A fault diagnosis method for rolling bearings under variable working conditions and complex noise environment

Abstract

The application discloses a fault diagnosis method of a rolling bearing under variable working conditions and a complex noise environment, which comprises two steps of obtaining a fault diagnosis algorithm model by using a historical data set and inputting real-time vibration data into the algorithm model after processing to diagnose and obtain a fault condition of the rolling bearing. The method has the advantages of low artificial experience requirement and low loss rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical fault diagnosis, in particular to the fault diagnosis of rolling bearings. BACKGROUND

[0002] The current rolling bearing fault diagnosis method mainly consists of three parts: the first part is the traditional signal processing method, the second part is the fault diagnosis method based on feature extraction and machine learning, and the third part is the intelligent fault diagnosis method based on deep learning.

[0003] There are mainly three kinds of traditional signal processing methods: time domain analysis, frequency domain analysis and time-frequency domain analysis. Time domain analysis: refers to the analysis of the stability, transient and steady-state performance of the system according to the statistical parameters of the vibration signal such as root mean square, kurtosis, average value, peak value, etc. in a certain time domain feature. Since time domain analysis is a method of directly analyzing the system in the time domain, time domain analysis has the advantages of intuition and accuracy. For example, the technical solutions given in CN111220388A and CN112857806A. Frequency domain analysis: vibration signals can be represented as a combination of different frequency sinusoidal signals. When different vibration signals (normal, fault) are subjected to different frequency sinusoidal functions, they will have different frequency characteristics, which can be used to distinguish whether the vibration signal belongs to normal or fault, and even what kind of fault. For example, the technical solutions given in CN107356432B and CN108956618B. Time-frequency domain analysis: since time domain analysis and frequency domain analysis have high requirements for signal stationarity, it is difficult to process non-stationary signals such as rolling bearing fault data. In order to solve this problem, time-frequency analysis methods have been proposed. Time-frequency analysis mainly represents the multi-dimensional characteristics of signals in time and frequency domains, so as to obtain the time and frequency characteristics of each component of the signal. For example, the technical solutions given in CN104655423A and CN1825082B.

[0004] There are also fault diagnosis methods based on feature extraction and machine learning in the prior art. The fault diagnosis method based on feature extraction and machine learning mainly uses traditional signal processing methods to extract features, and uses machine learning methods to distinguish the features, so as to achieve the purpose of fault diagnosis. The machine learning methods mainly include artificial neural network (ANN), support vector machine (Support Vector Machine (SVM) and expert system, etc. For example, the technical solutions given in CN102607845B and CN103542929A.

[0005] For the deep learning-based intelligent fault diagnosis method: deep learning, as a wave in the field of artificial intelligence in recent years is sweeping the globe, using deep learning method for fault diagnosis, its main use deep network to automatically mine out the representative features implied in the data set, the deep network will learn the features layer by layer down, ultimately get high-level, abstract representative features, save the artificial selection of features, also improve the representativeness and reliability of the features, so the fault diagnosis method based on deep learning and other aspects has caused the high attention of the experts in recent years. For example, the technical solution given in CN104616033A.

[0006] For the field, it is very important to establish a complete, systematic and comprehensive rolling bearing real-time state monitoring and fault diagnosis scheme, use Internet of Things technology, big data technology and artificial intelligence deep learning algorithm and other technologies, integrate time domain analysis, frequency domain analysis and time-frequency domain analysis and other analysis algorithms, carry out fault diagnosis on rolling bearings, ensure the accuracy of fault diagnosis, reduce unplanned downtime of industrial enterprises, and reduce the loss caused by unplanned downtime to industrial enterprises. SUMMARY

[0007] The purpose of the present application is to provide a fault diagnosis method for rolling bearings under variable working conditions and complex noise environment.

[0008] To achieve the above technical purpose, the following technical solutions will be adopted in the present application:

[0009] A fault diagnosis method for rolling bearings under variable working conditions and complex noise environment, comprising the following steps:

[0010] S1: generating a fault diagnosis algorithm model using a historical data set; the S1 step specifically comprises the following steps:

[0011] S11: extracting a rolling bearing equipment historical data set, the historical data set comprising a plurality of historical data;

[0012] S12: obtaining time domain analysis parameters by time domain analysis on all historical data in the historical data set in the S11 step;

[0013] S13: determining the first fault type of each historical data obtained in the S1 step according to the time domain analysis parameters of the S12 step;

[0014] S14: performing frequency domain analysis on each historical data according to the first fault type, and obtaining an energy feature vector;

[0015] S15: creating a data set of all historical data energy feature vectors obtained in the S15 step, and defining them as a training set, a validation set and a test set respectively;

[0016] S16: input the training set, the validation set and the test set in the step S15 into a one-dimensional convolutional neural network model for training, to obtain an algorithm model after training;

[0017] S2: input the real-time vibration data after processing into the algorithm model obtained in the step S1, to obtain a fault condition of the rolling bearing; the step S2 specifically includes the following steps:

[0018] S21: obtain real-time data of the rolling bearing, and obtain time-domain analysis parameters through time-domain analysis;

[0019] S22: determine a first fault type of the real-time data obtained in the step S21 according to the time-domain analysis parameters in the step S21;

[0020] S23: perform frequency-domain analysis on the real-time data according to the first fault type, to obtain an energy feature vector;

[0021] S24: perform operation on the energy feature vector obtained in the step S23 by using the algorithm model obtained in the step S16, to obtain a fault diagnosis result.

[0022] Further, the results of the time-domain analysis in the step S12 include one or more of a mean value, a standard deviation, a root mean square, a skewness, a kurtosis, a square root amplitude, a peak-to-peak value, a maximum value and an absolute average amplitude of the historical data, and further calculate one or more of a skewness factor, a kurtosis factor, a pulse factor, a waveform factor, a peak value factor and a margin factor.

[0023] Further, the first fault type in the step S13 is determined by the time-domain analysis parameters in the step S12.

[0024] Further, the frequency-domain analysis in the step S14 includes the following steps:

[0025] S141: perform wavelet transform on the signal to be processed;

[0026] S142: compare wavelet coefficients of the signal after the wavelet transform with a first threshold value, and set the wavelet coefficients less than the first threshold value to zero and keep the wavelet coefficients greater than the first threshold value unchanged;

[0027] S143: the signal obtained in the S143 step is subjected to 3-layer wavelet decomposition using db5 wavelet to obtain decomposition coefficients ca3, cd3, cd2, cd1 of each layer, and each layer of decomposition coefficients is reconstructed to obtain reconstructed signals a3, d3, d2, d1 and a synthesized reconstructed signal a0, and then the energy of the 3rd layer low-frequency wavelet reconstruction sequence a3 and each layer of high-frequency wavelet reconstruction sequence d3, d2, d1 is calculated; the energy of the 3rd layer low-frequency wavelet reconstruction sequence a3 and each layer of high-frequency wavelet reconstruction sequence d3, d2, d1 is subjected to complementary ensemble empirical mode decomposition to obtain each IMF energy feature vector, and the first n IMFs are extracted, wherein n is a natural number.

[0028] Further, the frequency domain analysis in the S15 step comprises the following steps:

[0029] S151: the obtained energy feature vector is used to create a data set and divide it into a training set, a validation set and a test set, and the division ratio is 7:2:1.

[0030] Further, the one-dimensional convolutional neural network model in the S16 step is constructed using a sequential function.

[0031] Further, the algorithm model training parameters are set as follows: batch size batch_size = 128, training rounds epochs = 60, number of fault types num_classes = 10, vibration signal length length = 1024, number of samples of each class number = 1000, the model has 5 convolutional layers and pooling layers, the first convolutional kernel size is 64*1, the size of the remaining convolutional kernels is 3*1 except the first layer, the hidden layer neurons are 100, and the SoftMax layer has 10 outputs.

[0032] Further, in the S21 step, time domain feature extraction is performed on real-time data through stream computing Flink real-time computing.

[0033] Further, in the time domain analysis in the S21 step, one or more types of parameters calculated in the S12 step are selected to obtain the time domain analysis results of the type.

[0034] Further, the method for determining the first fault type in the S22 step is the same as the method used in the S13 step.

[0035] Further, the method for obtaining the energy feature vector in the S23 step is the same as the method used in the S13 step.

[0036] To sum up, the beneficial effects of the present application are: 1. Overcome the problem of inaccurate judgment or low automaticity in the prior art; 2. The method adopted in the present application combines time domain method, frequency domain method, big data method, etc., effectively improving the accuracy and automation of the test, and reducing the dependence on artificial experience; 3. The loss of the model algorithm of the present application is low, and the closeness of the predicted value to the true value is high. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 Model platform architecture diagram

[0038] Figure 2 Algorithm center architecture diagram

[0039] Figure 3 Confusion matrix of model algorithm

[0040] Figure 4 ROC curve of model algorithm

[0041] Figure 5 Precision-recall curve of model algorithm

[0042] Figure 6 Loss curve of model algorithm

[0043] Figure 7 Accuracy curve of model algorithm

[0044] Figure 8 Internet of Things platform architecture DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] The fault diagnosis is divided into two stages, i.e. a training stage based on historical data of the data lake of the rolling bearing and fault data of the rolling bearing based on a large data platform and a rolling bearing fault diagnosis stage based on real-time data acquisition. The training stage mainly completes training, verification and testing of the model algorithm based on massive rolling bearing historical data, and ensures that the rolling bearing model algorithm has high accuracy. The fault diagnosis stage is to distribute the trained model algorithm to the edge, combine the Flink real-time computing framework to calculate the real-time collected vibration data and temperature data of the rolling bearing, then extract time domain features, frequency domain features and time-frequency domain features, and then use the wavelet packet decomposition algorithm for denoising, the feature data after denoising is extracted into an energy feature vector by a complementary ensemble empirical mode decomposition, and finally the feature vectors are pushed to the model algorithm on the edge side distributed by the platform for real-time diagnosis. After calculation by the model algorithm, a diagnosis result is generated and pushed to an online monitoring system and an operation and maintenance platform, and relevant staff are notified to check the diagnosis result in real time.

[0047] The method of the present application is divided into two parts, the first part is to train the model algorithm, and the second part is to diagnose. In the training model algorithm stage, the model algorithm training of the rolling bearing is mainly carried out through the following steps:

[0048] The health history data of the rolling bearing equipment is extracted from the big data platform. For the vibration data of the rolling bearing, the acceleration signal is obtained, the mean value, standard deviation, root mean square, skewness, kurtosis, root amplitude, peak-to-peak value, maximum value and absolute average amplitude are calculated through the time domain analysis module, and then the skewness factor, kurtosis factor, pulse factor, waveform factor, peak value factor and margin factor are calculated.

[0049] According to the calculated 15 time domain signals, the diagnostic signal classification is carried out through the diagnostic module, and the specific classification is shown in Table 1. The time domain analysis can quickly obtain preliminary fault diagnosis information. In the present application, the preliminary fault type is defined as the first fault type.

[0050]

[0051]

[0052] Table 1

[0053] For ordinary skilled persons in the art, the relationship between the above features and the first fault type is only exemplary. Ordinary skilled persons in the art can adjust the corresponding relationship between the above features and the first fault type according to the actual situation, including associating multiple features with the first fault type, or adding a new first fault type.

[0054] If the rolling bearing fault type and the fault type are obtained through time domain signal analysis, frequency domain analysis is performed on the vibration signal to obtain the amplitude spectrum, phase spectrum and power spectrum of the signal.

[0055] In the frequency domain analysis process, a complex signal composed of multiple signals is decomposed using wavelet transform. In order to support the rolling bearing fault under variable working conditions and harsh environments. For example, the inner ring fault signal is decomposed by wavelet transform. The wavelet coefficients generated by the signal after wavelet transform are larger, while the wavelet coefficients generated by the noise are smaller and less than the wavelet coefficients generated by the signal. By selecting a suitable threshold, the wavelet coefficients greater than the threshold are retained, while the wavelet coefficients less than the threshold are all set to zero, thereby realizing wavelet threshold denoising. The effect of wavelet threshold denoising can be judged according to the signal-to-noise ratio (SNR) and the root mean square error (RMSE). Generally, the larger the signal-to-noise ratio and the smaller the root mean square error, the better the denoising effect. Eight wavelets (db3, db5, db6, db8, sym3, sym5, sym6, sym8) and four thresholds (sqtwolog threshold, rigrsure threshold, heursure threshold, minimaxi threshold) are combined for denoising, and the corresponding signal-to-noise ratio and root mean square error are calculated. Then, select the wavelet and threshold with large signal-to-noise ratio and small root mean square error for denoising. The inner ring fault data is denoised by wavelet threshold with different combinations of wavelet and threshold, and the signal-to-noise ratio and root mean square error under different combinations of wavelet and threshold are obtained.

[0056] The signal denoised by wavelet threshold is decomposed by db5 wavelet to obtain the decomposition coefficients ca3, cd3, cd2, cd1 of each layer, and then the decomposition coefficients of each layer are reconstructed to obtain the reconstructed signals a3, d3, d2, d1 and the synthesized reconstructed signal a0. Then, the energy of the third layer low-frequency wavelet reconstruction sequence a3 and each layer high-frequency wavelet reconstruction sequence d3, d2, d1 is calculated.

[0057] The above technical terms are well known in the prior art and can be referred to many textbooks. The present application is to select and combine the above technical means to obtain the technical effects described in the present application. In the present application, a person skilled in the art can process the signal to be processed according to the description of the above steps.

[0058] Subsequently, the reconstructed energy signal is decomposed by complementary ensemble empirical mode decomposition (CEEMD, Complementary Ensemble Empirical Mode Decomposition) to obtain each IMF component, and the first 8 IMF components are extracted.

[0059] After the signal is processed by using the wavelet method and CEEMD above, the IMF value is obtained to realize accurate reconstruction of the original signal.

[0060] For each piece of historical data, the energy feature vector obtained from each piece of historical data in the historical data set is created into a data set, and the training set, the validation set and the test set are divided, and the division ratio is 7:2:1.

[0061] Then, a one-dimensional convolutional neural network model is constructed using a sequential function, and the sequential function can be completed using Keras. Keras is a high-level neural network API, and Keras is written in pure Python and based on Tensorflow, Theano and CNTK backends. Those skilled in the art can program the corresponding sequential function using Keras according to specific requirements.

[0062] The model training parameters are batch data size batch_size=128, training round number epochs=60, fault type number num_classes=10, vibration signal length length=1024, number of samples of each class number=1000, the model has 5 convolutional layers and pooling layers, the first convolutional kernel size is 64*1, the size of the remaining convolutional kernel is 3*1 except the first layer, the hidden layer neuron is 100, and the SoftMax layer has 10 outputs, corresponding to 10 fault states of the rolling bearing.

[0063] The data in the training set is imported into the neural network model, and through multiple rounds of training, the neural network model can obtain corresponding adaptive parameters to complete the training of the model algorithm.

[0064] Since the training set, the validation set and the test set have been divided in the application, the training set is used for training first, and the parameters are obtained. After determining the parameters of the model, the validation set is used to verify the model. After verification, the test set is used to test the model. The loss rate of the training set, the accuracy of the training set, the loss rate of the validation set and the accuracy of the validation set are evaluated, and after completing the specified number of training rounds, the preliminary model algorithm training is completed.

[0065] The second part of the method of the application is diagnosis.

[0066] First, real-time vibration data of the rolling bearing to be detected are acquired, and 25 time domain features are extracted through the Flink real-time computing technology; the 25 time domain signals calculated can be classified through the diagnosis module, and it is preliminarily judged whether the rolling bearing fault occurs.

[0067] Subsequently, according to the time domain analysis, the rolling bearing fault is calculated in real time, then the frequency domain analysis is performed on the vibration signal to obtain the amplitude spectrum, phase spectrum and power spectrum of the signal, and through the information in the frequency domain, the specific position of the fault can be diagnosed: outer ring, inner ring, rolling body or cage fault. Different fault positions require different processing parameters.

[0068] Next, wavelet threshold denoising is performed on the frequency domain features by using wavelet packet analysis. The signal after wavelet threshold denoising is subjected to 3-layer wavelet decomposition by using db5 wavelet to obtain the decomposition coefficients of each layer, and then the decomposition coefficients of each layer are reconstructed to obtain the reconstructed signal, and then the energy of the 3rd layer low-frequency wavelet reconstruction sequence a3 and each layer high-frequency wavelet reconstruction sequence d3, d2 and d1 is calculated; the reconstructed energy signal is subjected to complementary ensemble empirical mode decomposition (CEEMD) to obtain each IMF component, and finally the energy of each IMF is calculated to obtain the energy feature vector of the fault signal, and the first 8 IMF components are extracted.

[0069] Subsequently, the algorithm for obtaining the model in the first part is called to perform real-time fault diagnosis. The diagnosis process is to operate the IMF values of the real-time data in the diagnosis algorithm model to obtain the corresponding output value. The output value corresponds to the fault type, so as to obtain the fault diagnosis result. After obtaining the fault diagnosis result, the fault diagnosis result can be written into the database and pushed to the online monitoring system for relevant staff to master the health status of the rolling bearing in real time.

[0070] The application will be further described below in combination with the drawings.

[0071] Figure 1 The model platform architecture for rolling bearing fault diagnosis is shown, which is composed of data source, feature extraction, model center and model evaluation, etc.

[0072] The data source provides normal data and fault data of the rolling bearing required for model training, including all rolling bearing data on the big data platform and rolling bearing model algorithm training data from other academic and scientific research. According to the existing historical data, the set algorithm model structure can be trained, so as to be able to meet the algorithm model of the actual working condition data.

[0073] The feature extraction provides various feature models of the rolling bearing required for model training, such as time domain features, frequency domain features and time-frequency domain features.

[0074] The model center provides the support capability platform required for rolling bearing model training, including model design, training set data configuration, validation set data configuration, test set data configuration, model training, model verification, model testing, model export, real-time diagnosis and local diagnosis, etc.

[0075] The model evaluation provides a comprehensive evaluation of the model training results, including confusion matrix, ROC curve, precision recall curve, loss curve and accuracy curve, etc., to help the staff design, train and develop the model efficiently.

[0076] Figure 2 The architecture for the algorithm center is shown, including four parts: data source, feature extraction, algorithm center and algorithm evaluation. The data source provides rolling bearing training data and real-time fault diagnosis data for the algorithm center, rolling bearing simulation data from the data lake and other academic and scientific institutions, to help the early model and algorithm training.

[0077] Feature extraction provides time domain features, frequency domain features and time-frequency domain features for algorithm training.

[0078] The algorithm center provides time domain algorithms, frequency domain algorithms and time-frequency domain algorithms for time domain analysis, frequency domain analysis and time-frequency domain analysis, as well as supervised learning algorithms, unsupervised learning algorithms and deep learning algorithms, etc., which are the core of the entire algorithm center, providing a complete set of algorithm solutions for rolling bearing fault diagnosis.

[0079] Algorithm evaluation provides confusion matrix, ROC curve, precision recall curve, loss curve and accuracy curve, which are used to evaluate the training accuracy and loss rate, verification accuracy and loss rate, and test accuracy and loss rate of rolling bearing fault algorithms.

[0080] Figure 3 The confusion matrix result for model algorithm evaluation is shown, focusing on the diagonal part of the matrix. The values in the diagonal matrix represent the accuracy of the model algorithm. Each type of fault data represents a matrix in the diagonal matrix, and the values in the matrix represent the accuracy of the model algorithm training. The lower the value in the matrix, the more the model algorithm needs to be optimized. The optimization process runs through the entire process of feature extraction model algorithm tuning, feature denoising algorithm tuning, energy feature vector algorithm tuning, and convolutional neural network model and algorithm tuning.

[0081] Figure 4 The ROC curve of the model algorithm is shown. The legend in the lower right corner indicates the ROC area value of each type of rolling bearing fault training. Focus on the area enclosed by the curve. The closer the area is to 1, the higher the accuracy of the model algorithm. Otherwise, the model algorithm needs to be optimized. The optimization process runs through the entire process of feature extraction model algorithm tuning, feature denoising algorithm tuning, energy feature vector algorithm tuning, and convolutional neural network model and algorithm tuning.

[0082] Figure 5The precision recall curve of the model algorithm is shown, and the recall area value of each type of rolling bearing fault training is indicated in the lower left corner legend, the closer the area is to 1, the higher the accuracy of the model algorithm, otherwise the model algorithm needs to be optimized, and the optimization process runs through the whole process optimization of time domain feature, frequency domain feature and time-frequency domain feature extraction model algorithm optimization, feature denoising algorithm optimization, energy feature vector algorithm optimization and convolutional neural network model and algorithm.

[0083] Figure 6 The loss curve of the model algorithm is shown, and the closer the curve is to 0, the lower the loss rate of the model algorithm, otherwise the model algorithm needs to be optimized, and the optimization process runs through the whole process optimization of time domain feature, frequency domain feature and time-frequency domain feature extraction model algorithm optimization, feature denoising algorithm optimization, energy feature vector algorithm optimization and convolutional neural network model and algorithm.

[0084] Figure 7 The accuracy curve of the model algorithm is shown, and the closer the curve is to 1, the higher the accuracy of the model algorithm, otherwise the model algorithm needs to be optimized, and the optimization process runs through the whole process optimization of time domain feature, frequency domain feature and time-frequency domain feature extraction model algorithm optimization, feature denoising algorithm optimization, energy feature vector algorithm optimization and convolutional neural network model and algorithm.

[0085] The method utilizes historical fault data of rolling bearings for early model algorithm training, the model algorithm loss rate is 1.573033%, the model algorithm accuracy reaches 98.42697%, and real-time fault diagnosis is performed on the rolling bearings of key equipment in the rough rolling area and the finishing rolling area of the hot rolling production line of a certain steel plant in Hebei, the fault diagnosis accuracy reaches more than 95%, and the real-time fault diagnosis result delay can be controlled within seconds.

[0086] The above is only a specific embodiment of the present application, and those skilled in the art can make other improvements or modifications on the basis of the above embodiments under the above teaching of the present application. Those skilled in the art should understand that the above specific description is only to better explain the purpose of the present application, and the protection scope of the present application should be subject to the protection scope of the claims.

[0087] In addition, those skilled in the art can understand that although some embodiments described herein include certain features included in other embodiments rather than other features, the combination of features of different embodiments means that it is within the scope of the present application and forms different embodiments. For example, in the following claims, any one of the claimed embodiments can be used in any combination.

Claims

1. A method for diagnosing faults of rolling bearings under variable working conditions and complex noise environments, comprising the following steps: S1: generating a fault diagnosis algorithm model using a historical data set; the S1 step specifically comprises the following steps: S11: extracting a historical data set of rolling bearing equipment; the historical data set comprises a plurality of historical data; S12: obtaining time domain analysis parameters by time domain analysis on all historical data in the historical data set in the S11 step; S13: determining a first fault type of each historical data obtained in the S1 step according to the time domain analysis parameters in the S12 step; S14: performing frequency domain analysis on each historical data according to the first fault type determined in the S13 step to obtain an energy feature vector; S15: creating a data set of all energy feature vectors of the historical data obtained in the S14 step, and defining the data set as a training set, a validation set, and a test set, respectively; S16: inputting the training set, the validation set, and the test set in the S15 step into a one-dimensional convolutional neural network model for training to obtain a trained algorithm model; S2: inputting processed real-time vibration data into the algorithm model obtained in the S1 step to obtain a fault condition of the rolling bearing; the S2 step specifically comprises the following steps: S21: obtaining real-time data of the rolling bearing, and obtaining time domain analysis parameters by time domain analysis; S22: determining a first fault type of the real-time data obtained in the S21 step according to the time domain analysis parameters in the S21 step; S23: performing frequency domain analysis on the real-time data according to the first fault type to obtain an energy feature vector; S24: using the energy feature vector obtained in the S23 step to operate the algorithm model obtained in the S16 step to obtain a fault diagnosis result. 2.The method for diagnosing faults of rolling bearings under variable working conditions and complex noise environments according to claim 1, further comprising: wherein the results of the time domain analysis in the S12 step include a mean value, a standard deviation, a root mean square, a skewness, a kurtosis, a root amplitude, a peak-to-peak value, a maximum value, and an absolute average amplitude of the historical data, and one or more of a skewness factor, a kurtosis factor, a pulse factor, a waveform factor, a peak value factor, and a margin factor are calculated. 3.The method for diagnosing faults of rolling bearings under variable working conditions and complex noise environments according to claim 1, further comprising: wherein the first fault type in the S13 step is determined by the time domain analysis parameters in the S12 step. 4.The method for diagnosing faults of rolling bearings under variable working conditions and complex noise environments according to claim 1, further comprising: wherein the frequency domain analysis in the S14 step comprises the following steps: S141: performing wavelet transform on the signal to be processed; S142: comparing wavelet coefficients of the signal after the wavelet transform with a first threshold value, and setting the wavelet coefficients less than the first threshold value to zero and keeping the wavelet coefficients greater than the first threshold value unchanged. S143: The signal obtained in step S142 is subjected to 3-layer wavelet decomposition using db5 wavelet to obtain the decomposition coefficients ca3, cd3, cd2, cd1 of each layer, and then the decomposition coefficients of each layer are reconstructed to obtain the reconstructed sequences a3, d3, d2, d1 and the synthesized reconstructed sequence a0, and then the energy of the 3rd layer low-frequency wavelet reconstructed sequence a3 and the high-frequency wavelet reconstructed sequences d3, d2, d1 of each layer is calculated; the energy of the 3rd layer low-frequency wavelet reconstructed sequence a3 and the high-frequency wavelet reconstructed sequences d3, d2, d1 of each layer is subjected to complementary ensemble empirical mode decomposition to obtain each IMF energy feature vector, and the first n IMFs are extracted, where n is a natural number.

5. The fault diagnosis method of rolling bearings under variable working conditions and complex noise environment according to claim 1, further wherein the frequency domain analysis in step S15 comprises the following steps: S151: Create a data set with the obtained energy feature vectors and divide them into training set, validation set and test set, with a division ratio of 7:2:

1.

6. The fault diagnosis method of rolling bearings under variable working conditions and complex noise environment according to claim 1, further wherein the one-dimensional convolutional neural network model in step S16 is constructed using a sequential function.

7. The fault diagnosis method of rolling bearings under variable working conditions and complex noise environment according to claim 6, further wherein the algorithm model training parameters are set as batch_size = 128, epochs = 60, num_classes = 10, length = 1024, number = 1000, there are 5 convolutional layers and pooling layers in the model, the first convolutional kernel size is 64*1, the sizes of the remaining convolutional kernels are all 3*1, the hidden layer neurons are 100, and the SoftMax layer has 10 outputs.

8. The fault diagnosis method of rolling bearings under variable working conditions and complex noise environment according to claim 1, further wherein the time domain analysis in step S21 selects one or more types of parameters calculated in step S12 to obtain the time domain analysis results of the type.

9. The fault diagnosis method of rolling bearings under variable working conditions and complex noise environment according to claim 1, further wherein the method for judging the first fault type in step S22 is the same as the method used in step S13.

10. The fault diagnosis method of rolling bearings under variable working conditions and complex noise environment according to claim 1, further wherein the method for obtaining the energy feature vector in step S23 is the same as the method used in step S13.

Images