Visual identification-based vibration signal fault diagnosis method and device

Abstract

The invention discloses a vibration signal fault diagnosis method and device based on visual identification. The method comprises the following steps: acquiring an original vibration signal; preprocessing the original vibration signal to generate an RGB picture; carrying out model training by taking the RGB picture as the input of a neural network model; and performing fault diagnosis by using the trained model. On the basis of a traditional Morlet wavelet, the amplitude attenuation rate of a primary function is changed by increasing a waveform adjustment factor, and then the primary function with adjustable waveform characteristics is constructed. The method comprises the following steps: acquiring a signal RGB image, learning feature mapping from the signal RGB image by using a continuous convolutional layer, calculating a new feature mapping as a weighted average of the original feature mapping by adopting an attention mechanism, and in this way, learning that a model emphasizes an important region of the RGB image and suppresses a region weakly related to classification so as to more accurately capture abnormal fluctuation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of vibration signal fault diagnosis, and particularly relates to a vibration signal fault diagnosis method and device based on visual recognition. BACKGROUND

[0002] Common techniques for equipment fault diagnosis include signal processing, artificial intelligence and various methods:

[0003] 1. Diagnosis technology based on signal processing.

[0004] (1) Vibration analysis method: use acceleration sensor to collect vibration signal when equipment is running, and analyze the vibration signal in time domain and frequency domain. For example, frequency spectrum analysis can find out the characteristic frequency of vibration. If there is an abnormal peak at a specific frequency, it may indicate that the equipment has bearing failure, rotor imbalance and other problems.

[0005] (2) Current analysis method: monitor and analyze the stator current of the equipment. The current has certain waveform and spectrum characteristics when the equipment is running normally. When the equipment fails, such as rotor broken bar, there will be specific fault characteristic frequency components in the current spectrum. Through fast Fourier transform and other methods to process the current signal, the fault characteristics can be extracted.

[0006] 2. Diagnosis technology based on artificial intelligence: this method takes various operating parameters of the equipment, such as current, voltage, vibration, etc. as the input features of neural network, and takes the fault type of motor as the output. Through a large number of sample data, the neural network is trained to learn the feature mapping relationship under normal and fault conditions, so as to realize automatic diagnosis of equipment fault.

[0007] In recent years, artificial intelligence-based diagnostic techniques have made great progress. Studies have shown that deep learning-based fault diagnosis methods can overcome the problem of expert dependence and achieve higher diagnostic accuracy. Generally, there are three types of deep learning, namely deep belief networks, recurrent neural networks, and convolutional neural networks. Due to the solution of the gradient vanishing problem and the improvement of the performance of graphics processing units, deep learning has made significant progress. In the field of signal fault diagnosis, certain achievements have also been made. For example, a recurrent neural network can be used to directly capture the hidden patterns of vibration time series to diagnose device field voltage abnormalities and inter-turn short circuit faults (Reference: Zollanvari, A.; Kunanbayev, K.; Akhavan Bitaghsir, S.; Bagheri, M. Transformer Fault Prognosis Using Deep Recurrent Neural Network over Vibration Signals. IEEE Trans. Instrum. Meas. 2020, 70, 1-11). A convolutional neural network can also be used to identify transformed vibration images to identify three working states of the device (Reference: Hong, K.; Jin, M.; Huang, H. Transformer winding fault diagnosis using vibration image and deep learning. IEEE Trans. Power Deliv. 2021, 36, 676-685). A multi-scale fusion feature extraction model based on a convolutional neural network can also be designed, which has an attention mechanism and can identify the operating state of the device under different voltages and loads (Reference: Xiao, R.; Zhang, Z.; Wu, Y.; Jiang, P.; Deng, J. Multi-scale information fusion model for feature extraction of converter transformer vibration signal. Meas. J. Int. Meas. Confed. 2021, 180, 109555-109566).

[0008] The device fault diagnosis based on vibration signal analysis is more and more concerned due to its easy implementation and low cost, but the complex operation environment and load of the device also bring challenges to the fault diagnosis. In order to find out the fault information hidden in the vibration signal, the continuous wavelet transform is used for feature extraction, and the vibration signal is converted into a red-green-blue (RGB) image with time-frequency relationship. However, the traditional complex Morlet wavelet basis function has the disadvantage that its amplitude decay rate is constant, and has a fixed waveform feature, which will affect the matching ability of different abnormal vibration signals. When the visual model is used to classify the signal RGB image, the abnormal fluctuations are often in the minority, and how to accurately capture these fluctuations is also an urgent industry pain point. SUMMARY

[0009] In order to solve the above problems, the present application provides a vibration signal fault diagnosis method and device based on visual recognition. On the basis of the traditional Morlet wavelet, the amplitude decay rate of the basis function is changed by increasing the waveform adjustment factor, and then a basis function with adjustable waveform characteristics is constructed. And using continuous convolution layer to learn feature mapping from signal RGB image, using attention mechanism to calculate a new feature mapping as the weighted average of these original feature mappings, in this way, the model learns to emphasize the important areas of the RGB image, and suppresses the areas weakly related to classification, in order to more accurately capture abnormal fluctuations.

[0010] The technical scheme adopted by the present application is:

[0011] In a first aspect, the present application discloses a vibration signal fault diagnosis method based on visual recognition, comprising:

[0012] Collecting original vibration signals;

[0013] Pretreating the original vibration signals to generate an RGB picture;

[0014] Using the RGB picture as the input of a neural network model for model training;

[0015] Using the trained model for fault diagnosis.

[0016] As an optional technical scheme, the pretreatment comprises:

[0017] Performing complex Morlet wavelet transform on the original vibration signals to obtain wavelet transform coefficients W(a, b);

[0018] Extracting feature values of the wavelet transform coefficients W(a, b), the feature values including amplitude information and phase information;

[0019] Performing RGB mapping on the feature values to generate the RGB picture.

[0020] As an optional technical solution, the wavelet transform coefficient W(a, b) is calculated according to the following formula:

[0021]

[0022] Wherein, x(t) represents the original vibration signal, t is a time variable, ψ(·) is an improved complex Morlet wavelet function, ψ * (·) is the conjugate complex of ψ(·), a is a scale parameter, b is a translation parameter; the improved complex Morlet wavelet function is:

[0023]

[0024] Wherein, η is a normalization coefficient, e is a natural logarithm, ω0 is a center frequency parameter, j is an imaginary unit, β is a waveform adjustment factor, and β is calculated according to the following formula:

[0025]

[0026] Wherein, τ is a damping ratio parameter, and 0<τ<1.

[0027] As an optional technical solution, the neural network model comprises an original EfficientNet-B3 model and a GAP attention module, wherein the GAP attention module is connected to the output end of a target convolution kernel in the original EfficientNet-B3 model, and the target convolution kernel refers to the last convolution kernel with an expansion rate of 6 and a receptive field of 5x5 in the original EfficientNet-B3 model.

[0028] As an optional technical solution, the EfficientNet-B3 model comprises a GAP operation unit, and the input and output of the GAP operation unit satisfy the following relationship:

[0029]

[0030] Wherein, L k and L k-1 are the output layer and the input layer of the GAP function respectively, V j is a feature vector contained in the L k layer, and N is the number of feature vectors.

[0031] The GAP attention module comprises an improved GAP operation unit, and the input and output of the improved GAP operation unit satisfy the following relationship:

[0032]

[0033] Wherein, ω j is an attention weight learned automatically by the model.

[0034] As an optional technical solution, the loss function Loss of the neural network model is:

[0035] Loss = λ1L1 + λ2L2

[0036] Wherein, λ1 and λ2 are positive hyperparameters, L1 is the cross-entropy loss of the original EfficientNet-B3 model, L2 is the cross-entropy loss of the GAP attention module, and the cross-entropy loss is minimized using the back propagation algorithm:

[0037]

[0038] Wherein, n is the number of training samples, C is the number of classes, y ik is the predicted probability of sample i and class k, is the output of the last hidden layer, is the weight matrix of the hidden layer to the output layer; I(·) is an indicator function, which takes 1 if the statement is true, and 0 otherwise.

[0039] In the second aspect, the application discloses a vibration signal fault diagnosis device based on visual recognition, comprising:

[0040] A data acquisition module for acquiring original vibration signals;

[0041] An image generation module for preprocessing the original vibration signals to generate RGB pictures;

[0042] A model training module for training the model with the RGB pictures as the input of the neural network model;

[0043] A fault diagnosis module for fault diagnosis using the trained model.

[0044] As an optional technical solution, the image generation module comprises:

[0045] A wavelet transform unit for performing complex Morlet wavelet transform on the original vibration signals to obtain wavelet transform coefficients W(a, b);

[0046] A feature value extraction unit for extracting feature values of the wavelet transform coefficients W(a, b), the feature values including amplitude information and phase information;

[0047] A mapping unit for RGB mapping the feature values to generate the RGB pictures.

[0048] As an optional technical solution, the wavelet transform coefficients W(a, b) are calculated according to the following formula:

[0049]

[0050] wherein x(t) represents the original vibration signal, t is a time variable, ψ(·) is an improved complex Morlet wavelet function, ψ * (·) is a conjugate complex of ψ(·), a is a scale parameter, b is a translation parameter; the improved complex Morlet wavelet function is:

[0051]

[0052] wherein η is a normalization coefficient, e is a natural logarithm, ω0 is a center frequency parameter, j is an imaginary unit, and β is a waveform adjustment factor, and β is calculated according to the following formula:

[0053]

[0054] wherein τ is a damping ratio-like parameter, and 0 < τ < 1.

[0055] As an optional technical solution, the neural network model comprises an original EfficientNet-B3 model and a GAP attention module, wherein the GAP attention module is connected to the output end of a target convolution kernel in the original EfficientNet-B3 model, and the target convolution kernel refers to a last convolution kernel with an expansion rate of 6 and a receptive field of 5x5 in the original EfficientNet-B3 model. The EfficientNet-B3 model comprises a GAP operation unit, and the input and output of the GAP operation unit satisfy the following relationship:

[0056]

[0057] wherein L k and L k-1 are the output layer and the input layer of the GAP function respectively, V j is a feature vector contained in the L k layer, and N is the number of feature vectors; the GAP attention module comprises an improved GAP operation unit, and the input and output of the improved GAP operation unit satisfy the following relationship:

[0058]

[0059] wherein ω j is an attention weight learned automatically by the model.

[0060] The loss function Loss of the neural network model is:

[0061] Loss=λ1L1+λ2L2

[0062] wherein, λ1 and λ2 are positive hyperparameters, L1 is the cross-entropy loss of the original EfficientNet-B3 model, L2 is the cross-entropy loss of the GAP attention module, and the cross-entropy loss is minimized using the backpropagation algorithm:

[0063]

[0064] wherein, n is the number of training samples, C is the number of classes, y ik is the predicted probability of sample i and class k, is the output of the last hidden layer, is the weight matrix of the hidden layer to the output layer; I(·) is an indicator function, which takes 1 if the statement is true, and 0 otherwise.

[0065] The beneficial effects of the present application are:

[0066] 1. Accurate capture of abnormal fluctuations: The GAP attention module is introduced, which can calculate a new feature map as a weighted average of the original feature map, enabling the model to focus on important areas of the RGB image and suppress weakly related areas, thereby accurately capturing abnormal fluctuation signals hidden in the image and effectively overcoming the difficulty of accurately capturing a small number of abnormal fluctuations when using a visual model to classify signal RGB images.

[0067] 2. Joint training to improve accuracy: The GAP attention module is jointly trained with the classifier in the original EfficientNet-B3 model, and the outputs of the two independent branches correct each other's classification results, making the model more accurate in subsequent iterations and improving the accuracy of fault diagnosis.

[0068] 3. Strong generalization ability: The method has strong generalization ability and can handle fault signals of different devices, not limited to specific devices, and has a wider range of applications, such as airport logistics equipment fault diagnosis (such as conveyors, motors, etc.), factory equipment fault diagnosis (such as generators, sensors, etc.).

[0069] 4. Better adaptability of wavelet transform: The improved complex Morlet wavelet transform has a constant amplitude decay rate, which is changed by increasing the waveform adjustment factor to change the amplitude decay rate of the basis function, and a basis function with adjustable waveform characteristics is constructed. This makes it more capable of matching different waveform abnormal vibration signals, and the adjustment factor can be adjusted according to the specific scene, making it better adapt to various complex vibration signals in the feature extraction stage and improving the effectiveness of fault feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a flowchart of the vibration signal fault diagnosis method based on visual recognition.

[0071] Figure 2 is a structural schematic diagram of a neural network model of the embodiment.

[0072] Figure 3 is a flowchart of the method of the embodiment.

[0073] Figure 4 is an architectural diagram of a vibration signal fault diagnosis device based on visual recognition. DETAILED DESCRIPTION

[0074] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. 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 claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0075] EMBODIMENT

[0076] As shown in Figure 1 , the embodiment discloses a vibration signal fault diagnosis method based on visual recognition, comprising the following steps S1-S4.

[0077] S1, collecting an original vibration signal.

[0078] S2, preprocessing the original vibration signal to generate an RGB picture. Specifically, the preprocessing comprises: performing a complex Morlet wavelet transform on the original vibration signal to obtain a wavelet transform coefficient W(a, b); extracting eigenvalues of the wavelet transform coefficient W(a, b), the eigenvalues including amplitude information and phase information; and performing RGB mapping on the eigenvalues to generate the RGB picture.

[0079] As an optional implementation, the wavelet transform coefficient W(a, b) is calculated according to the following formula:

[0080]

[0081] wherein x(t) represents the original vibration signal, t is a time variable, ψ(·) is an improved complex Morlet wavelet function, ψ * (·) is a conjugate complex of ψ(·), a is a scale parameter, and b is a translation parameter; the improved complex Morlet wavelet function is:

[0082]

[0083] wherein, η is a normalization coefficient, e is a natural logarithm, ω0 is a center frequency parameter, j is an imaginary unit, and β is a waveform adjustment factor, and β is calculated according to the following formula:

[0084]

[0085] wherein, τ is a quasi-damping ratio parameter, and 0 < τ < 1.

[0086] S3, taking the RGB picture as an input of a neural network model to perform model training. The neural network model structure of the embodiment is as shown in the following figure: Figure 2 The GAP attention module is connected to the output end of a target convolution kernel in the original EfficientNet-B3 model, and the target convolution kernel refers to a last convolution kernel with an expansion rate of 6 and a receptive field of 5x5 in the original EfficientNet-B3 model.

[0087] As an optional implementation, the EfficientNet-B3 model comprises a GAP operation unit, and the input and output of the GAP operation unit satisfy the following relationship:

[0088]

[0089] wherein, L k and L k-1 are an output layer and an input layer of a GAP function respectively, V j is a feature vector contained in the L k layer, and N is the number of feature vectors; the GAP attention module comprises an improved GAP operation unit, and the input and output of the improved GAP operation unit satisfy the following relationship:

[0090]

[0091] wherein, ω j is an attention weight learned automatically by the model.

[0092] As an optional implementation, the loss function Loss of the neural network model is:

[0093] Loss = λ1L1 + λ2L2

[0094] wherein, λ1 and λ2 are positive hyperparameters, L1 is the cross-entropy loss of the original EfficientNet-B3 model, L2 is the cross-entropy loss of the GAP attention module, and the cross-entropy loss is minimized using the backpropagation algorithm:

[0095]

[0096] wherein, n is the number of training samples, C is the number of classes, y ik is the predicted probability of sample i and class k, is the output of the last hidden layer, is the weight matrix of the hidden layer to the output layer; I(·) is an indicator function, which takes 1 if the statement is true, and 0 otherwise.

[0097] S4, finally, using the trained model for fault diagnosis.

[0098] In order to better understand the present embodiment, the method of the present embodiment is further described below.

[0099] Reference Figure 3 , the method of the present embodiment is roughly divided into four stages of signal acquisition, data preprocessing, model training and fault diagnosis.

[0100] Data preprocessing stage: continuous wavelet transform (CWT, specifically complex Morlet wavelet transform) is selected to convert the original signal. In order to extract sufficient feature information from the original vibration signal, CWT is usually used to process the vibration signal. The time domain vibration signal is translated and expanded by the mother wavelet to convert it into an RGB image. However, the disadvantage of the traditional complex Morlet wavelet basis function is that its amplitude decay rate is constant, with fixed waveform characteristics, which will affect the matching ability for different waveform abnormal vibration signals. The expression of the traditional complex Morlet wavelet transform is:

[0101]

[0102] wherein, t is the time variable, ω0 is the center frequency parameter, which controls the oscillation frequency of the wavelet, and j is the imaginary unit. The improved wavelet basis function is expressed as:

[0103]

[0104] In the formula, η is the normalization coefficient; τ is the damping ratio parameter (0<τ<1), and the waveform adjustment factor is β:

[0105]

[0106] When ω0 is fixed according to specific scenarios, the value of the coefficient η can be determined by the wavelet basis function energy normalization condition. For a given original vibration signal x(t), the wavelet transform coefficient W(a, b) is calculated according to the formula of the complex Morlet wavelet transform:

[0107]

[0108] where ψ * is the conjugate complex of ψ(·), a is the scale parameter, and b is the translation parameter. By selecting different a and b values, the signal is analyzed at different scales and positions to obtain the wavelet transform coefficient W(a, b). Then the features of the wavelet transform coefficient are extracted, including amplitude information and phase information. The characteristic values are mapped to RGB, and for each scale parameter and translation parameter, a corresponding RGB value can be obtained. Arrange these RGB values according to certain rules (for example: sorting based on feature similarity, dynamic threshold screening arrangement, etc.), a two-dimensional image can be generated, where the horizontal coordinate can represent the translation parameter b, and the vertical coordinate represents the scale parameter a. The color of each pixel point is determined by the corresponding RGB value.

[0109] Then the model is trained: considering the accuracy and the limitation of available computing resources, the original EfficientNet-B3 is selected as the basic model, and the EfficientNet-B3 is a specific member of the EfficientNet model family (B0-B7), which is known for its excellent balance between computational efficiency, parameter quantity and accuracy. A global average pooling (GAP) attention module is added to the original EfficientNet-B3, as shown in Figure 2As shown. The model contains two independent branches, each with an output, which need to be jointly optimized. The GAP attention module is connected to the original EfficientNet-B3's 5×(MBCon6,k5×5,IRC), i.e., the target convolutional kernel mentioned above. Let the size of the input feature map of the GAP attention module be N×N×C, where N×N is the size of the two-dimensional feature map and C is the number of channels. The GAP attention module first uses two consecutive convolutional layers (i.e., Conv1×1+Swish) to compress the feature map to a size of N×N×16. Then, a locally connected2D layer (i.e., LocallyConnected2D) and a sigmoid activation function are used to learn N×N weights. Then, another convolutional layer is used to repeat the weights C times in the channel dimension, followed by a linear activation function (i.e., Conv1×1+Linear). The new feature map is then pooled into a feature vector of length C, and the result is scaled by dividing by the average weight vector. The output of the linear activation function is connected to the target convolutional kernel {i.e., ... Figure 2 The output of 5×(MBCon6,k5×5,IRC) is multiplied element-wise for weighted fusion. The resulting weighted fusion features are then fed into an improved GAP operation unit for processing. The result is then divided element-wise by the output of another improved GAP operation unit to adjust the feature ratio and normalize the data. Finally, the normalized features are fed into a fully connected layer (FC-Swish, where FC is short for Fully Connected Layer and Swish is the Swish activation function following the fully connected layer). The result of this fully connected layer is then fed into a fully connected layer with the softmax function applied (FC+softmax) to obtain the final output.

[0110] By learning to assign relative weights to corresponding feature vectors, the original EfficientNet-B3 model learns to focus on important regions of an image. The GAP operation in the original EfficientNet-B3 is shown below:

[0111]

[0112] Among them, L k and L k-1 These are the output and input layers of the GAP function, respectively. V j For L kwhere N is the number of feature vectors, and N is the number of feature vectors. Global average pooling (GAP) is a simple average operation with equal weights, which does not give any special attention to specific receptive fields or regions of the input image. In order to realize GAP attention, a GAP attention module is linked on the original EfficientNet-B3 model, and an improved GAP operation unit is set in the GAP attention module, which can calculate the weighted average value as follows:

[0113]

[0114] where ω j is the attention weight learned automatically by the model. A special network layer branch is added to the model to learn the most appropriate weight from each image and focus attention on the relevant area. The cross-entropy loss is minimized using the backpropagation algorithm:

[0115]

[0116] where n is the number of training samples, C is the number of classes, y ik is the predicted probability of sample i and class k, is the output of the last hidden layer, is the weight matrix from the hidden layer to the output layer; I(·) is an indicator function that takes 1 if the statement is true, otherwise 0.

[0117] Since there are two outputs, two cross-entropy losses (denoted as L1 and L2) are needed, and the optimal effect is achieved by optimizing the weighted sum of the two losses.

[0118] Loss=λ1L1+λ2L2

[0119] where λ1 and λ2 are positive hyperparameters that control the weight between the two branch outputs. In practical scenarios, λ1 and λ2 can both be 1 to indicate that L1 and L2 are equally important.

[0120] As shown in Figure 4 , the embodiment also discloses a vibration signal fault diagnosis device based on visual recognition, comprising:

[0121] A data acquisition module is configured to acquire original vibration signals.

[0122] An image generation module is configured to preprocess the original vibration signals to generate RGB pictures.

[0123] A model training module is configured to train a neural network model using the RGB pictures as input.

[0124] The fault diagnosis module is configured to perform fault diagnosis by using the trained model.

[0125] As an optional technical solution, the image generation module comprises:

[0126] The wavelet transform unit is configured to perform complex Morlet wavelet transform on the original vibration signal to obtain wavelet transform coefficients W(a, b).

[0127] The feature value extraction unit is configured to extract feature values of the wavelet transform coefficients w(a, b), the feature values including amplitude information and phase information.

[0128] The mapping unit is configured to perform RGB mapping on the feature values to generate the RGB picture.

[0129] As an optional implementation, the wavelet transform coefficients w(a, b) are calculated according to the following formula:

[0130]

[0131] wherein x(t) represents the original vibration signal, t is a time variable, ψ(·) is an improved complex Morlet wavelet function, ψ * (·) represents a conjugate complex number of ψ(·), a is a scale parameter, and b is a translation parameter; the improved complex Morlet wavelet function is:

[0132]

[0133] wherein η is a normalization coefficient, e is a natural logarithm, ω0 is a center frequency parameter, j is an imaginary unit, and β is a waveform adjustment factor, and β is calculated according to the following formula:

[0134]

[0135] wherein τ is a damping ratio-like parameter, and 0 < τ < 1.

[0136] As an optional implementation, the neural network model comprises an original EfficientNet-B3 model and a GAP attention module, wherein the GAP attention module is connected to a target convolution kernel in the original EfficientNet-B3 model. The EfficientNet-B3 model comprises a GAP operation unit, and the input and output of the GAP operation unit satisfy the following relationship:

[0137]

[0138] wherein L k and L k-1 are an output layer and an input layer of the GAP function respectively, V jFor L k characteristic vectors, N is the number of characteristic vectors; the GAP attention module comprises an improved GAP operation unit, and input and output of the improved GAP operation unit satisfy the following relationship:

[0139]

[0140] wherein ω j is an attention weight learned automatically by the model.

[0141] The loss function Loss of the neural network model is as follows:

[0142] Loss = λ1L1 + λ2L2

[0143] wherein λ1 and λ2 are positive hyperparameters, L1 is a cross-entropy loss of the original EfficientNet-B3 model, L2 is a cross-entropy loss of the GAP attention module, and the cross-entropy loss is minimized by using a back propagation algorithm:

[0144]

[0145] wherein n is the number of training samples, C is the number of classes, y ik is a predicted probability of the sample i and the class k, is the output of the last hidden layer, is a weight matrix of the hidden layer to the output layer; I(·) is an indicator function, which is 1 if the statement is true, and 0 otherwise. The specific working principle of the device can refer to the above method embodiments, which will not be repeated here.

[0146] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Any technical solution falling within the scope defined by the claims of the present application falls within the protection scope of the present application.

Claims

1. A vibration signal fault diagnosis method based on visual recognition, characterized in that, include: Acquire raw vibration signals; The original vibration signal is preprocessed to generate an RGB image; The RGB image is used as input to the neural network model for model training; Fault diagnosis is performed using the trained model.

2. The vibration signal fault diagnosis method based on visual recognition according to claim 1, characterized in that, The preprocessing includes: The original vibration signal is subjected to complex Morlet wavelet transform to obtain wavelet transform coefficients W(a,b); Extract the eigenvalues ​​of the wavelet transform coefficients W(a,b), wherein the eigenvalues ​​include amplitude information and phase information; The feature values ​​are mapped to RGB values ​​to generate the RGB image.

3. The vibration signal fault diagnosis method based on visual recognition according to claim 2, characterized in that, The wavelet transform coefficients W(a,b) are calculated according to the following formula: Where x(t) represents the original vibration signal, t is the time variable, and ψ(·) is the improved complex Morlet wavelet function. * (·) is the complex conjugate of ψ(·), a is the scaling parameter, and b is the translation parameter; the improved complex Morlet wavelet function is: Where η is the normalization coefficient, e is the natural logarithm, ω0 is the center frequency parameter, j is the imaginary unit, and β is the waveform adjustment factor, which is calculated according to the following formula: Where τ is a damping ratio parameter, and 0 < τ < 1.

4. The vibration signal fault diagnosis method based on visual recognition according to claim 1, characterized in that: The neural network model includes the original EfficientNet-B3 model and the GAP attention module. The GAP attention module is connected to the output of the target convolutional kernel in the original EfficientNet-B3 model. The target convolutional kernel refers to the last convolutional kernel in the original EfficientNet-B3 model with a scaling factor of 6 and a receptive field of 5×5.

5. The vibration signal fault diagnosis method based on visual recognition according to claim 4, characterized in that, The EfficientNet-B3 model includes a GAP operation unit, and the input and output of the GAP operation unit satisfy the following relationship: Among them, L k and L k-1 These are the output and input layers of the GAP function, respectively. j For L k The feature vectors contained in the layer, where N is the number of feature vectors; The GAP attention module includes an improved GAP operation unit, the input and output of which satisfy the following relationship: Where, ω j These are the attention weights that the model learns automatically.

6. The vibration signal fault diagnosis method based on visual recognition according to claim 4, characterized in that, The loss function Loss of the neural network model is: Loss=λ1L1+λ2L2 Where λ1 and λ2 are positive hyperparameters, L1 is the cross-entropy loss of the original EfficientNet-B3 model, and L2 is the cross-entropy loss of the GAP attention module. The cross-entropy loss is minimized using the backpropagation algorithm. Where n is the number of training samples, C is the number of classes, and y ik The predicted probabilities for sample i and class k, This is the output of the last hidden layer. The weight matrix from the hidden layer to the output layer; I(·) is an indicator function that takes the value 1 if the statement is true, and 0 otherwise.

7. A vibration signal fault diagnosis device based on visual recognition, characterized in that, include: The data acquisition module is used to acquire raw vibration signals; An image generation module is used to preprocess the original vibration signal to generate an RGB image; The model training module is used to train the neural network model by using the RGB image as input. The fault diagnosis module is used to diagnose faults using the trained model.

8. The vibration signal fault diagnosis device based on visual recognition according to claim 7, characterized in that, The image generation module includes: The wavelet transform unit is used to perform complex Morlet wavelet transform on the original vibration signal to obtain wavelet transform coefficients W(a,b); The feature extraction unit is used to extract the feature values ​​of the wavelet transform coefficients W(a,b), wherein the feature values ​​include amplitude information and phase information. A mapping unit is used to perform RGB mapping on the feature values ​​to generate the RGB image.

9. The vibration signal fault diagnosis device based on visual recognition according to claim 8, characterized in that, The wavelet transform coefficients W(a,b) are calculated according to the following formula: Where x(t) represents the original vibration signal, t is the time variable, and ψ(·) is the improved complex Morlet wavelet function. * (·) is the complex conjugate of ψ(·), a is the scaling parameter, and b is the translation parameter; the improved complex Morlet wavelet function is: Where η is the normalization coefficient, e is the natural logarithm, ω0 is the center frequency parameter, j is the imaginary unit, and β is the waveform adjustment factor, which is calculated according to the following formula: Where τ is a damping ratio parameter, and 0 < τ < 1.

10. The vibration signal fault diagnosis device based on visual recognition according to claim 7, characterized in that: The neural network model includes the original EfficientNet-B3 model and the GAP attention module. The GAP attention module is connected to the output of the target convolutional kernel in the original EfficientNet-B3 model. The target convolutional kernel refers to the last convolutional kernel in the original EfficientNet-B3 model with a scaling factor of 6 and a receptive field of 5×5. The EfficientNet-B3 model includes a GAP operation unit, and the input and output of the GAP operation unit satisfy the following relationship: Among them, L k and L k-1 These are the output and input layers of the GAP function, respectively. j For L k The feature vectors contained in the layer, where N is the number of feature vectors; The GAP attention module includes an improved GAP operation unit, the input and output of which satisfy the following relationship: Where, ω j These are the attention weights that the model learns automatically; The loss function Loss of the neural network model is: Loss=λ1L1+λ2L2 Where λ1 and λ2 are positive hyperparameters, L1 is the cross-entropy loss of the original EfficientNet-B3 model, and L2 is the cross-entropy loss of the GAP attention module. The cross-entropy loss is minimized using the backpropagation algorithm. Where n is the number of training samples, C is the number of classes, and y ik The predicted probabilities for sample i and class k, This is the output of the last hidden layer. The weight matrix from the hidden layer to the output layer; I(·) is an indicator function that takes the value 1 if the statement is true, and 0 otherwise.

Images