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

Abstract

The present invention discloses a method for constructing a bearing vibration signal processing model, a model and a signal classification method. By calculating the distances of multiple feature vectors, time-domain and frequency-domain loss functions are constructed, which are then used to train an encoder. By minimizing the difference in different distances in the time-domain and frequency-domain loss functions, the present invention can guide the encoder to extract similar feature representations in the signal, causing them to aggregate in the feature space while separating dissimilar features. When the trained encoder (a convolutional neural network with an attention mechanism) in the present invention processes vibration signals of unknown fault types, it can find that some feature representations in the signal are different from those of other signals, thus aggregating these feature representations in the feature space and distancing them from the feature representations of other signals, thereby effectively realizing the recognition of vibration signals of unknown fault types.

Description

Technical Field

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

[0002] Research shows that there is a close relationship between bearing faults and their vibration signals. The vibration signals of bearings can clearly obtain various information about the operation of the bearings, which helps to deeply analyze the bearing state and early predict the damage of the bearings. Therefore, analyzing the vibration signals of bearings is an effective means for detecting and analyzing bearing faults.

[0003] Using a deep learning model for bearing fault detection is an advanced technology. Through automated data collection, analysis, and feature extraction, it can effectively identify and classify different types of faults. This method relies on the mapping calculation of the neural network model for bearing vibration signals and can automatically learn and extract important patterns and features from the original data. Compared with traditional rule-based or shallow machine learning methods, the deep learning model performs excellently in 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 advantages of this technology lie in its high degree of automation and the ability to process complex data structures, which are suitable for efficient equipment monitoring and maintenance systems in industry.

[0004] For example, Chinese Patent Application CN 118423236A discloses a wind power gearbox defect detection system based on vibration time-domain and frequency-domain signal analysis. By combining time-domain and frequency-domain data analysis, a detection model is established and optimized through training and testing. Computational features are extracted from time-domain analysis and frequency-domain analysis for fusion analysis, and a detection model is established through a neural network. A training set with known defects is used to train and validate the detection model. 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. This method relies on experienced technicians and may not be able to comprehensively capture all the characteristics of the gearbox vibration signal, especially in the case of dealing with complex and non-linear signals. In addition, it uses a supervised learning method to train the defect detection model, which must rely on a large amount of labeled data. For wind power gearbox defect detection, it may be very difficult to obtain real and labeled defect data, especially the fault data of different types and degrees is relatively scarce. Its training module uses a training set with known defects to train and validate the detection model; its test optimization module also uses a test set with known defects to test and optimize the detection model. That is to say, the fault types in the training set need to be manually labeled, and the intelligence is insufficient. There are great limitations in accurately collecting labels for data of different fault types. Further, the ratio of normal data to defect data is often seriously unbalanced. Most of the data is in normal working conditions, and the defect data is relatively small. Such unbalanced data will cause the model to tend to predict normal working conditions, affecting the detection accuracy. In addition, the scarce abnormal data also increases the difficulty of model training, making it difficult for the model to effectively learn fault characteristics.

[0005] For another example, Chinese Patent CN 114722952B discloses an integrated method for bearing fault detection and classification based on representation learning, which solves the problems that it is very difficult to obtain a large number of bearing fault samples in the process of bearing fault diagnosis using existing data-driven methods, and there are great limitations in designing accurate labels for data of different fault types. The specific process is as follows: 1. Extract the vibration data of the bearing; 2. Feature extraction: Extract the features of the data from two perspectives of time domain and frequency domain; 3. Build a fault detection neural network: The fault detection neural network includes an encoder, noise introduction, a 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 achieve 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 achieve bearing fault classification. This technical method specifically proposes a transfer learning hypothesis that there is a similar feature distribution between the source domain (healthy data) and the target domain (fault data). However, if the data distribution of the target domain is quite different from the healthy data, for example, when the equipment or working conditions change significantly, the effect of transfer learning may be limited, and the generalization ability of the model will also be weakened. Moreover, this technical method depends on the sufficiency and representativeness of healthy data. The model is trained on healthy data to learn the feature distribution in the healthy state. However, the feature information in the healthy data may be relatively single and cannot cover potential fault features. Therefore, the model may be difficult to effectively identify newly emerging unknown faults. Summary of the Invention

[0006] 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 purpose, the present invention provides the following technical solutions:

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

[0009] Step S1, obtain the bearing vibration signal sample x i , form a time-domain signal sample set X ∈ {x1, x2, …, x n}}, and the length of each time-domain signal is L; use the fast Fourier transform to transform each of the bearing vibration signal samples x i into a bearing vibration frequency-domain signal sample to 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 every two 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 Euclidean distance between every two bearing vibration frequency-domain signal samples in the frequency-domain signal sample set X f where and and refer to any two frequency-domain signal samples;

[0012] Step S4: Respectively input the time-domain signal sample set X and the frequency-domain signal sample set X f into their respective bearing vibration signal processing encoders for mapping calculation to obtain a set P of time-domain signal sample feature representations with channel weights t ∈ {p1, p2, …, p n}, and a set of frequency-domain signal sample feature representations with channel weights

[0013] Step S5: Respectively calculate the Mahalanobis distance D t ∈ {p1, p2, …, p n} between every two bearing vibration time-domain signal sample feature representations in the time-domain signal sample feature representation set P t (p i , p j ) and the Mahalanobis distance between every two bearing vibration frequency-domain signal sample feature representations in the frequency-domain signal sample feature representation set

[0014] Step S6: Construct a time-domain signal loss function for training the bearing vibration signal processing encoder through the DTW distance S i and x j of the time-domain signal samples obtained by Step S2 and the Mahalanobis distance D t (x i , x j ) and the Mahalanobis distance between the time-domain signal samples x i and x j obtained by Step S5. This time-domain signal loss function is: t (p i , p j ). The time-domain signal loss function is:

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

[0016] where p i and p j are the time - domain signal sample feature representations of the time - domain signal samples x i and x j obtained after the mapping calculation by the encoder;

[0017] Similarly, the Euclidean distance between the frequency - domain signal samples and calculated by the step S3 and the Mahalanobis distance between the frequency - domain signal samples calculated by the step S5 and and are used to construct a frequency - domain signal loss function for training the bearing vibration signal processing encoder. The frequency - domain signal loss function is: where

[0018]

[0019] where are the frequency - domain signal sample feature representations of the frequency - domain signal samples and obtained after the mapping calculation by the encoder;

[0020] Step S7: Apply the calculated time - domain signal loss function and frequency - domain signal loss function to their respective bearing vibration signal processing encoders, and train to form a time - domain signal encoder and a frequency - domain signal encoder;

[0021] Step S8: Input the time - domain signal sample xi in the time - domain signal sample set X ∈ {x1, x2, …, x n} and the frequency - domain signal sample in into their respective trained time - domain signal encoder and frequency - domain signal encoder for mapping calculation, and obtain the corrected time - domain signal sample feature representation p i with channel weights and the frequency - domain signal sample feature representation p i f ;

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

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

[0024]

[0025] where x i , x j are two time-domain signal samples with a length of 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, the more similar the two time-domain signal samples x i , x j are; is the t-th data in the i-th sample, is the data in the i-th sample that is aligned with through the path π, and π(t) represents the index corresponding to t on the path π; is the distance metric between two points and ; is to weight the distance of each path, and α is a hyperparameter used to adjust the contribution degree of different paths to the final distance.

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

[0027]

[0028] where and are two frequency-domain signal samples, is the Euclidean distance between the two frequency-domain signal samples. The smaller the calculated Euclidean distance, the more similar the two frequency-domain signal samples are; M is the number of frequency intervals; and are the values of the i-th and j-th samples on the k-th frequency component respectively.

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

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

[0031]

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

[0033] Similarly, calculate the Mahalanobis distance between every two of the feature representations of the bearing vibration frequency-domain signal samples where are the feature representations of two frequency-domain signal samples. The smaller the calculated Mahalanobis distance, the more similar the two feature representations of frequency-domain signal samples are.

[0034] Furthermore, in the step S6, design a smooth function to calculate the loss between the two, and the formula is as follows,

[0035]

[0036] where x is the DTW distance S t (x i , x j ) and the difference between the 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 signals. The linear classifier is trained using the set of feature representations R ∈ {P fused1 , P fused2 , …, P fusedi} obtained by splicing according to the above method.

[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 an encoder (a convolutional neural network with an attention mechanism) on unlabeled samples. Therefore, the encoder can process various normal or abnormal vibration signals, enabling the encoder to learn a more extensive data distribution during training and improving the generalization ability of the model. On the other hand, the multi-dimensional feature space after feature fusion contains more information, enabling the model to analyze from more perspectives when encountering new or complex vibration signals, which is more conducive to improving the generalization ability of complex vibration signals.

[0041] 2. Enhance the recognition ability of vibration signals of unknown bearing fault types. By minimizing the differences between different distances D t and S t as well as between D f and S f in the time domain and frequency domain loss functions, it is possible to guide the encoder (a convolutional neural network with an attention mechanism) to extract similar feature representations in the signal, causing them to cluster in the feature space while separating dissimilar features. This process enables the encoder to learn how to classify similar signals. In this way, when the trained encoder (a convolutional neural network with an attention mechanism) processes vibration signals of unknown fault types, it can find that some feature representations in the signal are different from other signals, thus clustering these feature representations in the feature space and separating them 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 labeled samples. Fault labels of industrial equipment usually need to be labeled by engineers or experts with professional knowledge. However, professional labelers not only have high costs but also take a long time and are inefficient in the labeling process. The method proposed in the present invention only requires a small number of labeled samples when training the fault classifier, which can save a large amount of human and time resources. Brief Description of the Drawings

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

[0044] Figure 2 is a flow schematic diagram of the method for constructing a bearing vibration signal processing model of the present invention. Detailed Embodiments

[0045] The present invention discloses a method for constructing a bearing vibration signal processing model based on self-supervised learning, as well as a bearing vibration signal processing model constructed thereby. Automatic identification and diagnosis of unlabeled bearing faults based on similarity measurement using the idea of contrastive learning. When training the time-domain and frequency-domain encoders (convolutional neural network with attention mechanism), it is not necessary to use labeled samples, and only a small number of labeled samples are used when training the linear classifier through the model, so as to achieve the classification of bearing vibration signals, that is, for bearing fault detection.

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

[0047] First, the collected bearing vibration signal samples are transformed into frequency-domain signal samples through FFT, so as to form a frequency-domain signal sample set, and the original vibration signals form a time-domain signal sample set. The DTW distance between pairs of time-domain signal samples is calculated using softmin-DTW, and the Euclidean distance between pairs of frequency-domain signal samples is calculated using the Euclidean distance. The distance magnitude represents similarity. At the same time, these two sample sets are respectively input into the encoder (convolutional neural network with attention mechanism) for mapping calculation to output feature representations, and the Mahalanobis distance function is used to calculate the distances between pairs of time-domain and frequency-domain feature representations respectively to measure similarity. More importantly, the present invention adopts a unique loss function to minimize the distance between two signal feature representations and the distance between their corresponding actual signals. This loss function can minimize the difference between the similarity between the feature representations extracted by the encoder and the similarity between actual signal samples. Minimizing this difference can guide the encoder to extract similar feature representations in the signal, so that similar feature representations are clustered together in the feature space, while dissimilar ones are far apart. 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 through the formed model to effectively classify the bearing vibration signals.

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

[0049] S1. Obtain and collect bearing vibration signal samples x i from the sensor to form a sample set X ∈ {x1, x2, …, x n}, and the length of each signal is L. The vibration signals can be regarded as time-domain signals, so the time-domain sample set is X. Each sample x i is transformed into a frequency-domain signal sample using the fast Fourier transform (FFT) to form a frequency-domain signal sample set corresponding to the time-domain signal sample set

[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 likelihood that they belong to the same category. The traditional method directly uses the DTW algorithm to achieve this. However, the present invention proposes an improved algorithm that uses the softmin-DTW function to smooth the distance of the cumulative possible paths, avoiding the noise and possible local optimum problems that may be brought about by the selection of a single path in the traditional DTW algorithm, making the selection of the path smoother and more continuous, thereby improving the robustness of the model and its performance in gradient optimization. The specific formula is as follows.

[0051]

[0052] where x i , x j are two time-domain signal samples with length L, and S t (x i , x j ) is the DTW distance between these two time-domain signal samples. The smaller the calculated DTW distance, the more similar the two time-domain signal samples x i , x j are; is the t-th data in the i-th sample, is the data in the i-th sample that is aligned with through the path π, and π(t) represents the index corresponding to t on the path π; is the distance metric between two points and , and we use the square of the Euclidean distance to represent it. is the weighting of the distance for each path. α is a hyperparameter used to adjust the contribution degree of different paths to the final distance. When α takes a smaller value (close to 0), the contributions of all paths tend to be uniform. When α increases, the weight of the path with a smaller distance will increase.

[0053] Since this formula is proposed for the first time, the 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 , calculate the local distance between each data point using the Euclidean distance. The formula is as follows

[0056]

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

[0058]

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

[0060] The cumulative distance matrix C is an (L + 1)×(L + 1) matrix, and 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. Finally, the cumulative distance is stored at the position C(L, L). The formula for calculating C(i, j) using softmin for two signal samples 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] where the definition of the softmin function is:

[0066]

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

[0068]

[0069] where and are two frequency - domain signal samples, is the Euclidean distance between the two frequency - domain signal samples. The smaller the calculated Euclidean distance, 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 k - th frequency component respectively.

[0070] S4. Respectively input the time - domain signal sample set X and the frequency - domain signal sample set X f into their respective bearing vibration signal processing encoders for mapping calculations to obtain a set of time - domain signal sample feature representations P t ∈{p1, p2, …, p n}, and a set of frequency - domain signal sample feature representations 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 regions in images or other data, thereby improving the performance of the model and its ability to understand data.

[0072] Taking the time-domain signal as an example to describe the calculation process.

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

[0074] y i = x i * ω m (6)

[0075] 4-2. Perform the Squeeze operation on y i The formula is as follows, where z i is the average 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 performs global average pooling on each convolved signal y i to finally obtain a one-dimensional vector Z = [z1, z2,... z n .

[0076]

[0077] 4-3. Perform the Excitation operation. Regarding the vector Z as a set of channels, through two fully connected layers and a non-linear activation function, capture the dependency relationships between each channel, thereby adaptively adjusting the weights of the channels, enhancing the important feature channels and suppressing the unimportant feature channels. The formula is as follows. Among them, W1 and W2 are the weight parameters of the fully connected layers, δ is the ReLU activation function, and σ is the Sigmoid activation function. s i is the weight value of the i-th channel, and finally obtain the channel weight vector S = [s1, s2,... s n .

[0078] s i = σ(W2δ(W1z i )) (8)

[0079] 4-4. Perform the Reweighting operation. According to the channel weight vector S, calculate the predicted feature representation with weight information, and the formula is as follows. Finally, obtain the set P of predicted feature representations of the time-domain signal samples with channel weights t ∈{p1, p2, …, p n}.

[0080] p i = y i · s i (9)

[0081] Similarly, obtain the set of predicted feature representations of the frequency-domain signal samples

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

[0083] S5. Calculate the Mahalanobis distances D t ∈{p1, p2, …, p n} pairwise between the bearing vibration time-domain signal sample feature representations in the set P t (p i , p j ) and the Mahalanobis distances pairwise between the bearing vibration frequency-domain signal sample feature representations in the set of frequency-domain signal sample feature representations The 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, and plays an important role especially in anomaly detection, pattern recognition, and classification of high-dimensional data.

[0084] Describe the calculation process taking the time-domain signal as an example.

[0085] In the set P of time-domain signal sample feature representations t ∈{p1, p2, …, p n}, randomly select two time-domain signal sample feature representations p i , p j . It can be abstractly understood as two vectors. By calculating the Mahalanobis distance to measure the similarity between feature representations, 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 two time-domain signal sample feature representations; S -1 is the inverse matrix of the covariance matrix of the vector set. The covariance matrix S represents the linear correlation between the values in the vector. The inverse matrix S -1 is used to adjust the scale and correlation between the vector values. (p i -p j ) T is the transpose matrix and is also a vector.

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

[0089] 5-1. Assume that the length of the vectors p i , p j is R, and calculate the means of the vectors p i and p j using the following formula. Where

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

[0091]

[0092] 5-2. To eliminate the differences between the vector values, perform centralized data using the following formula. Subtract the mean from each value in the p i vector to obtain the centralized Similarly, obtain

[0093]

[0094] 5-3. Calculate the covariance matrix S using the following formula. Where is the variance of the vector p i , is the variance of the vector p j , is the covariance between the vectors p i and p j .

[0095]

[0096] Similarly, calculate the Mahalanobis distance between every two of the bearing vibration frequency-domain signal sample feature representations Where are two frequency-domain signal sample feature representations. The smaller the calculated Mahalanobis distance, the more similar the two frequency-domain signal sample feature representations are.

[0097] S6. The DTW distance S calculated through the step S2 t (x i , x j ) and the Mahalanobis distance D calculated through the step S5 t (p i , p j ) are used to construct 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 are the feature representations obtained after the time-domain signal samples x i and x j are mapped and calculated through the convolutional neural network with an attention mechanism, respectively.

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

[0102]

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

[0104] Similarly, the Euclidean distance calculated through the step S3 and the Mahalanobis distance calculated through the step S5 are used to construct a frequency-domain signal loss function for training the bearing vibration signal processing encoder. The frequency-domain signal loss function is:

[0105]

[0106] wherein are frequency-domain signal samples and the feature representation of the frequency-domain signal sample obtained after the mapping calculation by the encoder.

[0107] The objective of this loss function is to minimize the difference between the similarity of the two feature representations extracted by the model and the similarity of the 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 the similar feature representations are clustered together in the feature space, while the dissimilar ones are far apart. 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 the 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 the downstream classification task.

[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 train them to form a time-domain signal encoder and a frequency-domain signal encoder respectively.

[0109] S8. Input the time-domain signal sample xi in the time-domain signal sample set X∈{x1, x2, …, x n} and the frequency-domain signal samples in into their respective trained time-domain signal encoder and frequency-domain signal encoder for mapping calculation, and obtain the corrected time-domain signal sample feature representation p i with channel weights and the frequency-domain signal sample feature representation p i f ;

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

[0111] The present invention also discloses a method for classifying bearing vibration signals. A linear classifier is used to classify the bearing vibration signals. The linear classifier is trained using a set of feature representations R ∈ {P fused1 , P fused2 , …, P fusedi} obtained by splicing according to the above method, thereby realizing bearing fault detection and classification in industrial fields.

[0112] The training process of the linear classifier is briefly described below.

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

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

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

[0116] where 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 matching degree 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 W and b in the linear calculation formula through backpropagation to increase the predicted probability, making it as close to 1 as possible;

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

[0121] By comparing the similarities between different samples, the present invention extracts features useful for fault diagnosis from the original data and conducts bearing fault classification and prediction based on these features. Compared with traditional supervised learning methods, this method has the following main advantages:

[0122] 1. Reduce the data annotation cost:

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

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

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

[0126] 3. Adapt to complex environments:

[0127] In a complex industrial environment, the fault signals of bearings are often disturbed by various factors. The intelligent bearing fault diagnosis framework based on self-supervised learning, through the way of contrastive learning, can better handle these complex non-linear relationships, improving the accuracy and stability of fault diagnosis.

[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 label a large amount of data when extended to new fields.

[0130] The present invention greatly reduces the training parameters, saves computing resources, and improves the 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, and the bearing vibration signal processing model is constructed according to the construction method of the bearing vibration signal processing model described above in this article.

[0132] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and 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, It includes the following steps: Step S1: Obtain the bearing vibration signal sample x i , and form a time-domain signal sample set X ∈ {x1, x2, …, x n}, where the length of each time-domain signal is L; use the fast Fourier transform to transform each of the bearing vibration signal samples x i into a bearing vibration frequency-domain signal sample to form a frequency-domain signal sample set corresponding to the time-domain signal sample set Step S2: Calculate the DTW distance S between any two 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 Euclidean distance between every two of the bearing vibration frequency domain signal samples in the frequency domain signal sample set X f wherein the Euclidean distance is calculated between every two of the bearing vibration frequency domain signal samples in the frequency domain signal sample set X where and refer to any two frequency domain signal samples; Step S4: Respectively input the time-domain signal sample set X and the frequency-domain signal sample set X f into their respective bearing vibration signal processing encoders for mapping calculations to obtain a set P t ∈{p1, p2, …, p n} of time-domain signal sample feature representations with channel weights, and a set Step S5. Calculate the Mahalanobis distances D between the pairwise time-domain signal sample feature representations in the time-domain signal sample feature representation set P t ∈{p1, p2, …, p n}, and the Mahalanobis distances between the pairwise frequency-domain signal sample feature representations in the frequency-domain signal sample feature representation set t (p i , p j ), and the Mahalanobis distances between the pairwise bearing vibration frequency-domain signal sample features in Step S6. Calculate the DTW distance S i between the time-domain signal samples x j and x t (x i , x j ), and calculate the Mahalanobis distance D i between the time-domain signal samples x j and x t (p i , p j ) obtained in Step S5, and construct a time-domain signal loss function for training the bearing vibration signal processing encoder. The time-domain signal loss function is as follows: L t = smooth(D t (p i , p j ), S t (x i , x j )) where p i and p j are the time-domain signal sample feature representations of the time-domain signal samples x i and x j obtained after the calculation by the encoder mapping respectively; The smooth function calculates the loss between the two, and the formula is as follows: where x is the DTW distance S t (x i , x j ) and the Mahalanobis distance D t (p i , p j ) between the difference; Similarly, the Euclidean distance between the frequency-domain signal samples obtained by calculating through the step S3 and , and the Mahalanobis distance between the frequency-domain signal samples obtained by calculating through the step S5 and and are used to construct a frequency-domain signal loss function for training the bearing vibration signal processing encoder. The frequency-domain signal loss function is as follows: ​ wherein are respectively frequency-domain signal samples and the feature representation of the frequency-domain signal samples obtained after the mapping calculation by the encoder where x in the smooth function is the difference between the Euclidean distance and the Mahalanobis distance ; Step 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. Step S8. Input the time-domain signal sample x n in the time-domain signal sample set X ∈ {x1, x2, …, x i} and the frequency-domain signal sample xf i in their respective pre-trained time-domain signal encoder and frequency-domain signal encoder for mapping calculation, to obtain the corrected time-domain signal sample feature representation p i with channel weights and the frequency-domain signal sample feature representation p i f ; Step S9: Perform one-to-one correspondence fusion on the time-domain signal sample feature representation p i and the frequency-domain signal sample feature representation p i f The formula is: P fused = concat(p i , p i f ), where concat represents the feature concatenation operation, which realizes the concatenation of the corresponding dimensions of the two feature representations.

2. The method according to claim 1, wherein The specific operation of step S2 is to use the softmin-DTW function to smooth the distance of the cumulative possible path, and the specific formula of the softmin-DTW function is as follows: where x i , x j are two time-domain signal samples with length 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, the more similar the two time-domain signal samples x i , x j are; is the t-th data in the i-th sample, is the data in the i-th sample aligned with through the path π, and π(t) represents the index corresponding to t on the path π; is the distance metric between two points and ; is the weight for the distance of each path, and α is a hyperparameter used to adjust the contribution degree of different paths to the final distance.

3. The method according to claim 1, characterized in that, The calculation formula adopted in step S3 is wherein 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 k-th frequency component, respectively.

4. The method according to claim 1, wherein 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 adopted in step S5 is where p i , p j are the feature representations of two time-domain signal samples, and D t (p i , p j ) is the Mahalanobis distance between the two feature representations of the time-domain signal samples. The smaller the calculated Mahalanobis distance, the more similar the two feature representations of the time-domain signal samples are; (p i - p j ) T is the transposed matrix, and S -1 is the inverse matrix of the covariance matrix of the vector set; Similarly, calculate the Mahalanobis distance between every two of the sample feature representations of the bearing vibration frequency domain signals where are the sample feature representations of two frequency domain signals. The smaller the calculated Mahalanobis distance is, the more similar the two sample feature representations of the frequency domain signals are 6. A method for classifying bearing vibration signals, characterized in that, Use a linear classifier to classify the bearing vibration signals, where the linear classifier is trained using a set of feature representations \(R\in\{P fused1, P fused2 ,\ldots,P fusedi \}\) obtained by splicing according to the method described in any one of claims 1-5.

Images