Bearing fault diagnosis method based on multi-input and multi-channel parallel convolutional neural network
Through multi-input multi-channel parallel convolutional neural network in bearing fault diagnosis, overlapping interception and feature extraction methods are used to solve the problem of insufficient accuracy and reliability of bearing fault diagnosis in the existing technology, and achieve higher fault identification Accuracy and reliability.
Patent Information
- Application Number
- CN202411203800.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-30
AI Technical Summary
The accuracy and reliability of existing bearing fault diagnosis methods are limited, and a single type of data cannot fully and effectively express the characteristics of fault signals, resulting in a low accuracy of deep learning fault diagnosis methods.
Multi-input multi-channel parallel convolutional neural network is adopted to record vibration time series by installing vibration sensors on the bearing shell, perform overlapping interception and feature extraction, build a fusion abstract feature data set, and use one-dimensional and two-dimensional convolution modules and attention The mechanism module and the fully connected module perform fault identification and diagnosis.
The accuracy and reliability of bearing fault diagnosis are improved, and the generalization ability of the model and the reliability of fault feature indication are enhanced through a variety of training samples and feature extraction methods.
Smart Images

Figure CN118734190B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to a bearing fault diagnosis method based on a multi-input multi-channel parallel convolutional neural network. Background Art
[0002] Bearings are important parts that support the operation of mechanical equipment and transmit power. They are widely used in important fields such as vehicles, ships, and aerospace. Due to long-term exposure to alternating loads and erosion in harsh working environments, bearings are prone to plastic deformation, fatigue fracture, pitting, peeling and other faults. However, since the characteristics of bearing failures are weak and difficult to detect, it is very easy to ignore bearing failures, which can lead to mechanical equipment shutdown, damage or even serious accidents. It can be seen that real-time monitoring of the health status of bearings and high-precision fault diagnosis are of great significance to ensure the normal operation of mechanical equipment.
[0003] At present, there are two main types of bearing fault diagnosis methods: traditional methods and deep learning methods. Traditional fault diagnosis methods are mostly based on signal analysis and kinematic analysis. The algorithms are cumbersome and highly dependent on expert experience to judge the results, which limits the application of the methods. Deep learning fault diagnosis methods use deep learning network models to automatically and intelligently learn the characteristics of input data, classify data types based on the extracted high-dimensional abstract features, and judge the type of fault. Currently, the commonly used deep learning fault diagnosis methods usually directly learn fault features from a single original fault vibration signal and perform fault identification and classification. Although they can get rid of the limitations of expert experience and application occasions, due to the diversity of fault manifestations, a single type of data cannot fully and effectively express the characteristics of the fault signal, resulting in a low accuracy rate of deep learning fault diagnosis methods. Summary of the invention
[0004] In view of the above problems, the present invention provides a bearing fault diagnosis method based on a multi-input multi-channel parallel convolutional neural network, which solves the technical problems of limited bearing fault diagnosis accuracy and reliability in the prior art.
[0005] The present invention provides a bearing fault diagnosis method based on a multi-input multi-channel parallel convolutional neural network, comprising the following steps:
[0006] Step S1, installing a vibration sensor on the housing of the bearing to be diagnosed, and recording a vibration time series generated when the bearing to be diagnosed is running as an original vibration signal;
[0007] Step S2, overlapping and intercepting the original vibration signal to obtain a sample signal;
[0008] Step S3, extracting features from the sample signal to obtain time-domain vibration waveform features, statistical parameter features, and time-frequency domain image features; constructing a fused abstract feature data set based on the time-domain vibration waveform features, statistical parameter features, and time-frequency domain image features of the sample signal;
[0009] Step S4, dividing the fused abstract feature data set into a training set and a test set, training a multi-input multi-channel parallel convolutional neural network based on the training set, inputting the fused abstract features in the test set into the trained multi-input multi-channel parallel convolutional neural network, and obtaining the fault category of the bearing to be diagnosed.
[0010] Preferably, step S2 specifically includes: sliding a fixed time distance on the original vibration signal through a time window of preset length, intercepting the value of the original vibration signal at the position of the time window, obtaining each segment of the intercepted signal, the each segment of the intercepted signal overlaps, and determining the set of each segment of the intercepted signal as the sample signal.
[0011] Preferably, the intercept signal The way is as follows:
[0012]
[0013]
[0014] in, is the index of the intercepted signal obtained, For the Segment interception signal, is the original vibration signal, is the moving distance of the time window, represents the length of the time window, For time.
[0015] Preferably, step S3 specifically includes:
[0016] Step S3-1, performing variational modal decomposition on the sample signal to obtain frequency components of the sample signal, selecting and reconstructing the frequency components based on an optimal frequency component selection index, performing Hilbert transform on the reconstructed signal to obtain an envelope signal, and determining the envelope signal as the time domain vibration waveform feature;
[0017] Step S3-2, calculating the mean square value, variance, kurtosis, margin factor, permutation entropy and centroid frequency of the envelope signal as the statistical parameter features;
[0018] Step S3-3, performing time-frequency analysis on the envelope signal to obtain a frequency-time distribution diagram of the envelope signal as the time-frequency domain image feature;
[0019] Step S3-4: taking the set of the time-domain vibration waveform features, statistical parameter features and time-frequency domain image features of the sample signal as the fused abstract feature data set.
[0020] Preferably, step S3-1 specifically includes:
[0021] The constraint model of variational mode decomposition is as follows:
[0022]
[0023] in, represents the first i frequency components, i is the frequency component index, Indicates i The center frequency corresponding to the frequency component is represents the number of decomposed signals, represents the derivative symbol, represents the Dirichlet function, represents the imaginary unit, represents pi, * represents convolution operation, represents the L2 norm, It means to traverse each frequency component and the center frequency so that the expression in the brackets reaches the minimum value. For the Segment interception signal;
[0024] Through variational mode decomposition, the frequency components of the sample signal are obtained, and the optimal frequency component selection index is calculated for each frequency component. The value of is calculated as:
[0025]
[0026] in, represents the kurtosis value of the frequency component, represents the permutation entropy value of the frequency component, The entropy of the envelope spectrum amplitude representing the frequency component;
[0027] Select indicators The largest frequency component is used as the basis for signal reconstruction to obtain the reconstructed signal , for the reconstructed signal Perform Hilbert transform to obtain the envelope signal , the envelope signal Directly used as the time domain vibration waveform characteristics of the bearing to be diagnosed.
[0028] Preferably, in step S4:
[0029] The multi-input multi-channel parallel convolutional neural network is composed of a feature extractor and a classifier 110, wherein the feature extractor includes a one-dimensional convolution module 104, a two-dimensional convolution module 103, an attention mechanism module 106 and a fully connected module;
[0030] The set of time-domain vibration waveform features is taken as a one-dimensional feature set 102 and input into a one-dimensional convolution module 104; the set of time-frequency image features is taken as a two-dimensional feature set 101 and input into a two-dimensional convolution module 103; the outputs of the one-dimensional convolution module 104 and the two-dimensional convolution module 103 are calculated by the attention mechanism module 106 and then input into the fully connected module; the set of statistical parameter features is taken as a statistical parameter feature set 105 and input into the fully connected module; the output of the fully connected module is classified by the classifier 110 to obtain a fault category 111.
[0031] Preferably, the calculation process of the two-dimensional convolution module 103 is: the input is sequentially passed through a 3×3 convolution layer 211, a first normalization layer 212, a first activation function layer 213, a 3×3 convolution layer 211, a first normalization layer 212 and a first activation function layer 213 to obtain an output;
[0032] The calculation process of the one-dimensional convolution module 104 is: the input is sequentially passed through a 1×3 convolution layer 221, a second batch of normalization layers 222, a second activation function layer 223, a 1×3 convolution layer 221, a second batch of normalization layers 222, and a second activation function layer 223 to obtain an output.
[0033] Preferably, the calculation process of the attention mechanism module 106 is as follows: the input feature map 31 of the attention mechanism module 106 is composed of the outputs of the two-dimensional convolution module 103 and the one-dimensional convolution module 104, and each channel of the input feature map 31 is subjected to spatial dimension information aggregation 35 to obtain an aggregated vector 32, and then the aggregated vector 32 is subjected to a fully connected layer bottleneck calculation 36 to generate a weight vector 33, and based on the weight vector 33, the input feature map 31 is weighted, and the calculation of the proportional adjustment 37 is completed to obtain the calibrated feature map 34 that is finally output by the attention mechanism module 106.
[0034] Preferably, the fully connected module comprises a first fully connected layer 107, a second fully connected layer 108 and a third fully connected layer 109 which are connected sequentially, and the input of the fully connected module is calculated by the first fully connected layer 107, the second fully connected layer 108 and the third fully connected layer 109 in sequence, and outputs the abstract features;
[0035] The classifier 110 is a Softmax classifier, which inputs the abstract features, obtains the probability that the current bearing to be diagnosed belongs to each fault category, and takes the fault category with the highest probability as the output fault category.
[0036] Compared with the prior art, the present invention has at least the following beneficial effects:
[0037] (1) By intercepting signals in different time periods and using partially overlapping windows, the present invention can better capture the characteristics and changing trends of signals within a local time range, reduce the impact of noise and outliers, and generate more samples, thereby increasing the number and diversity of training samples, thereby improving the generalization ability of the model.
[0038] (2) After performing variational mode decomposition on the sample signal, the present invention provides an optimal frequency component index that is positively correlated with the kurtosis value and negatively correlated with the permutation entropy value and the envelope spectrum amplitude entropy for signal reconstruction, which can obtain a reconstructed signal with more obvious characteristics and improve the reliability of fault feature indication.
[0039] (3) The present invention provides a multi-input multi-channel parallel convolutional neural network, which utilizes the time-domain vibration waveform characteristics, statistical parameter characteristics and time-frequency image characteristics of the envelope signal of the input sample signal, combines the attention mechanism module, the fully connected layer and the softmax classifier, and performs fault identification and diagnosis based on fault features of different dimensions, thereby improving the accuracy and reliability of the diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] The drawings are only for the purpose of illustrating particular embodiments and are not to be construed as limiting the invention.
[0041] Figure 1 A flow chart of a bearing fault diagnosis method based on a multi-input and multi-channel parallel convolutional neural network provided by the present invention.
[0042] Figure 2 Schematic diagram of the multi-input multi-channel parallel convolutional neural network provided by the present invention.
[0043] Figure 3 Structural diagram of the one-dimensional and two-dimensional convolution modules provided by the present invention.
[0044] Figure 4 This is a structural diagram of the attention mechanism module provided by the present invention.
[0045] Figure 5 This is a schematic diagram of overlapping cuts provided by the present invention.
[0046] Figure 6 This is an example diagram of the time-frequency diagram provided by the present invention.
[0047] Figure numerals: 101-two-dimensional feature set, 102-one-dimensional feature set, 103-two-dimensional convolution module, 104-one-dimensional convolution module, 105-statistical parameter feature set, 106-attention mechanism module, 107-first fully connected layer, 108-second fully connected layer, 109-third fully connected layer, 110-classifier, 111-fault category, 21-two-dimensional convolution module structure, 211-3×3 convolution layer, 212-first batch of normalization layers, 213-first activation function layer, 22-one-dimensional convolution module structure, 221-1×3 convolution layer, 222-second batch of normalization layers, 223-second activation function layer, 31-input feature map, 32-aggregation vector, 33-weight vector, 34-calibrated feature map, 35-spatial dimension information aggregation, 36-fully connected layer bottleneck calculation, 37-proportion adjustment, 41-first sample, 42-second sample. DETAILED DESCRIPTION
[0048] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.
[0049] In order to illustrate the effectiveness of the method proposed by the present invention, the above technical solution of the present invention is described in detail below through a specific embodiment. Figure 1 As shown, a bearing fault diagnosis method based on a multi-input multi-channel parallel convolutional neural network is disclosed, and the specific implementation steps are as follows:
[0050] Step S1: Install a vibration sensor on the housing of the bearing to be diagnosed, and record a vibration time series generated when the bearing to be diagnosed is running as an original vibration signal.
[0051] The original vibration signal is the vibration data collected directly from the vibration sensor without being processed or analyzed. It can reflect the vibration characteristics generated by the bearing during operation, including the health status of the bearing, the inherent vibration mode of the mechanical structure, and possible fault information. The original vibration signal is usually expressed in the form of a time series, recording the change of vibration intensity over time. The original vibration signal is the basis for bearing fault diagnosis. By analyzing and processing the original vibration signal, the working status and potential faults of the bearing, such as wear, cracks, failure, etc., can be identified.
[0052] In this step, the original vibration signal can be obtained by installing a vibration sensor on the housing or nearby position of the bearing to be diagnosed in a laboratory or on site to monitor and record the vibration signal generated during the operation of the bearing in real time, or by obtaining the recorded bearing vibration signal from a publicly available vibration signal database.
[0053] In some embodiments, the original vibration data uses part of the data in the public bearing fault data set of Case Western Reserve University. The specific data used is shown in Table 1:
[0054] Table 1 Data examples in the bearing fault dataset
[0055]
[0056] Step S2: overlapping and intercepting the original vibration signal to obtain a sample signal.
[0057] The original vibration signal is expressed as ,in, is the time. For the original vibration signal The specific method of interception is to slide the original vibration signal for a fixed time distance through a time window of a specific length, intercept the value of the original vibration signal at the position of the time window, and obtain each segment of the intercepted signal. , the intercepted signals of each segment are obtained The way is as follows:
[0058]
[0059]
[0060] in, is the index of the intercepted signal obtained, For the Segment interception signal, is the original vibration signal, is the moving distance of the time window, represents the length of the time window, is the time. By the above method, the first Time series values to All values between the time series values are intercepted as The time window is continuously sliding in this way. When the length of the sliding window is Greater than the moving distance of the sliding window When the signal is intercepted, there is overlap between the intercepted signals.
[0061] In this way, intercepting the signal at different time periods can better capture the characteristics and changing trends of the signal within the local time range, and reduce the impact of noise and outliers on the overall signal analysis; and, since the windows can be partially overlapped, compared with non-overlapping interception, more samples can be generated from the same original signal. These samples contain different features and patterns, which significantly increases the number and diversity of training samples, and thus helps to improve the generalization ability of the model in the subsequent training process.
[0062] The process of overlapping interception is as follows Figure 5 As shown, the sliding window length is , the moving distance of the sliding window The sliding window overlap interception can obtain multiple samples such as the first sample 41 and the second sample 42. After the samples are expanded by sliding window overlap interception, the number of samples of each state of the bearing is shown in Table 2:
[0063] Table 2 Number of samples of bearings in different states
[0064]
[0065] Afterwards, the present invention intercepts each segment of the signal The set of samples is determined as ,Right now .
[0066] Step S3, extracting features from the sample signal to obtain time-domain vibration waveform features, statistical parameter features, and time-frequency domain image features; and constructing a fused abstract feature data set based on the time-domain vibration waveform features, statistical parameter features, and time-frequency domain image features of the sample signal.
[0067] Normally, the original signal is directly divided into samples, and the fault characteristics in the obtained samples are not obvious. In order to highlight the fault characteristics, the sample signal is firstly subjected to variational mode decomposition to construct the optimal frequency component selection index. , select the frequency component with the most obvious fault characteristics as the signal with fault characteristics. The constructed constraint model of variational mode decomposition is as follows:
[0068]
[0069] in represents the first i frequency components, i is the frequency component index, Indicates i The center frequency corresponding to the frequency component is represents the number of decomposed signals, represents the derivative symbol, represents the Dirichlet function, represents the imaginary unit, represents pi, * represents convolution operation, represents the L2 norm, It means to traverse each frequency component and center frequency so that the expression in the brackets reaches the minimum value. For the Segment interception signal.
[0070] Through variational mode decomposition, the frequency components of the sample signal are obtained, and the optimal frequency component selection index is calculated for each frequency component. The value of and select the indicator The largest frequency component is used as the basis for signal reconstruction to obtain a reconstructed signal with more obvious characteristics. .
[0071] Optimal frequency component selection index It is used to indicate the obviousness of the fault characteristics and is calculated as follows:
[0072]
[0073] in, represents the kurtosis value of the frequency component, represents the permutation entropy value of the frequency component, Represents the envelope spectrum amplitude entropy of the frequency component.
[0074] The kurtosis value reflects the degree of signal spikes. A higher kurtosis value usually means that there are obvious mutations or impact components in the signal, which is used to reflect fault characteristics; the permutation entropy value is used to measure the complexity and randomness of the signal. A lower permutation entropy value indicates that the signal has stronger regularity and certainty, which helps to highlight the fault characteristics; the envelope spectrum amplitude entropy reflects the complexity of the signal envelope spectrum. A lower envelope spectrum amplitude entropy means that the frequency components of the envelope spectrum are more concentrated, which helps to highlight the fault characteristic frequency. The optimal frequency component selection index of the present invention is positively correlated with the kurtosis value, and negatively correlated with the permutation entropy value and the envelope spectrum amplitude entropy. It can comprehensively reflect the obvious degree of the fault characteristics, and can balance the advantages and disadvantages of each index, avoid the errors and instability that may be caused by a single index, thereby improving the reliability of indicating the obvious degree of the fault characteristics.
[0075] Reconstructed signal Perform Hilbert transform to obtain the envelope signal , the Hilbert transform is as follows:
[0076]
[0077] in, represents the Hilbert transform, For time, is the integration variable.
[0078] According to the Hilbert demodulation principle, the envelope signal This is the bearing fault signal, so the envelope signal Directly used as the time domain vibration waveform feature set of the bearing to be diagnosed ;
[0079] Calculate the envelope signal The mean square value, variance, kurtosis, margin factor, permutation entropy and center of gravity frequency constitute the statistical parameter feature set of the bearing to be diagnosed. .
[0080] The statistical features constructed in this way by the present invention can reflect the potential fault characteristics in the signal. Among them, the kurtosis and margin factor can reveal the impact fault, the permutation entropy and the centroid frequency can reflect the complexity and frequency distribution characteristics of the signal, which helps to indicate the fault characteristics more clearly. These statistical parameters cover the amplitude, distribution form, dynamic characteristics and complexity of the signal and other aspects of information, which can provide multi-dimensional feature descriptions and help to fully capture the characteristics of the signal.
[0081] Statistical parameter feature set Some values are shown in Table 3:
[0082] Table 3 Examples of statistical parameter feature set values
[0083]
[0084] Using time-frequency analysis methods such as wavelet transform and short-time Fourier transform, the envelope signal Analyze and obtain the frequency-time distribution diagram of each frequency component of the envelope signal over time, which is used as the time-frequency image feature set of the bearing to be diagnosed. .
[0085] The frequency-time distribution diagram obtained by the time-frequency analysis method of the present invention can represent the signal in both time and frequency dimensions at the same time, can capture the transient characteristics and non-stationary characteristics of the signal, can more comprehensively reflect the complexity and changing laws of the signal, provide more accurate features for subsequent network models, and can improve the diagnostic performance of the model.
[0086] In some embodiments, the time-frequency image feature set is obtained The specific method can be: Based on the Python wavelet analysis toolkit PyWavelets for envelope signals Perform wavelet transform and obtain wavelet coefficient C. Figure 6The figure shows the frequency-time distribution diagram drawn based on the pcolormesh function in the matplotlib library according to the time t, the actual sequence frequency fs and the wavelet coefficient C, and then the time-frequency image feature set is obtained. Finally, the feature set of the bearing to be diagnosed is obtained. .
[0087] In some embodiments, the above feature extraction operation is performed on 2000 clips in the sample signal to obtain a feature set T_all of 2000 samples as the fused abstract feature data set.
[0088] Step S4, dividing the fused abstract feature data set into a training set and a test set, training a multi-input multi-channel parallel convolutional neural network based on the training set, inputting the fused abstract features in the test set into the trained multi-input multi-channel parallel convolutional neural network, and obtaining the fault category of the bearing to be diagnosed.
[0089] The structure of multi-input multi-channel parallel convolutional neural network is as follows Figure 2 As shown, it is composed of a feature extractor and a classifier 110, wherein the feature extractor includes a one-dimensional convolution module 104, a two-dimensional convolution module 103, an attention mechanism module 106 and a fully connected module;
[0090] The time domain vibration waveform feature set As a one-dimensional feature set 102, input into a one-dimensional convolution module 104; the time-frequency image feature set As the two-dimensional feature set 101, it is input into the two-dimensional convolution module 103; the outputs of the one-dimensional convolution module 104 and the two-dimensional convolution module 103 are calculated by the attention mechanism module 106 and then input into the fully connected module; the statistical parameter feature set 105 is input into the fully connected module, and the output of the fully connected module is classified by the classifier 110 to obtain the fault category 111.
[0091] like Figure 3 As shown, the calculation process of the two-dimensional convolution module 103 is: the input is sequentially passed through a 3×3 convolution layer 211, a first normalization layer 212, a first activation function layer 213, a 3×3 convolution layer 211, a first normalization layer 212 and a first activation function layer 213 to obtain an output. The activation function in the first activation function layer 213 is a ReLU function.
[0092] The calculation process of the one-dimensional convolution module 104 is: the input is sequentially passed through the 1×3 convolution layer 221, the second batch normalization layer 222, the second activation function layer 223, the 1×3 convolution layer 221, the second batch normalization layer 222 and the second activation function layer 223 to obtain the output. The activation function in the second activation function layer 223 is the ReLU function.
[0093] like Figure 4 As shown, the calculation process of the attention mechanism module 106 is as follows: the input feature map 31 of the attention mechanism module 106 is composed of the outputs of the two-dimensional convolution module 103 and the one-dimensional convolution module 104, and each channel of the input feature map 31 is subjected to spatial dimension information aggregation 35 to obtain an aggregation vector 32, and then the aggregation vector 32 is subjected to a fully connected layer bottleneck calculation 36 to generate a weight vector 33, and based on the weight vector 33, the input feature map 31 is weighted, and the feature map of each channel is assigned a corresponding weight, and the calculation of the ratio adjustment 37 is completed to obtain the calibrated feature map 34 that is finally output by the attention mechanism module 106.
[0094] The fully connected module includes a first fully connected layer 107, a second fully connected layer 108 and a third fully connected layer 109 which are connected sequentially. The input of the fully connected module is calculated by the first fully connected layer 107, the second fully connected layer 108 and the third fully connected layer 109 in sequence, and the extracted abstract features are output. .
[0095] The output of the fully connected module is the extracted abstract features ; The classifier 110 is a Softmax classifier, and the calculation function of the Softmax classifier is:
[0096]
[0097] in, is the probability value of the output bearing to be diagnosed belonging to category n, is the index of the current category, is the index of all possible categories, For input samples in category The score on For input samples in category The score on .
[0098] The Softmax classifier inputs the abstract features extracted by the feature extractor , output the probability that the current bearing to be diagnosed belongs to each fault category, and take the fault category with the highest probability as the network's diagnosis result for the current bearing to be diagnosed.
[0099] In some embodiments, for the feature set T_all of the 2000 samples, the feature set T_all is converted into a tensor and encapsulated based on the Dataset function in pytorch, and the tensor sample feature set is loaded using the DataLoader in pytorch to start multi-input multi-channel parallel convolutional neural network training.
[0100] The tensor sample feature set is divided into a training set Train and a test set Test in a ratio of 80%:20%. The model is trained on the training set Train, and the diagnostic accuracy of the network is checked on the test set Test. The multi-channel convolutional neural network training and parameter adjustment are based on the Adam optimization algorithm, and the loss during the network training process is calculated using the cross entropy loss function.
[0101] The multi-input with channel-parallel convolutional neural network (MCPCNN) bearing fault diagnosis method provided by the present invention is compared with the existing residual convolutional neural network (ResNet) bearing fault diagnosis method. Both methods perform network training and testing according to the ratio of training set to test set of 80%:20%, and both use Adam optimization algorithm for network training and parameter update. The loss in the network training process is calculated by cross entropy loss function.
[0102] The diagnostic accuracy of the two methods on the test set is shown in Table 4:
[0103] Table 4 Comparison of diagnostic accuracy
[0104]
[0105] It can be seen that the bearing fault diagnosis method based on the multi-input multi-channel parallel convolutional neural network provided by the present invention can achieve a higher fault classification accuracy.
[0106] Although the specific embodiments of the present invention depict various actions or steps in a specific order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in a sequential order, or requiring all illustrated actions or steps to be performed to obtain the desired results. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Certain features described in the context of a separate embodiment can also be implemented in a single implementation in combination. On the contrary, the various features described in the context of a single implementation can also be implemented in multiple implementations individually or in any suitable sub-combination. The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
[0107] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. A bearing fault diagnosis method based on multi-input multi-channel parallel convolutional neural network, characterized in that: The following steps are involved: Step S1, installing a vibration sensor on the housing of the bearing to be diagnosed, and recording a vibration time series generated when the bearing to be diagnosed is running as an original vibration signal; Step S2, overlapping and intercepting the original vibration signal to obtain a sample signal; Step S3, extracting features from the sample signal to obtain time-domain vibration waveform features, statistical parameter features, and time-frequency domain image features; Constructing a fused abstract feature data set based on the time-domain vibration waveform features, statistical parameter features, and time-frequency domain image features of the sample signal; Step S4, dividing the fused abstract feature data set into a training set and a test set, training a multi-input multi-channel parallel convolutional neural network based on the training set, inputting the fused abstract features in the test set into the trained multi-input multi-channel parallel convolutional neural network, and obtaining the fault category of the bearing to be diagnosed; Step S2 specifically includes: Sliding a fixed time distance on the original vibration signal through a time window of a preset length, intercepting the value of the original vibration signal at the position of the time window, and obtaining each segment of intercepted signal, wherein the each segment of intercepted signal overlaps, and determining the set of each segment of intercepted signal as the sample signal; Intercepting signals The expression is as follows: in, is the index of the intercepted signal obtained, For the Segment interception signal, is the original vibration signal, is the moving distance of the time window, represents the length of the time window, For time; Step S3 specifically includes: Step S3-1, performing variational modal decomposition on the sample signal to obtain frequency components of the sample signal, selecting and reconstructing the frequency components based on an optimal frequency component selection index, performing Hilbert transform on the reconstructed signal to obtain an envelope signal, and determining the envelope signal as the time domain vibration waveform feature; Step S3-2, calculating the mean square value, variance, kurtosis, margin factor, permutation entropy and centroid frequency of the envelope signal as the statistical parameter features; Step S3-3, performing time-frequency analysis on the envelope signal to obtain a frequency-time distribution diagram of the envelope signal as the time-frequency domain image feature; Step S3-4, taking the set of the time-domain vibration waveform features, statistical parameter features and time-frequency domain image features of the sample signal as the fused abstract feature data set; Step S3-1 specifically includes: The constraint model of variational mode decomposition is as follows: in, represents the first i frequency components, i is the frequency component index, Indicates i The center frequency corresponding to the frequency component is represents the number of decomposed signals, represents the derivative symbol, represents the Dirichlet function, represents the imaginary unit, represents pi, * represents convolution operation, represents the L2 norm, It means to traverse each frequency component and the center frequency so that the expression in the brackets reaches the minimum value. For the Segment interception signal; Through variational mode decomposition, the frequency components of the sample signal are obtained, and the optimal frequency component selection index is calculated for each frequency component. The value of is calculated as: in, represents the kurtosis value of the frequency component, represents the permutation entropy value of the frequency component, The entropy of the envelope spectrum amplitude representing the frequency component; Select indicators The largest frequency component is used as the basis for signal reconstruction to obtain the reconstructed signal , for the reconstructed signal Perform Hilbert transform to obtain the envelope signal , the envelope signal Directly used as the time domain vibration waveform characteristics of the bearing to be diagnosed.
2. The bearing fault diagnosis method based on multi-input multi-channel parallel convolutional neural network according to claim 1 is characterized in that: In step S4: The multi-input multi-channel parallel convolutional neural network is composed of a feature extractor and a classifier (110), wherein the feature extractor includes a one-dimensional convolution module (104), a two-dimensional convolution module (103), an attention mechanism module (106) and a fully connected module; The set of time-domain vibration waveform features is used as a one-dimensional feature set (102) and input into a one-dimensional convolution module (104); the set of time-frequency domain image features is used as a two-dimensional feature set (101) and input into a two-dimensional convolution module (103); the outputs of the one-dimensional convolution module (104) and the two-dimensional convolution module (103) are calculated by an attention mechanism module (106) and then input into the fully connected module; the set of statistical parameter features is used as a statistical parameter feature set (105) and input into the fully connected module; the output of the fully connected module is classified by a classifier (110) to obtain a fault category (111).
3. The bearing fault diagnosis method based on multi-input multi-channel parallel convolutional neural network according to claim 2 is characterized in that: The calculation process of the two-dimensional convolution module (103) is as follows: the input is sequentially passed through a 3×3 convolution layer (211), a first batch of normalization layers (212), a first activation function layer (213), a 3×3 convolution layer (211), a first batch of normalization layers (212) and a first activation function layer (213) to obtain an output; The calculation process of the one-dimensional convolution module (104) is: the input is sequentially passed through a 1×3 convolution layer (221), a second batch of normalization layers (222), a second activation function layer (223), a 1×3 convolution layer (221), a second batch of normalization layers (222) and a second activation function layer (223) to obtain an output.
4. The bearing fault diagnosis method based on multi-input multi-channel parallel convolutional neural network according to claim 3 is characterized in that: The calculation process of the attention mechanism module (106) is as follows: the input feature map (31) of the attention mechanism module (106) is composed of the outputs of the two-dimensional convolution module (103) and the one-dimensional convolution module (104), and each channel of the input feature map (31) is subjected to spatial dimension information aggregation (35) to obtain an aggregation vector (32), and then the aggregation vector (32) is subjected to full connection layer bottleneck calculation (36) to generate a weight vector (33), and based on the weight vector (33), the input feature map (31) is weighted, and the calculation of the ratio adjustment (37) is completed to obtain the calibrated feature map (34) finally output by the attention mechanism module (106).
5. The bearing fault diagnosis method based on multi-input multi-channel parallel convolutional neural network according to claim 4 is characterized in that: The fully connected module comprises a first fully connected layer (107), a second fully connected layer (108) and a third fully connected layer (109) which are connected in sequence, and the input of the fully connected module is calculated by the first fully connected layer (107), the second fully connected layer (108) and the third fully connected layer (109) in sequence, and outputs an abstract feature; The classifier (110) is a Softmax classifier, which inputs the abstract features, obtains the probability that the current bearing to be diagnosed belongs to each fault category, and uses the fault category with the highest probability as the output fault category.
