Complex mechatronic system multi-fault mode supervisory identification and adaptive learning modeling method

Abstract

The present application belongs to the technical field of digital twin technology and equipment fault prognostic health management, and discloses a complex mechatronic system multi-fault mode supervised identification and adaptive learning modeling method. By distinguishing normal and abnormal data, the accurate monitoring of the sub-health state of the system is realized, and the false negative confirmation scheme is adopted to avoid false negatives. In the cause identification layer, a single mode cause identification classifier and a fusion algorithm are constructed to accurately identify the sub-health mode in the abnormal data. The model is iteratively updated through self-learning based on the field operation and inspection results, ensuring that the model remains in the latest and optimal state. The optimized model can be distributed to multiple similar systems, and the new model can be updated and distributed according to the self-learning results. The present application adopts a general architecture and a self-learning mechanism to ensure the model transferability. By fully mining the data features, the algorithm judges the sub-health causes, and the accuracy and adaptability are improved through self-learning with data accumulation, realizing the liberation of manpower and cost reduction and efficiency improvement.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sub-health prediction of complex electromechanical systems, and particularly relates to a complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method. BACKGROUND

[0002] At present, complex electromechanical systems play a vital role in the fields of aerospace and rail transportation. During operation, complex electromechanical systems may encounter problems such as electrical faults, mechanical faults, control system faults, environmental factors, and equipment aging and wear, leading to operation delays and stoppages, safety accidents, and transportation system paralysis.

[0003] At present, the detection process and key parameters before product delivery cannot be effectively monitored; the causes of electromechanical system failures and the key working states before and after the failures cannot be obtained, and the sub-health state of key core products and equipment cannot be warned; when a fault occurs, it can only be found through on-site repair, and there is no mature repair process to guide maintenance. This also increases the potential for complex electromechanical system failures. Therefore, the design, maintenance, and management of complex electromechanical systems are of great significance to ensure the safe, efficient, and stable operation of aerospace and rail transportation systems.

[0004] The PHM system based on the industrial internet platform collects real-time data information of aerospace and rail transportation electromechanical systems online, and through algorithm statistics and model analysis, effectively discovers the health status and existing problems during the operation of aircraft or trains, and then arranges a reasonable maintenance plan to ensure the operation efficiency and safety of aerospace vehicles or trains. PHM technology uses sensor information, expert knowledge, and maintenance support information, combined with intelligent algorithms and reasoning models, to realize the monitoring, prediction, discrimination, and management of the operation state of aerospace vehicles or rail transportation systems. This intelligent maintenance method effectively reduces the false alarm rate of fault detection, making maintenance more efficient and ultimately replacing the traditional event or time-based maintenance method.

[0005] The current technology has realized fault diagnosis and sub-health prediction functions on some electromechanical systems. In terms of fault diagnosis, it can diagnose the fault types of the components of the above-mentioned complex electromechanical systems; in terms of sub-health prediction, it realizes the sub-health prediction of abnormal components. The existing technology has the following problems: the existing data-driven method has certain limitations in the field of complex electromechanical system sub-health monitoring, such as misjudgment of abnormal data, unreasonable data segmentation, unreasonable weight allocation, low accuracy, and insufficient transferability.

[0006] The difficulty of solving the above technical problems: processing complex and variable data, designing adaptive algorithms, managing and adjusting the flexibility of the rule base, optimizing the accuracy of weight distribution, dealing with human factors and environmental interference, and effectively utilizing the increased sample size.

[0007] The significance of solving the above technical problems: Through the implementation of the system, the traditional mode of alarm maintenance and regular manual maintenance after the failure of complex electromechanical systems of aerospace vehicles and rail transit will be broken, filling the technical gap in the field of complex electromechanical system failure prediction and health management at home and abroad, realizing intelligent maintenance of complex electromechanical systems such as aerospace vehicles and rail transit, thereby improving the operation reliability of complex electromechanical systems and reducing operation costs, and laying a technical foundation for improving the operation reliability of aerospace vehicles and rail transit and reducing operation costs. SUMMARY

[0008] In view of the problems existing in the prior art, the present application provides a complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method.

[0009] The present application is implemented as follows: a complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method, comprising:

[0010] Firstly, a complex electromechanical system sub-health monitoring layer is constructed, which distinguishes normal data from abnormal data from unknown data and confirms by using a normal model to realize monitoring of the sub-health state of the system. The abnormal missed report data is processed by using a missed report confirmation scheme to ensure that the abnormal data is correctly transmitted to the next layer for further processing;

[0011] Secondly, a complex electromechanical system sub-health reason identification layer is constructed, which receives abnormal data from the monitoring layer, processes in sections, and identifies and outputs specific sub-health modes in the abnormal data by constructing a single-mode reason identification classifier and an identification result fusion algorithm;

[0012] Thirdly, model updating and iteration are implemented, and the model is iteratively updated by self-learning using multiple complex electromechanical system operation and maintenance check results to ensure that the model always maintains the latest and optimal state. When the model makes a mistake, the first layer or the second layer model is adjusted accordingly according to whether the complex electromechanical system is abnormal;

[0013] Fourthly, model optimization and solidification are performed, and the optimized and solidified model is distributed to multiple similar complex electromechanical systems, and a new sub-health prediction model is redistributed according to the results of self-learning iterative updating;

[0014] Fifthly, different complex electromechanical systems are migrated through a relatively universal architecture, process and self-learning mechanism.

[0015] Further, the first step of constructing the sub-health monitoring layer of the complex mechatronic system includes the following steps:

[0016] (1) Obtain the complex mechatronic system test bench data (including normal operation data and sub-health operation data), the input signals include current signal, speed signal, and angle signal, and the friction force signal is calculated by the speed signal and the current signal;

[0017] (2) Extract features from the current signal, speed signal, angle signal, and friction force signal, respectively;

[0018] (3) Divide the obtained features into a training data set and a test data set. The training data set and the test data set are randomly extracted;

[0019] (4) Screen and fuse the extracted features;

[0020] (5) Take the feature data of the training data set as input, and construct a sub-health monitoring model to distinguish between normal data and sub-health data;

[0021] (6) Take the feature data of the test data set as input, and verify whether the accuracy of the algorithm model meets the index requirements.

[0022] (7) Review the normal data, calculate the root mean square error (RMSE) of the current signal, speed signal, angle signal, and friction force signal with the standard curve, take the RMSE as a feature to model and identify, and further distinguish between normal data and sub-health data.

[0023] Further, the second step of constructing the sub-health reason identification layer includes the following steps:

[0024] (1) Preprocess the complex mechatronic system test bench data (including normal operation data and sub-health operation data);

[0025] (2) Establish a sub-health reason identification model: establish a sub-health reason identification model for each mode, and each model only focuses on one mode;

[0026] (3) Extract sensitive intervals and calculate friction force: segment the original complete data using angle information, extract the sensitive intervals of each mode, and calculate the friction force of the mechatronic system according to expert experience and mechanism analysis;

[0027] (4) Feature extraction and screening: extract and screen the current, speed signal, and friction force in the sensitive data segment of each mode;

[0028] (5) Constructing a sub-health reason identification model: use machine learning algorithm to construct a sub-health reason identification model of each mode, set the corresponding sub-health sample abnormal data label as "1", set the remaining sub-health sample abnormal data label as "0", which is essentially a binary classification model;

[0029] (6) Calculate the accuracy and weight coefficient of the model: calculate the accuracy of each mode sub-health reason identification model through the training data, and calculate the weight coefficient of each model according to the accuracy;

[0030] (7) Calculate the sub-health reason discrimination candidate score: input a group of multiple test data, calculate the output probability value of each sub-health reason identification model, and combine the weight coefficient of each model to calculate the sub-health reason discrimination candidate score of each mode, and sort the sub-health reasons;

[0031] (8) Regularly update the weight coefficient of each mode sub-health reason identification model according to the actual accuracy of each model to improve the model performance and accuracy.

[0032] Further, the self-learning iterative update of the third step model comprises:

[0033] (1) Feature automatic screening: the algorithm automatically selects the most representative features according to the importance of the data and the contribution to the model;

[0034] (2) Dynamic adjustment of rule base: the algorithm dynamically adjusts the rule base according to the actual situation and the feedback of new data to maintain the timeliness and adaptability of the rules;

[0035] (3) Weight self-adaptive distribution: the algorithm self-adaptively distributes the weight, dynamically adjusts the weight of each feature according to the importance and change of the data;

[0036] (4) Intelligent segmentation of motion curve: the algorithm intelligently analyzes the motion curve data and intelligently segments according to the characteristics and changes of the curve;

[0037] (5) Accuracy improvement when the number of samples increases: the algorithm makes full use of the information and feedback of new samples as the number of samples increases, and continuously optimizes the model.

[0038] Another object of the present application is a supervised identification system for multiple fault modes of a complex electromechanical system, which comprises:

[0039] A data acquisition module is used to acquire multiple types of original data including current signals, speed signals, angle signals, etc. from the complex electromechanical system;

[0040] A feature extraction module is used to extract and screen features from the original data to generate a feature data set for training and testing the model;

[0041] a model construction module for constructing classification models of the sub-health monitoring layer and the sub-health cause identification layer based on the extracted feature data;

[0042] a self-learning iteration module for automatically updating and optimizing the parameters and structure of the model according to the operation and maintenance inspection results of the multiple complex mechanical and electrical systems, so as to maintain the latest and optimal state of the model;

[0043] a result output module for outputting the identification result of the model to a user interface or other systems to support fault prediction and health management of the complex mechanical and electrical system.

[0044] Another object of the present application is a supervised identification system for multiple fault modes of a complex mechanical and electrical system, which further comprises:

[0045] a data preprocessing module for performing preprocessing operations such as cleaning, standardization and normalization on the original data to improve data quality and model performance;

[0046] a model verification module for verifying and evaluating the constructed model to ensure the accuracy and reliability of the model;

[0047] a weight allocation module for adaptively adjusting the weights of various features according to the actual performance of the model and the characteristics of the data to improve the classification accuracy of the model;

[0048] a transfer learning module for supporting migration of the optimized and solidified model to other similar complex mechanical and electrical systems,

[0049] so as to realize reuse and sharing of the model.

[0050] Another object of the present application is an adaptive learning system for supervised identification of multiple fault modes of a complex mechanical and electrical system, which can:

[0051] automatically detect the operation data of the complex mechanical and electrical system and distinguish normal data from abnormal data;

[0052] identify and output specific causes leading to the sub-health state according to the characteristics of the abnormal data;

[0053] perform self-learning iteration update on the model by using the field operation and maintenance inspection results to improve the prediction ability and adaptability of the model;

[0054] after the model is updated, automatically distribute the new model to multiple similar complex mechanical and electrical systems to realize rapid deployment and application of the model.

[0055] Further, the system further comprises:

[0056] a rule base management module for storing and managing rules for identifying sub-health causes and supporting dynamic update and adjustment of the rules.

[0057] The intelligent segmentation processing module can automatically perform intelligent segmentation according to the motion curve of the electromechanical system,

[0058] Improve the efficiency and accuracy of data processing;

[0059] The accuracy improvement mechanism continuously improves the recognition accuracy and generalization ability of the model by increasing the number of samples and using new sample information.

[0060] Another object of the present application is a complex electromechanical system multi-fault mode supervised identification and self-learning integrated system, which integrates:

[0061] A plurality of fault mode identification engines are used for supervised identification of a plurality of fault modes of a complex electromechanical system;

[0062] A self-learning engine can automatically update and optimize the model parameters of the identification engine based on operation and maintenance inspection results;

[0063] A general interface module is used for data exchange and model distribution with other complex electromechanical systems;

[0064] A transfer learning management mechanism supports model transfer learning and sharing between different complex electromechanical systems, improving the migratability and expansibility of the system.

[0065] Another object of the present application is to provide a complex electromechanical fault detection system applying the complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method.

[0066] Another object of the present application is to provide a complex electromechanical information data processing terminal applying the complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method.

[0067] Compared with the prior art and method, the present application has the following advantages:

[0068] The complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method proposed by the present application organically combines sub-health monitoring and sub-health cause identification, and the prediction accuracy is improved to more than 80%. The algorithm has the following advantages: 1) increases the missed report reconfirmation, reduces the missed report rate; 2) proposes a data intelligent segmentation algorithm, improves the algorithm self-learning ability, and effectively solves the one-size-fits-all problem of data segmentation; 3) the weight is automatically distributed according to the data statistical value, overcoming the unreasonable weight distribution problem; 4) the algorithm scheme has greatly enhanced migratability; 5) effective features of different sub-health modes are formed, reducing the test workload by more than 50%; 6) reduces the interference of human factors and environmental factors.

[0069] Secondly, the complex mechatronic system multi-fault mode supervision identification and adaptive learning modeling method provided by the present application realizes accurate monitoring and cause identification of the sub-health state of the complex mechatronic system through a series of steps, and exhibits significant technical progress.

[0070] Firstly, the system can effectively distinguish normal and abnormal data from unknown data and confirm through a normal model to ensure the accuracy of the data. For the missed abnormal data, it can be found and transmitted to the sub-health cause identification layer in time, thereby improving the integrity and reliability of the monitoring.

[0071] Secondly, the system can accurately identify the specific sub-health mode from the abnormal data through the sub-health cause identification layer and output, providing strong support for system maintenance. This targeted identification method greatly improves the efficiency and accuracy of fault handling.

[0072] In addition, the system also has strong self-learning ability. Through the self-learning iteration and update of the operation and maintenance results of multiple complex mechatronic systems, the model can be continuously optimized and always maintain the latest and optimal state. This self-learning ability enables the model to adapt to different systems and different fault modes, improving the universality and adaptability of the model.

[0073] Finally, the system realizes migration between different complex mechatronic systems through a relatively universal architecture, process and self-learning mechanism. This enables the model to be easily applied to multiple similar systems, reducing development costs and improving the overall performance and reliability of the system.

[0074] In summary, the complex mechatronic system multi-fault mode supervision identification and adaptive learning modeling method provided by the present application exhibits significant technical progress in data monitoring, fault identification, self-learning and model migration, and provides an effective solution for fault prediction and health management of complex mechatronic systems. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 is the complex mechatronic system multi-fault mode supervision identification and adaptive learning modeling method flowchart provided by the embodiment of the present application.

[0076] Figure 2 is the overall architecture of the complex mechatronic system multi-fault mode supervision identification and adaptive learning modeling method provided by the embodiment of the present application.

[0077] Figure 3 is the system sub-health monitoring and determination flowchart provided by the embodiment of the present application.

[0078] Figure 4 is the sub-health cause identification flowchart provided by the embodiment of the present application.

[0079] Figure 5 is a speed, current intelligent segmentation flowchart based on LSTM provided by the embodiment of the application.

[0080] Figure 6 is a single fault mode sub-health cause judgment result schematic diagram provided by the embodiment of the application. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail. In view of the problems existing in the prior art, the present application provides a complex mechatronic system multi-fault mode supervised identification and adaptive learning modeling method. The present application will be described in detail below with reference to the accompanying drawings.

[0082] The following are two specific embodiments of the complex mechatronic system multi-fault mode supervised identification and adaptive learning modeling method provided by the present application:

[0083] Embodiment one:

[0084] In the operation and maintenance management of a wind power station, the complex mechatronic system multi-fault mode supervised identification and adaptive learning modeling method of the present application is adopted. First, in the sub-health monitoring layer, the system collects the data of the wind turbine generator set in real time through the sensor, and uses the machine learning algorithm to distinguish the normal data and the abnormal data from these data. When abnormal data is found, the system will trigger the false alarm confirmation scheme to further confirm whether there is an abnormal false alarm, so as to ensure that all potential problems can be accurately captured.

[0085] Next, the sub-health cause identification layer receives the abnormal data from the monitoring layer, uses the single mode cause identification classifier and the identification result fusion algorithm that have been constructed, and analyzes the abnormal data in depth to identify the specific causes of the sub-health of the wind turbine generator set, such as gear box wear, generator overheating, etc. These identification results will provide an important basis for subsequent operation and maintenance work.

[0086] As the operation and maintenance work of the wind power station continues, the system will continuously collect new field operation and maintenance inspection results, and use these data to iteratively update the model through self-learning. In this way, the model can continuously learn new fault modes and features, and maintain its latest and optimal state.

[0087] In addition, the optimized and solidified model can be distributed to other similar wind power stations to help them improve their fault prediction and health management capabilities. At the same time, according to the results of self-learning iterative update, the system can also redistribute new sub-health prediction models to ensure that each wind power station can enjoy the latest technology and algorithms.

[0088] Embodiment two:

[0089] The present application is also effectively applied in a subway train manufacturing and maintenance base. In the sub-health monitoring layer, the system realizes accurate monitoring of the sub-health state of the train by real-time monitoring and analyzing various data generated during the operation of the subway train, such as motor temperature, brake system pressure, etc. When the system detects abnormal data, it will also trigger a false alarm confirmation scheme to ensure the accuracy and reliability of the data.

[0090] In the sub-health reason identification layer, the system uses single-mode reason identification classifiers and identification result fusion algorithms to conduct in-depth analysis on abnormal data and identify the specific reasons for the sub-health of the subway train, such as wheel wear, circuit failure, etc. These information provides valuable reference for maintenance personnel, helping them quickly locate and solve problems.

[0091] As the subway train operation and maintenance work proceeds, the system will continuously collect new operation and maintenance inspection results for self-learning iterative updating of the model. In this way, the model can continuously adapt to new fault patterns and features, maintaining its prediction accuracy and reliability. At the same time, the optimized and solidified model can be distributed to other subway train manufacturing and maintenance bases to help them improve the level of fault prediction and health management. In addition, according to the results of self-learning iterative updating, the system can also redistribute new sub-health prediction models to ensure that each base can enjoy the convenience and benefits brought by the latest technology and algorithms.

[0092] These two embodiments demonstrate the application of the present application in complex electromechanical system operation and maintenance management such as wind power stations and subway train manufacturing and maintenance bases. Through the combination of supervised identification model system and self-learning system, accurate monitoring and reason identification of the sub-health state of the system are realized, improving the efficiency and accuracy of operation and maintenance work.

[0093] The application provides a complex electromechanical system multi-fault mode supervision identification and adaptive learning modeling method, a complex electromechanical system sub-health monitoring layer distinguishes normal data from abnormal data from unknown data, realizes monitoring of the sub-health state of the system, and adopts a false alarm confirmation scheme to process abnormal false alarm data; a complex electromechanical system sub-health cause identification layer receives abnormal data from the monitoring layer, identifies and outputs specific sub-health modes in the abnormal data by constructing a single mode cause identification classifier and an identification result fusion algorithm; the model is iteratively updated through self-learning by using the field operation and inspection results of multiple complex electromechanical systems, so as to ensure that the model always maintains the latest and optimal state; the optimized and solidified model is distributed to multiple similar complex electromechanical systems, and a new sub-health prediction model is redistributed according to the result of self-learning iteration and update; the model is ensured to be migratable through a relatively universal architecture, process and self-learning mechanism. The application can fully mine important features in input data, identify sub-health causes through an algorithm, improve the accuracy and adaptability of the algorithm through self-learning update as data accumulates, and finally achieve the goal of liberating manpower and reducing cost and increasing efficiency.

[0094] As shown in Figure 1 The complex electromechanical system multi-fault mode supervision identification and adaptive learning modeling method provided by the application comprises the following steps:

[0095] S101: The complex electromechanical system sub-health monitoring layer distinguishes normal data from abnormal data from unknown data, confirms the normal data by using a normal model, does not process the data confirmed as normal, and transmits the abnormal data found as false alarms to the sub-health cause identification layer;

[0096] S102: The sub-health cause identification layer identifies and outputs specific sub-health modes from the abnormal data;

[0097] S103: The single complex electromechanical system sub-health model is iteratively updated through self-learning by using the operation and inspection results of multiple complex electromechanical systems;

[0098] S104: The optimized and solidified single complex electromechanical system sub-health model is distributed to multiple similar complex electromechanical systems within a period of time, and a new sub-health prediction model is redistributed when the distributed model changes according to self-learning iteration and update;

[0099] S105: The model is migrated between different complex electromechanical systems through a relatively universal architecture, process and self-learning mechanism.

[0100] In the application, the step S101 that the system sub-health monitoring layer distinguishes normal data from abnormal data from unknown data is as follows:

[0101] Step one, the original data under different working conditions are grouped, and the main three monitoring parameters of the complex mechanical and electrical system: rotation angle, rotation speed and current original data are preprocessed. The missing data in the sampling is filled by the mean value of the left and right known data, and the outlier elimination is carried out by using the Relyada criterion method. In a running process, the friction force is calculated according to f=F-ma. Wherein f represents the friction force, F represents the driving force, a represents the acceleration, and the mass of the parts m;

[0102] Step two, respectively for current signal, rotation speed signal, rotation angle signal, friction force signal, 15 kinds of time domain features are extracted respectively, including: peak value (Peak), peak to peak value (peak to peak value, PPV), mean value (mean value, MV), mean amplitude (mean amplitude, MA), root square amplitude (root square amplitude, SRA), standard deviation (standard deviation, SD), effective value (root mean square, RMS), skewness (skewness, SKE), kurtosis (kurtosis, KUR), crest factor (crest factor, CRE), impulse factor (impulse factor, IMP), clearance factor (clearance factor, CLE), form factor, multiscale permutation entropy, multiscale diversity entropy;

[0103] Step three, 15 kinds of time domain features are extracted for current signal, rotation speed signal, rotation angle signal and friction force signal respectively, and 60 time domain features are obtained. The obtained 60 features are screened, the random forest importance is used for feature sorting, the first 15 features with the strongest sensitivity (i.e. separability) are selected as the input of the classifier;

[0104] Step four, the complex mechanical and electrical system sub-health reason recognition modeling data set is constructed, and the original data set of each sub-health mode data is randomly extracted according to the proportion of 7:2:1 to construct the training set, the verification set and the test set respectively;

[0105] Step five, for each sub-health mode, the class balanced training data is constructed on the training data set, that is, the sample number of the sub-health mode concerned and the sample number of other sub-health modes are in the proportion of 1:1. The specific features and algorithms are selected for model training;

[0106] Step six, model tuning is performed on the validation set, and finally the actual precision and reliability of the model are evaluated on the test set;

[0107] Step seven, the current signal, the speed signal and the rotation angle signal are input into the normal signal review model, and when calculating the RMSE of the signal, the standard curve is first calculated. The standard curve is obtained by taking the mean value of the normal data obtained on the test bench or in operation at each time. The RMSE of the signal is obtained through the following formula (1.1) and (1.2):

[0108] e(k) = |X 当前信号 (k) - X 标准曲线 (k) | (1.1)

[0109]

[0110] After obtaining the RMSE of the four signals, they are input into the KNN classifier as features for modeling and testing. When the model identifies a normal signal, 0 is output and no processing is performed; when the model identifies an abnormal signal, 1 is output and is transmitted to the sub-health reason identification model.

[0111] In the present application, the sub-health reason identification layer in step S102 identifies the specific sub-health mode from the abnormal data and outputs as follows:

[0112] Step one, input the speed and current signals, perform preprocessing, and use the Rida criterion method to remove outliers. The missing data in the sampling is filled by the mean value of the left and right known data, that is, the missing value is filled by the mean value of the left and right known data.

[0113] Step two, build an LSTM network, and specify the input as a sequence with a size of 2 (the number of features of the input data). Specify an LSTM layer containing 200 hidden units, and output the complete sequence. In the network, include a fully connected layer with a size of N, followed by a softmax layer and a classification layer to specify N segment categories. Specify the training options. Set the solver to 'adam'. Train for 60 rounds. Prevent gradient explosion by setting the gradient threshold to 2, and train for multiple rounds until the model converges. Input the new test data, the speed signal and the current signal, into the trained model to obtain the segment label.

[0114] Step three, the above segmented data is zero-mean and normalized, and the friction is calculated according to the sensitive segment.

[0115] Step four, 15 time domain features are extracted for current signal, speed signal, angle signal and friction force, including: peak value (Peak), peak-to-peak value (PPV), mean value (MV), mean amplitude (MA), root square amplitude (SRA), standard deviation (SD), root mean square (RMS), skewness (SKE), kurtosis (KUR), crest factor (CRE), impulse factor (IMP), clearance factor (CLE), form factor, multiscale permutation entropy, multiscale diversity entropy;

[0116] Step five, for different sub-health modes, machine learning algorithm is used for feature screening, and a sub-health identification model is established using random forest algorithm. The feature importance ranking and model accuracy change curve under different feature combinations are obtained by random forest algorithm. The relief algorithm is used to sort the features of the multi-classification model based on the inter-class distance and intra-class distance. The final feature ranking is obtained by combining the results of the various feature rankings, and the top-ranked features are selected as the candidate feature set.

[0117] Step six, through data-fault correlation verification analysis, the features consistent with the mechanism knowledge are added to the candidate feature set, and finally the optimal feature set for each sub-health mode is obtained;

[0118] Step seven, principal component analysis (PCA) algorithm is used to reduce the dimension of the optimal feature set, realize feature fusion and remove the linear correlation between features;

[0119] Step eight, the original data set of each sub-health mode data is randomly extracted by the ratio of 7:2:1 to construct the training set, the validation set and the test set respectively. For each sub-health mode, balanced training data is constructed on the training data set, that is, the ratio of the number of samples of the sub-health mode of interest to the number of samples of other sub-health modes is 1:1. Machine learning algorithm is used to construct the sub-health cause identification model of each mode, which is essentially a binary classification model. The corresponding sub-health sample abnormal data label is set to "1", and the remaining sub-health sample abnormal data label is set to "0". Each classifier is only sensitive to one mode.

[0120] Step nine, calculate the accuracy of each pattern sub-health reason identification model Xi by training data, i=1, 2, …N, Xi=Ki / M, wherein Ki is the number of samples determined correctly by the i-th model, M is the total number of input samples, N is the total number of pattern sub-health reasons to be identified, and the weight coefficient of each model is calculated according to Xi, ai=Xi / ∑Xi.

[0121] Step ten, input a group of test data, calculate the output probability value pi of each sub-health reason identification model, wherein n is the total number of samples in a group of test data, and m is the number of times that the model outputs the label "1". The candidate score si of each pattern sub-health reason is calculated by combining the weight coefficient ai of each model, wherein si=ai*pi. The candidate scores si of all pattern sub-health reasons are sorted, and the top three patterns and their probability values are output as the push result.

[0122] The data used in the embodiment of the application is collected from normal and abnormal data of parts of the rail transit electromechanical system experiment, and provides labeled data of actual operation of electromechanical system engineering as a training data set, and the data includes current, rotation angle and rotation speed signals.

[0123] The sub-health reason identification model is established for a single pattern, the data under the same working condition is selected to divide the training set and the test set, and the electromechanical system sub-health reason prediction is performed according to the steps described in the application, Figure 6 The results obtained by the method of the application.

[0124] According to the experimental results, it can be seen that the test results of the method proposed in the application: the accuracy rate of the single fault pattern sub-health reason judgment by the system sub-health reason identification model reaches 100%.

[0125] The above is only a specific embodiment of the application, but the protection scope of the application is not limited thereto, and any modification, equivalent replacement and improvement within the technical range disclosed in the application and within the spirit and principles of the application should be covered in the protection scope of the application.

Claims

1. A method for complex electromechanical system multi-fault mode supervisory identification and adaptive learning modeling, characterized in that, Comprise: The first step, the construction of complex electromechanical system sub-health monitoring layer, the layer is distinguished from the normal data and abnormal data in the unknown data, and the normal model is used for confirmation, the monitoring of the system sub-health state is realized; The abnormal data is processed by adopting the false alarm confirmation scheme to ensure that the abnormal data is correctly transmitted to the next layer for further processing; The second step, the construction of sub-health reason identification layer, the abnormal data from the monitoring layer is received, after the segmented processing, the specific sub-health mode in the abnormal data is identified and output by constructing single mode reason identification classifier and identification result fusion algorithm; The third step, the model updating and iteration is implemented, the self-learning iteration updating of the model is carried out by using the operation and maintenance check results of multiple complex electromechanical systems, so that the model always maintains the latest and optimal state; When the model is wrong, the first layer or the second layer model is adjusted accordingly according to whether the complex electromechanical system is abnormal; The fourth step, the model optimization and solidification is carried out, the optimized and solidified model is distributed to multiple similar complex electromechanical systems, and the new sub-health prediction model is redistributed according to the result of self-learning iteration updating; The fifth step, the migration of different complex electromechanical systems is carried out through the relatively universal architecture, process and self-learning mechanism; The first step of constructing complex electromechanical system sub-health monitoring layer comprises the following steps: (1) Obtain complex electromechanical system test data, input signals include current signal, speed signal, angle signal, friction force signal is calculated by speed signal and current signal, wherein the complex electromechanical system test data contains normal operation data and sub-health operation data; (2) Feature extraction is carried out on current signal, speed signal, angle signal and friction force signal respectively; (3) The obtained features are divided into training data set and test data set; Wherein, the training data set and test data set are obtained by random sampling; (4) The extracted features are screened and fused; (5) The feature data of the obtained training data set is taken as input, and a sub-health monitoring model is constructed to distinguish normal data from sub-health data; (6) The feature data of the test data set is taken as input to verify whether the accuracy of the algorithm model meets the index requirements; (7) The normal data is reviewed, the root mean square error RMSE of current signal, speed signal, angle signal and friction force signal and standard curve is calculated, the RMSE is taken as a feature for modeling and identification, and the normal data and sub-health data are further distinguished.

2. The complex electromechanical system multiple fault mode supervisory identification and adaptive learning modeling method of claim 1, wherein, The second step of constructing sub-health reason identification layer comprises the following steps: (1) Preprocess the complex electromechanical system test bench data, wherein the complex electromechanical system test bench data contains normal operation data and sub-health operation data; (2) Establish sub-health reason identification model: establish sub-health reason identification model for each mode, and each model only focuses on one mode; (3) Segment extraction of sensitive interval and calculation of friction force: the original complete operation data is segmented by using angle information, the sensitive interval of each mode is extracted, and the friction force of electromechanical system is calculated according to expert experience and mechanism analysis; (4) Feature extraction and screening: Extract and screen the current, speed signal, and friction force in each mode sensitive data segment; (5) Constructing a sub-health reason identification model: Use machine learning algorithms to construct a sub-health reason identification model for each mode. Set the corresponding sub-health sample abnormal data label to "1" and the remaining sub-health sample abnormal data label to "0". This is essentially a binary classification model; (6) Calculate the accuracy and weight coefficient of the model: Calculate the accuracy of each mode sub-health reason identification model using training data. Calculate the weight coefficient of each model based on the accuracy; (7) Calculate the sub-health reason discrimination candidate score: Input a set of test data. Calculate the output probability value of each sub-health reason identification model. Combine the weight coefficients of each model to calculate the sub-health reason discrimination candidate score for each mode. Sort the sub-health reasons; (8) Update the weight coefficients of each mode sub-health reason identification model based on the actual accuracy to improve model performance and accuracy.

3. The complex electromechanical system multiple fault mode supervisory identification and adaptive learning modeling method of claim 1, wherein, The self-learning iterative update of the third step model includes: (1) Automatic feature selection: The algorithm automatically selects the most representative features based on the importance of the data and its contribution to the model; (2) Dynamic adjustment of rule base: The algorithm dynamically adjusts the rule base based on actual conditions and feedback from new data to maintain the timeliness and adaptability of the rules; (3) Adaptive weight allocation: The algorithm adaptively allocates weights based on the importance and changes of the data to dynamically adjust the weights of each feature; (4) Intelligent segmentation of motion curve: The algorithm intelligently analyzes motion curve data and intelligently segments them based on their characteristics and changes; (5) Accuracy improvement with increasing sample size: The algorithm leverages the information and feedback from new samples as the sample size increases to continuously optimize the model.

4. A complex electromechanical system multiple fault mode supervised identification system, characterized in that, The system includes: A data acquisition module for acquiring raw data including current signals, speed signals, and angle signals from complex mechanical and electrical systems; A feature extraction module for extracting and screening features from raw data to generate a feature dataset for training and testing models; A model construction module that builds classification models for the sub-health monitoring layer and the sub-health reason identification layer based on the extracted feature data; A self-learning iteration module that automatically updates and optimizes the model parameters and structure based on the operation and maintenance results of multiple complex mechanical and electrical systems to maintain the latest and optimal state of the model; A result output module for outputting the model's identification results to a user interface or other systems to support fault prediction and health management of complex mechanical and electrical systems; The construction of the sub-health monitoring layer of the complex mechanical and electrical system includes the following steps: (1) Obtain complex mechanical and electrical system test data. Input signals include current signals, speed signals, and angle signals. Calculate the friction force signal from the speed signal and current signal. The complex mechanical and electrical system test data contains normal operation data and sub-health operation data; (2) Extract features from the current signal, speed signal, angle signal, and friction force signal; (3) The obtained features are divided into a training data set and a test data set; wherein the training data set and the test data set are obtained by random extraction; (4) The extracted features are screened and fused; (5) The feature data of the obtained training data set is taken as input to construct a sub-health monitoring model to distinguish normal data from sub-health data; (6) The feature data of the test data set is taken as input to verify whether the accuracy of the algorithm model meets the index requirements; (7) The normal data is reviewed, the root mean square error RMSE of the current signal, the speed signal, the angle signal and the friction signal with the standard curve is calculated, the RMSE is taken as a feature to model and identify, and the normal data and the sub-health data are further distinguished.

5. The complex electromechanical system multiple fault mode supervised identification system of claim 4, wherein, The system further comprises: A data preprocessing module for cleaning, standardizing and normalizing the original data to improve data quality and model performance; A model verification module for verifying and evaluating the constructed model to ensure the accuracy and reliability of the model; A weight distribution module that adaptively adjusts the weights of each feature according to the actual performance of the model and the characteristics of the data to improve the classification accuracy of the model; A transfer learning module that supports migrating the optimized and solidified model to other similar complex mechanical and electrical systems to realize model reuse and sharing.

6. An adaptive learning system for supervised identification of multiple failure modes in complex electromechanical systems, characterized in that, The system can: Automatically detect the running data of the complex mechanical and electrical system and distinguish normal data from abnormal data; According to the characteristics of the abnormal data, the specific reasons leading to the sub-health state are identified and output; The model is iteratively updated through self-learning based on the field operation and inspection results to improve the prediction ability and adaptability of the model; After the model is updated, the new model is automatically distributed to multiple similar complex mechanical and electrical systems to realize rapid deployment and application of the model; The construction of the sub-health monitoring layer of the complex mechanical and electrical system comprises the following steps: (1) Obtain the complex mechanical and electrical system test data, the input signals include current signal, speed signal and angle signal, and the friction signal is calculated from the speed signal and current signal, wherein the complex mechanical and electrical system test data contains normal running data and sub-health running data; (2) Feature extraction is performed on the current signal, speed signal, angle signal and friction signal; (3) The obtained features are divided into a training data set and a test data set; wherein the training data set and the test data set are obtained by random extraction; (4) The extracted features are screened and fused; (5) The feature data of the obtained training data set is taken as input to construct a sub-health monitoring model to distinguish normal data from sub-health data; (6) The feature data of the test data set is taken as input to verify whether the accuracy of the algorithm model meets the index requirements; (7) The normal data is reviewed, the root mean square error RMSE of the current signal, the speed signal, the angle signal and the friction signal with the standard curve is calculated, the RMSE is taken as a feature to model and identify, and the normal data and the sub-health data are further distinguished.

7. The adaptive learning system of claim 6, wherein, The system further comprises: A rule base management module for storing and managing rules for identifying sub-health reasons, and supporting dynamic updating and adjustment of the rules; The intelligent segmentation processing module can automatically perform intelligent segmentation according to the motion curve of the electromechanical system, thereby improving the efficiency and accuracy of data processing. The accuracy improvement mechanism continuously improves the recognition accuracy and generalization ability of the model by increasing the number of samples and utilizing new sample information.

8. A complex electromechanical system multiple fault mode supervised identification and self-learning integrated system, characterized in that, The system integrates: Multiple fault mode identification engines for supervised identification of multiple fault modes of complex electromechanical systems; Self-learning engine that can automatically update and optimize the model parameters of the identification engine based on operation and maintenance inspection results; General interface module for data exchange and model distribution with other complex electromechanical systems; Transfer learning management mechanism to support model transfer learning and sharing between different complex electromechanical systems, improving the migratability and expandability of the system. The construction of the complex electromechanical system sub-health monitoring layer includes the following steps: (1) Obtain complex electromechanical system test data, input signals include current signal, speed signal, angle signal, and friction force signal calculated from speed signal and current signal, wherein the complex electromechanical system test data contains normal operation data and sub-health operation data; (2) Extract features from current signal, speed signal, angle signal, and friction force signal; (3) Divide the obtained features into training data set and test data set; wherein the training data set and test data set are randomly extracted; (4) Screen and fuse the extracted features; (5) Use the feature data of the training data set as input to construct a sub-health monitoring model to distinguish normal data from sub-health data; (6) Use the feature data of the test data set as input to verify the accuracy of the algorithm model whether it meets the index requirements; (7) Review the normal data, calculate the root mean square error (RMSE) of the current signal, speed signal, angle signal, and friction force signal with the standard curve, and use the RMSE as a feature to model and identify, further distinguishing normal data from sub-health data.

9. A complex electromechanical fault detection system applying the complex electromechanical system multi-fault mode supervised identification and adaptive learning modeling method of any one of claims 1-3.

Images