equipment, the fault or normal time of the electronic equipment and the fault type of the electronic equipment, and converts the fault and normal time of the electronic equipment into a timestamp; the point process cyclic neural network module interprets the condition intensity function of the point process as nonlinear mapping, data preprocessed by the data preprocessing module is input into a time sequence and a fault event sequence and is sent into a network of an embedded mapping layer fusing information from two long-short term memory neural networks LSTM network, the LSTM adopts the time sequence with the length of T at uniform intervals to simulate the background intensity of the occurrence of the fault event, and learns the long-term event dependency relationship by using the fault event sequence with the length of N to train and recognize to obtain a test result, the time sequence and the fault event sequence of the test result are fused into an RNN embedded layer of the cyclic neural network, the RNN is used as a composite neural network for collaborative modeling, the composite neural network prediction model predicts the fault based on the condition intensity function of the point process of the cyclic neural network, and the predicted fault type is output through a prediction layer of the fault prediction module, A sub-type of the failure event and a predicted failure timestamp associated with the failure event, and storing the result of the predicted failure timestamp and the failure type in a second database.
The data preprocessing module extracts data acquired by the electronic equipment from a first database of the system, filters out other unnecessary information, leaves an ID number of the electronic equipment, the fault or normal time of the electronic equipment and the fault type of the electronic equipment, converts the fault and normal time of the electronic equipment into a timestamp, identifies the time at a certain moment, and performs digital coding according to different faults or normal types of the electronic equipment. If the electronic equipment has a fault, the electronic equipment is marked as 1, if the electronic equipment is normal, the electronic equipment is marked as 0, and if other fault subtypes occur, the electronic equipment continues to carry out coding from 2, such as locking, retaining and the like.
The time series recurrent neural network RNN can be regarded as an attribute characteristic, is similar to a background function in an intensity function, and can track spontaneous background in real time; while the fault event sequence recurrent neural network RNN represents a dependency on historical events.
Two Recurrent Neural Networks (RNN) of the point process recurrent neural network module are used as a composite neural network for collaborative modeling, wherein one Recurrent Neural Network (RNN) models a time sequence and is used for representing real-time continuously updated dynamic variables, such as continuous change data of electronic equipment temperature, voltage and the like; another recurrent neural network RNN models a sequence of fault events representing times of irregular occurrence, such as features of unequal intervals of occurrence of sudden failures of electronic devices.
The point process cyclic neural network module inputs data into the two long-short term memory neural networks LSTM networks, and connects the two long-short term memory neural networks LSTM networks together to an embedded mapping layer which fuses information from the two LSTM networks.
The long-short term memory neural network LSTM is a variation of RNN. RNN adopts long short-term memory neural network LSTM to train, uses evenly spaced time sequence with length T to simulate the background intensity of fault event, and learns long-term event dependency relationship by using fault event sequence with length N, wherein, represents the condition of event e under time stamp T.
The fault prediction module comprises a fault time prediction layer and a fault event prediction layer, after the output of the embedded layer is finished, the prediction layer is used for outputting the predicted main type, the sub-type of the fault event and the relevant time stamp of the fault event, for the fault event prediction, the fault event prediction layer uses a common classification measurement standard, measures the absolute difference value between a predicted time point and an actual time point by using Mean Absolute Error (MAE), and stores the result of the fault time prediction and the result of the fault type in a second database. The final output in the failure event prediction layer is a subtype of the failure event without hierarchy.
See fig. 2. In the collaborative modeling of the time sequence and the fault event sequence, the fault event type, the electronic equipment ID and the fault timestamp are simultaneously stored into a first database by using the electronic equipment time identifier and the coded identifier of the electronic equipment fault type, wherein the first column of the fault event sequence of the recurrent neural network RNN in the t-time axis represents the ID number of the electronic equipment, the second column represents the timestamp of the electronic equipment fault or normal, and the third column represents the fault type of the electronic equipment under the timestamp; each row is a time stamp indicating that a certain electronic device has failed under a certain failure type, and the time sequence of the occurrence of the failure event is used
Figure BDA0002657020910000071
To learn the dependency of long-term fault events.
See fig. 3. The LSTM neural network comprises an input gate, an output gate, a forgetting gate and three control gates, wherein the gates are a system for allowing information to pass selectively, the network is enabled to remember information and control a storage unit for storing information in the LSTM, and a weight is set at the edge of the storage unit consisting of a Sigmoid activation function and point-by-point multiplication and the other part of the neural network. The output gate learns when the activation state is transmitted out of the storage unit through training, and when the output value is 1, the other parts of the neural network write the content into the storage unit; the forgetting gate learns when the information of the storage unit at the previous moment is transmitted into the storage unit at the next moment, namely, the information of the previous unit is determined to be discarded; when the forget door is opened, the weight value of the system connection is 1, and the storage unit writes the content into the storage unit; when the forget gate output is 0, the storage unit deletes the previously stored content, and correspondingly, for the post-transfer, the output gate learns when to let the error information flow into the storage unit, and the input gate learns when to let the error information flow out of the storage unit.
The Sigmoid function output is a value between 0, which represents that no amount is allowed to pass, and 1, which represents that any amount is allowed to pass, describing how much of each part is allowed to pass. For forward propagation, the input gate determines when to pass the active state into the memory cell by training learning, and when its output value is 1, the other part of the neural network reads the memory cell.
The long short-term memory network LSTM (Long short-term memory) of the time sequence and the fault event sequence is a time Recursive Neural Network (RNN), the single hidden layer neural network is that a hidden layer is arranged between an input and an output, namely the output of the input layer is the input of the hidden layer, the product of the output of the hidden layer and the corresponding weight is the input of the output layer, and the output of the output layer is the final output. A hidden layer network is a layer of feature hierarchy, and each neuron can be similarly considered as a feature attribute. Old cell state C at t-1 input of long-short term memory network LSTMt-1Through gated cyclesThe ring unit circulation node transmits the Sigmoid activation function sigma vector to a forgetting gate ftAnd old cell state Ct-1Discarding part of information by multiplying old cell state point by point, then multiplying the output value of the tanh layer and the Sigmoid layer, and adding the vector sum needing to be updated
Figure BDA0002657020910000081
Obtaining a new cell state Ct。
Forget door ftReading previous sequence hidden layer output ht-1Vector x from the input of the present sequence at time ttAnd cell state Ct-1Obtaining the content to be discarded and retained in the cell state of the previous layer through a Sigmoid activation function, and inputting a gate itCreating a new candidate vector at the tanh level
Figure BDA0002657020910000082
Forget door ftAnd old cell state Ct-1Multiply and then add the data to be updated
Figure BDA0002657020910000083
Obtaining a new cell state Ct。
Outputting h through previous sequence of hidden layerst-1And the vector x of the input of the present sequence at time ttThe summed vector sum passes through the sigma layer through the output gate otMultiplying the output value to the output of the tanh layer to play a role of scaling, and outputting the hidden layer h at the time ttA value between 0 and 1, wherein 0 represents that no amount is allowed to pass and 1 represents that any amount is allowed to pass.
The long-short term memory network LSTM adopted in the embodiment calculates the forgetting gate f by the following calculation formulatAnd input gate itAnd an output gate otCell state CtHidden layer output ht：
ft＝σ(WfcCt-1+Wfhht-1+Wfxxt+bf)
it＝σ(WicCt-1+Wihht-1+Witxt+bi)
ot＝σ(WocCt+Wohht-1+Woxxt+bo)
Ct＝ftCt-1+it⊙tanh(Wchht-1+Wcxxt+bc)
ht＝ot⊙tanh(Ct)
Figure BDA0002657020910000084
The long-short term memory network LSTM is simplified as follows: (C)t,ht)＝LSTM(Ct-1,ht-1,xt)
Where σ is Sigmoid function
Figure BDA0002657020910000091
ht-1W represents ht-1，CtTo a new cell state, Ct-1Is old cell state,. indicates element-by-element multiplication, vector xtAnd a learnable parametric weight matrix of b represents bias terms for the respective gate and cell states.
The LSTM calculation formula adopted by the system is as follows:
ft＝σ(WfcCt-1+Wfhht-1+Wfxxt+bf)
it＝σ(WicCt-1+Wihht-1+Witxt+bi)
ot＝σ(WocCt+Wohht-1+Woxxt+bo)
Ct＝ftCt-1+it⊙tanh(Wchht-1+Wcxxt+bc)
ht＝ot⊙tanh(Ct)
the above LSTM is simplified to: (C)t,ht)＝LSTM(Ct-1,ht-1,xt) Wherein, the
Figure BDA0002657020910000092
xtFor input at time t, W represents a hidden layer ht-1New cell form CtAnd old cell form Ct-1B represents bias terms for the respective gates and cell states.
See fig. 4. The recurrent neural network module includes: embedding mapping layer EtThe input layer of the fault event sequence and the time sequence input layer acquire the fault event sequence and the time sequence data from the database and then input the data into two LSTMs of the time sequence LSTM and the fault event sequence LSTM, and the two LSTMs are connected to one embedded mapping layer EtMapping layer EtInformation from the two LSTMs is fused and sent to a fault prediction module, and a network point process intensity function is calculated through the following calculation formula:
Figure BDA0002657020910000093
Figure BDA0002657020910000094
Figure BDA0002657020910000095
wherein the content of the first and second substances,
Figure BDA0002657020910000096
indicating the state of the cell at event e,
Figure BDA0002657020910000097
indicating the state of the cell at the time stamp s, EtRepresents an embedded mapping layer, WeThe weight is represented by a weight that is,
Figure BDA0002657020910000098
and
Figure BDA0002657020910000099
output signals representing the time-series and event-series concealment layers at time t, respectively, etInput representing a sequence of fault events, stInput signal representing a time sequence, beRepresenting the bias term.
The fault prediction module comprises a fault time stamp prediction layer, a fault event main type prediction layer and a fault event subtype prediction layer which are sequentially connected, wherein the fault time prediction layer, the fault event main type prediction layer, the fault event subtype prediction layer and the fault time stamp prediction layer respectively output a predicted main type, a predicted fault event subtype and a fault event related time stamp, an input signal propagates to a hidden layer and an output layer by layer after passing through an action function from the input layer, and the fault prediction module finally outputs the fault event subtype without a layered structure in the fault event prediction layer after obtaining the output of an embedded mapping layer. The main calculation formula is as follows:
Mt＝sotfMax(WMEt+bM)
mt＝softMax(Wm[Et,Mt]+bm)
yt＝WyEt+by
wherein M istAnd mtIndicating the main type and subtype, y, respectively, of the fault event at time ttA timestamp associated with each fault event is indicated. softMax denotes the softMax function, WM、WmAnd WyWeights representing fault main type, fault sub type and fault time, respectively, bM、bmAnd byRespectively representing a fault main type, a fault sub-type and a faultOffset term of barrier time, EtRepresenting the output of the embedded layer at time t.
The system adopts an adam (adaptive mobility estimation) method of the Bert version to calculate the minimum value of an objective function, and the objective function is as follows:
Figure BDA0002657020910000101
wherein N is the number of training samples,
Figure BDA0002657020910000102
and
Figure BDA0002657020910000103
respectively representing the weights of the fault main type and the fault subtype of sample j,
Figure BDA0002657020910000104
and
Figure BDA0002657020910000105
respectively representing a sample j fault main type and a fault sub type,
Figure BDA0002657020910000106
is the time stamp of the current event and,
Figure BDA0002657020910000107
information representing historical events.
In order to classify the event types more accurately and improve that the corresponding failure event timestamp should be close to true, the system gives the actual timestamp of sample j
Figure BDA0002657020910000108
In the case of (2), calculating an output from the failure timestamp prediction layer
Figure BDA0002657020910000109
The formula is calculated by adopting a Gaussian penalty function as follows:
Figure BDA00026570209100001010
wherein σ2＝10，
Figure BDA00026570209100001011
Information that is representative of a historical event and,
Figure BDA00026570209100001012
representing the actual time to failure of sample j,
Figure BDA00026570209100001013
is a predicted time to failure result.
For fault event prediction, the usual classification metrics are used, while the result of the time to fault prediction uses Mean Absolute Error (MAE) to measure the absolute difference between the predicted and actual points in time. Finally, the predicted failure time stamp and the prediction result of the failure type are stored in a database.
Examples
When the system carries out fault diagnosis and prediction work, the method comprises the following steps:
firstly, extracting data collected by the equipment from a system database, filtering out other unnecessary information, and leaving the ID number of the equipment, the fault or normal time of the equipment and the fault type of the equipment. And after reading, converting the fault and normal time of the equipment into a time stamp, identifying the time at a certain moment, and carrying out digital coding according to different equipment faults or normal types. If the equipment has a fault, the result is marked as 1, if the equipment is normal, the result is marked as 0, and if other fault subtypes occur, the coding is continued from 2, such as locking, retaining and the like. Wherein, the first column represents the ID number of the equipment, the second column represents the time stamp of the equipment failure or normal, and the third column represents the failure type of the equipment under the time stamp; each row is a timestamp indicating that a device failed under a certain failure type, and the device ID, the device failure timestamp, and the failure event type are stored in a database, as shown in the following table:
TABLE 1 time to failure
Figure BDA0002657020910000111
In a second step, the data is entered into a time series and a fault event series, respectively. The LSTM representation sequence is adopted in the system, and other neural network methods can be replaced. After being computed separately by time series LSTM and fault event series LSTM, they are fused and connected together to an embedded mapping layer. A single layer LSTM of size 32 is provided in the present system, with Sigmoid and tanh used for the hidden representation. The embedding map layer is fully connected, uses the tanh function and outputs a 16-dimensional vector. For the time series RNN, the present system sets the length of each sub-window (i.e., evenly spaced time intervals) to 10 days, which for fault event dependencies can be any length. Acquiring weight information of each interlayer of the neural network by training the LSTM, and acquiring output information of two currently input LSTM sequences based on the trained interlayer weight information;
and thirdly, by embedding the output value of the mapping layer, the main type of the output fault of the prediction layer, the subtype of the output fault event and the associated fault timestamp are used, and the type of the output fault event and the result of the output fault timestamp are obtained and stored in a database.
While the invention has been described with reference to specific embodiments, any feature disclosed in this specification may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise; all of the disclosed features, or all of the method or process steps, may be combined in any combination, except mutually exclusive features and/or steps.

Claims (10)

1. An electronic device fault diagnosis system comprising: the data preprocessing module, the point process circulation neural network module and the failure prediction module that connect in series in order between two databases, its characterized in that: the data preprocessing module extracts acquired data from the first database, filters, screens and reads the ID number of the electronic equipment, the fault or normal time of the electronic equipment and the fault type of the electronic equipment, and converts the fault and normal time of the electronic equipment into a timestamp; the point process cyclic neural network module interprets the condition intensity function of the point process as nonlinear mapping, data preprocessed by the data preprocessing module is input into a time sequence and a fault event sequence and is sent into a network of an embedded mapping layer fusing information from two long-short term memory neural networks LSTM network, the LSTM adopts the time sequence with the length of T at uniform intervals to simulate the background intensity of the occurrence of the fault event, and learns the long-term event dependency relationship by using the fault event sequence with the length of N to train and recognize to obtain a test result, the time sequence and the fault event sequence of the test result are fused into an RNN embedded layer of the cyclic neural network, the RNN is used as a composite neural network for collaborative modeling, the composite neural network prediction model predicts the fault based on the condition intensity function of the point process of the cyclic neural network, and the predicted fault type is output through a prediction layer of the fault prediction module, A sub-type of the failure event and a predicted failure timestamp associated with the failure event, and storing the result of the predicted failure timestamp and the failure type in a second database.
2. The electronic device fault diagnosis system according to claim 1, characterized in that: the point process recurrent neural network module takes two recurrent neural networks RNNs as a composite neural network for collaborative modeling, wherein one recurrent neural network RNN models a time sequence and is used for representing real-time continuously updated dynamic variables, and the other recurrent neural network RNN models a fault event sequence and is used for representing irregularly occurring time; the point process cyclic neural network module inputs data into the two long-short term memory neural networks LSTM networks, and the two long-short term memory neural networks LSTM networks are connected together to an embedded mapping layer which fuses information from the two LSTM networks.
3. The electronic device fault diagnosis system according to claim 1, characterized in that: the failure prediction module comprises a failure time prediction layer and a failure event prediction layer, after the output of the embedded layer is finished, the prediction layer is used for outputting the predicted main type, the sub-type of the failure event and the time stamp related to the failure event, for the failure event prediction, the failure event prediction layer uses the classification measurement standard and the Mean Absolute Error (MAE) to measure the absolute difference value between the predicted time point and the actual time point, and the result of the failure time prediction and the result of the failure type are stored in a second database.
4. The electronic device fault diagnosis system according to claim 1, characterized in that: in the collaborative modeling of the time sequence and the fault event sequence, the fault event type, the electronic equipment ID and the fault timestamp are simultaneously stored in a first database by using the electronic equipment time identifier and the coding identifier of the fault type of the electronic equipment, the first column of the fault event sequence of the recurrent neural network RNN in the t-time axis represents the ID number of the electronic equipment, the second column represents the time stamp of the fault or normal occurrence of the electronic equipment, and the third column represents the fault type of the electronic equipment occurring under the time stamp; each row is a time stamp indicating that a certain electronic device has failed under a certain failure type, and the time sequence of the occurrence of the failure event is used
Figure FDA0002657020900000011
To learn the dependency of long-term fault events.
5. The electronic device fault diagnosis system according to claim 1, characterized in that: the LSTM neural network is provided with an input gate, an output gate and a forgetting gate, the three control gates are a system allowing information to pass in a selective mode, the network is allowed to remember information and control a storage unit for storing information in the LSTM, the edge of the storage unit consisting of a Sigmoid activation function and point-by-point multiplication and the edge of the other part of the neural network are set with a weight, the output gate learns when the activation state is transmitted out of the storage unit through training, and when the output value of the output gate is 1, the other part of the neural network writes the content into the storage unit; when the forget door is opened, the weight of the connection is 1, and the storage unit writes the content into the storage unit; when the forgetting gate output is 0, the storage unit deletes the previous content.
6. The electronic device fault diagnosis system according to claim 1, characterized in that: the long-short term memory network LSTM of the time sequence and the fault event sequence is a time Recursive Neural Network (RNN), the single hidden layer neural network is provided with a hidden layer between an input and an output, namely the output of the input layer is the input of the hidden layer, the product of the output of the hidden layer and the corresponding weight is the input of the output layer, and the output of the output layer is the final output.
7. The electronic device fault diagnosis system according to claim 1, characterized in that: old cell state C at t-1 input of long-short term memory network LSTMt-1Transmitting Sigmoid activation function sigma vector to forgetting gate f through gate control circulation unit circulation nodetAnd old cell state Ct-1Discarding part of information by multiplying old cell state point by point, then multiplying the output value of the tanh layer and the Sigmoid layer, and adding the vector sum needing to be updated
Figure FDA0002657020900000021
Obtaining a new cell state Ct(ii) a Forget door ftReading previous sequence hidden layer output ht-1Vector x from the input of the present sequence at time ttAnd cell state Ct-1Obtaining the content to be discarded and retained in the cell state of the previous layer through a Sigmoid activation function, and inputting a gate itCreating a new candidate vector at the tanh level
Figure FDA0002657020900000022
Forget door ftAnd old cell state Ct-1Multiply and then add the data to be updated
Figure FDA0002657020900000023
Obtaining a new cell state Ct(ii) a Outputting h through previous sequence of hidden layerst-1And the vector of the input of the present sequence at time txtThe summed vector sum passes through the sigma layer through the output gate otMultiplying the output value to the output of the tanh layer to play a role of scaling, and outputting the hidden layer h at the time ttA value between 0 and 1, wherein 0 represents that no amount is allowed to pass, and 1 represents that any amount is allowed to pass
8. The electronic device fault diagnosis system according to claim 7, characterized in that: long-short term memory network LSTM calculates forgetting gate f by following calculation formulatAnd input gate itAnd an output gate otCell state CtHidden layer output ht：
ft＝σ(WfcCt-1+Wfhht-1+Wfxxt+bf)
it＝σ(WicCt-1+Wihht-1+Witxt+bi)
ot＝σ(WocCt+Wohht-1+Woxxt+bo)
Ct＝ftCt-1+it⊙tanh(Wchht-1+Wcxxt+bc)
ht＝ot⊙tanh(Ct)
Figure FDA0002657020900000031
The long-short term memory network LSTM is simplified as follows: (C)t,ht)＝LSTM(Ct-1,ht-1,xt)
Where σ is Sigmoid function
Figure FDA0002657020900000032
ht-1W represents ht-1，CtTo a new cell state, Ct-1Is old cell state,. indicates element-by-element multiplication, vector xtAnd can learnB represents bias terms for the respective gate and cell states.
9. The electronic device fault diagnosis system according to claim 1, characterized in that: the recurrent neural network module includes: embedding mapping layer EtThe input layer of the fault event sequence and the time sequence input layer acquire the fault event sequence and the time sequence data from the database and then input the data into two LSTMs of the time sequence LSTM and the fault event sequence LSTM, and the two LSTMs are connected to one embedded mapping layer EtMapping layer EtInformation from the two LSTMs is fused and sent to a fault prediction module, and a network point process intensity function is calculated through the following calculation formula:
Figure FDA0002657020900000033
Figure FDA0002657020900000034
Figure FDA0002657020900000035
wherein the content of the first and second substances,
Figure FDA0002657020900000036
indicating the state of the cell at event e,
Figure FDA0002657020900000037
indicating the state of the cell at the time stamp s, EtRepresents an embedded mapping layer, WeThe weight is represented by a weight that is,
Figure FDA0002657020900000038
and
Figure FDA0002657020900000039
output signals representing the time-series and event-series concealment layers at time t, respectively, etInput representing a sequence of fault events, stInput signal representing a time sequence, beRepresenting the bias term.
10. The electronic device fault diagnosis system according to claim 1, characterized in that: the fault prediction module comprises a fault time stamp prediction layer, a fault event main type prediction layer and a fault event subtype prediction layer which are sequentially connected, wherein the fault time prediction layer, the fault event main type prediction layer, the fault event subtype prediction layer and the fault time stamp prediction layer respectively output a predicted main type, a predicted fault event subtype and a fault event related time stamp, an input signal propagates to a hidden layer and an output layer by layer after passing through an action function from the input layer, and the fault prediction module finally outputs the fault event subtype without a layered structure in the fault event prediction layer after obtaining the output of an embedded mapping layer. The main calculation formula is as follows:
Mt＝sotfMax(WMEt+bM)
mt＝softMax(Wm[Et,Mt]+bm)
yt＝WyEt+by
wherein M istAnd mtIndicating the main type and subtype, y, respectively, of the fault event at time ttA timestamp associated with each fault event is indicated. softMax denotes the softMax function, WM、WmAnd WyWeights representing fault main type, fault sub type and fault time, respectively, bM、bmAnd byBias terms representing fault main type, fault sub type and fault time, respectively, EtRepresenting the output of the embedded layer at time t.
CN202010891078.2A 2020-08-30 2020-08-30 Electronic equipment fault diagnosis system Pending CN112101431A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010891078.2A CN112101431A (en) 2020-08-30 2020-08-30 Electronic equipment fault diagnosis system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010891078.2A CN112101431A (en) 2020-08-30 2020-08-30 Electronic equipment fault diagnosis system

Publications (1)

Publication Number Publication Date
CN112101431A true CN112101431A (en) 2020-12-18

Family

ID=73756645

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010891078.2A Pending CN112101431A (en) 2020-08-30 2020-08-30 Electronic equipment fault diagnosis system

Country Status (1)

Country Link
CN (1) CN112101431A (en)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112818035A (en) * 2021-01-29 2021-05-18 湖北工业大学 Network fault prediction method, terminal equipment and storage medium
CN112948163A (en) * 2021-03-26 2021-06-11 中国航空无线电电子研究所 Method for evaluating influence of equipment on functional fault based on BP neural network
CN112988843A (en) * 2021-03-26 2021-06-18 桂林电子科技大学 SMT chip mounter fault management and diagnosis system based on SQL Server database
CN113537360A (en) * 2021-07-19 2021-10-22 中国人民解放军国防科技大学 Point-to-point classification fault detection method based on deep learning
CN113608163A (en) * 2021-09-10 2021-11-05 天目数据(福建)科技有限公司 Ammeter fault diagnosis method and device of stacked cyclic neural network
CN114564000A (en) * 2022-03-01 2022-05-31 西北工业大学 Active fault tolerance method and system based on fault diagnosis of intelligent aircraft actuator
CN116032359A (en) * 2022-12-27 2023-04-28 中国联合网络通信集团有限公司 Characteristic network data prediction method and system and electronic equipment
CN116303648A (en) * 2023-03-21 2023-06-23 国网山东省电力公司莱西市供电公司 Cable early warning method based on digital twinning
CN116467645A (en) * 2023-06-20 2023-07-21 中通服建设有限公司 Pollution source collection monitoring system
CN117131457A (en) * 2023-10-26 2023-11-28 杭州海兴泽科信息技术有限公司 AI model-based electric power big data acquisition and processing method and system

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108255656A (en) * 2018-02-28 2018-07-06 湖州师范学院 A kind of fault detection method applied to batch process
CN109033450A (en) * 2018-08-22 2018-12-18 太原理工大学 Lift facility failure prediction method based on deep learning
CN110580213A (en) * 2019-09-16 2019-12-17 浪潮软件股份有限公司 Database anomaly detection method based on cyclic marking time point process
CN110598851A (en) * 2019-08-29 2019-12-20 北京航空航天大学合肥创新研究院 Time series data abnormity detection method fusing LSTM and GAN
CN110781266A (en) * 2019-09-16 2020-02-11 北京航空航天大学 Urban perception data processing method based on time-space causal relationship
CN112083244A (en) * 2020-08-30 2020-12-15 西南电子技术研究所（中国电子科技集团公司第十研究所） Integrated avionics equipment fault intelligent diagnosis system
US20220100624A1 (en) * 2020-09-25 2022-03-31 Wuhan University Method and system of identifying and estimating complex analog circuit failure

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108255656A (en) * 2018-02-28 2018-07-06 湖州师范学院 A kind of fault detection method applied to batch process
CN109033450A (en) * 2018-08-22 2018-12-18 太原理工大学 Lift facility failure prediction method based on deep learning
CN110598851A (en) * 2019-08-29 2019-12-20 北京航空航天大学合肥创新研究院 Time series data abnormity detection method fusing LSTM and GAN
CN110580213A (en) * 2019-09-16 2019-12-17 浪潮软件股份有限公司 Database anomaly detection method based on cyclic marking time point process
CN110781266A (en) * 2019-09-16 2020-02-11 北京航空航天大学 Urban perception data processing method based on time-space causal relationship
CN112083244A (en) * 2020-08-30 2020-12-15 西南电子技术研究所（中国电子科技集团公司第十研究所） Integrated avionics equipment fault intelligent diagnosis system
US20220100624A1 (en) * 2020-09-25 2022-03-31 Wuhan University Method and system of identifying and estimating complex analog circuit failure

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
NOLABEL: "异常监测②——lstm时间序列预测&lstm简易原理", 《HTTPS://BLOG.CSDN.NET/QQ_33936417/ARTICLE/DETAILS/104062271》 *
庄雨璇: "基于深度学习的旋转轴承端到端故障诊断研究", 《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》 *
张祥: "基于LSTM和动态模型的化工过程混合故障诊断", 《中国优秀硕士学位论文全文数据库 (工程科技Ⅰ辑)》 *
池永为等: "基于LSTM-RNN的滚动轴承故障多标签分类方法", 《振动、测试与诊断》 *
王鑫等: "基于LSTM循环神经网络的故障时间序列预测", 《北京航空航天大学学报》 *
陈文豪等: "基于小波散射协同BiLSTM的输电线路故障诊断", 《国外电子测量技术》 *

Cited By (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112818035A (en) * 2021-01-29 2021-05-18 湖北工业大学 Network fault prediction method, terminal equipment and storage medium
CN112818035B (en) * 2021-01-29 2022-05-17 湖北工业大学 Network fault prediction method, terminal equipment and storage medium
CN112948163A (en) * 2021-03-26 2021-06-11 中国航空无线电电子研究所 Method for evaluating influence of equipment on functional fault based on BP neural network
CN112988843A (en) * 2021-03-26 2021-06-18 桂林电子科技大学 SMT chip mounter fault management and diagnosis system based on SQL Server database
CN112988843B (en) * 2021-03-26 2022-05-24 桂林电子科技大学 SMT chip mounter fault management and diagnosis system based on SQL Server database
CN112948163B (en) * 2021-03-26 2023-09-19 中国航空无线电电子研究所 Method for evaluating influence of equipment on functional failure based on BP neural network
CN113537360B (en) * 2021-07-19 2023-02-03 中国人民解放军国防科技大学 Point-to-point classification fault detection method based on deep learning
CN113537360A (en) * 2021-07-19 2021-10-22 中国人民解放军国防科技大学 Point-to-point classification fault detection method based on deep learning
CN113608163A (en) * 2021-09-10 2021-11-05 天目数据(福建)科技有限公司 Ammeter fault diagnosis method and device of stacked cyclic neural network
CN114564000A (en) * 2022-03-01 2022-05-31 西北工业大学 Active fault tolerance method and system based on fault diagnosis of intelligent aircraft actuator
CN114564000B (en) * 2022-03-01 2024-03-08 西北工业大学 Active fault tolerance method and system based on intelligent aircraft actuator fault diagnosis
CN116032359A (en) * 2022-12-27 2023-04-28 中国联合网络通信集团有限公司 Characteristic network data prediction method and system and electronic equipment
CN116303648A (en) * 2023-03-21 2023-06-23 国网山东省电力公司莱西市供电公司 Cable early warning method based on digital twinning
CN116303648B (en) * 2023-03-21 2023-10-24 国网山东省电力公司莱西市供电公司 Cable early warning method based on digital twinning
CN116467645A (en) * 2023-06-20 2023-07-21 中通服建设有限公司 Pollution source collection monitoring system
CN116467645B (en) * 2023-06-20 2023-09-19 中通服建设有限公司 Pollution source collection monitoring system
CN117131457A (en) * 2023-10-26 2023-11-28 杭州海兴泽科信息技术有限公司 AI model-based electric power big data acquisition and processing method and system
CN117131457B (en) * 2023-10-26 2024-01-26 杭州海兴泽科信息技术有限公司 AI model-based electric power big data acquisition and processing method and system

Similar Documents

Publication Publication Date Title
CN112101431A (en) Electronic equipment fault diagnosis system
Chen et al. Health indicator construction of machinery based on end-to-end trainable convolution recurrent neural networks
Frank et al. New developments using AI in fault diagnosis
CN111274737A (en) Method and system for predicting remaining service life of mechanical equipment
Javadpour et al. A fuzzy neural network approach to machine condition monitoring
Lo et al. Review of machine learning approaches in fault diagnosis applied to IoT systems
CN112131212A (en) Hybrid cloud scene-oriented time sequence data anomaly prediction method based on ensemble learning technology
CN107657250B (en) Bearing fault detection and positioning method and detection and positioning model implementation system and method
CN105825271B (en) Satellite failure diagnosis and prediction method based on evidential reasoning
CN112083244B (en) Integrated intelligent diagnosis system for faults of avionic equipment
CN112487058A (en) Numerical control machine tool fault monitoring and diagnosing system based on data mining
Xu et al. A novel health indicator for intelligent prediction of rolling bearing remaining useful life based on unsupervised learning model
Son et al. Deep learning-based anomaly detection to classify inaccurate data and damaged condition of a cable-stayed bridge
CN114519923A (en) Intelligent diagnosis and early warning method and system for power plant
CN116520806A (en) Intelligent fault diagnosis system and method for industrial system
CN116611523B (en) Method and system for predicting interpretable faults of turbofan engine
Bond et al. A hybrid learning approach to prognostics and health management applied to military ground vehicles using time-series and maintenance event data
CN116628869A (en) Analysis method for fault propagation mechanism of numerical control machine tool based on transfer entropy theory
CN114779739A (en) Fault monitoring method for industrial process under cloud edge end cooperation based on probability map model
Wang et al. Complex equipment diagnostic reasoning based on neural network algorithm
Cordier et al. Diagnosis and supervision: model-based approaches
Dash et al. A Comparison of Model-Based and Machine Learning Techniques for Fault Diagnosis
Frank et al. New developments using AI in fault diagnosis
Saddem et al. Machine learning-based approach for online fault Diagnosis of Discrete Event System
Hekmat et al. Real time fault detection and isolation: a comparative study

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WD01 Invention patent application deemed withdrawn after publication
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20201218