Motor fault diagnosis method based on fusion of vibration and sound sensors
By fusing the data of vibration and sound sensors in motor fault diagnosis, and using the SE attention mechanism and residual connection mechanism to build a feature fusion module and motor fault diagnosis model, the problem of low accuracy of fault diagnosis caused by single sensor signals is solved. , achieving more efficient and stable motor fault diagnosis.
Patent Information
- Application Number
- CN202411592572.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-08
- Publication Date
- 2025-02-18
AI Technical Summary
The existing motor fault diagnosis methods mainly rely on single sensor signals, resulting in low accuracy of fault diagnosis and limited anti-interference ability, making it difficult to achieve stable diagnostic results in complex environments.
The motor fault diagnosis method based on vibration and sound sensor fusion is adopted, and the multi-sensor data is acquired, and the feature fusion module is constructed using the SE attention mechanism and residual connection mechanism to output the fusion feature. Then, a motor fault diagnosis model based on dense blocks and convolutional blocks is constructed, the fusion features are diagnosed, the loss function is calculated using soft labels and the model parameters are updated.
Through the integration of multi-sensor data, the motor operating status can be more comprehensively reflected, the accuracy and stability of fault diagnosis can be improved, the anti-interference ability can be enhanced, and the accuracy of motor fault diagnosis can be significantly improved.
Smart Images
Figure CN119471360A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of fault diagnosis, and specifically relates to a motor fault diagnosis method based on the fusion of vibration and sound sensors, which can solve the problem of low fault diagnosis accuracy in the case of a single sensor and improve the fault diagnosis accuracy. Background Art
[0002] Motors are widely used in daily life due to their simple and reliable structure, low cost and easy control, playing an indispensable role in household and industrial society. Motors are usually in a long-term working state. After being corroded by harsh environment and worn out for a long time, they are prone to failure, leading to safety hazards or directly causing serious safety accidents.
[0003] Traditional fault diagnosis methods are mainly based on models and data-driven fault diagnosis. Model-based fault diagnosis methods are developed by analyzing signal features in different domains (such as time domain, frequency domain or time-frequency domain). Data-driven methods are implemented through machine learning algorithms, such as artificial neural networks, support vector machines, random forests, extreme learning machines, etc. However, this method relies on rich domain expert knowledge, which limits its promotion in practical applications. Deep learning has shown great application potential and advantages in natural language processing, speech recognition, image recognition and other solutions because it can automatically learn the hierarchical features and correlations between massive data, and has become a new solution to overcome its above-mentioned shortcomings. Compared with traditional machine learning algorithms, deep learning methods show superior performance when processing high-dimensional, nonlinear, and complex associated data. In addition, the high degree of automation and end-to-end learning capabilities of deep learning models enable them to better adapt to complex scenarios in practical applications, thereby significantly improving the accuracy and reliability of fault diagnosis.
[0004] However, most fault diagnosis methods only use single-channel signals as input to deep learning models. Single sensor signals can only reflect the local state of the motor system, and it is difficult to fully and accurately reveal the complex characteristics of motor faults. Secondly, the reliability and anti-interference ability of a single signal are limited, and it is easily affected by environmental noise, resulting in unstable diagnostic results, which in turn affects its application effect in actual scenarios. In order to improve the accuracy and robustness of fault diagnosis, researchers have gradually introduced the idea of multi-sensor information fusion into fault diagnosis.
[0005] The multi-sensor signal fusion method can more comprehensively reflect the running status of the motor by acquiring multiple physical signals from different parts of the motor, such as vibration, temperature, current, etc., and integrating multi-dimensional information. The fusion of multi-sensor data not only helps to improve the accuracy of feature extraction, but also enhances the anti-interference ability of fault diagnosis. The fusion method based on multi-sensor signals can make up for the shortcomings of single-channel signal diagnosis, thereby improving the accuracy and stability of motor fault diagnosis. At present, researchers are actively exploring the use of deep learning models to perform fusion analysis on multi-sensor signals in order to achieve more efficient and accurate fault diagnosis. Most existing data fusion methods integrate sensor data through simple feature splicing or weighted fusion, but this method fails to fully explore the complementary information between different sensors, resulting in the failure to efficiently utilize the features extracted by each sensor. In addition, existing methods usually lack effective modeling of the dependency relationship between multiple sensors, which makes it impossible for the feature information of each sensor to enhance each other, thereby weakening the effect of multi-sensor fusion in complex tasks. The attention mechanism can assign weights to different features to highlight important features while suppressing irrelevant features, but it mainly distributes weights based on local feature correlation, which easily ignores the global dependency between sensors. Summary of the invention
[0006] The purpose of the present invention is to provide a motor fault diagnosis method based on the fusion of vibration and sound sensors. The method can fuse the data of multiple sensors and introduce soft labels to improve the recognition between the collected data samples, making the entire motor fault diagnosis process more accurate and stable.
[0007] To achieve the above object, the technical solution adopted by the present invention is: A motor fault diagnosis method based on the fusion of vibration and sound sensors, the motor fault diagnosis method based on the fusion of vibration and sound sensors comprising: Acquire one-dimensional vibration data collected by a vibration sensor and one-dimensional sound data collected by a sound sensor, and convert the one-dimensional vibration data and the one-dimensional sound data into two-dimensional grayscale images respectively; A feature fusion module based on SE attention mechanism and residual connection mechanism is used to fuse the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, and output the fused features; Constructing a motor fault diagnosis model based on dense blocks and convolution blocks, and using the motor fault diagnosis model to output a diagnosis result for the fusion feature; Calculating the loss function of the current round using the soft labels generated in the previous round of training, and updating the parameters of the motor fault diagnosis model based on the loss function; Take the trained motor fault diagnosis model to perform motor fault diagnosis and output the diagnosis result.
[0008] Several optional methods are also provided below, but they are not intended to be additional limitations on the above-mentioned overall solution, but are merely further supplements or preferences. Under the premise that there are no technical or logical contradictions, each optional method can be combined with the above-mentioned overall solution separately, and multiple optional methods can also be combined.
[0009] Preferably, the converting of the one-dimensional vibration data and the one-dimensional sound data into two-dimensional grayscale images respectively comprises: The length of the one-dimensional vibration data obtained is , expressed as: in, represents one-dimensional vibration data, Represents the first data points, and ; The one-dimensional vibration data is divided into rows and columns to form a two-dimensional vibration matrix and reconstructed into a A two-dimensional vibration matrix, in which each element corresponds to a data point in the one-dimensional vibration data, is expressed as: Normalize the two-dimensional vibration matrix: in, is a two-dimensional vibration matrix Middle Line The elements of the column, and is a two-dimensional vibration matrix The maximum and minimum values in is the normalized two-dimensional vibration matrix Middle Line Elements of a column; The normalized two-dimensional vibration matrix Directly mapped to a two-dimensional vibration grayscale image, a two-dimensional vibration matrix An element in corresponds to a pixel in the two-dimensional vibration grayscale image. The two-dimensional vibration matrix An element value in corresponds to the pixel value of a pixel point in the two-dimensional vibration grayscale image; The length of the one-dimensional sound data obtained is , expressed as: in, represents one-dimensional sound data, Represents the first data points, and ; The one-dimensional sound data is divided into rows and columns to form a two-dimensional sound matrix, and reconstructed into a A two-dimensional sound matrix, in which each element corresponds to a data point in the one-dimensional sound data, is expressed as: Normalize the two-dimensional sound matrix: in, It is a two-dimensional sound matrix Middle Line The elements of the column, and It is a two-dimensional sound matrix The maximum and minimum values in is the normalized two-dimensional sound matrix Middle Line Elements of a column; The normalized two-dimensional sound matrix Directly mapped to a two-dimensional sound grayscale image, a two-dimensional sound matrix An element in corresponds to a pixel in the two-dimensional sound grayscale image. The two-dimensional sound matrix An element value in corresponds to the pixel value of a pixel point in the two-dimensional sound grayscale image.
[0010] Preferably, the feature fusion module constructed based on the SE attention mechanism and the residual connection mechanism is used to fuse the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, and output the fusion features, including: Take the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, respectively, and record them as images and ; Image and Input one-dimensional convolution and get: In the formula, represents one-dimensional convolution, Representing images The corresponding convolutional features are Representing images The corresponding convolution features; Using the SE attention mechanism to extract global features, we get: In the formula, Represents convolutional features The corresponding global features are Represents convolutional features The corresponding global features are Indicates the Squeeze operation. represents the size of a two-dimensional grayscale image, Represents convolutional features Middle Line The elements of the column, Represents convolutional features Middle Line Elements of a column; Perform Sigmoid function operation on the obtained global features to obtain normalized features: In the formula, Global feature The corresponding normalized features are Global feature The corresponding normalized features are Represents the Sigmoid function operation; Use normalized features as feature-level weights to enhance discriminative features: In the formula, Represents convolutional features The corresponding discriminant features are Represents convolutional features The corresponding discriminant features are represents the Hadamard product; The residual connection mechanism is introduced to obtain enhanced features, which can be expressed as: In the formula, Represents convolutional features The corresponding enhancement features, Represents convolutional features Corresponding enhancement features; Finally, the enhanced features are connected according to the channel dimension to obtain the fused features: In the formula, Represents fusion features.
[0011] Preferably, the structure of the motor fault diagnosis model is a convolutional layer, a first dense block, a second dense block, a first convolutional block, a third dense block, a second convolutional block, a fourth dense block, a global average pooling layer and a fully connected layer connected in sequence from the input side to the output side.
[0012] Preferably, the convolution kernel size of the convolution layer is 7x7, the stride=1, and the extended edge padding=3.
[0013] Preferably, the first dense block, the second dense block, the third dense block and the fourth dense block have the same structure, which is as follows: The first layer is a batch normalization layer; the second layer is a linear rectification layer; the third layer is a convolution layer with a kernel size of 3x3, stride=1, and padding=1; the fourth layer is a batch normalization layer; the fifth layer is a convolution layer with a kernel size of 3x3, stride=1, and no padding; the sixth layer is a linear rectification layer, which adds the initial input of the first layer and the output of the fifth layer in the channel direction and then inputs the sixth linear rectification layer.
[0014] Preferably, the first convolution block and the second convolution block have the same structure, which is as follows: First, it includes two branches, branch 1 and branch 2. For branch 1: the first layer is a batch normalization layer; the second layer is a linear rectification layer; the third layer is a convolution layer, the convolution kernel size is 3x3, the step size is stride=1, and the edge is expanded by padding=1; the fourth layer is a batch normalization layer; the fifth layer is a convolution layer, the convolution kernel size is 3x3, the step size is stride=1, and the edge is expanded by padding=1; for branch 2: the first layer is a convolution layer, the convolution kernel size is 1x1, the step size is stride=1, and the edge is not expanded. The second layer is a batch normalization layer; secondly, the outputs of branch 1 and branch 2 are added and input into the linear rectification layer.
[0015] Preferably, the calculation of the loss function of the current round using the soft labels generated by the previous round of training includes: In the In the round of training, Samples , calculate the loss function based on soft labels: In the formula, Indicates In the training round The soft-label-based loss function corresponding to each sample is: Indicates the total number of motor fault categories, Indicates the motor fault category index, Indicates Classes generated by round training Soft label, Indicates In the training round samples are predicted to belong to the class The probability of Indicates The number of samples correctly predicted in the training round, Indicates In the training round samples are predicted to belong to the class The probability of Then we get In the training round The total loss function corresponding to samples is: In the formula, Indicates In the training round The total loss function corresponding to samples is: Indicates In the training round The cross entropy loss function corresponding to samples is: is a hyperparameter.
[0016] The present invention provides a motor fault diagnosis method based on the fusion of vibration and sound sensors. The motor fault diagnosis network based on multiple sensors can fully learn the potential coupling characteristics between multiple sensors, thereby greatly improving the motor fault diagnosis accuracy and solving the problem of low fault diagnosis accuracy in existing motor fault diagnosis technologies. Compared with the prior art, the present invention normalizes and enhances the features of different sensors. Compared with traditional direct fusion and fusion using an attention module, the present invention is more flexible and simple, can better utilize the complementarity of each sensor, and uses residual connections to improve the feature recognition while ensuring that the original information is not lost. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a flow chart of the motor fault diagnosis method based on the fusion of vibration and sound sensors of the present invention; Figure 2 It is a structural schematic diagram of the feature fusion module of the present invention; Figure 3 It is a structural schematic diagram of the motor fault diagnosis model of the present invention. DETAILED DESCRIPTION
[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.
[0020] like Figure 1 As shown, this embodiment provides a motor fault diagnosis method based on vibration and sound sensor fusion, comprising the following steps: Step 1: Obtain one-dimensional vibration data collected by the vibration sensor and one-dimensional sound data collected by the sound sensor, and convert the one-dimensional vibration data and one-dimensional sound data into two-dimensional grayscale images respectively. If there are too many signals from the same sensor, PCA principal component analysis can be used to reduce their dimensionality.
[0021] Assume that the length of the obtained one-dimensional data is , expressed as: in, represents one-dimensional vibration data, Represents the first data points, and .
[0022] The one-dimensional vibration data is divided into rows and columns to form a two-dimensional vibration matrix and reconstructed into a A two-dimensional vibration matrix, in which each element corresponds to a data point in the one-dimensional vibration data, is expressed as: Normalize the two-dimensional vibration matrix: in, is a two-dimensional vibration matrix Middle Line The elements of the column, and is a two-dimensional vibration matrix The maximum and minimum values in is the normalized two-dimensional vibration matrix Middle Line Elements of a column.
[0023] The normalized two-dimensional vibration matrix Directly mapped to a two-dimensional vibration grayscale image, a two-dimensional vibration matrix An element in corresponds to a pixel in the two-dimensional vibration grayscale image. The two-dimensional vibration matrix An element value in corresponds to the pixel value of a pixel point in the two-dimensional vibration grayscale image.
[0024] Assume that the length of the obtained one-dimensional sound data is , expressed as: in, represents one-dimensional sound data, Represents the first data points, and .
[0025] The one-dimensional sound data is divided into rows and columns to form a two-dimensional sound matrix, and reconstructed into a A two-dimensional sound matrix, in which each element corresponds to a data point in the one-dimensional sound data, is expressed as: Normalize the two-dimensional sound matrix: in, It is a two-dimensional sound matrix Middle Line The elements of the column, and It is a two-dimensional sound matrix The maximum and minimum values in is the normalized two-dimensional sound matrix Middle Line Elements of a column.
[0026] The normalized two-dimensional sound matrix Directly mapped to a two-dimensional sound grayscale image, a two-dimensional sound matrix An element in corresponds to a pixel in the two-dimensional sound grayscale image. The two-dimensional sound matrix An element value in corresponds to the pixel value of a pixel point in the two-dimensional sound grayscale image.
[0027] In this embodiment, the one-dimensional data is converted into a two-dimensional grayscale image with a size of 128*128, and the image is used as training data.
[0028] Step 2: Use a feature fusion module based on SE attention mechanism (i.e. Squeeze-and-Excitation Attention, which is an attention mechanism based on feature channels) and residual connection mechanism to fuse the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, and output the fused features.
[0029] like Figure 2 As shown, the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor are taken and recorded as images and First, the image and Input one-dimensional convolution to generate new feature representation, and generate new feature representation without changing the feature space distribution through 1×1 convolution, and get:
[0030] In the formula, represents one-dimensional convolution, Representing images The corresponding convolutional features are Representing images Corresponding convolution features. After inputting the one-dimensional convolution, the local features of each sensor are extracted. Next, the dependency between different sensors is improved through the cross-modal enhancement strategy, and the Sigmoid function is operated on the features obtained through the convolution layer to generate a normalized feature map to represent the weight relationship between the sensors.
[0031] After extracting discriminative features, the SE attention mechanism is used to extract global features, mining important information in the features while suppressing irrelevant information, and obtaining: In the formula, Represents convolutional features The corresponding global features are Represents convolutional features The corresponding global features are Indicates the Squeeze operation. represents the size of a two-dimensional grayscale image, Represents convolutional features Middle Line The elements of the column, Represents convolutional features Middle Line Elements of a column; The obtained features are enhanced by cross-modal enhancement strategy to improve the dependency between different sensors, and the obtained global features are operated by Sigmoid function to obtain normalized feature mapping: In the formula, Global feature The corresponding normalized features are Global feature The corresponding normalized features are Represents the Sigmoid function operation.
[0032] The normalized features are then used as feature-level weight mapping to enhance the discriminant features. In this way, the common information between different sensors can be extracted and the correlation between different sensors can be improved.
[0033] In the formula, Represents convolutional features The corresponding discriminant features are Represents convolutional features The corresponding discriminant features are Represents Hadamard product; In this embodiment, by using normalized features as feature-level weight mapping, the module can effectively enhance the common features between different sensors, thereby improving the discrimination ability of each sensor and its mutual correlation.
[0034] In addition, the residual connection mechanism is introduced in the feature fusion module to obtain enhanced features, ensuring that the original information of each sensor can be retained during the fusion process, avoiding the loss of key information during feature transformation and processing. This mechanism effectively maintains the integrity of the input signal and enhances the model's ability to distinguish motor faults. It is expressed as:
[0035] In the formula, Represents convolutional features The corresponding enhancement features, Represents convolutional features Corresponding enhancement features; Finally, the enhanced features are connected according to the channel dimension to obtain fused features, which are then input into the subsequent network layer. This design can make full use of the complementary information between multiple sensors and improve the network's ability to discriminate multi-sensor data:
[0036] In the formula, Represents fusion features.
[0037] Step 3: construct a motor fault diagnosis model based on dense blocks (i.e., Dense Block) and convolution blocks (i.e., Conv Block), and use the motor fault diagnosis model to output a diagnosis result for the fusion feature.
[0038] The motor fault diagnosis model of this embodiment combines dense blocks and convolution blocks, and improves the efficiency of feature reuse through dense connections. Residual connections are used in convolution blocks to solve the gradient vanishing problem and enhance the deep learning ability of the model. At the same time, the number of parameters is reduced through multi-level feature extraction and global average pooling, which effectively improves the fault classification accuracy, training stability and generalization ability of the model, and performs well in processing complex fault models. The model includes 4 densely connected blocks, 2 convolution blocks, a global average pooling layer and an output layer, and the output layer uses a fully connected layer. The convolution block is composed of a series of convolution layers, which is designed to extract spatial features from multi-sensor data; the dense block is composed of dense connections, and each dense connection layer is connected to all previous layers, thereby realizing efficient transmission and reuse of features.
[0039] Define the learning rate as r, set r to 0.01, set the number of iterations variable to Num, input the obtained training data, and when the classification accuracy of the model is the highest in the iteration process, save the model parameters at this time to obtain a trained motor fault diagnosis model with motor fault diagnosis capabilities.
[0040] For dense blocks, the first layer is a batch normalization layer; the second layer is a linear rectification layer; the third layer is a convolutional layer with a kernel size of 3x3, a stride of 1, and padding of 1; the fourth layer is a batch normalization layer; the fifth layer is a convolutional layer with a kernel size of 3x3, a stride of 1, and no padding of 1; the initial input is added to the output of the fifth layer in the channel direction and then input into the sixth linear rectification layer to prevent the gradient from disappearing.
[0041] For the convolution block, the first layer of branch 1 is a batch normalization layer; the second layer is a linear rectification layer; the third layer is a convolution layer, the convolution kernel size is 3x3, the step size is stride=1, and the edge is expanded by padding=1; the fourth layer is a batch normalization layer; the fifth layer is a convolution layer, the convolution kernel size is 3x3, the step size is stride=1, and the edge is expanded by padding=1; further, for branch 2, the first layer is a convolution layer, the convolution kernel size is 1x1, the step size is stride=1, the edge is not expanded, and the second layer is a batch normalization layer; further, the outputs of branch 1 and branch 2 are added and input into the linear rectification layer.
[0042] like Figure 3As shown in the figure, based on the above scheme, the first layer: convolution layer (Conv), the input size is a grayscale image of 128*128*3, the convolution kernel size is 7x7, the step length stride=1, the expansion edge padding=3, 64 kernels are calculated respectively, the size of the feature map after convolution is 128*128*64, and then batch normalization is performed, and linear rectification is performed; The second layer: dense block, the size of the output feature map is 128*128*96; The third layer: dense block, the size of the output feature map is 128*128*128; The fourth layer: convolution block, the size of the output feature map is 128*128*128; The fifth layer: dense block, the size of the output feature map is 128*128*256; The sixth layer: convolution block, the size of the output feature map is 128*128*256; The seventh layer: dense block, the size of the output feature map is 128*128*512; The eighth layer: Global Avg-Pooling, which outputs a feature map of size 1*1*512 and has 512 neurons; The ninth layer: Fully Connected layer, outputs the diagnosis results.
[0043] The entire diagnostic process of the motor fault diagnosis model is to take a grayscale image of size 128*128, pass it through the first layer of convolution (the convolution kernel size is 7x7, the step size is stride=1, and the padding=3), perform batch normalization, perform linear rectification, and then input it into the dense block, dense block, convolution block, dense block, convolution block, dense block, global average pooling layer in sequence, and finally output the diagnosis result through the fully connected layer.
[0044] Step 4: Calculate the loss function of the current round using the soft labels generated in the previous round of training, and update the parameters of the motor fault diagnosis model based on the loss function.
[0045] In the first epoch of training, since the predicted labels cannot be obtained directly, a uniform distribution is used to initialize each category. Soft label : in, Indicates the total number of motor fault categories.
[0046] This embodiment adopts an online label smoothing strategy to update parameters during training. In the round epoch, given the Samples ,use Types of epoch generation Soft label To supervise the training of the entire network model, in the In the round of training, Samples , the loss function based on soft labels is expressed as: In the formula, For a given sample , its true category is , Indicates the motor fault category index, indicating a specific type of fault. Indicates In the training round samples are predicted to belong to the class This embodiment uses both hard labels and soft labels for supervision. In the training round The total loss function corresponding to samples is:
[0047] in, Indicates In the training round The total loss function corresponding to samples is: Indicates In the training round The cross entropy loss function corresponding to samples. The larger the cross entropy loss, the greater the difference between the model's prediction and the true label, and the worse the model performance; A hyperparameter to balance hard label loss and soft label loss.
[0048] At the end of each epoch, the soft label is updated using the predicted probability of the input sample. In the round epoch, for the correctly classified samples , the soft label of the corresponding category The update formula is: In the formula, Indicates In the training round samples are predicted to belong to the class The probability of . Except for the first round, at the beginning of each iteration, the soft label is initialized to 0, and then accumulated and updated according to the training situation. At the end of each epoch, the accumulated soft labels are normalized to ensure that the sum of the soft labels of each category is 1:
[0049] In the formula, Indicated in In the round epoch, for the correctly classified samples , the corresponding category Soft label.
[0050] Based on the updated soft labels of all correctly classified samples, the loss function can be further rewritten as: in, Indicates The number of samples correctly predicted in a round of training. This embodiment dynamically adjusts the soft labels of each category through online updating to better capture the similarities between categories and improve classification performance.
[0051] Step 5: Take the trained motor fault diagnosis model to perform motor fault diagnosis and output the diagnosis result.
[0052] In the inference application stage, the one-dimensional vibration data and one-dimensional sound data to be detected are converted into two-dimensional grayscale images, the two-dimensional grayscale images are input into the feature fusion module to obtain fusion features, and the fusion features are input into the trained motor fault diagnosis model to perform motor fault diagnosis and obtain diagnosis results.
[0053] It should be noted that the present embodiment mainly performs motor fault diagnosis on vibration and sound. In the actual implementation process, more dimensional data may be added, such as temperature data. The preprocessing method is the same as that of vibration and sound data. When fusing features, the fusion may be performed by combining two by two and fusing them in sequence, and then sent to the motor fault diagnosis model for processing.
[0054] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0055] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.
Claims
1. A motor fault diagnosis method based on vibration and sound sensor fusion, characterized in that: The motor fault diagnosis method based on vibration and sound sensor fusion includes: Acquire one-dimensional vibration data collected by a vibration sensor and one-dimensional sound data collected by a sound sensor, and convert the one-dimensional vibration data and the one-dimensional sound data into two-dimensional grayscale images respectively; A feature fusion module based on SE attention mechanism and residual connection mechanism is used to fuse the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, and output the fused features; Constructing a motor fault diagnosis model based on dense blocks and convolution blocks, and using the motor fault diagnosis model to output a diagnosis result for the fusion feature; Calculating the loss function of the current round using the soft labels generated in the previous round of training, and updating the parameters of the motor fault diagnosis model based on the loss function; Take the trained motor fault diagnosis model to perform motor fault diagnosis and output the diagnosis result.
2. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 1 is characterized in that: The step of converting the one-dimensional vibration data and the one-dimensional sound data into two-dimensional grayscale images comprises: The length of the one-dimensional vibration data obtained is , expressed as: in, represents one-dimensional vibration data, Represents the first data points, and ; The one-dimensional vibration data is divided into rows and columns to form a two-dimensional vibration matrix and reconstructed into a A two-dimensional vibration matrix, in which each element corresponds to a data point in the one-dimensional vibration data, is expressed as: Normalize the two-dimensional vibration matrix: in, is a two-dimensional vibration matrix Middle Line The elements of the column, and is a two-dimensional vibration matrix The maximum and minimum values in is the normalized two-dimensional vibration matrix Middle Line Elements of a column; The normalized two-dimensional vibration matrix Directly mapped to a two-dimensional vibration grayscale image, a two-dimensional vibration matrix An element in corresponds to a pixel in the two-dimensional vibration grayscale image. The two-dimensional vibration matrix An element value in corresponds to the pixel value of a pixel point in the two-dimensional vibration grayscale image; The length of the one-dimensional sound data obtained is , expressed as: in, represents one-dimensional sound data, Represents the first data points, and ; The one-dimensional sound data is divided into rows and columns to form a two-dimensional sound matrix, and reconstructed into a A two-dimensional sound matrix, each element in the two-dimensional sound matrix corresponds to a data point in the one-dimensional sound data, expressed as: Normalize the two-dimensional sound matrix: in, It is a two-dimensional sound matrix Middle Line The elements of the column, and It is a two-dimensional sound matrix The maximum and minimum values in is the normalized two-dimensional sound matrix Middle Line Elements of a column; The normalized two-dimensional sound matrix Directly mapped to a two-dimensional sound grayscale image, a two-dimensional sound matrix An element in corresponds to a pixel in the two-dimensional sound grayscale image. The two-dimensional sound matrix An element value in corresponds to the pixel value of a pixel point in the two-dimensional sound grayscale image.
3. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 1 is characterized in that: The feature fusion module constructed based on the SE attention mechanism and the residual connection mechanism is used to fuse the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, and output the fusion features, including: Take the two-dimensional grayscale images corresponding to the vibration sensor and the sound sensor, respectively, and record them as images and ; Image and Input one-dimensional convolution and get: In the formula, represents one-dimensional convolution, Representing images The corresponding convolutional features are Representing images The corresponding convolution features; Using the SE attention mechanism to extract global features, we get: In the formula, Represents convolutional features The corresponding global features are Represents convolutional features The corresponding global features are Indicates the Squeeze operation. represents the size of a two-dimensional grayscale image, Represents convolutional features Middle Line The elements of the column, Represents convolutional features Middle Line Elements of a column; Perform Sigmoid function operation on the obtained global features to obtain normalized features: In the formula, Global feature The corresponding normalized features are Global feature The corresponding normalized features are Represents the Sigmoid function operation; Use normalized features as feature-level weights to enhance discriminative features: In the formula, Represents convolutional features The corresponding discriminant features are Represents convolutional features The corresponding discriminant features are represents the Hadamard product; The residual connection mechanism is introduced to obtain enhanced features, which can be expressed as: In the formula, Represents convolutional features The corresponding enhancement features, Represents convolutional features Corresponding enhancement features; Finally, the enhanced features are connected according to the channel dimension to obtain the fused features: In the formula, Represents fusion features.
4. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 1 is characterized in that: The structure of the motor fault diagnosis model is a convolution layer, a first dense block, a second dense block, a first convolution block, a third dense block, a second convolution block, a fourth dense block, a global average pooling layer and a fully connected layer connected in sequence from the input side to the output side.
5. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 4 is characterized in that: The convolution kernel size of the convolution layer is 7x7, the stride is 1, and the padding is 3.
6. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 4 is characterized in that: The first dense block, the second dense block, the third dense block and the fourth dense block have the same structure, which is as follows: The first layer is a batch normalization layer; the second layer is a linear rectification layer; the third layer is a convolutional layer with a kernel size of 3x3, stride=1, and padding=1; the fourth layer is a batch normalization layer; the fifth layer is a convolutional layer with a kernel size of 3x3, stride=1, and no padding; The sixth layer is a linear rectification layer. The initial input of the first layer and the output of the fifth layer are added in the channel direction and then input into the sixth linear rectification layer.
7. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 4 is characterized in that: The first convolution block and the second convolution block have the same structure, which is as follows: First, it includes two branches, branch 1 and branch 2. For branch 1: the first layer is a batch normalization layer; the second layer is a linear rectification layer; the third layer is a convolution layer, the convolution kernel size is 3x3, the step size is stride=1, and the edge is expanded by padding=1; the fourth layer is a batch normalization layer; the fifth layer is a convolution layer, the convolution kernel size is 3x3, the step size is stride=1, and the edge is expanded by padding=1; for branch 2: the first layer is a convolution layer, the convolution kernel size is 1x1, the step size is stride=1, and the edge is not expanded. The second layer is a batch normalization layer; secondly, the outputs of branch 1 and branch 2 are added and input into the linear rectification layer.
8. The motor fault diagnosis method based on vibration and sound sensor fusion according to claim 1 is characterized in that: The calculation of the loss function of the current round using the soft labels generated by the previous round of training includes: In the In the round of training, Samples , calculate the loss function based on soft labels: In the formula, Indicates In the training round The soft-label-based loss function corresponding to each sample is: Indicates the total number of motor fault categories, Indicates the motor fault category index, Indicates Classes generated by round training Soft label, Indicates In the training round samples are predicted to belong to the class The probability of Indicates The number of samples correctly predicted in the training round, Indicates In the training round samples are predicted to belong to the class probability; Then we get In the training round The total loss function corresponding to samples is: In the formula, Indicates In the training round The total loss function corresponding to samples is: Indicates In the training round The cross entropy loss function corresponding to samples is: is a hyperparameter.