Bearing fault diagnosis method, device and equipment and storage medium

Abstract

The invention provides a bearing fault diagnosis method and device, equipment and a storage medium, and the method comprises the steps: obtaining the real-time vibration signal data of a to-be-detected bearing; inputting the real-time vibration signal data into a fault diagnosis model to obtain a fault type of the to-be-detected bearing output by the fault diagnosis model; wherein the fault diagnosis model is obtained by training a residual mechanism and a multi-dimensional convolutional neural network. The fault diagnosis model obtained through training of the scheme of the invention is high in bearing fault diagnosis precision, and is more practical.

Description

Technical Field

[0001] The present invention relates to the technical field of fastener quality and reliability research, and particularly to a bearing fault diagnosis method, device, equipment, and storage medium. Background Art

[0002] As an important component of the mechanical overall mechanism, bearings are widely used in high-end equipment such as aerospace, aviation, weapons, and ships, intelligent devices, and other fields. In addition, the processing and manufacturing technology is continuously developing towards large-scale, intelligent, and standardized, so the requirements for the working environment of bearings are gradually increasing. Once a bearing fails, other components will also be damaged, and the mechanical whole cannot operate, and in severe cases, production accidents may even occur. Therefore, real-time monitoring of the operating state of bearings, monitoring relevant signals to diagnose the fault type, so as to better repair and replace bearings, is of great significance to the development of the processing and manufacturing industry.

[0003] Fault diagnosis of bearings can timely detect the damaged conditions of bearings during operation, and can timely replace and perform equipment maintenance. However, the existing bearing fault diagnosis is mainly carried out through a fault diagnosis model, but the diagnostic accuracy and practicability of the existing bearing fault diagnosis model are poor, and further research is needed. Summary of the Invention

[0004] Embodiments of the present invention provide a bearing fault diagnosis method, device, equipment, and storage medium to solve the problem that the diagnostic accuracy and practicability of the existing fault diagnosis model are poor.

[0005] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] Embodiments of the present invention provide a bearing fault diagnosis method, including:

[0007] Obtain real-time vibration signal data of the bearing to be detected;

[0008] Input the real-time vibration signal data into a fault diagnosis model to obtain the fault type of the bearing to be detected output by the fault diagnosis model;

[0009] Wherein, the fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network.

[0010] Optionally, the method further includes:

[0011] Obtain training vibration signal data of bearing wear degradation;

[0012] Obtain experimental data according to the fault label corresponding to the bearing wear degradation and the training vibration signal data;

[0013] The experimental data is trained using the residual mechanism and the multi-dimensional convolutional neural network to obtain the fault diagnosis model.

[0014] Optionally, the obtaining of the training vibration signal data of bearing wear degradation includes:

[0015] Obtain the historical vibration signal data of bearing wear degradation;

[0016] The empirical mode decomposition (EMD) algorithm is used to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

[0017] Optionally, using the empirical mode decomposition (EMD) algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data includes:

[0018] Obtain the maximum envelope data and the minimum envelope data in the historical vibration signal data;

[0019] According to the maximum envelope data and the minimum envelope data, obtain the first difference data corresponding to the maximum envelope data and the second difference data corresponding to the minimum envelope data;

[0020] In the case where the first difference data does not meet the first condition and the second difference data does not meet the second condition, update the first difference data to the maximum envelope data and update the second difference data to the minimum envelope data, and return to the step of obtaining the first difference data corresponding to the maximum envelope data and the second difference data corresponding to the minimum envelope data according to the maximum envelope data and the minimum envelope data, until the first difference data meets the first condition and the second difference data meets the second condition, and use the first difference data and the second difference data as the intrinsic mode function (IMF);

[0021] Obtain the first residual signal data according to the historical vibration signal data, the first difference data and the second difference data;

[0022] If the first residual signal data is a monotonic function, use the first residual signal data as the training vibration signal data;

[0023] If the first residual signal data is not a monotonic function, update the first residual signal data to the historical vibration signal data, and return to the step of obtaining the maximum envelope data and the minimum envelope data in the historical vibration signal data, until the first residual signal data is a monotonic function, and then use the first residual signal data as the training vibration signal data.

[0024] Optionally, the first condition includes at least one of the following:

[0025] The absolute value of the difference between the number of zeros and the number of poles in the maximum envelope data is less than a first threshold;

[0026] The mean value of the first envelope data in the maximum envelope data is zero, and the mean value of the second envelope data in the maximum envelope data is zero, where the first envelope data is envelope data determined according to the maximum values in the maximum envelope data, and the second envelope data is envelope data determined according to the minimum values in the maximum envelope data;

[0027] The second condition includes at least one of the following:

[0028] The absolute value of the difference between the number of zeros and the number of poles in the minimum envelope data is less than a second threshold;

[0029] The mean value of the third envelope data in the minimum envelope data is zero, and the mean value of the fourth envelope data in the minimum envelope data is zero, where the third envelope data is envelope data determined according to the maximum values in the minimum envelope data, and the fourth envelope data is envelope data determined according to the minimum values in the minimum envelope data.

[0030] Optionally, obtaining the maximum envelope data and the minimum envelope data in the historical vibration signal data includes:

[0031] Determining the maximum value data and the minimum value data in the historical vibration signal data;

[0032] Using the cubic spline curve fitting method to construct curve segments for all maximum value data to obtain the maximum envelope data;

[0033] Using the cubic spline curve fitting method to construct curve segments for all minimum value data to obtain the minimum envelope data.

[0034] Optionally, using the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data includes:

[0035] Performing time-domain analysis on the historical vibration signal data to obtain a time-domain analysis result;

[0036] Performing frequency-domain analysis on the historical vibration signal data to obtain a frequency-domain analysis result;

[0037] In the case that the time-domain analysis result and / or the frequency-domain analysis result indicates that the historical vibration signal data is non-stationary vibration signal data, the empirical mode decomposition (EMD) algorithm is used to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

[0038] Optionally, experimental data is obtained according to the fault label corresponding to the bearing wear degradation and the training vibration signal data, including:

[0039] The training vibration signal data corresponding to different bearings is sorted into a unified length to obtain the processed training vibration signal data;

[0040] The processed training vibration signal data is corresponded with the fault label corresponding to the bearing wear degradation to obtain the experimental data.

[0041] Optionally, the fault diagnosis model is obtained by training the experimental data using the residual mechanism and the multi-dimensional convolutional neural network, including:

[0042] The experimental data is divided according to a preset ratio to obtain training data and validation data;

[0043] The training data is trained using the residual mechanism and the multi-dimensional convolutional neural network to obtain a first diagnosis model;

[0044] The training vibration signal data in the validation data is input into the first diagnosis model to obtain the predicted fault label output by the first diagnosis model;

[0045] In the case that the difference rate between the predicted fault label and the fault label in the validation data is less than or equal to a third threshold, the first diagnosis model is determined as the fault diagnosis model;

[0046] In the case that the difference rate between the predicted fault label and the fault label in the validation data is greater than the third threshold, the first diagnosis model is trained using the residual mechanism, the multi-dimensional convolutional neural network and the training data to obtain a second training model, the second training model is updated as the first diagnosis model, and the step of inputting the training vibration signal data in the validation data into the first diagnosis model to obtain the predicted fault label output by the first diagnosis model is returned until the difference rate between the predicted fault label and the fault label in the validation data is less than or equal to the third threshold, and the first diagnosis model is determined as the fault diagnosis model.

[0047] Optionally, the training data is trained using the residual mechanism and the multi-dimensional convolutional neural network to obtain a first diagnosis model, including:

[0048] Input the training data into the first multi-dimensional convolutional neural network to obtain first processed data output by the first multi-dimensional convolutional neural network;

[0049] Perform feature fusion processing on the first processed data to obtain second processed data;

[0050] Perform residual processing on the second processed data to obtain third processed data;

[0051] Input the third processed data into the second multi-dimensional convolutional neural network to obtain fourth processed data output by the second multi-dimensional convolutional neural network;

[0052] Perform feature fusion processing on the fourth processed data to obtain fifth processed data;

[0053] Perform classification processing on the fifth processed data to obtain the first diagnostic model.

[0054] Optionally, inputting the training data into the first multi-dimensional convolutional neural network to obtain first processed data output by the first multi-dimensional convolutional neural network includes:

[0055] Divide the training data into first sub-data, second sub-data, and third sub-data;

[0056] Perform convolution processing, batch normalization processing, and activation function processing on the first sub-data in sequence to obtain first sub-processed data, perform convolution processing, batch normalization processing, and activation function processing on the second sub-data in sequence to obtain second sub-processed data, and perform convolution processing, batch normalization processing, and activation function processing on the third sub-data in sequence to obtain third sub-processed data;

[0057] Perform max pooling processing on the first sub-processed data to obtain fourth sub-processed data, and perform max pooling processing on the third sub-processed data to obtain fifth sub-processed data;

[0058] Obtain the first processed data according to the second sub-processed data, the fourth sub-processed data, and the fifth sub-processed data.

[0059] Optionally, performing feature fusion processing on the first processed data to obtain second processed data includes:

[0060] Perform feature extraction on the first processed data to obtain feature data;

[0061] Perform data screening and data dimensionality reduction processing on the feature data to obtain processed feature data;

[0062] Perform fusion processing on the processed feature data to obtain the second processed data.

[0063] Optionally, perform residual processing on the second processed data to obtain third processed data, including:

[0064] Input the second processed data into two preset residual networks in sequence to obtain first residual data output by the two preset residual networks;

[0065] Input the second processed data into one preset residual network in sequence to obtain second residual data output by the one preset residual network;

[0066] Obtain the third processed data according to the first residual data and the second residual data.

[0067] Optionally, perform classification processing on the fifth processed data to obtain the first diagnostic model, including:

[0068] Input the fifth processed data into a classification neuron to obtain classification data output by the classification neuron;

[0069] Process the classification data by using a Dropout function to obtain processed classification data;

[0070] Input the processed classification data into a Softmax classifier to obtain training fault labels output by the Softmax classifier;

[0071] Obtain the first diagnostic model according to the training vibration signal data and the training fault labels.

[0072] An embodiment of the present invention further provides a bearing fault diagnosis device, including:

[0073] A first acquisition module, configured to acquire real-time vibration signal data of a bearing to be detected;

[0074] A first processing module, configured to input the real-time vibration signal data into a fault diagnosis model to obtain a fault type of the bearing to be detected output by the fault diagnosis model;

[0075] Wherein, the fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network.

[0076] An embodiment of the present invention further provides a bearing fault diagnosis device, including: a processor, a memory, and a program stored on the memory and executable on the processor, and when the program is executed by the processor, the steps of the bearing fault diagnosis method described in any one of the above are implemented.

[0077] An embodiment of the present invention further provides a readable storage medium, on which a program is stored, and when the program is executed by a processor, the steps in the bearing fault diagnosis method described in any one of the above are implemented.

[0078] The beneficial effects of the present invention are as follows:

[0079] In the solution of the present invention, a fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network. The real-time vibration signal data of the bearing to be detected is input into the above-trained fault diagnosis model, and the fault type of the bearing to be detected is obtained through the fault diagnosis model of the bearing. The fault diagnosis model trained in the above manner has high diagnosis accuracy for bearing faults and is more practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It represents a flowchart of the bearing fault diagnosis method provided by the embodiment of the present invention;

[0081] Figure 2 It represents an overall structure diagram of the model network of the bearing fault diagnosis model provided by the embodiment of the present invention;

[0082] Figure 3 It represents a structural schematic diagram of the residual block provided by the embodiment of the present invention;

[0083] Figure 4 It represents a specific flowchart of the bearing fault diagnosis method provided by the embodiment of the present invention;

[0084] Figure 5 It represents a structural schematic diagram of the bearing fault diagnosis device provided by the embodiment of the present invention;

[0085] Figure 6 It represents a structural schematic diagram of the bearing fault diagnosis equipment provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0086] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0087] Before the description of the specific embodiments, the following explanations are made first:

[0088] The bearing fault diagnosis mainly includes three steps: signal collection, feature extraction, and fault classification. To achieve a higher fault diagnosis rate, corresponding method research needs to be carried out in the entire process. At present, the research on bearing fault diagnosis has been deeply studied by many scholars and significant results have been achieved. However, there is still room for further research in terms of the diagnosis rate accuracy and practicality of the model. The current research still has the following deficiencies:

[0089] (1) Most research focuses on analyzing and studying the correct rate of fault diagnosis of the model, but does not emphasize and study signal acquisition and feature extraction. Therefore, there is a large research space for improving the overall fault diagnosis rate.

[0090] (2) The research focuses on the theoretical research in the three steps of signal analysis, feature extraction, and fault classification, ignoring the experimental conditions under actual working conditions. As a result, the proposed method has good effects in cases, but its effects are worrying under actual conditions.

[0091] In view of the problem that there is no systematic reliability evaluation method for the anti-loosening fastening system in the prior art, the present invention provides a reliability evaluation method, device, and equipment for the anti-loosening fastening system.

[0092] As Figure 1 shown, an embodiment of the present invention provides a bearing fault diagnosis method, including:

[0093] Step 101: Obtain the real-time vibration signal data of the bearing to be detected.

[0094] Optionally, after obtaining the real-time vibration signal data of the bearing to be detected, use the Empirical Mode Decomposition (EMD) algorithm to perform adaptive decomposition processing on the real-time vibration signal data to obtain the processed vibration signal data.

[0095] Step 102: Input the real-time vibration signal data into the fault diagnosis model to obtain the fault type of the bearing to be detected output by the fault diagnosis model; wherein, the fault diagnosis model is trained using a residual mechanism and a multi-dimensional convolutional neural network.

[0096] In this step, first use a residual mechanism and a multi-dimensional convolutional neural network to train a fault diagnosis model, that is, the trained model is a multi-dimensional convolutional neural network bearing fault diagnosis model with a residual mechanism.

[0097] Input the real-time vibration signal data into the fault diagnosis model (optionally, input the above-mentioned processed vibration signal data into the fault diagnosis model), and then the fault type of the bearing to be detected predicted by the fault diagnosis model can be obtained. The accuracy of the prediction result is high, and the model is more practical.

[0098] The following specifically describes the training process of the fault diagnosis model:

[0099] In some embodiments, the method further includes:

[0100] Obtain the training vibration signal data of bearing wear degradation. It should be noted that this training vibration signal data is the basic data set for training the fault diagnosis model;

[0101] Experimental data are obtained based on the fault labels corresponding to the bearing wear degradation and the training vibration signal data, where different faults caused by the bearing wear degradation correspond to different fault labels, and different fault labels and the training vibration signal data corresponding to the fault labels are put into one-to-one correspondence to form an experimental data set;

[0102] The fault diagnosis model is obtained by training the experimental data using the residual mechanism and the multi-dimensional convolutional neural network.

[0103] Further, the obtaining of the training vibration signal data of the bearing wear degradation includes:

[0104] Obtain the historical vibration signal data of the bearing wear degradation, that is, collect the periodic vibration signal data during the bearing wear degradation process. Specifically, use a vibration signal sensor to collect the periodic vibration signal data during the bearing wear degradation process, where the sensor collects vibration signal data at equal time intervals. Through steps such as reading and converting the signals collected by the sensor, the vibration signal data in digital format are arranged in sequence according to the time series and initially used as the basic data set for model training;

[0105] Use the empirical mode decomposition (EMD) algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data. It should be noted that the EMD algorithm is determined according to the characteristics of the historical vibration signal data. Specifically, using the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data includes: performing time-domain analysis on the historical vibration signal data to obtain a time-domain analysis result; performing frequency-domain analysis on the historical vibration signal data to obtain a frequency-domain analysis result; in the case where the time-domain analysis result and / or the frequency-domain analysis result indicates that the historical vibration signal data is non-stationary vibration signal data, using the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data. That is, first perform time-domain analysis and frequency-domain analysis on the historical vibration signal data, determine that the historical vibration signal data is non-linear or non-stationary signal data according to the time-domain analysis result and / or the frequency-domain analysis result, and select a suitable EMD algorithm to process the historical vibration signal data according to the non-linearity and non-stationarity of the historical vibration signal data to obtain the training vibration signal data.

[0106] It should also be noted that the specific process of adaptively decomposing the real-time vibration signal data using the EMD algorithm to obtain the processed vibration signal data is basically the same as the process of adaptively decomposing the historical vibration signal data using the empirical mode decomposition (EMD) algorithm to obtain the training vibration signal data. It is also to perform time-domain analysis and frequency-domain analysis on the real-time vibration signal data, determine that the real-time vibration signal data is a non-linear or non-stationary signal data according to the time-domain analysis results and / or frequency-domain analysis results, and use the EMD algorithm to process the real-time vibration signal data to obtain the processed vibration signal data.

[0107] Among them, when performing time-domain analysis on the vibration signal data (including historical vibration signal data and real-time vibration signal data), the signal data is visually analyzed in chronological order, and the overall change trends of time-domain characteristics such as the mean, variance, and standard deviation of the signal data are observed from the time dimension to determine whether the vibration signal data is non-linear or non-stationary signal data according to the change trend.

[0108] When performing frequency-domain analysis on the vibration signal data (including historical vibration signal data and real-time vibration signal data), the Python language is used to change the vibration signal data distributed in time series to vibration signal data distributed in the frequency domain through Fourier transform, and visual analysis is performed to observe the change characteristics, energy distribution, change period, peak distribution, etc. of the vibration signal data in the frequency domain, and determine whether the vibration signal data is non-linear or non-stationary signal data according to the change characteristics.

[0109] In this embodiment, when using the EMD algorithm to adaptively decompose the vibration signal data (including historical vibration signal data and real-time vibration signal data), the basis function of the EMD algorithm can be obtained according to the characteristics of the vibration signal data itself, avoiding the problem of difficult selection of the basis function in the traditional wavelet transform. The EMD algorithm decomposes the non-linear and non-stationary vibration signal data into intrinsic mode function (IMF) components representing the time scales of signal characteristics.

[0110] As an optional embodiment, for the historical vibration signal data x(t), the steps of performing EMD decomposition specifically include, that is, using the empirical mode decomposition (EMD) algorithm to adaptively decompose the historical vibration signal data to obtain the training vibration signal data, specifically including:

[0111] The first step: Obtain the maximum envelope data and minimum envelope data in the historical vibration signal data;

[0112] Among them, obtaining the maximum envelope data and minimum envelope data in the historical vibration signal data includes:

[0113] Determine the maximum value data and minimum value data in the historical vibration signal data, that is, search for all the maximum value point data and minimum value point data in the data segment of the historical vibration signal data x(t).

[0114] Use the cubic spline curve fitting method to construct curve segments for all the maximum value data to obtain the maximum value envelope data, and use the cubic spline curve fitting method to construct curve segments for all the minimum value data to obtain the minimum value envelope data, that is, use the cubic spline curve fitting method to construct curve segments for all the maximum value point data and minimum value point data respectively as the maximum value envelope data and the minimum value envelope data.

[0115] According to the maximum value envelope data and the minimum value envelope data, obtain the first difference data corresponding to the maximum value envelope data and the second difference data corresponding to the minimum value envelope data. Specifically, solve the mean values of the maximum value envelope data and the minimum value envelope data respectively to obtain the corresponding average value curve data s a (t), and use Formula 1 to obtain the first difference data corresponding to the maximum value envelope data and the second difference data corresponding to the minimum value envelope data:[[]]END]]

[0116] w 1 (t) = x(t) - s a (t) Formula 1

[0117] where w 1 (t) represents the first difference data corresponding to the maximum value envelope data or represents the second difference data corresponding to the minimum value envelope data. In the case where w 1 (t) represents the first difference data corresponding to the maximum value envelope data, x(t) represents the maximum value envelope data, and s a (t) represents the average value curve data corresponding to the maximum value envelope data. In the case where w 1 (t) represents the second difference data corresponding to the minimum value envelope data, x(t) represents the minimum value envelope data, and s a (t) represents the average value curve data corresponding to the minimum value envelope data.

[0118] Step 2: In the case where the first difference data does not satisfy the first condition and the second difference data does not satisfy the second condition, update the first difference data to the maximum envelope line data and update the second difference data to the minimum envelope line data, and return to the step of obtaining the first difference data corresponding to the maximum envelope line data and the second difference data corresponding to the minimum envelope line data according to the maximum envelope line data and the minimum envelope line data, until the first difference data satisfies the first condition and the second difference data satisfies the second condition, and use the first difference data and the second difference data as the intrinsic mode function IMF;

[0119] Wherein, the first condition includes at least one of the following:

[0120] The absolute value of the difference between the number of zeros and the number of poles in the maximum envelope line data is less than a first threshold. Optionally, the first threshold is 1, that is, the number of zeros and the number of poles in the entire signal data of the maximum envelope line data are equal or differ by at most 1;

[0121] The mean value of the first envelope line data in the maximum envelope line data is zero, and the mean value of the second envelope line data in the maximum envelope line data is zero. Wherein, the first envelope line data is the envelope line data determined according to the maximum values in the maximum envelope line data, that is, the first envelope line data is the envelope line data determined according to the local maximum data in the maximum envelope line data, and the second envelope line data is the envelope line data determined according to the minimum values in the maximum envelope line data, that is, the second envelope line data is the envelope line data determined according to the local minimum data in the maximum envelope line data. That is, in this condition, the entire signal data of the maximum envelope line data is locally symmetric about the time axis.

[0122] The second condition includes at least one of the following:

[0123] The absolute value of the difference between the number of zeros and the number of poles in the minimum envelope line data is less than a second threshold. Optionally, the second threshold is 1, that is, the number of zeros and the number of poles in the entire signal data of the minimum envelope line data are equal or differ by at most 1;

[0124] The mean value of the third envelope data in the minimum envelope data is zero, and the mean value of the fourth envelope data in the minimum envelope data is zero. Among them, the third envelope data is the envelope data determined according to the maximum value in the minimum envelope data, that is, the third envelope data is the envelope data determined according to the local maximum value data in the minimum envelope data. The fourth envelope data is the envelope data determined according to the minimum value in the minimum envelope data, that is, the fourth envelope data is the envelope data determined according to the local minimum value data in the minimum envelope data. That is to say, under this condition, the entire segment of signal data of the minimum envelope data is locally symmetric about the time axis.

[0125] Specifically, when w 1 (t) does not meet the conditions of the IMF (that is, the first difference data does not meet the first condition and the second difference data does not meet the second condition), update the first difference data w 1 (t) to the maximum envelope data and update the second difference data w 1 (t) to the minimum envelope data, repeat the above formula one, and obtain the first difference data and the second difference data again until the first difference data meets the first condition and the second difference data meets the second condition, and the calculation ends to obtain the first-order IMF component, denoted as k 1 (t). The first-order IMF component includes the first difference data and the second difference data.

[0126] The first step: Obtain the first residual signal data according to the historical vibration signal data, the first difference data, and the second difference data. Specifically, subtract the first difference data from the historical vibration signal data to obtain the first sub-residual signal data, subtract the second difference data from the historical vibration signal data to obtain the second sub-residual signal data, and obtain the first residual signal data according to the first sub-residual signal data and the second sub-residual signal data. That is, the first residual signal data includes the first sub-residual signal data and the second sub-residual signal data. The specific formula is as formula two:

[0127] r 1 (t) = x(t) - k 1 (t) Formula two

[0128] Among them, r 1 (t) represents the first residual signal data (which can also be called the first-order residual signal), x(t) represents the historical vibration signal data, and k 1 (t) represents the first difference data and the second difference data.

[0129] The first-order residual signal r 1(t) contains components with long periods and low frequencies. If the first residual signal data is a monotonic function, then use the first residual signal data as the training vibration signal data;

[0130] If the first residual signal data is not a monotonic function, update the first residual signal data to the historical vibration signal data, and return to the step of obtaining the maximum envelope data and the minimum envelope data in the historical vibration signal data until the first residual signal data is a monotonic function. Then use the first residual signal data as the training vibration signal data, that is, use the first residual signal data as the original target signal data to continue EMD decomposition, repeat the first step and the second step, and obtain the second-order IMF component k 2 (t), the third-order IMF component k 3 (t), ……, the nth-order IMF component k n (t) and the residual signal r n (t), as shown in Equation Three:

[0131] r 1 (t) - k 1 (t) = r 2 (t) Equation Three

[0132] When r n (t) is a monotonic function, the EMD decomposition process ends, and Equation Four is obtained:

[0133]

[0134] Equation Four can be understood as: the sum of the IMF and the residual component is the original signal data.

[0135] In an alternative embodiment, experimental data is obtained according to the fault label corresponding to the bearing wear degradation and the training vibration signal data, including:

[0136] Organize the training vibration signal data corresponding to different bearings into a unified length to obtain the processed training vibration signal data. That is, the signal-to-noise ratio of the training vibration signal data obtained after EMD decomposition of the historical vibration signal data is high, and the data quality is further improved. However, since the noise information contained in each group of training vibration signal data is not equal in amount, when noise reduction processing is performed on the training vibration signal data, the lengths of the noise-reduced training vibration signal data are not equal, and it is impossible to smoothly input them into the fault diagnosis model for iterative training. In order to unify the dimension and length of the data, it is necessary to perform unified sorting and division, that is, organize the training vibration signal data into a unified length;

[0137] Correspond the processed training vibration signal data with the fault labels corresponding to bearing wear degradation to obtain the experimental data, that is, correspond the processed data with the fault category labels one by one to form an experimental data set.

[0138] The process of dividing and sorting the experimental data, assigning labels, and constructing the experimental data set is completed.

[0139] Furthermore, input the experimental data set into the bearing fault diagnosis model of the multi-dimensional convolutional neural network with a residual mechanism for iterative training. The overall structure of the model network of the bearing fault diagnosis model of the multi-dimensional convolutional neural network with a residual mechanism is as Figure 2 shown. The fault diagnosis model includes a multi-dimensional convolutional module, a feature fusion module, a residual module, and a classification module.

[0140] The following specifically describes the process of using the residual mechanism and the multi-dimensional convolutional neural network to train the experimental data to obtain the fault diagnosis model:

[0141] Divide the experimental data according to a preset ratio to obtain training data and validation data. Optionally, the preset ratio is 7:3, that is, randomly divide the experimental data into a training set and a validation set according to a ratio of 7:3. The Adam optimization method is used in the training process, the learning rate is set to 0.0001, and the amount of data read in at one time during training is 32.

[0142] Use the residual mechanism and the multi-dimensional convolutional neural network to train the training data to obtain a first diagnosis model. Input the training vibration signal data in the validation data into the first diagnosis model to obtain the predicted fault labels output by the first diagnosis model. When the difference rate between the predicted fault labels and the fault labels in the validation data is less than or equal to a third threshold, determine the first diagnosis model as the fault diagnosis model. Here, the third threshold is a threshold determined according to the processing conditions, that is, iteratively train the fault diagnosis model. When the fault diagnosis rate of the model reaches the threshold specified by the current processing conditions, save the trained fault diagnosis model and apply it to online bearing fault diagnosis.

[0143] When the difference rate between the predicted fault label and the fault label in the verification data is greater than a third threshold, the first diagnostic model is trained using the residual mechanism, the multi-dimensional convolutional neural network, and the training data to obtain a second training model. The second training model is updated to the first diagnostic model, and the step of inputting the training vibration signal data in the verification data into the first diagnostic model to obtain the predicted fault label output by the first diagnostic model is returned until the difference rate between the predicted fault label and the fault label in the verification data is less than or equal to the third threshold, and the first diagnostic model is determined as the fault diagnostic model.

[0144] Further, training the training data using the residual mechanism and the multi-dimensional convolutional neural network to obtain a first diagnostic model includes:

[0145] Input the training data into the first multi-dimensional convolutional neural network to obtain first processed data output by the first multi-dimensional convolutional neural network;

[0146] Perform feature fusion processing on the first processed data to obtain second processed data;

[0147] Perform residual processing on the second processed data to obtain third processed data;

[0148] Input the third processed data into the second multi-dimensional convolutional neural network to obtain fourth processed data output by the second multi-dimensional convolutional neural network;

[0149] Perform feature fusion processing on the fourth processed data to obtain fifth processed data;

[0150] Perform classification processing on the fifth processed data to obtain the first diagnostic model.

[0151] That is, as Figure 2 The overall structure of the model network shown includes two multi-dimensional convolutional modules, two feature fusion modules, one residual module, and one classification module. The training data passes through the first multi-dimensional convolutional neural network, feature fusion processing, residual processing, the second multi-dimensional convolutional neural network, feature fusion processing, and classification processing in sequence to obtain the first diagnostic model.

[0152] Further, inputting the training data into the first multi-dimensional convolutional neural network to obtain the first processed data output by the first multi-dimensional convolutional neural network includes:

[0153] Divide the training data into first sub-data, second sub-data, and third sub-data. It should be noted that the multi-dimensional convolution module consists of a convolution layer (Conv), a batch normalization layer (BN), an activation function layer (ReLu), and a max pooling layer (Maxpooling). The training data is divided into three parts and convolution calculations are learned through three channels respectively. Specifically, perform convolution processing, batch normalization processing, and activation function processing on the first sub-data in sequence to obtain first sub-processed data, perform convolution processing, batch normalization processing, and activation function processing on the second sub-data in sequence to obtain second sub-processed data, and perform convolution processing, batch normalization processing, and activation function processing on the third sub-data in sequence to obtain third sub-processed data;

[0154] Perform max pooling processing on the first sub-processed data to obtain fourth sub-processed data, and perform max pooling processing on the third sub-processed data to obtain fifth sub-processed data, that is, the output data of the first channel and the third channel undergo feature dimensionality reduction through the max pooling layer;

[0155] Obtain the first processed data according to the second sub-processed data, the fourth sub-processed data, and the fifth sub-processed data.

[0156] Specifically, the convolution layer mines the features of regional data by sharing weights of the convolution kernel, which can reduce the complexity of the network and improve the learning efficiency of the network. The convolution calculation process is shown in Equation Five:

[0157]

[0158] Among them, M j represents the input data set (first sub-data, second sub-data, third sub-data), l represents the l-th layer of the network, x represents the activation function value of the feature vector, k represents the convolution kernel vector of the feature vector; * represents the convolution operation, b represents the bias of the feature vector, and f represents the activation function.

[0159] In the convolution layer network structure, it is generally necessary to add a pooling layer to reduce the dimensionality of the features and reduce the number of network parameters. Max pooling calculation is used in this module, and the pooling calculation process is shown in Equation Six:

[0160]

[0161] Among them, down(·) represents the downsampling function, β represents the connection weight, b represents the bias of this layer, and f(·) represents the activation function of the pooling layer.

[0162] The batch normalization layer calculates the mean and variance of each batch of information data, and then performs normalization processing to reduce the differences between different batches, accelerate the convergence speed, and reduce the overfitting of the model to a certain extent. The calculation process is shown in Formula 7:

[0163]

[0164] Among them, represents the training batch, and x (k) represents the training data (the first sub-data, the second sub-data, the third sub-data), E(x (k) ) represents the mean of the training data (the first sub-data, the second sub-data, the third sub-data), and var(x (k) ) represents the training data variance.

[0165] After that, the first feature extraction is performed on the first processed data. Optionally, feature fusion processing is performed on the first processed data to obtain second processed data, including:

[0166] Performing feature extraction on the first processed data to obtain feature data, that is, the information processed by the multi-dimensional convolution module, and inputting it into the feature fusion module, and the feature fusion module obtains the feature data;

[0167] Performing data screening and data dimensionality reduction processing on the feature data to obtain processed feature data, that is, performing data screening and data dimensionality reduction on the feature data, and further refining the feature information that can truly represent the real degradation information;

[0168] Performing fusion processing on the processed feature data to obtain the second processed data, that is, performing a fusion operation on the processed feature data.

[0169] Optionally, performing residual processing on the second processed data to obtain third processed data, including:

[0170] Sequentially inputting the second processed data into two preset residual networks to obtain first residual data output by the two preset residual networks;

[0171] Sequentially inputting the second processed data into one preset residual network to obtain second residual data output by the one preset residual network;

[0172] Obtaining the third processed data according to the first residual data and the second residual data.

[0173] Specifically, the second processed data obtained after being processed by the feature fusion module is input into the residual module. The residual network in the residual module calculates and processes data through the stacking of residual blocks. The structure of one residual block is as Figure 3As shown, where x is the output of the first several layers of the network (Weight layer) and the input of the subsequent several layers of the network (Weight layer), and F(x) is the output of the network after non-linearly transforming x.

[0174] Through its special network calculation method, when the number of network layers reaches the optimal effect, the residual network structure enables the remaining network layers to achieve identity mapping, avoiding the occurrence of network degradation in deep learning networks and ensuring the operation effect of the network; in addition, by changing the calculation method, the residual network structure changes "continuous multiplication" to "continuous addition", and also avoids the problem of gradient disappearance in the case of a large number of network parameters; the residual network calculation is as shown in Formula VIII:

[0175] F(x) = F main (x) + F res (x) Formula VIII

[0176] Among them, F main (x) represents the output of two layers of the network in the residual block, that is, the first residual data, and F res (x) represents the output of the residual block after one layer of network calculation, that is, the second residual data, and F(x) represents the third processed data.

[0177] Furthermore, the third processed data is input into the second multi-dimensional convolutional neural network to obtain the fourth processed data output by the second multi-dimensional convolutional neural network, that is, the information after residual processing is input into the second multi-dimensional convolutional module for in-depth feature extraction. This part mainly includes the extraction of detailed features and hidden features. The network structure of the second multi-dimensional convolutional module is exactly the same as that of the first multi-dimensional convolutional module. That is, the process of inputting the third processed data into the second multi-dimensional convolutional neural network to obtain the fourth processed data output by the second multi-dimensional convolutional neural network is the same as the process of inputting the training data into the first multi-dimensional convolutional neural network to obtain the first processed data output by the first multi-dimensional convolutional neural network, which will not be elaborated here.

[0178] Furthermore, the fourth processed data is subjected to feature fusion processing to obtain the fifth processed data, that is, the information processed by the second multi-dimensional convolutional module is input into the second feature fusion module. The second feature fusion module will perform data fusion on the detailed features and hidden features of the fourth processed data, adjust the dimension size of the data, refine the features that can truly represent the real degradation information, and improve the calculation speed of the model to ensure the timeliness of the real-time fault classification task. That is, the process of subjecting the fourth processed data to feature fusion processing to obtain the fifth processed data is basically the same as the process of subjecting the first processed data to feature fusion processing to obtain the second processed data, which will not be elaborated here.

[0179] Further, classify the fifth processed data to obtain the first diagnostic model.

[0180] Specifically, the classification module includes a fully connected layer, a Dropout layer, and a Softmax classifier layer, and the three network layers are connected in sequence.

[0181] Therefore, optionally, classifying the fifth processed data to obtain the first diagnostic model includes:

[0182] Input the fifth processed data into the classification neuron to obtain the classification data output by the classification neuron;

[0183] Use the Dropout function to process the classification data to obtain the processed classification data;

[0184] Input the processed classification data into the Softmax classifier to obtain the training fault label output by the Softmax classifier;

[0185] According to the training vibration signal data and the training fault label, obtain the first diagnostic model.

[0186] To perform more accurate fault identification, the model considers adding the same number of neurons as the number of fault categories (i.e., classification neurons) as the fully connected output layer (Fullconnection layer). The output layer uses the concatenated feature data of the previous layers (i.e., the fifth processed data) as the input, and matches and analyzes the output data after passing through the feature fusion module with the fault categories to achieve the fault classification task. The calculation of the fully connected layer is shown in Formula Nine:

[0187] O = f(b o + ω o f v ) Formula Nine

[0188] where O represents the output of the fully connected layer, b o represents the bias vector, ω o represents the weight matrix, and f v represents the feature vector.

[0189] After that, add the Dropout function of the Dropout layer to prevent overfitting in the final calculation stage of the model, improve the generalization performance of the model, and then output the bearing fault category through the Softmax classifier.

[0190] Next, in combination with Figure 4 , specifically illustrate the specific process of the bearing fault diagnosis method provided by the embodiments of the present invention:

[0191] Collect the periodic vibration signals during the bearing wear degradation process;

[0192] After preprocessing the data of the periodic vibration signal, perform time-domain analysis and frequency-domain analysis on the periodic vibration signal, and select appropriate empirical mode decomposition (EMD) according to the characteristics to process the periodic vibration signal with non-linearity and non-stationarity;

[0193] Perform feature fusion processing on the signal after EMD decomposition, divide the experimental group, and construct an experimental data set;

[0194] Construct a bearing fault diagnosis model of a multi-dimensional convolutional neural network with a residual mechanism (ResNet);

[0195] Obtain the fault category obtained by the bearing fault diagnosis model;

[0196] Conduct model evaluation. After several iterative trainings, when the model diagnosis error is within the specified reliable range, save the model parameters with the best effect as the model parameters for the bearing fault diagnosis of the subsequent similar sample set;

[0197] Real-time collect the vibration sensor signal during the rotation of the bearing. After the above steps, the model outputs the current fault category of the bearing, that is, online real-time diagnosis of the bearing fault.

[0198] The bearing fault diagnosis method based on EMD modal decomposition and multi-dimensional convolutional neural network with a residual mechanism provided by the embodiment of the present invention can perform modal decomposition on the bearing vibration signal characteristics when constructing the bearing fault diagnosis model, provide experimental data with high signal-to-noise ratio for the input of the model, and can automatically extract the key features during the bearing wear and degradation process, providing good preconditions for realizing fault diagnosis, and can avoid the occurrence of gradient disappearance and network degradation problems in the deep network to accurately identify the fault types occurring in the bearing. The bearing fault diagnosis model has high recognition accuracy and good applicability, and has good promotion value in the actual application of bearings. Compared with the prior art, the beneficial effects of the embodiment of the present invention are as follows:

[0199] (1) The bearing fault diagnosis rate involved in this embodiment is high, and it can accurately identify the fault category of the bearing in real-time online, effectively providing strong support for fault elimination;

[0200] (2) The bearing fault diagnosis model proposed in this embodiment starts from the source of the fault diagnosis task and focuses on the two steps of signal analysis and feature extraction, improving the robustness and generalization of the model;

[0201] (3) This embodiment provides technical support for the intelligent development of bearing fault diagnosis. Users can accurately identify the fault category occurring in the bearing through this bearing fault diagnosis method, providing convenient conditions for the maintenance of the equipment.

[0202] Such asFigure 5 As shown in the figure, an embodiment of the present invention further provides a bearing fault diagnosis device, including:

[0203] A first acquisition module 501, configured to acquire real-time vibration signal data of a bearing to be detected;

[0204] A first processing module 502, configured to input the real-time vibration signal data into a fault diagnosis model, and obtain a fault type of the bearing to be detected output by the fault diagnosis model;

[0205] Wherein, the fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network.

[0206] Optionally, the device further includes:

[0207] A second acquisition module, configured to acquire training vibration signal data of bearing wear degradation;

[0208] A second processing module, configured to obtain experimental data according to a fault label corresponding to the bearing wear degradation and the training vibration signal data;

[0209] A third processing module, configured to train the experimental data by using the residual mechanism and the multi-dimensional convolutional neural network to obtain the fault diagnosis model.

[0210] Optionally, the second acquisition module includes:

[0211] A first acquisition unit, configured to acquire historical vibration signal data of bearing wear degradation;

[0212] A first processing unit, configured to perform adaptive decomposition processing on the historical vibration signal data by using an empirical mode decomposition (EMD) algorithm to obtain the training vibration signal data.

[0213] Optionally, the first acquisition unit is specifically configured to:

[0214] Acquire maximum envelope data and minimum envelope data in the historical vibration signal data;

[0215] According to the maximum envelope data and the minimum envelope data, obtain first difference data corresponding to the maximum envelope data and second difference data corresponding to the minimum envelope data;

[0216] In the case where the first difference data does not satisfy the first condition and the second difference data does not satisfy the second condition, update the first difference data to the maximum envelope line data and update the second difference data to the minimum envelope line data, and return the step of obtaining the first difference data corresponding to the maximum envelope line data and the second difference data corresponding to the minimum envelope line data according to the maximum envelope line data and the minimum envelope line data, until the first difference data satisfies the first condition and the second difference data satisfies the second condition, and use the first difference data and the second difference data as the intrinsic mode function IMF;

[0217] Obtain first residual signal data according to the historical vibration signal data, the first difference data and the second difference data;

[0218] If the first residual signal data is a monotonic function, use the first residual signal data as the training vibration signal data;

[0219] If the first residual signal data is not a monotonic function, update the first residual signal data to the historical vibration signal data, and return the step of obtaining the maximum envelope line data and the minimum envelope line data in the historical vibration signal data, until the first residual signal data is a monotonic function, and use the first residual signal data as the training vibration signal data.

[0220] Optionally, the first condition includes at least one of the following:

[0221] The absolute value of the difference between the number of zeros and the number of poles in the maximum envelope line data is less than a first threshold;

[0222] The mean value of the first envelope line data in the maximum envelope line data is zero, and the mean value of the second envelope line data in the maximum envelope line data is zero, where the first envelope line data is the envelope line data determined according to the maximum values in the maximum envelope line data, and the second envelope line data is the envelope line data determined according to the minimum values in the maximum envelope line data;

[0223] The second condition includes at least one of the following:

[0224] The absolute value of the difference between the number of zeros and the number of poles in the minimum envelope line data is less than a second threshold;

[0225] The mean value of the third envelope data in the minimum value envelope data is zero, and the mean value of the fourth envelope data in the minimum value envelope data is zero, where the third envelope data is the envelope data determined according to the maximum value in the minimum value envelope data, and the fourth envelope data is the envelope data determined according to the minimum value in the minimum value envelope data.

[0226] Optionally, the first acquisition unit is specifically configured to:

[0227] Determine the maximum value data and the minimum value data in the historical vibration signal data;

[0228] Use the cubic spline curve fitting method to construct curve segments for all the maximum value data to obtain the maximum value envelope data;

[0229] Use the cubic spline curve fitting method to construct curve segments for all the minimum value data to obtain the minimum value envelope data.

[0230] Optionally, the first processing unit is specifically configured to:

[0231] Perform time-domain analysis on the historical vibration signal data to obtain a time-domain analysis result;

[0232] Perform frequency-domain analysis on the historical vibration signal data to obtain a frequency-domain analysis result;

[0233] In the case where the time-domain analysis result and / or the frequency-domain analysis result indicates that the historical vibration signal data is non-stationary vibration signal data, use the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

[0234] Optionally, the second processing module includes:

[0235] A second processing unit, configured to organize the training vibration signal data corresponding to different bearings into a unified length to obtain processed training vibration signal data;

[0236] A third processing unit, configured to correspond the processed training vibration signal data with the fault labels corresponding to bearing wear degradation to obtain the experimental data.

[0237] Optionally, the third processing module includes:

[0238] A fourth processing unit, configured to divide the experimental data according to a preset ratio to obtain training data and validation data;

[0239] A fifth processing unit, configured to use the residual mechanism and the multi-dimensional convolutional neural network to train the training data to obtain a first diagnostic model;

[0240] A sixth processing unit, configured to input the training vibration signal data in the verification data into the first diagnosis model, and obtain a predicted fault label output by the first diagnosis model;

[0241] A seventh processing unit, configured to determine the first diagnosis model as the fault diagnosis model when a difference rate between the predicted fault label and a fault label in the verification data is less than or equal to a third threshold;

[0242] An eighth processing unit, configured to, when the difference rate between the predicted fault label and the fault label in the verification data is greater than the third threshold, train the first diagnosis model by using the residual mechanism, the multi-dimensional convolutional neural network, and the training data to obtain a second training model, update the second training model as the first diagnosis model, and return to the step of inputting the training vibration signal data in the verification data into the first diagnosis model to obtain the predicted fault label output by the first diagnosis model, until the difference rate between the predicted fault label and the fault label in the verification data is less than or equal to the third threshold, and determine the first diagnosis model as the fault diagnosis model.

[0243] Optionally, the fifth processing unit is specifically configured to:

[0244] Input the training data into a first multi-dimensional convolutional neural network to obtain first processed data output by the first multi-dimensional convolutional neural network;

[0245] Perform feature fusion processing on the first processed data to obtain second processed data;

[0246] Perform residual processing on the second processed data to obtain third processed data;

[0247] Input the third processed data into a second multi-dimensional convolutional neural network to obtain fourth processed data output by the second multi-dimensional convolutional neural network;

[0248] Perform feature fusion processing on the fourth processed data to obtain fifth processed data;

[0249] Perform classification processing on the fifth processed data to obtain the first diagnosis model.

[0250] Optionally, the fifth processing unit is specifically configured to:

[0251] Divide the training data into first sub-data, second sub-data, and third sub-data;

[0252] Perform convolution processing, batch normalization processing, and activation function processing on the first sub-data in sequence to obtain first sub-processed data. Perform convolution processing, batch normalization processing, and activation function processing on the second sub-data in sequence to obtain second sub-processed data. Moreover, perform convolution processing, batch normalization processing, and activation function processing on the third sub-data in sequence to obtain third sub-processed data;

[0253] Perform max pooling processing on the first sub-processed data to obtain fourth sub-processed data. Moreover, perform max pooling processing on the third sub-processed data to obtain fifth sub-processed data;

[0254] Obtain the first processed data according to the second sub-processed data, the fourth sub-processed data, and the fifth sub-processed data.

[0255] Optionally, the fifth processing unit is specifically configured to:

[0256] Extract features from the first processed data to obtain feature data;

[0257] Perform data screening and data dimensionality reduction processing on the feature data to obtain processed feature data;

[0258] Perform fusion processing on the processed feature data to obtain the second processed data.

[0259] Optionally, the fifth processing unit is specifically configured to:

[0260] Input the second processed data into two layers of preset residual networks in sequence to obtain first residual data output by the two layers of preset residual networks;

[0261] Input the second processed data into one layer of preset residual network in sequence to obtain second residual data output by the one layer of preset residual network;

[0262] Obtain the third processed data according to the first residual data and the second residual data.

[0263] Optionally, the fifth processing unit is specifically configured to:

[0264] Input the fifth processed data into a classification neuron to obtain classification data output by the classification neuron;

[0265] Process the classification data by using a Dropout function to obtain processed classification data;

[0266] Input the processed classification data into a Softmax classifier to obtain a training fault label output by the Softmax classifier;

[0267] Based on the training vibration signal data and the training fault labels, the first diagnostic model is obtained.

[0268] It should be noted that the bearing fault diagnosis device provided by the embodiments of the present invention is a device capable of executing the above-mentioned bearing fault diagnosis method. Therefore, all embodiments of the above-mentioned bearing fault diagnosis method are applicable to this device and can achieve the same or similar technical effects.

[0269] As Figure 6 shown, the embodiments of the present invention also provide a bearing fault diagnosis device, including: a processor 600; and a memory 610 connected to the processor 600 through a bus interface. The memory 610 is used to store the programs and data used by the processor 600 when executing operations, and the processor 600 calls and executes the programs and data stored in the memory 610.

[0270] Among them, the receiving end device further includes a transceiver 620. The transceiver 620 is connected to the bus interface and is used to receive and send data under the control of the processor 600;

[0271] Specifically, the processor 600 is used for:

[0272] Obtain the real-time vibration signal data of the bearing to be detected;

[0273] Input the real-time vibration signal data into the fault diagnosis model to obtain the fault type of the bearing to be detected output by the fault diagnosis model;

[0274] Among them, the fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network.

[0275] Optionally, the processor 600 is further used for:

[0276] Obtain the training vibration signal data of bearing wear degradation;

[0277] Obtain experimental data according to the fault labels corresponding to the bearing wear degradation and the training vibration signal data;

[0278] Use the residual mechanism and the multi-dimensional convolutional neural network to train the experimental data to obtain the fault diagnosis model.

[0279] Optionally, the processor 600 is specifically used for:

[0280] Obtain the historical vibration signal data of bearing wear degradation;

[0281] Use the empirical mode decomposition (EMD) algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

[0282] Optionally, the processor 600 is specifically configured to:

[0283] Obtain the maximum envelope data and the minimum envelope data in the historical vibration signal data;

[0284] According to the maximum envelope data and the minimum envelope data, obtain the first difference data corresponding to the maximum envelope data and the second difference data corresponding to the minimum envelope data;

[0285] In the case where the first difference data does not satisfy the first condition and the second difference data does not satisfy the second condition, update the first difference data to the maximum envelope data and update the second difference data to the minimum envelope data, and return to the step of obtaining the first difference data corresponding to the maximum envelope data and the second difference data corresponding to the minimum envelope data according to the maximum envelope data and the minimum envelope data, until the first difference data satisfies the first condition and the second difference data satisfies the second condition, and use the first difference data and the second difference data as the intrinsic mode function IMF;

[0286] Obtain the first residual signal data according to the historical vibration signal data, the first difference data and the second difference data;

[0287] If the first residual signal data is a monotonic function, use the first residual signal data as the training vibration signal data;

[0288] If the first residual signal data is not a monotonic function, update the first residual signal data to the historical vibration signal data, and return to the step of obtaining the maximum envelope data and the minimum envelope data in the historical vibration signal data, until the first residual signal data is a monotonic function, and use the first residual signal data as the training vibration signal data.

[0289] Optionally, the first condition includes at least one of the following:

[0290] The absolute value of the difference between the number of zeros and the number of poles in the maximum envelope data is less than the first threshold;

[0291] The mean value of the first envelope data in the maximum envelope data is zero, and the mean value of the second envelope data in the maximum envelope data is zero, where the first envelope data is the envelope data determined according to the maximum values in the maximum envelope data, and the second envelope data is the envelope data determined according to the minimum values in the maximum envelope data;

[0292] The second condition includes at least one of the following:

[0293] The absolute value of the difference between the number of zeros and the number of poles in the minimum envelope data is less than a second threshold;

[0294] The mean value of the third envelope data in the minimum envelope data is zero, and the mean value of the fourth envelope data in the minimum envelope data is zero, where the third envelope data is envelope data determined according to the maximum values in the minimum envelope data, and the fourth envelope data is envelope data determined according to the minimum values in the minimum envelope data.

[0295] Optionally, the processor 600 is specifically configured to:

[0296] Determine the maximum value data and the minimum value data in the historical vibration signal data;

[0297] Use the cubic spline curve fitting method to construct curve segments for all the maximum value data to obtain the maximum envelope data;

[0298] Use the cubic spline curve fitting method to construct curve segments for all the minimum value data to obtain the minimum envelope data.

[0299] Optionally, the processor 600 is specifically configured to:

[0300] Perform time-domain analysis on the historical vibration signal data to obtain a time-domain analysis result;

[0301] Perform frequency-domain analysis on the historical vibration signal data to obtain a frequency-domain analysis result;

[0302] In the case where the time-domain analysis result and / or the frequency-domain analysis result indicates that the historical vibration signal data is non-stationary vibration signal data, use the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

[0303] Optionally, the processor 600 is specifically configured to:

[0304] Organize the training vibration signal data corresponding to different bearings into a unified length to obtain the processed training vibration signal data;

[0305] Correspond the processed training vibration signal data with the fault labels corresponding to bearing wear degradation to obtain the experimental data.

[0306] Optionally, the processor 600 is specifically configured to:

[0307] Divide the experimental data according to a preset ratio to obtain training data and validation data;

[0308] Train the training data by using the residual mechanism and the multi-dimensional convolutional neural network to obtain a first diagnostic model;

[0309] Input the training vibration signal data in the verification data into the first diagnostic model to obtain a predicted fault label output by the first diagnostic model;

[0310] When the difference rate between the predicted fault label and the fault label in the verification data is less than or equal to a third threshold, determine that the first diagnostic model is the fault diagnostic model;

[0311] When the difference rate between the predicted fault label and the fault label in the verification data is greater than the third threshold, train the first diagnostic model by using the residual mechanism, the multi-dimensional convolutional neural network and the training data to obtain a second training model, update the second training model to the first diagnostic model, and return to the step of inputting the training vibration signal data in the verification data into the first diagnostic model to obtain the predicted fault label output by the first diagnostic model until the difference rate between the predicted fault label and the fault label in the verification data is less than or equal to the third threshold, and determine that the first diagnostic model is the fault diagnostic model.

[0312] Optionally, the processor 600 is specifically configured to:

[0313] Input the training data into a first multi-dimensional convolutional neural network to obtain first processed data output by the first multi-dimensional convolutional neural network;

[0314] Perform feature fusion processing on the first processed data to obtain second processed data;

[0315] Perform residual processing on the second processed data to obtain third processed data;

[0316] Input the third processed data into a second multi-dimensional convolutional neural network to obtain fourth processed data output by the second multi-dimensional convolutional neural network;

[0317] Perform feature fusion processing on the fourth processed data to obtain fifth processed data;

[0318] Perform classification processing on the fifth processed data to obtain the first diagnostic model.

[0319] Optionally, the processor 600 is specifically configured to:

[0320] Divide the training data into first sub-data, second sub-data and third sub-data;

[0321] Perform convolution processing, batch normalization processing, and activation function processing on the first sub-data in sequence to obtain first sub-processed data. Perform convolution processing, batch normalization processing, and activation function processing on the second sub-data in sequence to obtain second sub-processed data. And perform convolution processing, batch normalization processing, and activation function processing on the third sub-data in sequence to obtain third sub-processed data;

[0322] Perform max pooling processing on the first sub-processed data to obtain fourth sub-processed data. And perform max pooling processing on the third sub-processed data to obtain fifth sub-processed data;

[0323] Obtain the first processed data according to the second sub-processed data, the fourth sub-processed data, and the fifth sub-processed data.

[0324] Optionally, the processor 600 is specifically configured to:

[0325] Extract features from the first processed data to obtain feature data;

[0326] Perform data screening and data dimensionality reduction processing on the feature data to obtain processed feature data;

[0327] Perform fusion processing on the processed feature data to obtain the second processed data.

[0328] Optionally, the processor 600 is specifically configured to:

[0329] Input the second processed data into two layers of preset residual networks in sequence to obtain first residual data output by the two layers of preset residual networks;

[0330] Input the second processed data into one layer of preset residual network in sequence to obtain second residual data output by the one layer of preset residual network;

[0331] Obtain the third processed data according to the first residual data and the second residual data.

[0332] Optionally, the processor 600 is specifically configured to:

[0333] Input the fifth processed data into a classification neuron to obtain classification data output by the classification neuron;

[0334] Process the classification data by using a Dropout function to obtain processed classification data;

[0335] Input the processed classification data into a Softmax classifier to obtain training fault labels output by the Softmax classifier;

[0336] Based on the training vibration signal data and the training fault labels, the first diagnosis model is obtained.

[0337] Among them, in Figure 6 The bus architecture can include any number of interconnected buses and bridges, specifically various circuits of one or more processors represented by the processor 600 and the memory represented by the memory 610 are linked together. The bus architecture can also link together various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art, so they will not be further described herein. The bus interface provides the user interface 630. The transceiver 620 can be multiple components, that is, including a transmitter and a receiver, and provides a unit for communicating with various other devices on the transmission medium. The processor 600 is responsible for managing the bus architecture and general processing, and the memory 610 can store the data used by the processor 600 when performing operations.

[0338] In addition, a specific embodiment of the present invention further provides a readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps in the bearing fault diagnosis method described in any one of the above are implemented.

[0339] The above are the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements can be made without departing from the principles described in the present invention, and these improvements and refinements are also within the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method, characterized in that, it includes: Obtain the real-time vibration signal data of the bearing to be detected; Input the real-time vibration signal data into the fault diagnosis model to obtain the fault type of the bearing to be detected output by the fault diagnosis model; wherein, the fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network.

2. The method according to claim 1, characterized in that, the method further includes: Obtain the training vibration signal data of bearing wear degradation; Obtain experimental data according to the fault label corresponding to the bearing wear degradation and the training vibration signal data; Use the residual mechanism and the multi-dimensional convolutional neural network to train the experimental data to obtain the fault diagnosis model.

3. The method according to claim 2, characterized in that, the obtaining of the training vibration signal data of bearing wear degradation includes: Obtain the historical vibration signal data of bearing wear degradation; Use the empirical mode decomposition (EMD) algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

4. The method according to claim 3, characterized in that, using the empirical mode decomposition (EMD) algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data includes: Obtain the maximum envelope data and the minimum envelope data in the historical vibration signal data; According to the maximum envelope data and the minimum envelope data, obtain the first difference data corresponding to the maximum envelope data and the second difference data corresponding to the minimum envelope data; In the case where the first difference data does not meet the first condition and the second difference data does not meet the second condition, update the first difference data to the maximum envelope data and update the second difference data to the minimum envelope data, and return to the step of obtaining the first difference data corresponding to the maximum envelope data and the second difference data corresponding to the minimum envelope data according to the maximum envelope data and the minimum envelope data, until the first difference data meets the first condition and the second difference data meets the second condition, and use the first difference data and the second difference data as the intrinsic mode function (IMF); Obtain the first residual signal data according to the historical vibration signal data, the first difference data and the second difference data; If the first residual signal data is a monotonic function, use the first residual signal data as the training vibration signal data; If the first residual signal data is not a monotonic function, update the first residual signal data to the historical vibration signal data, and return to the step of obtaining the maximum envelope data and the minimum envelope data in the historical vibration signal data, until the first residual signal data is a monotonic function, and use the first residual signal data as the training vibration signal data.

5. The method according to claim 4, characterized in that, the first condition includes at least one of the following: The absolute value of the difference between the number of zeros and the number of poles in the maximum envelope data is less than a first threshold; The mean value of the first envelope data in the maximum envelope data is zero, and the mean value of the second envelope data in the maximum envelope data is zero, where the first envelope data is envelope data determined according to the maximum values in the maximum envelope data, and the second envelope data is envelope data determined according to the minimum values in the maximum envelope data; The second condition includes at least one of the following: The absolute value of the difference between the number of zeros and the number of poles in the minimum envelope data is less than a second threshold; The mean value of the third envelope data in the minimum envelope data is zero, and the mean value of the fourth envelope data in the minimum envelope data is zero, where the third envelope data is envelope data determined according to the maximum values in the minimum envelope data, and the fourth envelope data is envelope data determined according to the minimum values in the minimum envelope data.

6. The method according to claim 4, wherein, Obtaining the maximum envelope data and the minimum envelope data in the historical vibration signal data includes: Determining the maximum value data and the minimum value data in the historical vibration signal data; Using the cubic spline curve fitting method to construct curve segments for all the maximum value data to obtain the maximum envelope data; Using the cubic spline curve fitting method to construct curve segments for all the minimum value data to obtain the minimum envelope data.

7. The method according to claim 3, wherein, Using the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data includes: Performing time-domain analysis on the historical vibration signal data to obtain a time-domain analysis result; Performing frequency-domain analysis on the historical vibration signal data to obtain a frequency-domain analysis result; In the case where the time-domain analysis result and / or the frequency-domain analysis result indicates that the historical vibration signal data is non-stationary vibration signal data, using the EMD algorithm to perform adaptive decomposition processing on the historical vibration signal data to obtain the training vibration signal data.

8. The method according to claim 2, wherein, Obtaining experimental data according to the fault label corresponding to the bearing wear degradation and the training vibration signal data includes: Sorting the training vibration signal data corresponding to different bearings into a unified length to obtain processed training vibration signal data; Corresponding the processed training vibration signal data with the fault label corresponding to the bearing wear degradation to obtain the experimental data.

9. The method according to claim 2, wherein, Using the residual mechanism and the multi-dimensional convolutional neural network to train the experimental data to obtain the fault diagnosis model includes: Dividing the experimental data according to a preset ratio to obtain training data and validation data; Using the residual mechanism and the multi-dimensional convolutional neural network to train the training data to obtain a first diagnosis model; Input the training vibration signal data in the verification data into the first diagnostic model to obtain the predicted fault label output by the first diagnostic model; When the difference rate between the predicted fault label and the fault label in the verification data is less than or equal to the third threshold, determine that the first diagnostic model is the fault diagnostic model; When the difference rate between the predicted fault label and the fault label in the verification data is greater than the third threshold, use the residual mechanism, the multi-dimensional convolutional neural network, and the training data to train the first diagnostic model to obtain a second training model, update the second training model to the first diagnostic model, and return to the step of inputting the training vibration signal data in the verification data into the first diagnostic model to obtain the predicted fault label output by the first diagnostic model until the difference rate between the predicted fault label and the fault label in the verification data is less than or equal to the third threshold, and determine that the first diagnostic model is the fault diagnostic model.

10. The method according to claim 9, wherein, Training the training data using the residual mechanism and the multi-dimensional convolutional neural network to obtain a first diagnostic model includes: Input the training data into the first multi-dimensional convolutional neural network to obtain the first processed data output by the first multi-dimensional convolutional neural network; Perform feature fusion processing on the first processed data to obtain second processed data; Perform residual processing on the second processed data to obtain third processed data; Input the third processed data into the second multi-dimensional convolutional neural network to obtain the fourth processed data output by the second multi-dimensional convolutional neural network; Perform feature fusion processing on the fourth processed data to obtain fifth processed data; Perform classification processing on the fifth processed data to obtain the first diagnostic model.

11. The method according to claim 10, wherein, Inputting the training data into the first multi-dimensional convolutional neural network to obtain the first processed data output by the first multi-dimensional convolutional neural network includes: Divide the training data into first sub-data, second sub-data, and third sub-data; Perform convolution processing, batch normalization processing, and activation function processing on the first sub-data in sequence to obtain first sub-processed data, perform convolution processing, batch normalization processing, and activation function processing on the second sub-data in sequence to obtain second sub-processed data, and perform convolution processing, batch normalization processing, and activation function processing on the third sub-data in sequence to obtain third sub-processed data; Perform max pooling processing on the first sub-processed data to obtain fourth sub-processed data, and perform max pooling processing on the third sub-processed data to obtain fifth sub-processed data; Obtain the first processed data according to the second sub-processed data, the fourth sub-processed data, and the fifth sub-processed data.

12. The method according to claim 10, wherein, Performing feature fusion processing on the first processed data to obtain second processed data includes: Extract features from the first processed data to obtain feature data; Perform data screening and data dimensionality reduction processing on the feature data to obtain processed feature data; Perform fusion processing on the processed feature data to obtain the second processed data.

13. The method according to claim 10, wherein, Performing residual processing on the second processed data to obtain third processed data, including: Sequentially inputting the second processed data into two preset residual networks to obtain first residual data output by the two preset residual networks; Sequentially inputting the second processed data into one preset residual network to obtain second residual data output by the one preset residual network; Obtaining the third processed data according to the first residual data and the second residual data.

14. The method according to claim 10, wherein, Performing classification processing on the fifth processed data to obtain the first diagnostic model, including: Inputting the fifth processed data into a classification neuron to obtain classification data output by the classification neuron; Processing the classification data by using a Dropout function to obtain processed classification data; Inputting the processed classification data into a Softmax classifier to obtain a training fault label output by the Softmax classifier; Obtaining the first diagnostic model according to the training vibration signal data and the training fault label.

15. A bearing fault diagnosis device, wherein, including: A first acquisition module for acquiring real-time vibration signal data of a bearing to be detected; A first processing module for inputting the real-time vibration signal data into a fault diagnosis model to obtain a fault type of the bearing to be detected output by the fault diagnosis model; wherein, the fault diagnosis model is trained by using a residual mechanism and a multi-dimensional convolutional neural network.

16. A bearing fault diagnosis device, wherein, including: A processor, a memory, and a program stored on the memory and executable on the processor, and when the program is executed by the processor, the steps of the bearing fault diagnosis method according to any one of claims 1 to 14 are implemented.

17. A readable storage medium, wherein, A program is stored on the readable storage medium, and when the program is executed by a processor, the steps in the bearing fault diagnosis method according to any one of claims 1 to 14 are implemented.

Images