Data-driven adaptive fault diagnosis method based on multi-evaluation index traction

Abstract

The invention belongs to the technical field of fault diagnosis, and discloses a data-driven adaptive fault diagnosis method based on multi-evaluation index traction, and the method comprises the following steps: S1, inputting system data defined in a standard format; s2, establishing a data processing algorithm model and a fault diagnosis algorithm model; s3, comprehensively evaluating the index APR; and S4, based on an adaptive fault diagnosis method, obtaining an optimal algorithm combination and a final fault diagnosis result, and performing reverse verification. According to the data-driven adaptive fault diagnosis method based on multi-evaluation index traction, the requirements of different indexes are comprehensively measured based on the comprehensive performance evaluation index APR, the adaptability of different system data to the current model library is reversely evaluated, unnecessary manual research time consumption is eliminated, and the fault diagnosis efficiency is improved. And a model library can be expanded in a directional manner.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a data-driven adaptive fault diagnosis method based on multi-evaluation index traction. Background Art

[0002] At present, machine learning methods are the mainstream data-driven fault diagnosis methods. The diagnosis process basically includes feature extraction, data dimension reduction, noise reduction, balancing, normalization, selection of different algorithm models, optimization, etc. For different systems and different types of data, matching suitable fault diagnosis methods usually requires research and comparative analysis of different algorithm models. Based on the existing popular algorithms, special fault diagnosis models are developed through different parameter optimization and data processing to obtain models that meet the requirements of performance evaluation indicators. However, this has the disadvantages of relatively long development time and poor versatility.

[0003] For many existing studies, for example, a single algorithm optimization training model is prone to overfitting, which is suitable for relatively optimal fault diagnosis of simple systems and special data; an ensemble learning method is used to integrate multiple homogeneous or heterogeneous base classifiers, but if there are many weak classifiers, the evaluation performance index will be reduced, and it will take a certain amount of time to achieve the optimal base classifier combination, which is relatively complex; based on adaptive learning methods such as genetic algorithms and particle swarm algorithms, whether there are better solutions from other perspectives such as data processing is not considered; the fusion of data processing and fault diagnosis algorithms and the fusion methods of different algorithms, if they are not adapted to matching, will bring worse performance indicators, which also requires a lot of time and energy, etc. Therefore, it is difficult to meet the needs with a single accuracy rate as an evaluation indicator.

[0004] In summary, the existing fault diagnosis methods lack systematic resource integration and have "technical island barriers". Different technicians who want to obtain the optimal fault diagnosis method for different system data basically have to spend a long time to conduct research on data processing, fault diagnosis methods and their relationship. Moreover, the evaluation indicators are relatively single and insufficient for comprehensive evaluation. There is a lack of an adaptive fault diagnosis method based on the integration of existing system resources. Summary of the invention

[0005] The purpose of the present invention is to provide a data-driven adaptive fault diagnosis method based on multiple evaluation indicators. Based on the comprehensive performance evaluation index APR of fault diagnosis accuracy A, fault detection precision rate P, and fault detection completeness rate R, the needs of different indicators are comprehensively weighed; for different system data, whether data processing such as dimension reduction / noise reduction is needed is adaptively evaluated and different fault classification diagnosis models are matched, and the optimal algorithm model or combination and fusion model is output to obtain the optimal fault diagnosis result based on the APR indicator threshold evaluation. At the same time, according to the obtained optimal APR indicator, the adaptability of different system data to the current model library is reversely evaluated, unnecessary time-consuming manual research is eliminated, and the model library can be expanded in a directional manner.

[0006] To achieve the above object, the present invention provides a data-driven adaptive fault diagnosis method based on multiple evaluation indicators, comprising the following steps: Step S1, inputting system data defined in a standard format; Step S2, establishing a data processing algorithm model and a fault diagnosis algorithm model; Step S3, comprehensive evaluation index APR; Step S4: Based on the adaptive fault diagnosis method, the optimal algorithm combination and the final fault diagnosis result are obtained, and reverse verification is performed.

[0007] Preferably, in step S1, the input system data is specified in a unified standard format, each row is specified to store a type of normal / failure mode feature data, the first column is specified as a classification label, and the other columns are different feature data; Among them, the data format supports .xlsx / .xls / .txt / .csv, and can be expanded to recognize different data types as needed.

[0008] Preferably, in step S1, the input system data is discrete feature data stored according to the format requirements, and the next step of adaptive fault diagnosis is performed directly; for time-continuous data, typical time domain features and frequency domain features are adaptively extracted, and the time domain features and frequency domain feature data are automatically saved locally according to the discrete data format requirements, and then the next step of adaptive fault diagnosis is performed.

[0009] Preferably, in step S2, for different feature data, a library of classic or improved data processing and fault classification algorithms is integrated to support adaptive random fusion calling and matching of different data processing and fault classification algorithm combinations to evaluate performance indicators.

[0010] Preferably, in step S3, the comprehensive evaluation index APR includes accuracy A, precision P and recall R; (1) Accuracy A represents the ratio of the number of predicted classification labels that are consistent with the true classification labels to the total number of labels; (2) Precision P: It indicates how many of the test samples predicted as other classes are actually other class samples, as shown below: (1); in, TN X Indicates the number of predicted classes that are actually other classes. FN X Indicates that the prediction is other classes and the actual X The number of classes; (3) Recall rate R: It indicates how many other types of faults in the test sample are predicted correctly, as shown below: (2); in, FP X It is predicted to be X The class is actually the number of other classes; Since precision and recall are often a set of contradictory indicators, F1 is used as the evaluation indicator; F1 X It is defined as the weighted harmonic mean of precision and recall, assuming that the weights of all indicators are 1, as shown below: (3); Among them, F1 X The indicators are also contradictory, so the overall F1 indicator value is defined as the F1 of different categories of data X The weighted harmonic mean of the indicator values ​​is as follows: (4); Therefore, the APR metric is defined as the weighted harmonic mean of F1 and accuracy as follows: (5).

[0011] Preferably, in step S4, based on the adaptive fault diagnosis method, the final fault diagnosis result is obtained and reversely verified, and the specific steps are as follows: Step S41, selecting a genetic algorithm to implement adaptive fault diagnosis, and determining the highest algorithm combination number dim according to existing data processing and model algorithm library; Step S42, using a genetic algorithm to generate a chromosome with dim binary bits for random encoding, generating a given population number num, and giving a termination genetic generation number gen; Step S43, decoding the population according to the pre-defined meaning of each bit code of the chromosome corresponding to different algorithm enabling; Step S44, respectively calculating the evaluation index APR for the algorithm combination corresponding to the chromosomes in different populations, the original algorithm combination and each single classification algorithm therein, and then comparing the APR index obtained by the original algorithm combination with the algorithm combination after eliminating the weak single algorithm whose APR index does not meet the standard, and finally optimizing and retaining the algorithm combination with the largest APR index; Step S45: Based on the diagnosis results of the combination of multiple algorithms and combined with the idea of ​​ensemble learning, a voting decision-making method is adopted; Step S46, the parent population is screened according to the pre-given evaluation index APR threshold to retain the excellent population; if there is no chromosome that meets the APR threshold in the population, the chromosome that obtains the optimal APR index threshold in the population is retained as the excellent parent chromosome; Step S47, while generating offspring populations according to crossover and mutation of parent populations, fusing excellent parent populations with offspring populations, and then generating a new generation population according to the evaluation index APR competition and selection, iterating repeatedly until the given evaluation index APR threshold is reached, terminating the iteration or satisfying the given genetic algebra condition; Step S48: Finally, the output algorithm combination is decoded from the chromosome of the output optimal APR index to obtain the final fault diagnosis result, and the adaptability of the current system input data is determined based on whether the optimal APR index meets the given APR threshold.

[0012] Therefore, the present invention adopts the above-mentioned data-driven adaptive fault diagnosis method based on multiple evaluation indexes, and the beneficial effects are as follows: (1) The comprehensive evaluation performance index APR method proposed in the present invention is suitable for comprehensive consideration of multi-objective optimization problems. Other different evaluation indicators can be constructed as optimization objectives in the same way.

[0013] (2) The present invention can reversely evaluate whether the input system data is suitable for the current data processing and algorithm library by obtaining a comprehensive evaluation performance index APR to see whether it meets the threshold requirements. It can directionally expand the existing different types of feature extraction, data processing and machine learning algorithm databases according to needs, which is conducive to the integration of technical resources and avoidance of "technical island barriers".

[0014] (3) The present invention aims to construct a universal system data-driven adaptive fault diagnosis method to reduce the time-consuming cost of repeated cross-research, comparison and development of special models by different technicians on different system data. It can lay a diagnostic foundation for auxiliary decision optimization and health management evaluation of different systems, and help to construct an integrated operation and maintenance platform for different systems.

[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a principle diagram of a data-driven adaptive fault diagnosis method based on multiple evaluation index traction according to the present invention; Figure 2 The present invention is a workflow diagram of a data-driven adaptive fault diagnosis method based on multiple evaluation index traction. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0018] like Figure 1 As shown, the present invention is a data-driven adaptive fault diagnosis system based on multiple evaluation indexes, taking the comprehensive performance evaluation index APR of fault diagnosis accuracy A, fault detection precision rate P, and fault detection rate R as the optimization target, and adopting an adaptive learning method to perform combined encoding of data processing algorithm model and fault diagnosis algorithm model. It only needs to input discrete feature data or time continuous data defined in a standard format, and can adaptively evaluate and screen a single or multiple combinations of algorithm models according to the encoding and decoding system, obtain the optimal algorithm model combination that meets the comprehensive evaluation index threshold and the diagnosis results of multi-measurement point and multi-classification fault data, and can reversely evaluate the adaptability of input system data according to the APR index threshold.

[0019] In this invention, the connotation of feature extraction, data processing, and machine learning algorithms is not elaborated too much. The present invention is a data-driven adaptive fault diagnosis method based on multiple evaluation indicators, focusing on the integration and utilization of various algorithm model resources that have been optimized and improved, and adaptive evaluation selection. Its workflow is as follows: Figure 2 shown.

[0020] Step S1: input system data defined in a standard format.

[0021] In order to meet the universality of automatic identification of system data, a unified standard format is specified for input system data. Each row is specified to store a type of normal / fault mode characteristic data. The first column is specified as the classification label, and the other columns are different characteristic data. The data format supports .xlsx / .xls / .txt / .csv, and different data types can be expanded as needed to identify them.

[0022] When inputting system data, you need to select discrete feature data or time-continuous data type; for discrete feature data stored according to the format requirements, you can directly proceed to the next step of adaptive fault diagnosis; for time-continuous data, typical time domain features or frequency domain features will be adaptively extracted, and the time domain features and frequency domain feature data will be automatically saved locally according to the discrete data format requirements before proceeding to the next step of adaptive fault diagnosis.

[0023] Step S2: Establish a data processing algorithm model and a fault diagnosis algorithm model.

[0024] For different feature data, it integrates classic or improved data processing and fault classification algorithm libraries, supports adaptive random fusion calls to match different data processing and fault classification algorithm combinations, and evaluates performance indicators to obtain the optimal algorithm combination.

[0025] Through comparative analysis, it is found that the performance indicators can be improved by integrating and calling certain fault classification algorithms after dimensionality reduction processing. This is because the principal component analysis method forms independent principal components after transforming the characteristic variables of the original data indicators. If the correlation between the features is strong, it will have a greater impact on the diagnosis and prediction of certain algorithms. For example, it has been found through research that the performance indicators can be significantly improved by calling the KNN algorithm after dimensionality reduction.

[0026] Step S3: Comprehensive evaluation index APR.

[0027] The evaluation of fault diagnosis results is generally inseparable from accuracy, precision, and recall, and fault diagnosis often focuses more on "fault-type" samples. For the classification test results of different algorithm models, for a certain category X, there are the following concentrated categories: TP X Indicates the number of predicted X categories that are actually X categories; FP X Indicates the number of predicted X categories that are actually other categories, TN X Indicates the number of predicted classes that are actually other classes, FN X Indicates the number of classes X that are predicted to be other classes.

[0028] Assuming that the total number of fault categories (including normal categories) is n, we can get: Accuracy A: It indicates the ratio of the number of predicted classification labels that are consistent with the true classification labels to the total number of labels.

[0029] The precision rate P of a certain type of fault: indicates how many of the test samples predicted as other classes are actually other class samples, as shown below: (1); The recovery rate of a certain type of fault R: indicates how many other types of faults in the test sample are predicted correctly, as shown below: (2); Since precision and recall are often a set of contradictory indicators, when it is uncertain which indicator is more important, or when both indicators are considered equally important, F1 is generally used as the evaluation indicator. X It is defined as the weighted harmonic mean of precision and recall, taking the weight of each indicator as 1 as an example, as shown below: (3); When evaluating and comparing the fitness of different algorithm combinations, the comprehensive index APR is used as the optimization target. APR is a comprehensive evaluation index that considers multi-objective optimization of fault classification accuracy A, fault detection precision P, and fault detection rate R. X Indicators are often contradictory, so the overall F1 indicator value is defined as the F1 of different categories of data. X The weighted harmonic mean of the indicator values ​​is as follows: (4); Therefore, the APR metric is defined as the weighted harmonic mean of F1 and accuracy as follows: (5).

[0030] Step S4: Based on the adaptive fault diagnosis method, obtain the final fault diagnosis result and perform reverse verification.

[0031] Combining the different characteristics and requirements of data processing and fault classification algorithm model combination or fusion, the adaptive algorithm model combination optimization is achieved by machine learning to further improve the performance indicators through the proper algorithm model fusion selected.

[0032] Based on the adaptive combination optimization of evaluation indicators, under the existing resource integration, there is no need for different technical personnel to repeatedly conduct time-consuming cross-research, and the model combination or optimal combination that meets the indicator threshold can be adaptively evaluated and selected. The specific implementation process is as follows: Step S41, selecting a genetic algorithm to implement adaptive fault diagnosis in the test verification, and determining the highest algorithm combination number dim according to the existing data processing and model algorithm library; Step S42, using a genetic algorithm to generate a chromosome with dim binary bits for random encoding, generating a given population number num and a given termination genetic generation number gen; Step S43, decoding the population according to the pre-defined meaning of each bit code of the chromosome corresponding to different algorithm enabling; Step S44, respectively calculating the evaluation index APR for the algorithm combination corresponding to the chromosomes in different populations, the original algorithm combination and each single classification algorithm therein, and then comparing the APR index obtained by the original algorithm combination with the algorithm combination after eliminating the weak single algorithm whose APR index does not meet the standard, and finally optimizing and retaining the algorithm combination with the largest APR index; Step S45: Based on the diagnosis results of the combination of multiple algorithms and combined with the idea of ​​ensemble learning, a voting decision-making method is adopted; Step S46, the parent population is screened according to the pre-given evaluation index APR threshold to retain the excellent population; if there is no chromosome that meets the APR threshold in the population, the chromosome that obtains the optimal APR index threshold in the population is retained as the excellent parent chromosome; Step S47, while generating offspring populations according to crossover and mutation of parent populations, fusing excellent parent populations with offspring populations, and then generating a new generation population according to the evaluation index APR competition and selection, iterating repeatedly until the given evaluation index APR threshold is reached, terminating the iteration or satisfying the given genetic algebra condition; Step S48: Finally, the output algorithm combination is decoded from the chromosome of the output optimal APR index to obtain the final fault diagnosis result, and the adaptability of the current system input data is determined based on whether the optimal APR index meets the given APR threshold.

[0033] Example The test verification of a data-driven adaptive fault diagnosis method based on multiple evaluation indexes in the present invention is as follows: a) Discrete feature data and time-continuous data.

[0034] The test verification of this embodiment is divided into inputting discrete feature data and time-continuous data, starting with inputting standard format system data.

[0035] Discrete feature data is trained and tested with 4 types of data, including normal (category 1), single point failure (category 2), too large impedance (category 3) and too small impedance (category 4) of self-extracted CAN network data. There are 1502 training samples: 700 samples of category 1, 160 samples of category 2, 284 samples of category 3, and 358 samples of category 4. There are 926 test samples: 88 samples of category 1, 222 samples of category 2, 156 samples of category 3, and 460 samples of category 4. Each sample contains the maximum amplitude, minimum amplitude, overshoot and other characteristic values ​​of the 15-dimensional network, and the data format is stored in .xlsx. The first column stores the data classification label, the first row stores the data feature name corresponding to each column of data, and each other column is the corresponding 15-dimensional feature.

[0036] The time-continuous data tested in this embodiment takes the rolling bearing standard fault data set CWRU disclosed by the University of the Western Reserve University in the United States as an example. The extracted data includes normal data and single-point fault data under 4 different loads (speeds). The single-point fault data takes the single-point drive end fault data as an example, including the faults of the bearing roller, the inner ring of the bearing, and the outer ring of the bearing. The bearing roller fault, the bearing inner ring fault, and the bearing outer ring fault respectively include a minor fault (fault diameter 0.007mils), a moderate fault (fault diameter 0.014mils), and a severe fault (fault diameter 0.021mils), that is, each load data contains 10 types of fault category data (including normal data).

[0037] In order to keep the extracted features unified, two sensor measurement point data are retained for each fault category data, including the time-continuous data of the DE drive end acceleration data and the FE fan end acceleration data; the data sampling frequency is 12kHz.

[0038] For the time-continuous data tested in this embodiment, the data format is stored in .xlsx when input into the system test evaluation. The first row is the classification label for each column of time-continuous data, and the DE drive end acceleration data and FE fan end acceleration data collected in the same category are placed in two adjacent columns.

[0039] Taking the training data of this embodiment as an example, each category of data includes 2 columns of adjacent time-continuous data, and 10 types of fault category data (including normal data) have a total of 20 columns of time-continuous data.

[0040] b) Feature extraction of time-continuous data.

[0041] For time-continuous data input into the system according to the standard format requirements, by inputting the number of measurement points contained in each type of data (each column of time-continuous data is 1 measurement point data), sampling frequency (12kHz in this case), and sampling period (5s in this case, that is, feature extraction is performed every 60,000 points), it is possible to automatically extract 10-dimensional typical time domain features such as maximum value, minimum value, peak-to-peak value, mean value, and kurtosis of each column of time-continuous data, as well as 12-dimensional typical frequency domain features such as mean frequency and frequency center, for a total of 22-dimensional features, and automatically fuse the data features of multiple measurement points that represent the same type of data.

[0042] Each type of data in the test data of this embodiment contains acceleration data of 2 measuring points, and 44-dimensional fusion features of each type of data can be obtained. The data of this embodiment extracts 70 groups of training feature data containing 10 types of data (10 groups of each type of data) under a certain load and 70 groups of test feature data containing 10 types of data (10 groups of each type of data) under other 3 loads. The extracted feature data is saved in .xlsx format according to the standard format of discrete feature data.

[0043] c) Establish data processing and fault diagnosis algorithm model.

[0044] During the test and verification of this embodiment, the data processing model selection objects include two commonly used methods: kernel principal component analysis (KPCA) and normalized data processing. The machine learning fault diagnosis algorithm model selection objects include seven machine learning methods, including support vector machine (SVM), K-nearest neighbor (KNN), BP neural network algorithm, random forest method (RF), naive Bayes (NB), improved logistic regression (LR), and improved K-means clustering.

[0045] d) Test and verify the results.

[0046] Through the test of this embodiment, for the input discrete system data, the adaptive genetic algorithm sets the highest possible algorithm combination number corresponding to the number of chromosome individual bits to 9, the number of generated populations to 18, the iterative termination genetic generation to 3, the APR input threshold to 0.9, and the APR iterative termination index to 0.99. A variety of algorithm combinations with an APR index of 0.9 can be obtained, and the algorithm model combination with the best APR index of 0.99875 (A=0.99676, P=0.99839, R=0.9991, F1=0.99875) and the final fault diagnosis result can be obtained in the first generation population. Among them, the adaptive evaluation output of a group of optimal algorithm model combinations is "001111101". According to the algorithm combination decoding, no data processing is required, and the SVM base classifier, KNN base classifier, BP neural network base classifier, RF random forest base classifier, LR logistic regression base classifier, and K-means base classifier are selected for voting decision fault classification. Finally, according to the optimal APR index of 0.99875 meeting the 0.9 threshold requirement, the adaptability of the input discrete system data to the current system solution was reversely verified.

[0047] For the time-continuous data input in this embodiment, the adaptive genetic algorithm sets the maximum possible algorithm combination number corresponding to the number of chromosome individual bits to 9, the number of generated populations to 18, the iterative termination genetic generation to 3, the APR input threshold to 0.9, and the APR iterative termination index to 0.99. The algorithm model combination and the final fault diagnosis result with the optimal APR index reaching 1 (A=1, P=1, R=1, F1=1) can still be obtained in the first generation population. Among them, the adaptive evaluation output of a set of optimal algorithm model combinations is "100000100". According to the algorithm combination decoding, the KPCA algorithm is used to reduce the dimension of the data and the LR logistic regression base classifier is selected for voting decision fault diagnosis. Finally, according to the optimal APR index 1 that meets the 0.9 threshold requirement, the adaptability of the input time-continuous data to the current system solution is reversely verified.

[0048] Therefore, the present invention adopts the above-mentioned data-driven adaptive fault diagnosis method based on multiple evaluation indicators, and comprehensively weighs the needs of different indicators based on the comprehensive performance evaluation indicator APR of fault diagnosis accuracy A, fault detection precision rate P, and fault detection rate R; for different system data, through adaptive evaluation of whether data processing such as dimension reduction / noise reduction is needed and matching different fault classification diagnosis models, the optimal algorithm model or combination and fusion model is output to obtain the optimal fault diagnosis result based on the APR indicator threshold evaluation; at the same time, according to the obtained optimal APR indicator, the adaptability of different system data to the current model library is reversely evaluated, unnecessary time-consuming manual research is eliminated, and the model library can be expanded in a directional manner.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A data-driven adaptive fault diagnosis method based on multiple evaluation indicators, characterized in that: The following steps are involved: Step S1, inputting system data defined in a standard format; Step S2, establishing a data processing algorithm model and a fault diagnosis algorithm model; Step S3, comprehensive evaluation index APR; Step S4: Based on the adaptive fault diagnosis method, the optimal algorithm combination and the final fault diagnosis result are obtained, and reverse verification is performed.

2. The data-driven adaptive fault diagnosis method based on multiple evaluation indexes according to claim 1 is characterized in that: In step S1, the input system data is specified in a unified standard format, each row is specified to store a type of normal / failure mode feature data, the first column is specified as a classification label, and the other columns are different feature data; Among them, the data format supports .xlsx / .xls / .txt / .csv, and can be expanded to recognize different data types as needed.

3. The data-driven adaptive fault diagnosis method based on multiple evaluation indexes according to claim 2 is characterized in that: In step S1, the input system data is discrete feature data stored according to the format requirements, and the next step of adaptive fault diagnosis is directly performed; for time-continuous data, typical time domain features and frequency domain features are adaptively extracted, and the time domain features and frequency domain feature data are automatically saved locally according to the discrete data format requirements, and then the next step of adaptive fault diagnosis is performed.

4. The data-driven adaptive fault diagnosis method based on multiple evaluation indexes according to claim 1 is characterized in that: In step S2, for different feature data, a library of classic or improved data processing and fault classification algorithms is integrated to support adaptive random fusion calling and matching of different data processing and fault classification algorithm combinations to evaluate performance indicators.

5. The data-driven adaptive fault diagnosis method based on multiple evaluation indexes according to claim 1 is characterized in that: In step S3, the comprehensive evaluation index APR includes accuracy A, precision P and recall R; (1) Accuracy A represents the ratio of the number of predicted classification labels that are consistent with the true classification labels to the total number of labels; (2) Precision P: It indicates how many of the test samples predicted as other classes are actually other class samples, as shown below: （1）； in, TN X Indicates the number of predicted classes that are actually other classes. FN X Indicates that the prediction is other classes and the actual X The number of classes; (3) Recall rate R: It indicates how many other types of faults in the test sample are predicted correctly, as shown below: （2）； in, FP X It is predicted to be X The class is actually the number of other classes; Since precision and recall are often a set of contradictory indicators, F1 is used as the evaluation indicator; F1 X It is defined as the weighted harmonic mean of precision and recall, assuming that the weights of all indicators are 1, as shown below: （3）； Among them, F1 X The indicators are also contradictory, so the overall F1 indicator value is defined as the F1 of different categories of data X The weighted harmonic mean of the indicator values ​​is as follows: （4）； Therefore, the APR metric is defined as the weighted harmonic mean of F1 and accuracy as follows: （5）。 6. The data-driven adaptive fault diagnosis method based on multiple evaluation indexes according to claim 1 is characterized in that: In step S4, based on the adaptive fault diagnosis method, the final fault diagnosis result is obtained and reversely verified. The specific steps are as follows: Step S41, selecting a genetic algorithm to implement adaptive fault diagnosis, and determining the highest algorithm combination number dim according to existing data processing and model algorithm library; Step S42, using a genetic algorithm to generate a chromosome with dim binary bits for random encoding, generating a given population number num, and giving a termination genetic generation number gen; Step S43, decoding the population according to the pre-defined meaning of each bit code of the chromosome corresponding to different algorithm enabling; Step S44, respectively calculating the evaluation index APR for the algorithm combination corresponding to the chromosomes in different populations, the original algorithm combination and each single classification algorithm therein, and then comparing the APR index obtained by the original algorithm combination with the algorithm combination after eliminating the weak single algorithm whose APR index does not meet the standard, and finally optimizing and retaining the algorithm combination with the largest APR index; Step S45: Based on the diagnosis results of the combination of multiple algorithms and combined with the idea of ​​ensemble learning, a voting decision-making method is adopted; Step S46, the parent population is screened according to the pre-given evaluation index APR threshold to retain the excellent population; if there is no chromosome that meets the APR threshold in the population, the chromosome that obtains the optimal APR index threshold in the population is retained as the excellent parent chromosome; Step S47, while generating offspring populations according to crossover and mutation of parent populations, fusing excellent parent populations with offspring populations, and then generating a new generation population according to the evaluation index APR competition and selection, iterating repeatedly until the given evaluation index APR threshold is reached, terminating the iteration or satisfying the given genetic algebra condition; Step S48: Finally, the output algorithm combination is decoded from the chromosome of the output optimal APR index to obtain the final fault diagnosis result, and the adaptability of the current system input data is determined based on whether the optimal APR index meets the given APR threshold.

Images