Bearing vibration signal processing model construction method, model and signal classification method

Abstract

The invention discloses a bearing vibration signal processing model construction method, a model and a signal classification method, which are used for constructing time domain and frequency domain loss functions by calculating a plurality of feature vector distances so as to train an encoder. According to the method, by minimizing the difference value of different distances in the time domain and frequency domain loss functions, an encoder can be guided to extract similar feature representations in signals, so that the similar feature representations are gathered in a feature space, and dissimilar features are separated. When a trained encoder (a convolutional neural network with an attention mechanism) processes a vibration signal of an unknown fault type, some feature representations in the signal are different from those of other signals, so that the feature representations are gathered in a feature space and are far away from the feature representations of other signals, and the fault detection accuracy is improved. Therefore, the recognition of the vibration signal of the unknown fault type is effectively realized.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing vibration analysis, and specifically to a method for constructing a bearing vibration signal processing model, and further to a processing model for feature extraction and classification of bearing vibration signals and a bearing vibration signal classification method based on the model. Background Art

[0002] Research shows that bearing failure is closely related to its vibration signal. The vibration signal of the bearing can clearly obtain various information about the bearing operation, which is helpful for in-depth analysis of the bearing status and early prediction of bearing damage. Therefore, analyzing the vibration signal of the bearing is an effective means to detect and analyze bearing failures.

[0003] The use of deep learning models for bearing fault detection is an advanced technology that can effectively identify and classify different types of faults through automated data collection, analysis, and feature extraction. This method relies on the mapping calculation of the bearing vibration signal by a neural network model, which can automatically learn and extract important patterns and features from the raw data. Compared with traditional rule-based or shallow machine learning methods, deep learning models perform well when processing large-scale data and complex signals. It can process data in real time or near real time, making online monitoring and early fault detection possible. The main advantage of this technology is its high degree of automation and ability to handle complex data structures, which is suitable for efficient equipment monitoring and maintenance systems in industry.

[0004] For example, Chinese patent application CN 118423236A discloses a wind turbine gearbox defect detection system based on vibration time domain and frequency domain signal analysis. The detection model is established by combining time domain and frequency domain data analysis, and optimized through training and testing. The calculation features are extracted from the time domain analysis and frequency domain analysis for fusion analysis, and the detection model is established through a neural network; the detection model is trained and verified using a training set of known defects. However, the time domain and frequency domain features (such as kurtosis, meshing index, angular similarity, etc.) in this bearing vibration detection model are extracted manually, which relies on experienced technicians and may not be able to fully capture all the characteristics of the gearbox vibration signal, especially when dealing with complex and nonlinear signal conditions. In addition, it uses supervised learning methods to train the defect detection model, which must rely on a large amount of labeled data. For wind turbine gearbox defect detection, it may be very difficult to obtain real, labeled defect data, especially when fault data of different types and degrees are relatively scarce. Its training module uses a training set of known defects to train and verify the detection model; its test optimization module also uses a test set of known defects to test and optimize the detection model. In other words, the fault types in the training set need to be manually labeled, which is not intelligent enough, and there are great limitations in collecting accurate labels for data of different fault types. Furthermore, the ratio of normal data to defective data is often seriously unbalanced, with most data being normal working conditions and relatively few defective data. Such unbalanced data will cause the model to tend to predict normal working conditions, affecting detection accuracy. In addition, scarce abnormal data also increases the difficulty of model training, making it difficult for the model to effectively learn fault characteristics.

[0005] For example, Chinese patent CN 114722952B discloses an integrated method for bearing fault detection and classification based on representation learning, which solves the problem that it is very difficult to obtain a large number of bearing fault samples in the existing data-driven method for bearing fault diagnosis, and there are great limitations in designing accurate labels for data of different fault types. The specific process is: 1. Extract bearing vibration data; 2. Feature extraction: extract data features from two perspectives: time domain and frequency domain; 3. Build a fault detection neural network: the fault detection neural network includes an encoder, noise introduction, self-attention mechanism layer and a decoder; 4. Train the fault detection neural network; 5. Build a fault classification neural network; 6. Train the fault classification neural network; 7. Online fault detection and fault classification. This technical method does not require any fault label data and can realize fault detection under unsupervised conditions. At the same time, under the condition of small sample faults, only a few fault samples are used for offline training to realize bearing fault classification. This technical method specifically proposes a transfer learning hypothesis that the source domain (healthy data) and the target domain (fault data) have similar feature distributions. However, if the data distribution of the target domain is significantly different from the healthy data, such as when the equipment or working conditions change significantly, the effect of transfer learning may be limited and the generalization ability of the model will be weakened. In addition, this technical method relies on the sufficiency and representativeness of healthy data. The model is trained on healthy data to learn the feature distribution in a healthy state. However, the feature information in the healthy data may be relatively simple and cannot cover potential fault characteristics, so the model may be difficult to effectively identify new unknown faults. Summary of the invention

[0006] In order to solve the problems in the prior art, the purpose of the present invention is to provide a method for constructing a bearing vibration signal processing model, a bearing vibration signal processing model and a bearing vibration signal classification method.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for constructing a bearing vibration signal processing model comprises the following steps:

[0009] Step S1: Obtain bearing vibration signal sample x i , forming a time domain signal sample set X∈{x1,x2,…,x n}, the length of each time domain signal is L; each bearing vibration signal sample x is transformed using fast Fourier transform i Transformed into bearing vibration frequency domain signal samples Form a frequency domain signal sample set corresponding to the time domain signal sample set X

[0010] Step S2: Calculate the DTW distance S between each pair of the bearing vibration time domain signal samples in the time domain signal sample set X. t (x i ,x j ), where x i ,x j refers to any two time-domain signal samples;

[0011] Step S3: Calculate the frequency domain signal sample set X f The Euclidean distance between the bearing vibration frequency domain signal samples in in and refers to any two frequency domain signal samples;

[0012] Step S4: respectively convert the time domain signal sample set X and the frequency domain signal sample set X f Input the respective bearing vibration signal processing encoders for mapping calculation to obtain the time domain signal sample feature representation set P with channel weights t ∈{p1,p2,…,p n}, and the frequency domain signal sample feature representation set with channel weights

[0013] Step S5: Calculate the time domain signal sample feature representation set P respectively t ∈{p1,p2,…,p n The sample characteristics of the bearing vibration time domain signal in} represent the Mahalanobis distance D between each pair. t (p i ,p j ) and the frequency domain signal sample feature representation set The characteristic representation of the bearing vibration frequency domain signal sample in the Mahalanobis distance between two

[0014] Step S6: The time domain signal sample x calculated in step S2 i and x j The DTW distance S t (x i ,x j ) and step S5 calculates the time domain signal sample x i and x j The Mahalanobis distance D t (p i ,p j ) constructs a time domain signal loss function for training the bearing vibration signal processing encoder, the time domain signal loss function is:

[0015] L t =smooth(D t (pi ,p j ),S t (x i ,x j ))

[0016] where p i and p j They are respectively the time domain signal samples x i and x j The time domain signal sample feature representation obtained after the encoder mapping calculation;

[0017] Similarly, the frequency domain signal samples calculated by step S3 are and Euclidean distance The frequency domain signal samples calculated in step S5 are and The Mahalanobis distance A frequency domain signal loss function for training the bearing vibration signal processing encoder is constructed, and the frequency domain signal loss function is:

[0018]

[0019] in They are frequency domain signal samples and The frequency domain signal sample feature representation obtained after the encoder mapping calculation;

[0020] Step S7, applying the calculated time domain signal loss function and frequency domain signal loss function to the respective bearing vibration signal processing encoders, and correspondingly training to form a time domain signal encoder and a frequency domain signal encoder;

[0021] Step S8: Substitute the time domain signal sample set X∈{x1,x2,…,x n} in the time domain signal samples xi and Frequency domain signal samples in Input the trained time domain signal encoder and frequency domain signal encoder to perform mapping calculation, and obtain the modified time domain signal sample feature representation p with channel weights i And the frequency domain signal sample feature representation p i f ;

[0022] Step S9: represent the time domain signal sample feature p i And the frequency domain signal sample feature representation p i f One-to-one fusion is performed, and the formula is: P fused =concat(p i ,pi f ), where concat represents the feature concatenation operation, which realizes the concatenation of two feature representations in corresponding dimensions.

[0023] Furthermore, the step S2 is specifically to use the softmin-DTW function to smooth the accumulated distance of possible paths. The specific formula of the softmin-DTW function is as follows:

[0024]

[0025] Among them, x i ,x j are two time domain signal samples with lengths L and S t (x i ,x j ) is the DTW distance between the two time domain signal samples. The smaller the calculated DTW distance is, the better the two time domain signal samples x are. i ,x j The more similar; is the tth data in the ith sample, is the number of samples in the ith sample that pass through path π and For the aligned data, π(t) represents the index corresponding to t on the path π; There are two points and The distance measure between It is the distance weighting of each path, and α is a hyperparameter used to adjust the contribution of different paths to the final distance.

[0026] Furthermore, the calculation formula used in step S3 is:

[0027]

[0028] in and are two frequency domain signal samples, is the Euclidean distance between the two frequency domain signal samples. The smaller the calculated Euclidean distance is, the more similar the two frequency domain signal samples are. M is the number of frequency intervals. and are the values ​​of samples i and j at the kth frequency component respectively.

[0029] Furthermore, in step S4, the encoder is a convolutional neural network with an attention mechanism.

[0030] Furthermore, the calculation formula used in step S5 is:

[0031]

[0032] where p i ,p j are the feature representations of two time domain signal samples, D t (p i ,p j ) is the Mahalanobis distance between the feature representations of the two time domain signal samples. The smaller the calculated Mahalanobis distance is, the more similar the feature representations of the two time domain signal samples are; (p i -p j ) T is the transposed matrix, S -1 is the inverse of the covariance matrix of the vector set;

[0033] Similarly, the Mahalanobis distance between the two sample feature representations of the bearing vibration frequency domain signal is calculated: in It is the characteristic representation of two frequency domain signal samples. The smaller the calculated Mahalanobis distance is, the more similar the characteristic representation of the two frequency domain signal samples is.

[0034] Furthermore, in step S6, a smooth function is designed to calculate the losses of both, and the formula is as follows:

[0035]

[0036] Where x is the DTW distance S t (x i ,x j ) and Mahalanobis distance D t (p i ,p j )

[0037] The present invention also discloses a bearing vibration signal processing model, which is constructed according to the construction method of the bearing vibration signal processing model as described above.

[0038] The present invention also discloses a bearing vibration signal classification method, which uses a linear classifier to classify the bearing vibration signal. The linear classifier uses a feature representation set R∈{P fused1 ,P fused2 ,…,P fusedi} to train.

[0039] The beneficial effects of the present invention are mainly reflected in:

[0040] 1. Improve generalization ability. On the one hand, the present invention adopts a self-supervised learning method to train the encoder (convolutional neural network with attention mechanism) on unlabeled samples, so that the encoder can process various normal or abnormal vibration signals, so that the encoder can learn a wider range of data distribution during the training process, and improve the generalization ability of the model. On the other hand, the feature multi-dimensional space after feature fusion contains more information, so that when the model encounters new or complex vibration signals, it can be analyzed from more angles, which is more conducive to improving the generalization ability of complex vibration signals.

[0041] 2. Enhance the ability to identify vibration signals of unknown bearing fault types. By minimizing the different distances D in the time domain and frequency domain loss functions t With S t and D f With S f The difference between them can guide the encoder (convolutional neural network with attention mechanism) to extract similar feature representations in the signal, aggregate them in the feature space, and separate dissimilar features. This process enables the encoder to learn how to classify similar signals. In this way, when the trained encoder (convolutional neural network with attention mechanism) processes vibration signals of unknown fault types, it can find that some feature representations in the signal are different from other signals, so that these feature representations are aggregated in the feature space and separated from the feature representations of other signals, thereby realizing the recognition of vibration signals of unknown fault types.

[0042] 3. Reduce the difficulty and cost of manually labeling samples. The fault labels of industrial equipment usually need to be labeled by engineers or experts with professional knowledge. However, professional labelers are not only costly, but also time-consuming and inefficient. The method proposed in the present invention only requires a small number of labeled samples when training the fault classifier, which can save a lot of manpower and time resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the framework of the method for constructing a bearing vibration signal processing model of the present invention;

[0044] Figure 2 It is a flow chart of the method for constructing a bearing vibration signal processing model according to the present invention. DETAILED DESCRIPTION

[0045] The present invention discloses a method for constructing a bearing vibration signal processing model based on self-supervised learning, and a bearing vibration signal processing model constructed thereby, which uses the idea of ​​contrastive learning to automatically identify and diagnose unlabeled bearing faults based on similarity measurement. When training time domain and frequency domain encoders (convolutional neural networks with attention mechanisms), it is not necessary to use labeled samples. Only a small amount of labeled samples are used when training linear classifiers through the model to achieve classification of bearing vibration signals, that is, for bearing fault detection.

[0046] The framework of the model building method of the present invention is as follows Figure 1 shown.

[0047] First, the collected bearing vibration signal samples are converted into frequency domain signal samples by FFT, thereby forming a frequency domain signal sample set, and the original vibration signal is in the form of a time domain signal sample set. Softmin-DTW is used to calculate the DTW distance between the time domain signal samples, and the Euclidean distance is used to calculate the Euclidean distance between the frequency domain signal samples. The size of the distance represents the similarity. At the same time, the two sample sets are respectively input into the encoder (convolutional neural network with attention mechanism) for mapping calculation, and the feature representation is output. The Mahalanobis distance function is used to calculate the distance between the time domain and frequency domain feature representations respectively to measure the similarity. More importantly, the present invention adopts a unique loss function to minimize the distance between the two signal feature representations and the distance between their corresponding actual signals. The loss function can minimize the difference between the similarity between the feature representations extracted by the encoder and the similarity between the actual signal samples. Minimizing this difference can guide the encoder to extract similar feature representations in the signal, so that similar feature representations are gathered together in the feature space, while dissimilar ones are far away. This process enables the encoder to learn how to organize data in the feature space for better classification and detection. Finally, the classifier is trained by the formed model to effectively classify the bearing vibration signal.

[0048] The specific method for constructing the bearing vibration signal processing model is as follows: Figure 2 shown.

[0049] S1. Collect bearing vibration signal samples x from the sensor i , forming a sample set X∈{x1,x2,…,x n}, the length of each signal is L. The vibration signal can be regarded as a time domain signal, so the time domain sample set is X. Use the fast Fourier transform (FFT) to transform each sample x i Converted into frequency domain signal samples Form a frequency domain signal sample set corresponding to the time domain signal sample set X

[0050] S2. For the time domain signal sample set X, measure the distance between each sample and the rest of the samples. The smaller the distance, the more similar the two samples are, and the greater the possibility that they belong to the same category. The traditional method directly uses the DTW algorithm to implement it, but the present invention proposes an improved algorithm that uses the softmin-DTW function to smoothly accumulate the distance of possible paths, avoiding the noise and local optimal problems that may be caused by the traditional DTW algorithm selecting a single path, making the path selection smoother and continuous, thereby improving the robustness of the model and its performance in gradient optimization. The specific formula is as follows,

[0051]

[0052] Among them, x i ,x j are two time domain signal samples with lengths L and S t (x i ,x j ) is the DTW distance between the two time domain signal samples. The smaller the calculated DTW distance is, the better the two time domain signal samples x are. i ,x j The more similar; is the tth data in the ith sample, is the number of samples in the ith sample that pass through path π and For the aligned data, π(t) represents the index corresponding to t on the path π; There are two points and We use the square of the Euclidean distance to measure the distance between them. It is a distance weighting for each path. α is a hyperparameter used to adjust the contribution of different paths to the final distance. When α takes a small value (close to 0), the contribution of all paths tends to be uniform. When α increases, the weight of paths with smaller distances will increase.

[0053] Since this formula is first proposed, its key calculation steps are explained as follows:

[0054] 2-1. Calculate the local distance matrix D

[0055] For the time series x i ,x j , the Euclidean distance is used to calculate the local distance between each data point. The formula is as follows

[0056]

[0057] Then construct the local distance matrix. The length of the signal sample is L, so D is an L×L matrix.

[0058]

[0059] 2-2. Initialize the cumulative distance matrix C

[0060] The cumulative distance matrix C is a (L+1)×(L+1) matrix. Each element C(i,j) in the matrix represents the minimum cumulative distance from the starting point to the current point. The boundaries of the matrix are initialized as: C(0,0)=0, C(i,0)=∞,

[0061] C(0,j)=∞.

[0062]

[0063] 2-3. Fill the cumulative distance matrix C. When filling the elements of the cumulative distance matrix, all possible paths are considered at each step and these paths are weighted. The final cumulative distance is stored in the position C(L,L), that is, the softmin calculation formula of two signal samples C(i,j) is as follows,

[0064] C(i,j)=D(i,j)+softmin(C(i-1,j),C(i,j-1),C(i-1,j-1)) (3)

[0065] Among them, the softmin function is defined as:

[0066]

[0067] S3. For the frequency domain signal sample set X f , also measures the similarity of samples, the formula is as follows:

[0068]

[0069] in and are two frequency domain signal samples, is the Euclidean distance between the two frequency domain signal samples. The smaller the calculated Euclidean distance is, the more similar the two frequency domain signal samples are. M is the number of frequency intervals. and are the values ​​of samples i and j at the kth frequency component respectively.

[0070] S4. respectively divide the time domain signal sample set X and the frequency domain signal sample set X f Input the respective bearing vibration signal processing encoders for mapping calculation to obtain the time domain signal sample feature representation set P with channel weights t ∈{p1,p2,…,p n}, and the frequency domain signal sample feature representation set with channel weights

[0071] The encoder is a convolutional neural network with an attention mechanism, which is a deep learning model that combines a convolutional neural network (CNN) with an attention mechanism. Based on the traditional convolutional neural network, by introducing the attention mechanism, it can focus more on the key information areas in the image or other data, thereby improving the performance of the model and the ability to understand the data.

[0072] The calculation process is described using the time domain signal as an example.

[0073] 4-1. Perform convolution calculation, the formula is as follows, where y i is the output value of the convolution operation, which can be regarded as the predicted feature representation, x i is the input value of the time domain signal, * represents the convolution operation, ω m is the convolution kernel parameter, size is 1×m, ω m Can be adjusted during training.

[0074] y i =x i *ω m (6)

[0075] 4-2. For y i Perform the Squeeze operation, the formula is as follows, where z i is the tie pooling, y i (c) represents the value of the i-th signal at the c-th position, and L represents the length (dimension) of the signal. The Squeeze operation is performed on each convolution signal y i Perform global average pooling and finally obtain a one-dimensional vector Z = [z1, z2, … z n ].

[0076]

[0077] 4-3. Perform Excitation operation. Consider vector Z as a set of channels, and capture the dependencies between each channel through two fully connected layers and nonlinear activation functions, so as to adaptively adjust the weights of the channels, so that important feature channels are enhanced and unimportant feature channels are suppressed. The formula is as follows. Where W1 and W2 are the weight parameters of the fully connected layer, δ is the ReLU activation function, and σ is the Sigmoid activation function. i is the i-th channel weight value, and finally the channel weight vector S = [s1, s2, …s n ].

[0078] s i =σ(W2δ(W1z i )) (8)

[0079] 4-4. Perform the Reweighting operation and calculate the prediction feature representation with weight information according to the channel weight vector S. The formula is as follows. Finally, the prediction feature representation set P of the time domain signal sample with channel weight is obtained. t ∈{p1,p2,…,p n}.

[0080] p i =y i ·s i (9)

[0081] Similarly, we get the frequency domain signal sample prediction feature representation set

[0082] In the present invention, steps 4-2 to 4-4 are a channel attention algorithm (SeNet). By analyzing and processing the relationship between feature channels, SENet selectively learns important feature information to obtain the importance of each feature channel. According to this importance, useful features are promoted and features that are not very useful for the current task are suppressed, thereby effectively improving the diagnostic classification accuracy of the model.

[0083] S5. Calculate the time domain signal sample feature representation set P respectively t ∈{p1,p2,…,p n The sample characteristics of the bearing vibration time domain signal in} represent the Mahalanobis distance D between each pair. t (p i ,p j ) and the frequency domain signal sample feature representation set The characteristic representation of the bearing vibration frequency domain signal sample in the Mahalanobis distance between two Mahalanobis distance is a method that can be used to measure the distance of high-dimensional vectors. It captures the linear correlation between features through the covariance matrix, making the calculated distance more real and reliable, especially in anomaly detection, pattern recognition and classification of high-dimensional data.

[0084] The calculation process is described using the time domain signal as an example.

[0085] In the time domain signal sample feature representation set P t ∈{p1,p2,…,p n}, any two time domain signal sample feature representations p i ,p j , can be abstractly understood as two vectors. The similarity between feature representations is measured by calculating the Mahalanobis distance. The smaller the distance, the greater the similarity. The formula is as follows,

[0086]

[0087] Where Dt (p i ,p j ) is the Mahalanobis distance between the feature representations of the two time domain signal samples; S -1 is the inverse matrix of the covariance matrix of the vector set. The covariance matrix S represents the linear dependence between the values ​​in the vector. The inverse matrix S -1 Used to adjust the scale and correlation between vector values. (p i -p j ) T is the transposed matrix, which is also a vector.

[0088] The steps to calculate the covariance matrix S are as follows:

[0089] 5-1. Assume that the vector p i ,p j The length of R is R, calculate the vector p i and p j The mean of is as follows.

[0090] are the means of two vectors, and p(i) is the i-th value in the vector.

[0091]

[0092] 5-2. In order to eliminate the differences between the vector values, the data is centered. The formula is as follows. i Subtract the mean from each value in the vector to get the centered value. Similarly,

[0093]

[0094] 5-3. Calculate the covariance matrix S, the formula is as follows. is the vector p i The variance of is the vector p j The variance of is the vector p i and p j The covariance between .

[0095]

[0096] Similarly, the Mahalanobis distance between the two sample feature representations of the bearing vibration frequency domain signal is calculated: in It is the characteristic representation of two frequency domain signal samples. The smaller the calculated Mahalanobis distance is, the more similar the characteristic representation of the two frequency domain signal samples is.

[0097] S6. The DTW distance S calculated in step S2 t (x i ,x j ) and the Mahalanobis distance D calculated in step S5 t (p i ,p j ) constructs a time domain signal loss function for training the bearing vibration signal processing encoder.

[0098] The present invention defines the time domain signal loss function of the training model as:

[0099] L t =smooth L1 (D t (p i ,p j ),S t (x i ,x j )) (16)

[0100] where p i and p j They are respectively the time domain signal samples x i and x j Feature representation obtained after mapping and calculation by a convolutional neural network with an attention mechanism.

[0101] D t (p i ,p j ) is the Mahalanobis distance between two feature representations, S t (x i ,x j ) are two signal samples x i and x j The DTW distance between them. A smooth function is designed to calculate the loss of the two. The formula is as follows:

[0102]

[0103] Where x is the DTW distance S t (x i ,x j ) and Mahalanobis distance D t (p i ,p j )

[0104] Similarly, the Euclidean distance calculated in step S3 is and the Mahalanobis distance calculated in step S5 A frequency domain signal loss function for training the bearing vibration signal processing encoder is constructed, and the frequency domain signal loss function is:

[0105]

[0106] in They are frequency domain signal samples and The frequency domain signal sample feature representation obtained after the encoder mapping calculation.

[0107] The goal of this loss function is to minimize the difference between the similarity of two feature representations extracted by the model and the similarity of two actual signal samples. The design of the loss function of the present invention uses the idea of ​​contrastive learning. Minimizing this difference can guide the model to extract similar feature representations in the signal and map them to the feature space, so that similar feature representations are clustered together in the feature space, while dissimilar ones are far away. This process enables the encoder to learn how to organize data in the feature space to better classify similar signals. When the trained model processes new unlabeled signals, the mapping calculation of the model can automatically identify similar features in the signal and map these features to the feature space, so that the feature representations of similar signal samples are clustered in the special space, which is beneficial to downstream classification tasks.

[0108] S7. Apply the calculated time domain signal loss function and frequency domain signal loss function to their respective bearing vibration signal processing encoders, and correspondingly train to form a time domain signal encoder and a frequency domain signal encoder.

[0109] S8. The time domain signal sample set X∈{x1,x2,…,x n} in the time domain signal samples xi and Frequency domain signal samples in Input the trained time domain signal encoder and frequency domain signal encoder to perform mapping calculation, and obtain the modified time domain signal sample feature representation p with channel weights i And the frequency domain signal sample feature representation p i f ;

[0110] S9. Represent the time domain signal sample feature p i And the frequency domain signal sample feature representation p i f One-to-one fusion is performed, and the formula is: P fused =concat(p i ,p i f ), where concat represents the feature concatenation operation, which realizes the concatenation of two feature representations in corresponding dimensions.

[0111] The present invention also discloses a bearing vibration signal classification method, which uses a linear classifier to classify the bearing vibration signal. The linear classifier uses a feature representation set R∈{P fused1 ,P fused2 ,…,P fusedi} to train, and then realize bearing fault detection and classification in industrial sites.

[0112] The following is a brief description of the training process of the linear classifier.

[0113] 10-1. First, select a small number of initial vibration signal samples and manually label them. Each sample belongs to a category y∈{1,2…,C}, where C is the total number of categories;

[0114] 10-2. Calculate using the linear calculation formula: z = WP fused +b, where W is the weight matrix, b is the bias vector, and z is the score for each category, i.e., the unnormalized probability;

[0115] 10-3. Use the Softmax function to convert the linear output score into a probability distribution for each category

[0116] in is the predicted probability of category i; z i is the score of category i; z j is the score of category j;

[0117] 10-4. Use the cross entropy loss function to calculate the degree of match between the predicted category probability and the true label

[0118] where y i is the one-hot encoding of the true label of category i, is the predicted probability of category i;

[0119] 10-5. Optimize the linear calculation of W and b through back propagation to improve The predicted probability of is as close to 1 as possible;

[0120] 10-6. Select the category corresponding to the maximum probability as the output result,

[0121] The present invention extracts features useful for fault diagnosis from the original data by comparing the similarities between different samples, and classifies and predicts bearing faults based on these features. Compared with traditional supervised learning methods, this method has the following advantages:

[0122] 1. Reduce data annotation costs:

[0123] Traditional labeled methods require a large amount of labeled data to train the model, but labeled data is often difficult to obtain and costly. The intelligent bearing fault diagnosis framework based on self-supervised learning does not require a large amount of labeled data through self-supervised learning, reducing the cost and difficulty of data labeling.

[0124] 2. Improve the generalization ability of the model:

[0125] Contrastive learning learns feature representation by capturing the similarities and differences between positive and negative samples. This learning method enables the model to learn more general features, thereby improving the generalization ability of the model. In contrast, traditional labeled methods may be more likely to overfit specific patterns in the training data, resulting in poor performance on unseen data.

[0126] 3. Adapt to complex environment:

[0127] In complex industrial environments, bearing fault signals are often interfered by multiple factors. The intelligent bearing fault diagnosis framework based on self-supervised learning can better handle these complex nonlinear relationships and improve the accuracy and stability of fault diagnosis through comparative learning.

[0128] 4.Flexibility and Scalability:

[0129] The contrastive learning model can be easily extended to new fault types and new application scenarios. Since it does not require a large amount of labeled data, it can be more easily adapted to new data and tasks. In contrast, traditional labeled methods may need to re-collect and annotate a large amount of data when expanding to new areas.

[0130] The present invention greatly reduces training parameters, saves computing resources, improves response time, and is suitable for application in scenarios of mobile equipment and online real-time monitoring.

[0131] The present invention also discloses a bearing vibration signal processing model, which is constructed by the bearing vibration signal processing model construction method described above in this article.

[0132] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a bearing vibration signal processing model, characterized in that: The following steps are included: Step S1, obtaining bearing vibration signal sample x i , forming a time domain signal sample set X∈{x1,x2,…,x n }, the length of each time domain signal is L; each bearing vibration signal sample x is transformed using fast Fourier transform i Transformed into bearing vibration frequency domain signal samples Form a frequency domain signal sample set corresponding to the time domain signal sample set X Step S2: Calculate the DTW distance S between each pair of the bearing vibration time domain signal samples in the time domain signal sample set X. t (x i ,x j ), where x i ,x j refers to any two time-domain signal samples; Step S3: Calculate the frequency domain signal sample set X f The Euclidean distance between the bearing vibration frequency domain signal samples in in and refers to any two frequency domain signal samples; Step S4: respectively convert the time domain signal sample set X and the frequency domain signal sample set X f Input the respective bearing vibration signal processing encoders for mapping calculation to obtain the time domain signal sample feature representation set P with channel weights t ∈{p1,p2,…,p n }, and the frequency domain signal sample feature representation set with channel weights Step S5: Calculate the time domain signal sample feature representation set P respectively t ∈{p1,p2,…,p n The sample characteristics of the bearing vibration time domain signal in} represent the Mahalanobis distance D between each pair. t (p i ,p j ) and the frequency domain signal sample feature representation set The characteristic representation of the bearing vibration frequency domain signal sample in the Mahalanobis distance between two Step S6: The time domain signal sample x calculated in step S2 i and x j The DTW distance S t (x i ,x j ) and step S5 calculates the time domain signal sample x i and x j The Mahalanobis distance D t (p i ,p j ) constructs a time domain signal loss function for training the bearing vibration signal processing encoder, the time domain signal loss function is: L t =smooth(D t (p i ,p j ),S t (x i ,x j )) where p i and p j They are respectively the time domain signal samples x i and x j The time domain signal sample feature representation obtained after the encoder mapping calculation; Similarly, the frequency domain signal samples calculated by step S3 are and Euclidean distance The frequency domain signal samples calculated in step S5 are and The Mahalanobis distance A frequency domain signal loss function for training the bearing vibration signal processing encoder is constructed, and the frequency domain signal loss function is: in They are frequency domain signal samples and The frequency domain signal sample feature representation obtained after the encoder mapping calculation; Step S7, applying the calculated time domain signal loss function and frequency domain signal loss function to the respective bearing vibration signal processing encoders, and correspondingly training to form a time domain signal encoder and a frequency domain signal encoder; Step S8: Substitute the time domain signal sample set X∈{x1,x2,…,x n } in the time domain signal samples xi and Frequency domain signal samples in Input the trained time domain signal encoder and frequency domain signal encoder to perform mapping calculation, and obtain the modified time domain signal sample feature representation p with channel weights i And the frequency domain signal sample feature representation p i f ; Step S9: represent the time domain signal sample feature p i And the frequency domain signal sample feature representation p i f One-to-one fusion is performed, and the formula is: P fused =concat(p i ,p i f ), where concat represents the feature concatenation operation, which realizes the concatenation of two feature representations in corresponding dimensions.

2. The method according to claim 1, characterized in that The step S2 specifically uses the softmin-DTW function to smoothly accumulate the distance of possible paths. The specific formula of the softmin-DTW function is as follows: Among them, x i ,x j are two time domain signal samples with lengths L and S t (x i ,x j ) is the DTW distance between the two time domain signal samples. The smaller the calculated DTW distance is, the better the two time domain signal samples x are. i ,x j The more similar; is the tth data in the ith sample, is the number of samples in the ith sample that pass through path π and For the aligned data, π(t) represents the index corresponding to t on the path π; There are two points and The distance measure between It is the distance weighting of each path, and α is a hyperparameter used to adjust the contribution of different paths to the final distance.

3. The method according to claim 1, characterized in that The calculation formula used in step S3 is: in and are two frequency domain signal samples, is the Euclidean distance between the two frequency domain signal samples. The smaller the calculated Euclidean distance is, the more similar the two frequency domain signal samples are. M is the number of frequency bins; and are the values ​​of samples i and j at the kth frequency component respectively.

4. The method according to claim 1, characterized in that: In step S4, the encoder is a convolutional neural network with an attention mechanism.

5. The method according to claim 1, characterized in that The calculation formula used in step S5 is: where p i ,p j are the feature representations of two time domain signal samples, D t (p i ,p j ) is the Mahalanobis distance between the feature representations of the two time domain signal samples. The smaller the calculated Mahalanobis distance is, the more similar the feature representations of the two time domain signal samples are; (p i -p j ) T is the transposed matrix, S -1 is the inverse of the covariance matrix of the vector set; Similarly, the Mahalanobis distance between the two sample feature representations of the bearing vibration frequency domain signal is calculated: in It is the characteristic representation of two frequency domain signal samples. The smaller the calculated Mahalanobis distance is, the more similar the characteristic representation of the two frequency domain signal samples is.

6. The method according to claim 1, characterized in that In step S6, a smooth function is designed to calculate the losses of both, and the formula is as follows: Where x is the DTW distance S t (x i ,x j ) and Mahalanobis distance D t (p i ,p j ) 7. A bearing vibration signal processing model, characterized in that: The bearing vibration signal processing model is constructed by the construction method of the bearing vibration signal processing model according to any one of claims 1-6.

8. A method for classifying bearing vibration signals, characterized in that: The bearing vibration signal is classified by a linear classifier, wherein the linear classifier uses a feature representation set R∈{P fused1 ,P fused2 ,…,P fusedi } to train.

Images