Motor bearing fault diagnosis method and system based on multi-source information fusion

Abstract

The invention discloses a motor bearing fault diagnosis method and system based on multi-source information fusion, relates to the technical field of bearing fault diagnosis, and solves the technical problem that the existing bearing fault diagnosis technology cannot effectively extract the characteristics of a composite signal and is difficult to adapt to uncertain factors, so that the diagnosis effect is poor. The method comprises the following steps: synchronously acquiring vibration signals and sound signals of a motor bearing through multiple channels, dynamically distributing multi-channel signal weights by using an entropy weight method, and performing data fusion on the same kind of signals of the multiple channels; respectively extracting low-frequency features of the vibration signals and high-frequency features of the sound signals; and finally, inputting the low-frequency features of the vibration signals and the high-frequency features of the sound signals into a double-flow depth separable convolutional network to extract spatio-temporal features, and outputting a diagnosis result of a fault category. A multi-source dynamic fusion mechanism is used, and adaptive weight distribution of the vibration and sound signals of a data-level fusion layer is realized based on an entropy weight method; manual deviation setting is avoided, and the features of the composite signal are effectively extracted.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a motor bearing fault diagnosis method system based on multi-source information fusion. Background Art

[0002] As core transmission components of rotating machinery, the health of motor bearings directly impacts equipment safety and energy efficiency. Traditional diagnostic methods rely on feature extraction and pattern recognition of single physical quantity sensor signals (such as vibration, sound, or temperature). However, under complex operating conditions, the nonlinear dynamic characteristics caused by electromechanical coupling make it difficult to fully characterize fault characteristics with a single signal. This is especially true for early, weak faults, where signal energy is low and susceptible to interference from environmental noise and load fluctuations. This leads to bottlenecks in traditional methods, such as high false positive rates and poor robustness.

[0003] Existing research on multi-source information fusion technology primarily employs a multi-sensor data parallel processing framework. Early approaches, such as weighted average fusion strategies based on fixed weights, can simply superimpose multi-channel signals but ignore the differences in signal quality between different sensors under dynamic conditions. These static fusion mechanisms are unable to dynamically adjust weight distribution based on the signal-to-noise ratio, resulting in high redundancy in fused features and a dilution of effective information.

[0004] As a core method for processing non-stationary signals, time-frequency analysis directly impacts the extraction of weak impact features. In early-stage bearing failures, characteristic fault frequencies often fall within the 1-5 kHz range, while ambient noise often covers lower frequencies. Existing methods struggle to achieve high-resolution feature extraction across a wide frequency band, resulting in the energy in the fault-sensitive frequency band being drowned out by noise.

[0005] In summary, existing bearing fault diagnosis technology faces three technical challenges in complex industrial scenarios: First, single signal feature extraction has low robustness, and there is a lack of feature extraction technology for composite signals. Second, the multi-source signal fusion layer lacks a dynamic weight allocation mechanism, making it difficult to adapt to uncertainties such as sensor performance degradation and environmental interference. Third, time-frequency analysis methods face a resolution bottleneck in broadband feature extraction, which restricts the enhancement and separation of weak impact components. Existing technologies for motor bearing fault diagnosis based on multi-source information fusion have significant technical gaps, and there is an urgent need to develop new fault diagnosis methods with high fusion efficiency, high accuracy, and strong classification capabilities. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a motor bearing fault diagnosis method and system based on multi-source information fusion, which solves the technical problems that the existing bearing fault diagnosis technology cannot effectively extract the characteristics of composite signals and is difficult to adapt to uncertain factors, resulting in poor diagnostic effect.

[0007] A motor bearing fault diagnosis method based on multi-source information fusion includes: synchronously collecting vibration signals and sound signals of motor bearings through multiple channels, dynamically allocating multi-channel signal weights using an entropy weight method, and fusing multi-channel similar signals; then, extracting low-frequency features of the vibration signal and high-frequency features of the sound signal through parameter-optimized short-time Fourier transform and Mel spectrogram; finally, inputting the low-frequency features of the vibration signal and the high-frequency features of the sound signal into a dual-stream deep separable convolutional network to extract spatiotemporal features and then outputting a diagnosis result of the fault category, wherein the fault category includes outer race fault, inner race fault, rolling element fault, and cage fault.

[0008] Furthermore, after signal acquisition, the sliding window method is used to segment the data, and then adaptive filtering and normalization are used to process the segmented data to eliminate noise. The sliding window is then used to divide the processed data set into samples, and finally a stratified sampling strategy is used to divide the training set and the test set.

[0009] Furthermore, the entropy weight method is used to dynamically allocate the weights of multi-channel signals. The method includes: first extracting three types of statistical features for each channel: standard deviation, mean, and peak-to-peak value, and constructing the three types of statistical features into an m×3 feature matrix, where m is the number of channels, extracting three statistical features for each of the m channels, and constructing an m×3 feature matrix; using the entropy weight method to calculate the statistical feature weight w j (j=1,2,3); then use the statistical feature weights to calculate the channel score:

[0010]

[0011] Among them, s i is the weighted score of the i-th channel, f ij is the normalized value of the jth statistical feature of the i-th channel (derived from the entropy weight normalization process);

[0012]

[0013] Among them, α i is the final weight of the i-th channel, used for multi-source signal fusion, is the sum of the weighted scores of all m channels;

[0014] Finally, the final channel weight is used to fuse the data to obtain the fused signal:

[0015]

[0016] Among them, x fused is the fused signal, which is the result of weighted combination of multi-channel signals, x iis the original signal (time domain waveform) of the i-th channel, the signal that has been preprocessed (filtered and normalized) but not fused, and m is the total number of sensor channels, including all valid channels of vibration and sound sensors.

[0017] The entropy weight method includes first normalization processing to eliminate the dimension effect. In the probability distribution calculation, the normalized value is converted into a probability distribution to avoid zero value interference; then the information entropy calculation is performed to evaluate the degree of chaos of the feature distribution. The lower the entropy value, the higher the feature discrimination; followed by weight allocation, generating feature weights based on the entropy value difference.

[0018] Furthermore, extracting the low-frequency features of the vibration includes calculating the standard deviation of the Gaussian window so that the standard deviation is dynamically adjusted with the frequency. The formula is as follows:

[0019] Dynamic adjustment, the specific adjustment formula is as follows:

[0020]

[0021] Among them, σ(f k ) is the standard deviation, f k is the frequency, β is the proportional coefficient, and the optimal proportional coefficient is determined by grid search;

[0022] The low frequency band uses a wide window to improve the frequency resolution, and the high frequency band uses a narrow window to enhance the time resolution. For each frequency component f k The window function is calculated separately, and then the local spectrum is extracted through convolution operation. The standard deviation of the wide window is greater than the threshold, and the standard deviation of the narrow window is less than the threshold. The threshold is set in advance based on engineering experience.

[0023] Furthermore, extracting the high-frequency features of the sound involves designing a Mel filter to cover the 20Hz-8kHz frequency band, optimizing energy compression and dynamic range, logarithmically compressing the Mel band energy, and finally, normalizing the time-frequency graph, uniformly scaling the Mel spectrogram, and using bilinear interpolation to preserve high-frequency details.

[0024] Furthermore, the dual-stream depthwise separable convolutional network includes a vibration signal branch, a sound signal branch, a feature fusion layer and a classifier layer; the vibration signal branch and the sound signal branch both include an input layer, a first depthwise separable convolutional layer, a first maximum pooling layer, a second depthwise separable convolutional layer, and a second maximum pooling layer arranged in sequence.

[0025] Furthermore, the convolution kernel used in the depthwise separable convolution layer in the vibration signal branch is larger than the convolution kernel used in the depthwise separable convolution layer in the sound signal branch.

[0026] A motor bearing fault diagnosis system based on multi-source information fusion includes a data acquisition module, a data processing module and a fault diagnosis module. The data acquisition module includes a vibration sensor for collecting vibration signals and a sound sensor for collecting sound signals; the data processing module is used to preprocess the vibration signal and the sound signal and then extract the low-frequency features of the vibration signal and the high-frequency features of the sound signal respectively; the fault diagnosis module includes a trained two-stream deep separable convolutional network, which is used to extract spatiotemporal features based on the low-frequency features of the vibration signal and the high-frequency features of the sound signal and then output the diagnosis result of the fault category.

[0027] Furthermore, the data processing module includes a data division submodule, a data fusion submodule, and a feature extraction submodule. The data division submodule is used to divide the collected signals through a sliding window method. The data fusion submodule uses an entropy weight method to dynamically allocate multi-channel signal weights and fuse multi-channel similar signals. The feature extraction submodule extracts the low-frequency features of the vibration signal and the high-frequency features of the sound signal through short-time Fourier transform and Mel spectrum after parameter tuning.

[0028] The beneficial effects of the present invention include:

[0029] (1) First, a multi-source dynamic fusion mechanism is used to achieve adaptive weight distribution of vibration and sound signals at the data-level fusion layer based on the entropy weight method, avoiding manual setting bias and solving the problem of traditional methods relying on prior knowledge;

[0030] (2) Secondly, an improved time-frequency analysis algorithm is used, and the parameter-tuned STFT transform and Mel spectrogram are used to collaboratively extract broadband features, providing a solution for the joint extraction of low-frequency and high-frequency features, which is suitable for non-stationary signal analysis and significantly improves the model accuracy and classification ability;

[0031] (3) Then, using a lightweight dual-stream architecture, depth-wise separable convolution and heterogeneous kernel design, we reduced the number of parameters while ensuring accuracy, significantly reducing computational overhead and improving operational efficiency.

[0032] (4) Finally, the end-to-end optimization process of this method, from signal preprocessing to network training, optimizes the entire link and adapts to the real-time and reliability requirements of industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of a motor bearing fault diagnosis method based on multi-source information fusion involved in an embodiment of the present application.

[0034] Figure 2 This is a schematic diagram of the multi-channel original signal diagram and data sample division of a normal bearing involved in an embodiment of the present application.

[0035] Figure 3 This is an example diagram of sample division for each bearing fault type involved in the embodiments of the present application.

[0036] Figure 4 : is a schematic diagram of a channel fusion sample involved in an embodiment of the present application, wherein: Figure 4 (a) Corresponding to the fusion of normal bearing samples, Figure 4 (b) Fusion of bearing samples corresponding to outer ring faults, Figure 4 (c) Fusion of bearing samples corresponding to inner ring faults, Figure 4 (d) Fusion of bearing samples corresponding to rolling element faults, Figure 4 (e) Fusion of bearing samples corresponding to cage failure.

[0037] Figure 5 Schematic diagram of the time-frequency diagram of the fused vibration and sound signal conversion involved in the embodiment of the present application, wherein: Figure 5 (a) Time-frequency conversion of the corresponding normal bearing fusion sample, Figure 5 (b) Time-frequency conversion of the fusion sample of the bearing corresponding to the outer ring fault, Figure 5 (c) Time-frequency conversion of the bearing fusion sample corresponding to the inner ring fault, Figure 5 (d) Time-frequency conversion of the fusion sample of the corresponding rolling element fault bearing, Figure 5 (e) Time-frequency conversion of the fusion sample of the bearing corresponding to the cage failure.

[0038] Figure 6 This is a schematic diagram of a dual-stream deep separable convolutional network model involved in an embodiment of the present application.

[0039] Figure 7 This is the training set loss function and test set accuracy change curve of the model involved in the embodiment of this application.

[0040] Figure 8 This is a confusion matrix diagram for model fault classification involved in an embodiment of the present application.

[0041] Figure 9 This is a schematic diagram of the Conv1-Fused layer feature space involved in an embodiment of the present application.

[0042] Figure 10 This is a schematic diagram of the Conv2-Fused layer feature space involved in an embodiment of the present application.

[0043] Figure 11 This is a schematic diagram of the Final-Fused layer feature space involved in an embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0045] Example 1

[0046] The following is combined with Figure 1 The specific embodiments of the present invention are described in detail;

[0047] A motor bearing fault diagnosis method based on multi-source information fusion, such as Figure 1 As shown, the method includes: synchronously collecting vibration signals and sound signals of motor bearings through multiple channels, dynamically allocating weights of multi-channel signals using the entropy weight method, and fusing data of multi-channel similar signals; then extracting low-frequency features of vibration signals and high-frequency features of sound signals through short-time Fourier transform and Mel spectrogram after parameter tuning; finally, inputting the low-frequency features of vibration signals and high-frequency features of sound signals into a dual-stream deep separable convolutional network to extract spatiotemporal features and then outputting the diagnosis results of fault categories, which include outer ring fault, inner ring fault, rolling element fault and cage fault.

[0048] The test motor was an M2BAX 90LA 4 model from ABB Shanghai Electric Co., Ltd., with a rated power of 1.5 kW, a rated frequency of 50 Hz, and a maximum speed of 1754 rpm. Data was collected for healthy motor bearings, bearings with outer ring faults, bearings with inner ring faults, bearings with rolling element faults, and bearings with cage faults. All five bearing types used the same model number, 6205-22 / CC.

[0049] The vibration and sound signals of the motor bearings are collected synchronously through a four-channel sensor. The sampling frequency is set to 10kHz and the rotation speed is 850rpm to ensure that the main frequency bands of the bearing fault characteristics are covered.

[0050] The vibration signal is acquired by two acceleration sensors at different positions in orthogonal directions (X / Y direction), and the sound signal is collected by two microphones respectively from the near and far ends of the bearing.

[0051] In another embodiment, after signal acquisition, the data is segmented using a sliding window method, and then the segmented data is processed using adaptive filtering and normalization to eliminate noise. The processed data set is then divided into samples using a sliding window, and finally a stratified sampling strategy is used to divide the training set and the test set.

[0052] The original continuous signal needs to be segmented to adapt to the model input. The sliding window method is used to segment the data. The window length is 1024 points (corresponding to 0.1024 seconds of time domain signal) and the step size (slip) is set to 512 points to achieve a 50% overlap rate.

[0053] This design takes into account both time domain continuity and sample utilization, and can generate more than 1,000 valid samples, each of which contains four-channel data (shape 1024×4) of vibration and sound signals.

[0054] The sliding window sampling strategy adopts a 50% overlapping sliding window segmentation algorithm for the characteristics of long time series signals. Assume that the length of the original signal data point is L p , the number of split windows N s (number of samples) is calculated as:

[0055]

[0056] In formula (1), L p is the length of the original signal data point; L w is the sample length, i.e. the window length; S represents the step length (slip); N s Indicates the number of split windows (number of samples).

[0057] It is important to note that each sample itself is used only once and samples in one set are not shared with other sets.

[0058] For each label of the original faulty bearing signal, the data point length L p is fixed.

[0059] First, the sample length L w Set it to 1024. Then, try adjusting the step size (slip) until the number of available samples N s Not less than 1000.

[0060] The experimental samples are collected from the available samples of each signal state, with no less than 1000 samples for each bearing state and 1024 data points for each sample.

[0061] Table 1 shows the sample distribution table of the data set for the fault diagnosis test. Each bearing type has 1,048,575 original signal data points, so 2,046 samples with a length of 1,024 can be obtained under 512 slip lengths.

[0062] In order to maintain the statistical consistency of data distribution, stratified sampling strategy is used to divide the training set and test set. The specific implementation is as follows:

[0063] Assume that the original data set D contains N samples, and the category label set C = {c1, c2, ..., c K}, where each category c k The number of samples is n k . Ensure that the training set D train and the test set D test The proportion of samples in each category is consistent with the original data:

[0064]

[0065] Finally, we get 1000 sets of samples, and the ratio of training set to test set is 7:3, so the number of training set samples is 700, and the number of test set samples is 300. The sample length is 1024, and the overlap rate is 50%, that is, 512 data points are slid each time.

[0066] Then the 1000 samples collected for each state are divided into training set and test set in a ratio of 7:3.

[0067] Table 1 Data set sample distribution table of fault diagnosis test

[0068]

[0069] Four channels of raw signals are collected for each bearing type, namely vibration channel 1, vibration channel 2, sound channel 1, and sound channel 2 from top to bottom. The collected multi-channel raw signal graph is a time domain graph with the horizontal axis representing the data point length (also the acquisition time) and the vertical axis representing the amplitude.

[0070] For example, among the five bearing types, the normal bearing multi-channel original signal diagram and data sample division diagram are as follows: Figure 2 As shown, an example enlarged view of sample 28 obtained after dividing the multi-channel original signal data samples of a normal bearing is shown.

[0071] By continuously truncating and dividing the multi-channel original signal data, we will obtain more than 1,000 groups of samples for each bearing type. Figure 3 An example graph showing the sample partitioning for each bearing type is shown.

[0072] To eliminate high-frequency noise and sensor drift in industrial environments, a sliding average filter is performed on the original signal. The filter adopts a non-causal structure with a window length of 5 points (0.5 milliseconds). A mirror filling strategy is used at the signal boundary to avoid truncation effect.

[0073] The filtered signal is normalized independently through the channel to eliminate the dimension difference. The normalization formula is:

[0074]

[0075] in, represents the normalized signal output of the cth channel, x (c) Represents the original signal input of the c-th channel, min(X (c) ) represents the minimum value of all data points in the cth channel, max(X (c) ) represents the maximum value of all data points in the cth channel, c represents the channel index, ∈ is a very small constant, ∈=10 -12 Used to prevent division by zero errors.

[0076] This operation linearly maps each channel to the interval [0,1], preserving the dynamic range of the signal while enhancing the convergence stability of the model.

[0077] Noise reduction is performed using a 5-point sliding average filter, as shown in the following formula. The normalization step involves normalizing each channel independently to the range [0, 1] to eliminate sensor dimension differences.

[0078]

[0079] Among them, x filtered represents the output value of the filtered signal at time point t, represents the summation operation of index k from -2 to 2, covering 5 consecutive time points, x t+k represents the input value of the original signal at time point t+k, Represents the normalized coefficient after summing the 5 points in the window.

[0080] In another embodiment, the entropy weight method is used to dynamically assign weights to multi-channel signals, including: first, extracting three types of statistical features for each channel: standard deviation, mean, and peak-to-peak value, and the calculation formulas are:

[0081] Standard Deviation:

[0082] Mean:

[0083] Peak-to-peak value: PP c =max(x)-min(x)+∈ (7)

[0084] Among them, σ c is the standard deviation of the c-th channel signal, x i is the signal value of the i-th sampling point of the c-th channel, μ c is the mean value of the cth channel signal, N is the total number of sampling points of a single channel, ∈ is a minimum constant, ∈=10-6 , PP c is the peak-to-peak value of the c-th channel signal, max(x) is the maximum value of the c-th channel signal in the window, and min(x) is the minimum value of the c-th channel signal in the window.

[0085] The three types of statistical features are constructed as an m×3 feature matrix, where m is the number of channels, σ m is the standard deviation of the mth channel signal, μ m is the mean value of the mth channel signal, PP c,m is the peak-to-peak value of the mth channel signal, and the characteristic matrix construction formula is as follows:

[0086]

[0087] The objective weight of each statistical feature is then calculated using the entropy weight method, which includes normalization processing to eliminate the dimension effect. The formula is as follows:

[0088]

[0089] Where x′ ij is the normalized eigenvalue, ranging from [0,1], x ij is the original value of the jth statistical feature of the i-th channel, min(x j ) is the minimum value of the jth statistical feature in all channels, max(x j ) is the maximum value of the jth statistical feature in all channels, ∈ is a minimum constant, ∈=10 -12 .

[0090] In probability distribution calculations, normalized values ​​are converted to probability distributions to avoid zero value interference:

[0091]

[0092] Among them, p ij is the probability value of the jth statistical feature of the i-th channel, is the sum of the normalized values ​​of all m channels on the jth feature, ∈ is a very small constant, ∈ = 10 -12 .

[0093] Then, information entropy calculation is performed to evaluate the degree of disorder of feature distribution. The lower the entropy value, the higher the feature discrimination. The information entropy calculation formula is as follows:

[0094]

[0095] Among them, e j is the information entropy of the j-th statistical feature, is the normalization coefficient, where m is the total number of channels.

[0096] This is followed by weight distribution, which generates feature weights based on the entropy value difference. The formula is:

[0097]

[0098] Among them, w j is the weight of the jth statistical feature, where j∈{1,2,3} corresponds to the standard deviation, mean, and peak-to-peak value, respectively, and 1-e j is the difference of the feature (or information utility value), is the sum of the differences of all three statistical features.

[0099] Each channel calculates the weighted scores of its three statistical features through feature weights, and normalizes the weighted scores to obtain the final channel weight:

[0100]

[0101] Among them, s i is the weighted score of the i-th channel, f ij is the normalized value of the jth statistical feature of the i-th channel (from the entropy weight normalization process), α i is the final weight of the i-th channel, used for multi-source signal fusion, is the sum of the weighted scores of all m channels.

[0102] The final fusion signal is a weighted combination of the signals from each channel. The weighted fusion formula is as follows. This method can effectively reduce the risk of misjudgment caused by single-channel anomalies, sensor failure, or environmental interference.

[0103]

[0104] Among them, x fused is the fused signal, which is the result of weighted combination of multi-channel signals, x i is the original signal (time domain waveform) of the i-th channel, the signal that has been preprocessed (filtered and normalized) but not fused, and m is the total number of sensor channels, including all valid channels of vibration and sound sensors.

[0105] The output format is a single-channel fusion signal. The length of the fused signal remains unchanged, with 1024 data points per sample. The 2-channel vibration signal and 2-channel sound signal of each bearing type are fused separately. The sample example of the fused signal is as follows: Figure 4 As shown, Figure 4 (a) Corresponding to the fusion of normal bearing samples, Figure 4 (b) Fusion of bearing samples corresponding to outer ring faults, Figure 4 (c) Fusion of bearing samples corresponding to inner ring faults, Figure 4 (d) Fusion of bearing samples corresponding to rolling element faults, Figure 4 (e) Fusion of bearing samples corresponding to cage failure.

[0106] In another embodiment, extracting low-frequency features of vibration includes: in the process of performing time-frequency transformation on the fused vibration signal to extract features, addressing the defect of fixed time-frequency resolution of traditional short-time Fourier transform (STFT), a parameter-tuned STFT algorithm based on a dynamic Gaussian window is proposed to analyze the low-frequency fault characteristics of the vibration signal.

[0107] In specific implementation, for each frequency component f k Calculate window functions separately:

[0108]

[0109] Among them, w(t,f k ) is the frequency component f k The Gaussian window function value at time t, is the normalization coefficient of the Gaussian function.

[0110] Standard deviation σ(f k ) directly determines the time width of the window. The effective width of the time domain window is 3σ(f k ), frequency resolution and σ(f k ) is positively correlated, and the time resolution is related to σ(f k ) negative correlation; standard deviation σ(f k ) with frequency f k Dynamic adjustment, the specific adjustment formula is as follows:

[0111]

[0112] Where β is the proportional coefficient. The optimal proportional coefficient is determined by grid search. Experimental data show that in bearing fault diagnosis, the comprehensive performance is optimal when β = 0.8. In actual application, it needs to be adjusted according to the sensor sampling rate. By subtracting the standard deviation σ(f k ) and frequency f k The inversely proportional mathematical design realizes adaptive wide / narrow window adjustment with clear physical meaning. This design uses a wide window (such as σ = 0.02s at 50Hz) in the low-frequency band to improve frequency resolution, and a narrow window (such as σ = 0.0005s at 2kHz) in the high-frequency band to enhance time resolution. Compared with the traditional fixed window method, the accuracy of feature extraction in bearing fault diagnosis can be improved.

[0113] Specifically, the engineering definition of the high / low frequency bands and wide / narrow windows sets thresholds based on the characteristic frequency band of bearing faults (1-5kHz), as shown in Table 2:

[0114] Table 2 σ threshold setting

[0115] Frequency band division σ threshold Window Type Design Goals <500Hz σ≥0.002 seconds Wide window Improve frequency resolution ≥500Hz σ<0.002 seconds Narrow window Enhanced temporal resolution

[0116] Then the local spectrum is extracted through convolution operation, the formula is:

[0117]

[0118] Among them, S(t,f k ) is the signal at time t and frequency f k The time-frequency spectrum value (complex number) at time t is t, τ is the time offset (unit: number of sampling points), the range is [-L / 2, L / 2]; x(t+τ) is the sampling value of the original signal at time t+τ (preprocessed), w(τ,f k ) is the frequency f k The corresponding value of the Gaussian window function at the offset τ, is the complex exponential term of Fourier transform (Euler formula expansion), j is the imaginary unit, and j satisfies 2 =-1.

[0119] The window length is L = 128, and the overlap step is set to 64 points to achieve a 75% overlap rate. The parameter configuration settings are: FFT points are set to 256, window length is 128 points, overlap rate is 75% (step size is 64 points), and Hamming window function is set to reduce spectral leakage.

[0120] The specific implementation process steps are as follows: first, the fused vibration signal is framed, with 128 points per frame; second, a Gaussian window is calculated frequency by frequency, and an FFT is performed to obtain the time-frequency spectrum S(t,f); then, the amplitude is converted to decibels, and finally, bilinear interpolation is performed to scale the signal to a 64×64 size. The amplitude-to-decibel conversion formula is as follows:

[0121] S dB =20log 10 (|S(t,f)|+∈) (19)

[0122] Among them, S dB is the time-frequency spectrum amplitude in decibels (dB), |S(t,f)| is the amplitude (modulus) of the complex time-frequency spectrum S(t,f), ∈ is a minimum constant, ∈=10 -12 .

[0123] In this example, the sound signal contains high-frequency bearing friction and impact information (such as broadband acoustic emissions caused by rolling element spalling). After performing data-level fusion of similar sound signals using the entropy weight method, Mel spectrograms are used to extract features that conform to the human ear's auditory characteristics. The key technologies used include the following three points:

[0124] The first step is to design the Mel filter bank. The 128 triangular filters cover the 20Hz-8kHz frequency band. The Mel scale mapping formula is:

[0125]

[0126] Where Mel(f) is the Mel scale value corresponding to frequency f (unit: Hz), 2595 is the global scaling factor used to adjust the range of the Mel scale, f is the input linear frequency (unit: Hz), and 700 is the frequency adjustment factor that controls the degree of nonlinearity of the Mel scale.

[0127] Secondly, energy compression and dynamic range optimization are performed, and the Mel band energy is logarithmically compressed. The formula is:

[0128] M dB =10log 10 (|X a | 2 +∈) (21)

[0129] Among them, M dB is the energy of the ath Mel band (in decibels, dB), X a is the original energy value of the ath Mel band (linear scale), |X a | 2 is the square of the energy of the ath Mel band (if Xa is a complex number). In actual implementation, if the Mel filter bank has output real energy, the square operation can be omitted and the formula is simplified to: dB =10log 10 (M mel +∈).

[0130] ∈=10 -12 Prevent taking logarithm of zero and control the dynamic range within 60dB after compression.

[0131] Finally, the time-frequency graph is normalized, the Mel spectrogram size is uniformly scaled to 64×64 pixels, and bilinear interpolation is used to retain high-frequency details.

[0132] The specific implementation process is as follows: first, STFT calculates the short-time amplitude spectrum, the FFT number of points is 256, and the step size is 64 points; then, the Mel filter bank is weighted, and the weighting formula is as follows:

[0133]

[0134] Among them, M mel is the weighted energy (linear scale) of the mth Mel band, |X(k)| 2 is the time-frequency spectrum energy at frequency point k (linear scale), φ m (k) is the weight value of the mth Mel filter at frequency point k (φ m (k)∈[0,1]), k is the discrete frequency index, corresponding to the frequency point number after FFT.

[0135] Next, the Mel band energy is logarithmically compressed, and the logarithmic compression formula has been given above. dB =10log 10 (M mel Finally, the time-frequency graph is also normalized, the Mel spectrogram size is uniformly scaled to 64×64 pixels, and bilinear interpolation is used to retain high-frequency details. Figure 5 The schematic diagram of the time-frequency diagram of the fused vibration and sound signal conversion is shown, where: Figure 5 (a) Time-frequency conversion of the corresponding normal bearing fusion sample, Figure 5 (b) Time-frequency conversion of the fusion sample of the bearing corresponding to the outer ring fault, Figure 5 (c) Time-frequency conversion of the bearing fusion sample corresponding to the inner ring fault, Figure 5 (d) Time-frequency conversion of the fusion sample of the corresponding rolling element fault bearing, Figure 5 (e) Time-frequency conversion of the corresponding bearing fusion sample with cage failure. The horizontal and vertical coordinates of the time-frequency graph in the figure are time, the vertical coordinate is Hertz (0-8192Hz), and the color scale from bottom to top is -80dB, -70dB, -60dB, -50dB, -40dB, -30dB, -20dB, -10dB, and 0dB.

[0136] In another embodiment, the dual-stream depthwise separable convolutional network includes a vibration signal branch, a sound signal branch, a feature fusion layer and a classifier layer; the vibration signal branch and the sound signal branch both include an input layer, a first depthwise separable convolutional layer, a first maximum pooling layer, a second depthwise separable convolutional layer, and a second maximum pooling layer arranged in sequence.

[0137] In the mathematical expression of the depth-wise separable convolution module, the depth-wise convolution formula is as follows:

[0138]

[0139] in, is the depth convolution result of the c-th output channel, To sum up in the 5×5 spatial range of the convolution kernel (i, j∈{-2,-1,0,1,2}), is the weight value of the depth convolution kernel of the cth channel at position (i, j), X (c) (x+i,y+j) is the value of the cth channel of the input feature map at position (x+i,y+j), and (i,j) is the offset of the convolution kernel in the spatial dimension (relative to the center point (x,y)).

[0140] The point-by-point convolution formula is as follows:

[0141]

[0142] Among them, Y pw is the output feature map of point-by-point convolution (single channel), W pw (c) is the weight value (scalar) of the point-by-point convolution kernel on the c-th input channel, It is the output feature map of the cth channel after depth convolution (the spatial dimension is consistent with the input).

[0143] Feature stitching:

[0144] F fusion =Concat(F vib ,F Acoustic )∈R 128×16×16

[0145] Fully connected layer design:

[0146] enter:

[0147] 128×16×16 → Flatten is 32,768 dimensions

[0148] Hidden layer:

[0149] h1=ReLU(W1·F flat +b1)

[0150] h2=Dropout(h1,p=0.5)

[0151] Output layer:

[0152] y=Softmax(W2·h2+b2)

[0153] The specific network architecture design is shown in Table 3:

[0154] Table 3 Two-stream depth-wise separable convolutional network model architecture

[0155]

[0156] The specific schematic diagram of the two-stream depth-separable convolutional network model is as follows Figure 6 shown.

[0157] The depthwise separable convolution module consists of a series of depthwise convolution and pointwise convolution, which reduces the number of parameters by 94.44% compared to traditional convolution. Depthwise convolution is a spatial convolution performed independently on each input channel, with a parameter count of C in ×K 2 The point-by-point convolution is 1×1 convolution to integrate channel information, and the number of parameters is C in ×C out .

[0158] Computational efficiency comparison: When the input and output channels are 32→64 and the convolution kernel is 5×5, the number of traditional convolution parameters is 32×64×25=51200, while the depthwise separable convolution only requires 32×25+32×64=2848, reducing model storage and computational overhead.

[0159] The mathematical principle of depth-wise separable convolution is as follows. Depth-wise convolution: channel-by-channel spatial convolution, parameter C in ×K 2 For example, in this method: depthwise convolution input 32 channels, 5×5 kernel → parameter amount 32×25=800. Pointwise convolution: 1×1 convolution fusion channel, parameter amount C in ×C out Point-by-point convolution has 32 input channels and 64 output channels → 32×64=2048 parameters. Total parameters: 800+2048=2848 (traditional convolution parameters: 32×64×25=51200). FLOPs reduction rate:

[0160]

[0161] The parameter comparison is shown in Table 4:

[0162] Table 4 Comparison of model parameters

[0163]

[0164] The advantages of the dual-stream structure are: first, it adopts a heterogeneous core design. The vibration branch uses a 5×5 large core to capture low-frequency and long-period features, and the sound branch uses a 3×3 small core to extract high-frequency local details; second, parallel computing acceleration. The dual-stream structure supports GPU parallel computing, which increases the inference speed by 1.8 times compared to the serial network; in terms of feature decoupling capability, experiments show that the dual-stream network is more robust to cross-modal noise (such as electromagnetic interference in vibration signals) and can reduce the false alarm rate.

[0165] In terms of adaptive optimization, the adaptive moment estimation optimization algorithm (Adam) is used, and its parameter update rule is:

[0166] m t =β1m t-1 +(1-β1)g t

[0167]

[0168] Where β1=0.9,β2=0.999,∈=10 -8 Adam combines the advantages of Momentum and RMSProp, and can adaptively adjust the learning rate of each parameter. It is particularly suitable for dealing with the gradient sparsity problem between multimodal signal features.

[0169] In terms of learning rate scheduling, a step-by-step learning rate scheduling step decay strategy is used. The initial learning rate is set to 0.001 and decays by 50% every 5 epochs. The formula is:

[0170]

[0171] Among them, η t is the learning rate for the tth round, η0 is the initial learning rate, γ is the decay factor, T = 5 is the decay period, and t is the current epoch number. This design maintains a high learning rate in the early stages of training to accelerate convergence, and then reduces the learning rate to fine-tune the parameters later. This can theoretically lead the loss function to converge to a flatter minimum, improving the model's generalization ability.

[0172]

[0173] This strategy maintains a large learning rate in the early stages of training to accelerate convergence, and reduces the step size in the later stages to fine-tune parameters.

[0174] Set the early stopping condition to monitor the sliding average change of the test set accuracy. When the accuracy improvement for three consecutive epochs is less than 0.1%, the training is terminated to prevent overfitting. The mathematical expression is:

[0175]

[0176] Among them, ΔA is the sliding average change of the test set accuracy (the average of three consecutive epochs), A k is the test set accuracy of the kth epoch (A k ∈[0,1]), t is the index of the current epoch (training round number), (A k -A k-1 ) is the accuracy increment between adjacent epochs (the improvement of the k-th epoch compared to the k-1-th epoch), and 0.001 is the early stopping trigger threshold (0.1% accuracy improvement).

[0177] For multi-classification tasks, cross entropy loss is used as the optimization objective. A weighted cross entropy function is used to alleviate the class imbalance problem, giving higher weights to minority classes (such as cage failure):

[0178]

[0179] where y ic ∈{0,1} is the true category label of sample i, p ic is the category probability predicted by the model. The cross entropy loss combined with the Softmax activation function can effectively measure the KL divergence between the predicted probability distribution and the true distribution, which is especially suitable for fault diagnosis scenarios where categories are mutually exclusive.c Calculated based on the category distribution of the training set, the maximum weight ratio is 2:1.

[0180] L2 regularization introduces weight decay (coefficient 0.001) in the fully connected layer to suppress the overfitting tendency.

[0181] In terms of data enhancement and noise-resistant training, random cropping in the time domain randomly cuts 90% of the length of the input time-frequency graph on the time axis to simulate the incomplete signal acquisition scenario; frequency domain masking randomly masks a 3×3 frequency band area on the Mel spectrum with a probability of 20% to enhance the model's robustness to the missing of some frequency bands.

[0182] Preprocessed time-frequency graph data is stored in HDF5 format, with zero-copy reads achieved through memory mapping. An index file records the storage offset and length of each sample, supporting concurrent multi-threaded access. This improves data loading speed and reduces memory usage. FP16 mixed-precision training converts model weights and activation values ​​to 16-bit floating-point numbers, accelerating inference on NVIDIA GPUs.

[0183] The diagnostic accuracy of the proposed model reached an average accuracy of 99.8% on the laboratory bearing dataset, an improvement of 6.1% over the traditional CNN (93.7%); the model has strong convergence ability and achieved high-performance convergence within a limited training cycle (10 epochs); the confusion matrix accurately presents the bearing fault classification performance, and all categories show perfect classification performance, indicating that the model has a 100% recognition accuracy for the five types of bearing conditions, without any cross-category misjudgment; in terms of the comparison of the number of calculation parameters, the model has 5184 parameters, which is 92.56% less than the traditional two-stream CNN (69632).

[0184] Specifically, the experimental verification process of the two-stream depth-separable convolutional network model is as follows:

[0185] (I) During the model training process, the batch size was set to 32 and the epoch was set to 50 by continuously optimizing the parameters. The following shows the training set loss function and test set accuracy curve of this model. Figure 7 shown.

[0186] The performance evolution of this model over 50 training epochs is clearly shown in the two-axis subplots. The left axis is the training loss scale (0-5), and the right axis is the test set accuracy scale (0.984-1.000).

[0187] The training set loss function initially had a value of approximately 1.6. It experienced a steep decline between epochs 0 and 10, with an average decrease of 0.12 per epoch, before rapidly converging to 0.4. After epoch 10, it entered a period of gentle convergence, with the loss value continuously decreasing at a rate of approximately 0.02 per epoch, ultimately stabilizing around 0.2. The entire curve was smooth and oscillatory (maximum fluctuation <0.05), indicating that the optimizer parameters were appropriately set and batch training was stable.

[0188] The test set accuracy exhibited a unique evolutionary pattern, initially reaching a high of 0.984 (98.4%). It gradually increased over epochs 0-10, then briefly dipped to 0.986 before recovering through parameter adjustments. After epoch 10, it entered a stable upward fluctuation channel, with continued optimization at a rate of approximately 0.0008 per epoch, ultimately converging to a high plateau of 0.998 (99.8%). The accuracy curve showed slight fluctuations between epochs 20 and 30 (standard deviation ≈ 0.0012), reflecting the model's adaptive adjustments to difficult samples during the test phase.

[0189] When the training loss exceeded the 0.4 threshold (around the 10th round), the corresponding test accuracy rebounded from the bottom, reflecting the model's transition from coarse-tuning to fine-tuning learning stages. Although the test accuracy briefly dropped initially, its final value (0.998) significantly exceeded the initial level (0.984), and the hyperbola maintained a synchronous progressive trend in the later stage (maximum phase difference <3 rounds), verifying its generalization ability. In terms of convergence characteristics, the main performance improvement was concentrated in the first 20% of training cycles (0-10 rounds), which contributed 67% of the total accuracy gain and 75% of the total loss reduction.

[0190] In terms of experimental effectiveness verification, the learning rate strategy scheduling is normal, the inflection point of the loss curve at 10 rounds matches the typical piecewise constant decay strategy, and the change in the sign of the second-order derivative verifies the effectiveness of dynamic parameter adjustment; in overfitting protection, the ratio of the slope of the hyperbola terminal segment is 0.93 (training loss slope: -0.015, test accuracy slope: +0.0007), which meets the robustness standard of |λ|<1.

[0191] This visualization shows that the model achieves high performance convergence within a limited number of training cycles.

[0192] (2) The confusion matrix diagram shows the classification ability of this model, such as Figure 8 shown.

[0193] The confusion matrix, organized in a 5×5 structure, accurately demonstrates bearing fault classification performance. Rows represent the true classes (Ball, Cage, Inner, Normal, and Outer), while columns correspond to the predicted classes. All classes exhibit perfect classification performance, with the corresponding cell values ​​for Ball, Cage, Inner, Normal, and Outer all being 300, indicating that the model achieves 100% accuracy for all five bearing conditions. All off-diagonal elements are zero, indicating no cross-class misclassifications.

[0194] All 300 samples of rolling element faults (Ball) were correctly classified, with no misjudgments. All 300 samples of cage faults (Cage) were accurately identified, eliminating the common cage-outer race confusion. All 300 samples of inner race faults (Inner) achieved zero misjudgment, validating the effectiveness of time-frequency domain feature extraction. All 300 samples of normal bearings (Normal) were correctly classified, eliminating the risk of confusion between normal conditions and early-stage faults. All 300 samples of outer race faults (Outer) were perfectly identified, overcoming the multi-harmonic interference problem of outer race faults in traditional methods. The system eliminates three major sources of confusion in traditional fault diagnosis: cage-outer race, rolling element-normal, and inner race-outer race misjudgments.

[0195] Among the global performance indicators, the overall accuracy is 1500 / 1500=100% (total sample size 5 categories × 300=1500), the micro-average F1-score is 1.000 (standard deviation 0.000), and the Kappa coefficient is 1.000, indicating that the model classification performance has reached a statistically perfect level.

[0196] (3) In order to verify the classification ability of the model, the data distribution of the learned features at different layers is mapped to the three-dimensional space through t-Distributed random Neighbor Embedding (t-SNE).

[0197] The visualization study systematically reveals the model's feature decoupling process for bearing condition classification using three sets of three-dimensional scatter plots. The three images correspond to classification diagrams for the fused convolutional layer 1 (Conv1-Fused layer), the fused convolutional layer 2 (Conv2-Fused layer), and the model's final fully connected layer (Final-Fused layer), which validate the model's feature extraction performance. The charts use a five-color labeling system: dark blue (rolling element fault), green (cage fault), gray (normal), brown (inner race fault), and light blue (outer race fault).

[0198] The x-axis (-25, 15), y-axis (-15, 15), and z-axis (-15, 10) are the projection dimensions used by the t-SNE algorithm to map the high-dimensional feature space into three-dimensional space. The values ​​along these three axes have no direct physical meaning. Their core value lies in reflecting the similarity of the original high-dimensional space through the Euclidean distance between adjacent points, and in reflecting the separability of the underlying features through the local neighborhood structure (clustering).

[0199] Conv1-Fused layer feature space: The features extracted by Conv1 of the vibration branch and the sound branch are merged in the channel dimension. The dimensionality reduction visualization includes global average pooling of the merged 64-channel features to obtain a 64-dimensional feature vector, which is then projected into the 3D space through t-SNE, as shown in the following figure: Figure 9 As shown:

[0200] The five types of samples exhibit a highly mixed, cloud-like distribution, spanning [-20, 12] along the X-axis, [-10, 10] along the Y-axis, and [-12, 8] along the Z-axis. Rolling element faults (dark blue) and normal conditions (gray) form overlapping clusters in the X>0 region. Cage faults (green) are stretched along the Z-axis. Inner race faults (brown) and outer race faults (light blue) intersect in the Y=5 plane, with the maximum confusion density occurring in the (X=0, Y=5, Z=0) region. After the first layer of convolution and pooling, some fault feature points are initially clustered, but there is still some fuzzy boundary between them.

[0201] Conv2-Fused layer feature space: The features extracted by Conv2 of the vibration branch and the sound branch are merged in the channel dimension. The dimensionality reduction visualization includes global average pooling of the merged 64-channel features to obtain a 64-dimensional feature vector, which is then projected into the 3D space through t-SNE, as shown in the following figure: Figure 10 As shown:

[0202] The feature distribution range shrinks to X[-15, 10], Y[-8, 8], and Z[-10, 6]. Outer-ring faults (light blue) first form dense spherical clusters (3.2 units in diameter) in the X<-5 region. Cage faults (green) establish a linear distribution along the Y=7 axis, but still overlap with the normal state (gray) by 4 units on the Z axis. Inner-ring faults (brown) exhibit a bimodal structure, indicating the deep network's ability to separate features. In the second layer of convolutional pooling, the entropy of the fault features is significantly reduced, and they gradually form their own groups, with a certain degree of intermixing.

[0203] Final-Fused layer feature space, such as Figure 11 As shown:

[0204] The five types of features are geometrically decoupled, forming mutually exclusive decision regions. Rolling element faults (dark blue) occupy the X>8 quadrant, condensing into a hypersphere with a diameter of 2.5 units. Cage faults (green) construct a thin lamellar structure with a thickness of only 0.8 units along the Y=10 plane. The normal state (gray) establishes a cubic cluster in the X<-10 region (with a margin of ±1.2 units). Inner and outer race faults are completely separated along the Z axis (inner race Z>3, outer race Z<-2). The minimum inter-class distance is extended to 8.7 units, a 412% improvement compared to the Conv1 layer. The final layer feature space exhibits perfect geometric separability, and the five types of samples form a clear cluster distribution in three-dimensional space. The results show that the model has good convergence and classification performance.

[0205] Example 2

[0206] A motor bearing fault diagnosis system based on multi-source information fusion, it should be understood that the system is similar to the above-mentioned Figure 1 The method embodiment corresponds to the following Figure 1 The various steps involved in the method embodiment and the specific functions of the system can be found in the description above. To avoid repetition, detailed description is appropriately omitted here.

[0207] A motor bearing fault diagnosis system based on multi-source information fusion includes a data acquisition module, a data processing module and a fault diagnosis module. The data acquisition module includes a vibration sensor for collecting vibration signals and a sound sensor for collecting sound signals; the data processing module is used to preprocess the vibration signal and the sound signal and then extract the low-frequency features of the vibration signal and the high-frequency features of the sound signal respectively; the fault diagnosis module includes a trained two-stream deep separable convolutional network, which is used to extract spatiotemporal features based on the low-frequency features of the vibration signal and the high-frequency features of the sound signal and then output the diagnosis result of the fault category.

[0208] In another embodiment, the data processing module includes a data division submodule, a data fusion submodule, and a feature extraction submodule. The data division submodule is used to divide the collected signals through a sliding window method. The data fusion submodule uses an entropy weight method to dynamically allocate multi-channel signal weights and fuse multi-channel similar signals. The feature extraction submodule extracts the low-frequency features of the vibration signal and the high-frequency features of the sound signal through short-time Fourier transform and Mel spectrum after parameter tuning.

[0209] The above-described embodiments merely represent specific implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of the present application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the technical concept of the present application, and all such variations and improvements fall within the scope of protection of the present application.

Claims

1. A motor bearing fault diagnosis method based on multi-source information fusion, characterized in that: include: The vibration and sound signals of the motor bearings are collected synchronously through multiple channels. The entropy weight method is used to dynamically assign weights to the multi-channel signals, and data of the same multi-channel signals is fused. The low-frequency features of the vibration signal and the high-frequency features of the sound signal are extracted respectively through short-time Fourier transform and Mel spectrogram after parameter tuning. Finally, the low-frequency features of the vibration signal and the high-frequency features of the sound signal are input into a dual-stream deep separable convolutional network to extract spatiotemporal features and then output the diagnosis results of the fault categories, which include outer ring fault, inner ring fault, rolling element fault and cage fault.

2. The motor bearing fault diagnosis method based on multi-source information fusion according to claim 1 is characterized in that: After signal acquisition, the sliding window method is used to segment the data, and then adaptive filtering and normalization are used to process the segmented data to eliminate noise. Then the sliding window is used to divide the processed data set into samples, and finally a stratified sampling strategy is used to divide the training set and the test set.

3. The motor bearing fault diagnosis method based on multi-source information fusion according to claim 1 is characterized in that: The entropy weight method is used to dynamically allocate the weights of multi-channel signals. First, three types of statistical features are extracted for each channel: standard deviation, mean, and peak-to-peak value. The three types of statistical features are constructed into an m×3 feature matrix, where m is the number of channels. Three statistical features are extracted for each of the m channels to construct an m×3 feature matrix. The entropy weight method is used to calculate the statistical feature weight w. j (j=1,2,3); then use the statistical feature weight to weight the statistical features in the channel to obtain the channel score s i , and then score the channel s i Normalization processing obtains the final channel weight α i , and finally use the final channel weight α i Perform data fusion to obtain a fused signal.

4. The motor bearing fault diagnosis method based on multi-source information fusion according to claim 1, characterized in that: Extracting the low-frequency characteristics of vibration involves calculating the standard deviation of the Gaussian window so that the standard deviation is dynamically adjusted with frequency. The formula is as follows: Among them, σ(f k ) is the standard deviation, f k is the frequency, β is the proportional coefficient, and the optimal proportional coefficient is determined by grid search; The low frequency band uses a wide window to improve the frequency resolution, and the high frequency band uses a narrow window to enhance the time resolution. For each frequency component f k The window function is calculated separately, and then the local spectrum is extracted through convolution operation. The wide window refers to a standard deviation greater than a threshold, and the narrow window refers to a standard deviation less than a threshold. The threshold is set in advance based on engineering experience.

5. The motor bearing fault diagnosis method based on multi-source information fusion according to claim 1, characterized in that: Extracting the high-frequency features of the sound involves designing a Mel filter to cover the 20Hz-8kHz frequency band, optimizing energy compression and dynamic range, logarithmically compressing the Mel band energy, and finally, normalizing the time-frequency graph, uniformly scaling the Mel spectrogram, and using bilinear interpolation to preserve high-frequency details.

6. The motor bearing fault diagnosis method based on multi-source information fusion according to claim 1, characterized in that: The dual-stream depthwise separable convolutional network includes a vibration signal branch, a sound signal branch, a feature fusion layer and a classifier layer; the vibration signal branch and the sound signal branch both include an input layer, a first depthwise separable convolutional layer, a first maximum pooling layer, a second depthwise separable convolutional layer, and a second maximum pooling layer, which are arranged in sequence.

7. The motor bearing fault diagnosis method based on multi-source information fusion according to claim 6, characterized in that: The convolution kernel used in the depthwise separable convolution layer in the vibration signal branch is larger than the convolution kernel used in the depthwise separable convolution layer in the sound signal branch.

8. A motor bearing fault diagnosis system based on multi-source information fusion, characterized in that: It includes a data acquisition module, a data processing module and a fault diagnosis module. The data acquisition module includes a vibration sensor for collecting vibration signals and a sound sensor for collecting sound signals. The data processing module is used to pre-process the vibration signal and the sound signal and then extract the low-frequency characteristics of the vibration signal and the high-frequency characteristics of the sound signal respectively. The fault diagnosis module includes a trained two-stream deep separable convolutional network, which is used to extract spatiotemporal features based on the low-frequency features of the vibration signal and the high-frequency features of the sound signal and then output the diagnosis result of the fault category.

9. The motor bearing fault diagnosis system based on multi-source information fusion according to claim 8, characterized in that: The data processing module includes a data division submodule, a data fusion submodule, and a feature extraction submodule. The data division submodule is used to divide the collected signals through the sliding window method. The data fusion submodule uses the entropy weight method to dynamically allocate multi-channel signal weights and fuse multi-channel similar signals. The feature extraction submodule extracts the low-frequency features of the vibration signal and the high-frequency features of the sound signal through the short-time Fourier transform and Mel spectrum after parameter tuning.

Images