CN117591949A - Equipment abnormality identification method, equipment and medium - Google Patents

Abstract

The embodiment of the specification discloses a method, a device and a medium for identifying device abnormality, comprising: acquiring a real-time health state key characterization parameter of the appointed equipment; acquiring operation data of the appointed equipment; inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment; performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment; and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

Description

Equipment abnormality identification method, equipment and medium

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to a method, an apparatus, and a medium for identifying an abnormality of an apparatus.

Background

With the development and application of technology, equipment abnormality recognition of manufacturing enterprises has been greatly progressed in recent years. However, for most manufacturing enterprises, failure data acquisition is difficult, so that the construction of the prediction model lacks corresponding data support. In this case, the manufacturing company cannot accurately recognize the abnormality of the apparatus.

Disclosure of Invention

One or more embodiments of the present disclosure provide a method, an apparatus, and a medium for identifying an abnormality of an apparatus, which are used to solve the technical problem set forth in the background art.

One or more embodiments of the present disclosure adopt the following technical solutions:

one or more embodiments of the present disclosure provide a device anomaly identification method, including:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

One or more embodiments of the present specification provide an apparatus abnormality recognition apparatus including:

at least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

One or more embodiments of the present description provide a non-volatile computer storage medium storing computer-executable instructions that, when executed by a computer, enable:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

The above-mentioned at least one technical scheme that this description embodiment adopted can reach following beneficial effect:

according to the embodiment of the specification, the state of the equipment can be known in time by acquiring the real-time health state key characterization parameters, the problem of difficult data acquisition can be solved by using the prediction model, the predicted health state key characterization parameters of the equipment are obtained through the prediction model, and the possible abnormality is found out by comparing the difference between the real-time health state and the predicted health state. When the residual analysis result exceeds a preset threshold, an indication that an abnormality is likely to exist is made, and by identifying the operation data, it can be further determined which subsystem is abnormal, so that the abnormality is located. In addition, the embodiment of the specification has self-adaptability by setting the preset threshold value, and can adjust the standard for judging the abnormality according to the real-time condition.

Drawings

In order to more clearly illustrate the embodiments of the present description or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described below, it being obvious that the drawings in the following description are only some of the embodiments described in the present description, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art. In the drawings:

FIG. 1 is a schematic flow diagram of a method for identifying device anomalies according to one or more embodiments of the present disclosure;

FIG. 2 is a flow diagram of a method for identifying device anomalies provided in one or more embodiments of the present disclosure;

FIG. 3 is a flow diagram of data dimension reduction provided in one or more embodiments of the present disclosure;

FIG. 4 is a flow diagram of a deep learning algorithm for constructing a predictive model provided in one or more embodiments of the present disclosure;

fig. 5 is a schematic structural diagram of an apparatus abnormality recognition apparatus according to one or more embodiments of the present disclosure.

Detailed Description

The embodiment of the specification provides a device abnormality identification method, device and medium.

With the development and application of technologies such as sensors, the internet of things, "internet+", cloud computing, artificial intelligence and the like, data-driven fault prediction and health management technologies have been greatly developed in recent years. Through advanced sensor technology, obtain operation data that temperature, noise, voltage, electric current, vibration etc. are closely correlated with equipment health state, realize the collection storage to relevant data through the thing networking. The method comprises the steps of evaluating the running state of equipment by combining technologies and methods such as signal processing, feature extraction, data mining, machine learning and the like, monitoring early faults of the equipment, qualitatively or quantitatively evaluating the fault degree, revealing the degradation rule of equipment performance, predicting the health state and the residual service life of the equipment at the future moment, and realizing self-adaptive fault-tolerant control of the equipment by combining spare part inventory management information such as fault cost analysis, purchasing, storing and the like on the basis of historical running information, maintenance record and future predicted use condition of the equipment, improving resource management efficiency and optimizing running maintenance strategy of the equipment.

In this context, there have been advances in the field of early abnormality identification of devices, but some drawbacks remain. Related scientific research institutions and production equipment providers can artificially manufacture various types of faults by arranging sensors as many as possible, acquire a large number of complete data of various types under various operating conditions through long-term accumulation, and train corresponding prediction models based on the complete data. But for most common manufacturing enterprises, the following real problems still exist:

1. the placement of the sensors increases costs and may also affect the production process, making it difficult to obtain a complete data set.

2. For different kinds of equipment, the data acquisition terminal configuration, the data transmission format, the model training method, the abnormality judgment basis and the like lack of standardization and normalization, so that the universality and the mobility of the abnormality identification method are limited.

3. In the production process, fault data only occupy a very small part, and the phenomenon of data unbalance is serious, so that the prediction model obtained by the traditional training mode has poor generalization capability and inaccurate performance evaluation, and the phenomenon of fitting is easy to occur.

There are three main implementation paths for current fault prediction and health management. The fault prediction technology based on the model has the advantages of being capable of penetrating into the nature of the object system essence and realizing real-time fault prediction by integrating a physical model and a random process modeling, but in practice, a model capable of accurately describing the complex system behavior cannot be constructed; the fault prediction technology based on the statistical reliability performs fault prediction from the statistical characteristic angle of past fault history data, and is easy to realize but lower in accuracy; the fault prediction technology based on data driving is based on collected data, and the implicit information in the collected data is mined through various data analysis and processing methods to perform prediction operation, so that the defects of the fault prediction technology based on models and statistical reliability are overcome, and the fault prediction method based on the model and statistical reliability becomes a practical fault prediction method.

However, for most manufacturers planning to introduce fault prediction and health management systems, the problems of difficult data acquisition, small fault data occupation ratio and data unbalance are prominent, and the construction of a prediction model lacks data support.

In order to better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be clearly and completely described below with reference to the drawings in the embodiments of the present specification, and it is obvious that the described embodiments are only one subsystem embodiment of the present specification, not all embodiments. All other embodiments, which can be made by one of ordinary skill in the art based on the embodiments herein without making any inventive effort, shall fall within the scope of the present disclosure.

Fig. 1 is a schematic flow diagram of a device anomaly identification method according to one or more embodiments of the present disclosure, where the flow may be executed by a device anomaly identification system. Some input parameters or intermediate results in the flow allow for manual intervention adjustments to help improve accuracy.

The method flow steps of the embodiment of the present specification are as follows:

s102, acquiring real-time health state key characterization parameters of the appointed equipment.

In embodiments of the present description, real-time health status key characterization parameters of a specified device may be used to monitor the health status of the specified device. The key characterization parameters of the health state of the specified equipment can be determined according to the type of the specified equipment, and also can be determined in a pre-measurement mode. When the real-time health state key characterization parameters of the specified equipment are obtained, corresponding equipment can be selected for obtaining according to the type of the health state key characterization parameters, for example, a temperature sensor can be used for obtaining when the health state key characterization parameters are temperatures, a vibration sensor can be used for measuring the vibration of an object when the health state key characterization parameters are amplitudes, and therefore vibration signals can be obtained.

Further, before acquiring the real-time health state key characterization parameters of the specified equipment, the analysis result of the specified equipment can be obtained by performing fault tree analysis and fault mode influence analysis on the specified equipment: determining the hazard degree and characterization parameters of each fault cause in the designated equipment according to the analysis result; and finally, determining the key characterization parameters of the health state of the appointed equipment according to the damage degree and the characterization parameters of each fault cause.

It should be noted that, the key health status characterization parameter of the specified device may represent the health status of the entire specified device. Fault tree analysis (Fault Tree Analysis, FTA) may be used to identify fundamental events and logical relationships between them that may lead to device failure, and fault pattern impact analysis (Failure Mode Effect and Criticality Analysis, FMECA) may be used to evaluate the impact of various fault patterns on device performance and safety

The analysis results may include a fault tree analysis result and a fault pattern influence analysis result, and the fault tree analysis result may include a relationship between an event tree and a basic event for analyzing a possibility of occurrence of a fault. The failure mode influence analysis result may include information such as influence degree, hazard degree, etc. of each failure mode.

According to the analysis result, when determining the hazard degree and the characterization parameter of each fault cause in the designated equipment, summarizing the analysis result of the fault tree, wherein the analysis result can comprise the relation between an event tree and basic events and the fault cause related to each basic event; the results of the fault mode influence analysis are summarized, and information such as influence degree, hazard degree and the like of each fault mode can be included. And making standards or indexes for evaluating the damage degree so as to quantitatively evaluate different fault reasons. Quantitative indicators such as percentage of the extent of influence, cost of loss, etc. may be considered for use. And evaluating the hazard degree according to the formulated evaluation standard aiming at each fault cause. Factors such as the probability of the cause of the failure, the degree of influence on the performance of the device, etc. can be considered. And determining the characterization parameters related to each fault cause. The characterization parameters can be obtained through monitoring and measurement to reflect the influence of fault reasons on the health state of the equipment. Finally, a mapping relationship between the fault causes and the characterization parameters can be established, and the mapping relationship can be a table or a mapping diagram, so that each fault cause and the corresponding characterization parameters are clearly displayed.

Further, when the key characterization parameters of the health state of the specified equipment are determined, the correlation of the hazard degree evaluation and the characterization parameters can be integrated, and the key characterization parameters of the health state of the specified equipment can be determined. The characterization parameters should be available in real-time monitoring, such as amplitude, temperature, current, etc.

It should be noted that, the damage degree evaluation criteria established previously may be used to quantitatively evaluate each cause of the fault, which may be a quantitative score, percentage or other quantifiable index. For each failure cause, the most critical characterization parameters in terms of their hazard level are identified. The importance of the characterization parameters is considered, as well as the sensitivity of these parameters to health effects when a fault cause occurs. And then the key characterization parameters are ranked according to the importance of the key characterization parameters in the hazard degree evaluation, so as to form a priority list. Prioritization facilitates centralized resource monitoring and collection of parameters that are most informative, and one or more health-state-critical characterization parameters that are highest in priority may represent the health state of the designated device as a whole. Furthermore, the combined effect between different characterization parameters may also be considered, sometimes a common variation of multiple parameters may indicate a more serious health problem. It is contemplated that some statistical methods or models may be employed to integrate the plurality of characterization parameters.

S104, acquiring the operation data of the designated equipment.

In the embodiment of the specification, the operation data can be divided into two types, such as temperature, voltage, current and the like, and the instantaneous value of the sensing terminal can be directly collected as a subsequent analysis basis, so that the data is generally strong in real-time performance and can be used for monitoring the basic operation state of the equipment in real time; the other type of the data processing method is vibration, noise signals and the like, the data in a certain period of time are required to be acquired at a designated acquisition frequency, an automatic data processing tool is constructed in combination with actual requirements, characteristics of amplitude, frequency, variance, mean value, kurtosis and the like of original signals are extracted, and the data processing aim is to convert complex original signals into characteristics with actual analysis value, and the characteristics can help understand the health condition of equipment operation.

In addition, the collected and processed sensor data can be transmitted through the Internet of things and stored in a time sequence database.

S106, inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the predicted health state key characterization parameters of the appointed equipment.

In the embodiment of the present disclosure, data in a time sequence database may be read first, and then a prediction model may be constructed by using a deep learning algorithm, which specifically includes the following steps:

(1) Neural network structural design

The neural network structure is designed by selecting the hypothesis space of the model, namely the relation set which can be expressed by the model. The neural network architecture includes the number of layers of the network, the number of neurons per layer, and the choice of activation functions. The structural design may take into account that the model is able to capture key features of the device's health status.

(2) Training configuration

The find solution algorithm, i.e., the loss function, used by the set model may be gradient descent, adam, etc., and specifies computing resources, including GPU acceleration, etc., to accelerate the training process.

(3) Data reading and processing

And reading the required data from the time sequence database, checking the data set and converting the format to adapt to the input requirement of the neural network, dividing the data into a training set and a testing set, and generally dividing the training set and the testing set according to a certain proportion to evaluate the generalization capability of the model.

(4) Model training

The training set is read, the training process is circularly invoked in batches, the data of each batch is ensured to be used for parameter updating of the model, and each round comprises three steps of forward calculation (prediction of the model), loss function and backward propagation (gradient calculation and parameter updating).

(5) Training process analysis

And drawing a loss change trend in the training process so as to observe the change of the model performance in the training process, and analyzing the loss change trend to know whether the model is converged or not and whether adjustment is needed or not.

(6) Model evaluation

And (3) calling the test set test model prediction accuracy, analyzing the evaluation model in combination with the training process, and when the requirements are not met, adjusting the neural network structure and the training parameters, and executing training, analysis and evaluation again until the requirements are met.

(7) Model preservation

The trained model is saved in a fixed format for recall in performing the predictive task.

Further, in the data stored in the time series database, some features are necessarily without effective information (such as signal noise), or some features are repeatedly used (such as some features may be linearly related), and the model construction process can be optimized through data dimension reduction. Meanwhile, the additional arrangement of sensors can be reduced as much as possible by using equipment, and the on-site implementation is facilitated.

Before the operation data of the specified equipment is input into a pre-trained prediction model, the embodiment of the specification can determine a characteristic subset corresponding to the operation data of the specified equipment according to a preset analysis mode, namely, the data of a time-lapse database can be traversed, and the characteristic subset is split by calculating F statistics, pearson correlation coefficients and distance correlation coefficients and combining a recursive characteristic elimination tool; inputting the characteristic subset into a pre-trained dimension reduction model, outputting operation data of the specified equipment after dimension reduction, reducing the dimension of the characteristic subset through Principal Component Analysis (PCA) or t-distribution neighborhood embedding (t-SNE), and inputting the operation data of the specified equipment after dimension reduction into the prediction model.

S108, carrying out residual analysis according to the real-time health state key characterization parameters of the appointed equipment and the predicted health state key characterization parameters of the appointed equipment to obtain a residual analysis result of the appointed equipment.

In the embodiment of the specification, the key characterization parameters of the health state of the specified equipment can be obtained in real time through the corresponding sensors. These parameters may include temperature, pressure, vibration, current, voltage, etc. The predicted health state key characterization parameters can be obtained from the prediction model. For each subsystem, subtracting the real-time health state key characterization parameters from the predicted health state key characterization parameters to calculate a residual error. The residual is the difference between the actual observed state of health parameter and the predicted value.

S110, if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

In the embodiment of the present specification, a preset threshold value may be set for abnormality recognition. This preset threshold is a limitation of the residual, and if the residual exceeds the threshold, the designated device will be identified as anomalous. Anomaly related information is recorded, including a timestamp and a residual value. And the operation data of the fault period can be intercepted for analysis, and the subsystem with the abnormality of the appointed equipment is identified.

If the subsystem with the abnormality of the appointed equipment gives false alarm or fails to give alarm, the false alarm and failure to give alarm of the appointed equipment are recorded, which conditions can include that the system gives false alarm under which conditions and that the system fails to give alarm under which conditions. And carrying out detailed analysis on the false alarm and the missing alarm so as to determine the reason thereof. There may be various reasons for false positives and false negatives, such as data noise, model inaccuracy, unreasonable threshold settings, etc. And according to the reasons of false alarm and missing alarm, making a self-adaptive adjustment strategy. The adaptive adjustment strategy may include the following aspects:

dynamic threshold: and dynamically adjusting the threshold according to the change and the historical performance of the real-time data.

Model improvement: the prediction model is updated or improved to increase accuracy and robustness.

Data cleaning: noise in the data is reduced, and the quality of the input data is ensured.

If the reason for false alarm or missing alarm is that the threshold is set unreasonably, the adjustment policy may include: and carrying out adaptive adjustment on the preset threshold value.

If the cause of the false alarm or the missing alarm is the data noise, the adjustment strategy may include: and replacing the operation data corresponding to the false report or the missing report in the data set, and retraining the prediction model through the updated data set.

If the cause of the false alarm or the missing alarm is the inaccuracy of the model, the adjustment strategy may include: acquiring historical operation data of the appointed equipment, and determining failure rate according to the historical operation data; and determining the update period of the prediction model according to the failure rate and the actual running condition of the appointed equipment, so that the training prediction model is updated according to the latest running data in each update period.

After the adaptive adjustment is implemented, testing and verification can be performed to ensure false positives and false negatives are reduced. Once the adaptive adjustment is in effect, a continuous monitoring mechanism can be established to track the performance of the anomaly detection system in real time and further adjust as needed.

Further, the health status of the entire designated device is an integration of the health status of the individual subsystems. Interactions and integration between subsystems are considered in order to more accurately assess the health of the overall device. The real-time health state key characterization parameters of the specified equipment in the embodiment of the specification can comprise the real-time health state key characterization parameters of all subsystems of the specified equipment; the operation data of the designated device may include operation data of each subsystem of the designated device; the predicted health state critical characterizing parameters of the specified device may include predicted health state critical characterizing parameters of each subsystem of the specified device.

It should be noted that the real-time health status key characterization parameters include not only parameters of the whole equipment level but also parameters of each subsystem level. The health status of a given device can be integrated from the status of the individual subsystems, and therefore subsystem-level characterization parameters need to be considered. The operation data is not only operation data of the entire apparatus but also operation data of each subsystem. The operating data of the subsystem may include parameters of temperature, pressure, amplitude, etc., which have a significant impact on the health of the subsystem. The predicted state of health key characterization parameters are also subdivided into the layers of the individual subsystems. Predicting the critical characterization parameters of the health status of each subsystem can help to pre-warn in advance and take corresponding maintenance measures.

Further, based on the above analysis, embodiments may include:

acquiring real-time health state key characterization parameters of subsystems of the appointed equipment;

acquiring operation data of each subsystem of the appointed equipment;

inputting the operation data of each subsystem of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of each subsystem of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of each subsystem of the appointed equipment and the predicted health state key characterization parameters of each subsystem of the appointed equipment to obtain residual analysis results of each subsystem of the appointed equipment;

and if the residual analysis result of the specific subsystem of the designated equipment exceeds a preset threshold, identifying the specific subsystem of the designated equipment as the subsystem with the abnormality.

Further, before acquiring the real-time health state key characterization parameters of the specified equipment, analyzing the fault tree and the fault mode influence of each subsystem of the specified equipment to obtain an analysis result of each subsystem of the specified equipment: determining the hazard degree and characterization parameters of each fault cause in each subsystem of the designated equipment according to the analysis result; and determining the key characterization parameters of the health state of each subsystem of the appointed equipment according to the hazard degree and the characterization parameters of each fault cause.

It should be noted that, in the embodiment of the present specification, by introducing residual analysis, the industrial mechanism analysis determines the key characterization parameter of the health state of the device, and uses this as the residual analysis object; simplifying a data acquisition flow and a residual analysis object prediction model construction flow on the premise of ensuring the accuracy of a prediction model by using a dimension reduction algorithm; and carrying out residual analysis on the actual values of the parameters and the predicted values of the model in the running process, determining an initial threshold value through historical data and production experience, adjusting the threshold value according to the actual field in the running process of the system, and carrying out early abnormality judgment on equipment when the residual exceeds the threshold value.

It should be noted that, the embodiment of the specification can identify the early abnormality of the equipment by comparing the actual value of the key characterization parameter of the health state of the equipment with the predicted value of the model based on the healthy operation sample which is relatively easy to obtain by the manufacturing enterprises, so that the data requirements required by the construction of the fault prediction and health management system are reduced, the generalization capability of the early abnormality identification system of the equipment is improved, and the construction feasibility of the system is improved.

The flow chart of the method for identifying the abnormality of the equipment is shown in fig. 2, and the content includes:

1. device industry mechanism analysis

During the operation of the device, the health state of the device can be characterized by parameters of a certain or certain specific parts. Before an early abnormality recognition system of equipment is built, fault tree analysis (Fault Tree Analysis, FTA) and fault mode influence analysis (failure mode effect and criticalityanalysis, FMECA) are carried out from the composition, structure and operation principle of the equipment, so that reasons corresponding to equipment faults and combinations thereof are combed, the influence of the fault modes and the fault reasons on the equipment is quantitatively processed, and the degree of severity of the influence of the fault modes and the fault reasons on the subsystems of the equipment is analyzed, so that the degree of harm of each subsystem on the whole equipment is determined.

In combination with FTA and FMECA, the device industry mechanism analysis was performed based on the following logic:

(i) The fault-characterization parameter is strongly related to whether the device is faulty.

(ii) The model input parameters are strongly related to the cause of the fault.

(iii) The probability that the equipment fails and cannot normally operate due to a certain failure reason j of the subsystem i is recorded asThe probability of damage to subsystem i due to failure cause j is denoted as beta _ij Then->And beta _ij The larger the model input parameters are, the more easily the system faults are caused, and the +.>The extent of damage can be considered.

Therefore, key characterization parameters of the health state of the equipment (the key characterization parameters of the health state) are determined, and initial theoretical basis is provided for design and implementation of a data acquisition system.

2. Data acquisition and processing

Based on industrial mechanism analysis, the sensor terminal is configured and required original data is acquired in combination with field reality. Raw data can be divided into two classes according to its availability. For example, temperature, voltage, current and the like, the instantaneous value of the sensing terminal can be directly obtained as a follow-up analysis basis; for example, vibration and noise signals, data in a certain period of time need to be acquired at a designated acquisition frequency, an automatic data processing tool is constructed in combination with actual requirements, and characteristics such as amplitude, frequency, variance, mean value, kurtosis and the like of an original signal are extracted. The acquired and processed sensor data are transmitted through the Internet of things and stored in a time sequence database.

3. Data dimension reduction

In the data stored in the time series database, some features are necessarily without effective information (such as signal noise), or some features are repeatedly carried by information and other features (such as some features may be linearly related), and the model construction process can be optimized through data dimension reduction. Meanwhile, existing parameters of the equipment can be used as much as possible, additional arrangement of sensors is reduced, and the on-site implementation is facilitated.

The data dimension reduction flow chart is shown in fig. 3, and the steps are as follows:

(1) The traversing device monitors the parameter dataset and splits the feature subset from it by computing the F statistic, pearson correlation coefficient, distance correlation coefficient, in combination with a recursive feature elimination tool.

(2) And training a neural network model according to the characteristic subset by using the training set data.

(3) Evaluating the model by using the test set data, and if the model does not meet the requirements, re-splitting the characteristic subset and performing training evaluation

(4) And outputting the feature subset after the dimension reduction if the requirement is met.

4. Prediction model construction

According to the data dimension reduction result, reading the data in the time sequence database, constructing a prediction model by using a deep learning algorithm, wherein a flow chart of the construction prediction model by using the deep learning algorithm is shown in fig. 4, and the steps are as follows:

(1) Neural network structural design

The neural network structure is designed by selecting the hypothesis space of the model, namely the relation set which can be expressed by the model.

(2) Training configuration

And setting a solution searching algorithm adopted by the model, namely a loss function, and designating computing resources.

(3) Data reading and processing

And reading the required data from the time sequence library, checking the data set and converting the format, and dividing the data into a training set and a testing set.

(4) Model training

The training set is read, the training process is circularly invoked in batches, and each round comprises three steps of forward calculation, loss function (optimization target) and backward propagation.

(5) Training process analysis

And drawing a loss change trend in the training process, and analyzing the training process.

(6) Model evaluation

And (3) calling the test set test model prediction accuracy, analyzing the evaluation model in combination with the training process, and when the requirements are not met, adjusting the neural network structure and the training parameters, and executing training, analysis and evaluation again until the requirements are met.

(7) Model preservation

The trained model is saved in a fixed format and invoked when the predictive task is performed.

5. Residual analysis and early abnormality determination

In the running process of the equipment, key characterization parameter values of the health state of the equipment are monitored in real time, residual analysis is carried out by combining the output values of the prediction model, if the residual error is out of tolerance, fault period signals are intercepted for analysis, early abnormal judgment results of the equipment are given, and the early abnormal judgment results are pushed to personnel with relevant authorities.

6. Model update and optimization

In order to ensure the accuracy of the prediction model and improve the utilization rate of the model, the model is updated and optimized by adopting the following three strategies.

(1) When the equipment early-stage abnormal report missing error report occurs, the residual error threshold value is actually adjusted in combination with the site so as to meet the abnormal identification requirement.

(2) When the abnormal recognition model cannot be perfected by adjusting the residual error threshold, extracting data corresponding to a prediction error period, updating the data set, repeating the prediction model construction process, and optimizing the prediction model.

(3) The distribution or characteristics of the equipment operation data can be gradually changed along with the accumulation of the equipment operation time length, and the data drift phenomenon occurs. Thus, during the period of use of the device from the time of the device being put to a stop, the data set retraining model is updated at regular intervals in combination with its failure rate curve. Meanwhile, the original model is stored in a model library for the subsequent equipment with the same type to be used in the corresponding stage, and guidance is provided for the model training process of the equipment with the same type.

The flowchart of the method for constructing the abnormal recognition system of the gear hobbing machine is shown in fig. 4, and the method comprises the following steps:

1. analysis of gear hobbing machine industrial mechanism

And determining that the Y-direction amplitude of the gear box is a key characterization parameter of the health state of the gear hobbing machine based on equipment information, related academic research data and equipment use process records provided by equipment suppliers and combining fault tree analysis and fault mode influence analysis.

2. Sensor arrangement

According to the analysis result of the industrial mechanism, arranging a temperature sensor and a three-way strain sensor at the large bracket of the hob box; three-way acceleration sensors and temperature sensors are arranged on the side face of the gear box, the side face of the workbench and the X-axis screw rod mounting support; a three-way acceleration sensor is arranged on the Z-axis lead screw mounting support; temperature sensors are arranged on the upper end face and the side face close to the bottom of the lubricating oil tank; a temperature sensor, a three-way strain sensor and a three-way acceleration sensor are arranged in the middle of the left/right lathe bed; a displacement sensor is arranged on the end face of the left/right lathe bed; and a noise sensor and a temperature sensor are arranged at the upper left of the operating platform.

3. Data acquisition and processing

For each sensor data acquisition, 39 pieces of raw data are acquired, wherein the vibration signal and the noise signal acquired by the acceleration sensor and the noise sensor must intercept continuous data for a period of time at a certain frequency to be able to perform subsequent analysis. In combination with field practice, an acquisition strategy was determined that intercepts the signal for 0.5s at a frequency of 10kHz in each acquisition cycle per minute. And constructing a data acquisition program, executing an acquisition strategy, extracting signals containing 39 types of original data, converting the signals into an xlsx-format file, and naming and storing the file by a time stamp. And (3) constructing a data processing program, reading the corresponding file at regular time, processing to obtain a signal average value or amplitude, and pushing the signal average value or amplitude to a time sequence library through a websocket interface.

4. Data dimension reduction

Based on the accumulated data, the data set is simplified using a developed data dimension reduction algorithm. In this embodiment, the gearbox Y-direction amplitude is the predicted value and the other 38 parameters are the predicted model input values. And the input value is reduced to 24 items through a data dimension reduction algorithm, so that the training efficiency of the amplitude prediction model is improved. Meanwhile, when the early abnormality recognition system of the same type or the same type of equipment is constructed, the data dimension reduction result can provide reference for the sensor arrangement stage, and the data acquisition and processing flow is simplified.

5. Amplitude prediction model construction

And training an amplitude prediction model by taking the Y-direction amplitude of the gear box of the gear hobbing machine as a model output value and taking 24 parameters obtained by data dimension reduction as input values. And extracting the accumulated data set from the time sequence library, dividing the accumulated data set into a training set and a test set, training an amplitude prediction convolutional neural network model by using the training set, enabling the test set to test the model prediction effect, and if the model meets the requirement, storing the model. If the model does not meet the requirements, the size of the convolution kernel, the step length of the convolution kernel, the size of the pooling window, the step length of the pooling window, the number of full connection layers and the like of the model are adjusted, and the training process is repeated until the model meets the requirements.

6. Amplitude residual analysis

And in the running process of the gear hobbing machine, calculating residual errors of the gear box Y-direction amplitude actual value and the amplitude prediction model prediction value, and extracting fault signals for subsequent processing when the residual errors exceed a set threshold value.

7. Vibration signal analysis

And processing a corresponding Y-direction vibration signal of the gear box by using a developed vibration signal analysis program, extracting main characteristic frequency, comparing the main characteristic frequency with fault characteristic frequency, determining an abnormal position, and pushing the result to related personnel to realize early abnormal identification of the gear hobbing machine.

8. Model update and optimization

When false alarm and missing alarm occur, the amplitude residual error threshold is adjusted, if the abnormal identification requirement is not met, the data set is updated, the amplitude prediction model construction flow is repeated, and therefore the optimization model is updated. In addition, according to the failure rate curve of the gear hobbing machine, the prediction model is updated regularly and the old model is stored in a model library for the subsequent equipment with the same type to call.

Fig. 5 is a schematic structural diagram of an apparatus anomaly identification apparatus according to one or more embodiments of the present disclosure, including:

at least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

One or more embodiments of the present description provide a non-volatile computer storage medium storing computer-executable instructions that, when executed by a computer, enable:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

In this specification, each embodiment is described in a progressive manner, and the same and similar subsystems of each embodiment are referred to each other, where each embodiment focuses on differences from other embodiments. In particular, for device, non-volatile computer storage medium embodiments, since they are substantially similar to method embodiments, the description is relatively simple, with reference to subsystem descriptions of method embodiments being sufficient.

The foregoing describes specific embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

The foregoing is merely one or more embodiments of the present description and is not intended to limit the present description. Various modifications and alterations to one or more embodiments of this description will be apparent to those skilled in the art. Any modification, equivalent replacement, improvement, or the like, which is within the spirit and principles of one or more embodiments of the present description, is intended to be included within the scope of the claims of the present description.

Claims (10)

1. A method for identifying device anomalies, the method comprising:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

2. The method of claim 1, wherein prior to the obtaining the real-time health status key characterization parameter for the specified device, the method further comprises:

analyzing the designated equipment through fault tree analysis and fault mode influence analysis to obtain an analysis result of the designated equipment:

determining the hazard degree and characterization parameters of each fault cause in the designated equipment according to the analysis result;

and determining the critical characterization parameters of the health state of the appointed equipment according to the hazard degree and the characterization parameters of each fault cause.

3. The method of claim 1, wherein the real-time health-related critical characterizing parameters of the designated device comprise real-time health-related critical characterizing parameters of subsystems of the designated device; the operation data of the appointed equipment comprises the operation data of each subsystem of the appointed equipment; the predicted health state key characterization parameters of the specified equipment comprise predicted health state key characterization parameters of all subsystems of the specified equipment;

before the acquiring the real-time health state key characterization parameters of the specified equipment, the method further comprises:

analyzing the subsystem of the appointed equipment through fault tree analysis and fault mode influence analysis to obtain an analysis result of the subsystem of the appointed equipment:

determining the hazard degree and characterization parameters of each fault cause in each subsystem of the designated equipment according to the analysis result;

and determining the key characterization parameters of the health state of each subsystem of the appointed equipment according to the hazard degree and the characterization parameters of each fault cause.

4. The method of claim 1, wherein before said entering the operational data of the specified device into the pre-trained predictive model, the method further comprises:

determining a characteristic subset corresponding to the operation data of the appointed equipment according to a preset analysis mode;

and inputting the characteristic subset into a pre-trained dimension reduction model, outputting operation data of the designated equipment after dimension reduction, and inputting the operation data of the designated equipment after dimension reduction into the prediction model.

5. The method of claim 1, wherein if the specified device has an abnormal subsystem false alarm or false miss, the method further comprises:

analyzing reasons of false alarm or missing alarm, wherein the reasons of false alarm or missing alarm comprise data noise, model inaccuracy and unreasonable threshold setting;

and determining an adjustment strategy according to the reasons of the false alarm or the missing alarm.

6. The method of claim 5, wherein the cause of the false alarm or the missing alarm is that the threshold setting is not reasonable, and the adjusting policy comprises:

and carrying out adaptive adjustment on the preset threshold value.

7. The method of claim 5, wherein the cause of the false alarm or missing alarm is the data noise, and wherein the adjustment strategy comprises:

and replacing the operation data corresponding to the false report or the missing report in the data set, and retraining the prediction model through the updated data set.

8. The method of claim 5, wherein the cause of the false positive or false negative is the model inaccuracy, the adjusting strategy comprising:

acquiring historical operation data of the appointed equipment, and determining failure rate according to the historical operation data;

and determining the update period of the prediction model according to the failure rate and the actual running condition of the appointed equipment, so that the training prediction model is updated according to the latest running data in each update period.

9. An apparatus abnormality recognition apparatus, characterized by comprising:

at least one processor; the method comprises the steps of,

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.

10. A non-transitory computer storage medium storing computer executable instructions that when executed by a computer enable:

acquiring a real-time health state key characterization parameter of the appointed equipment;

acquiring operation data of the appointed equipment;

inputting the operation data of the appointed equipment into a pre-trained prediction model to obtain the key characterization parameters of the predicted health state of the appointed equipment;

performing residual analysis according to the real-time health state key characterization parameters of the specified equipment and the predicted health state key characterization parameters of the specified equipment to obtain a residual analysis result of the specified equipment;

and if the residual analysis result of the specified equipment exceeds a preset threshold, identifying a subsystem with abnormality of the specified equipment according to the operation data of the specified equipment.