Bearing degradation starting point detection method based on auto-encoder and AOMM

Abstract

The invention relates to the technical field of bearing fault diagnosis, and discloses a bearing degradation starting point detection method based on an auto-encoder and an AOMM, and the method comprises the steps: collecting a real-time vibration signal of a bearing; obtaining a characteristic sequence of the real-time vibration signal by using an auto-encoder; obtaining a symbolized feature sequence of the real-time vibration signal according to an optimal symbol number obtained by pre-training; obtaining a plurality of sliding windows according to an optimal window size obtained by pre-training; an AOMM method is applied to adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, and a state stationary probability matrix is obtained; and inputting the state stationary probability matrix and the label data into a pre-trained adaptive enhancement classifier for prediction, and outputting a predicted degradation starting point. According to the method, the degradation starting point is identified more accurately, and the problems that the feature extraction of the degradation starting point is not comprehensive and is greatly influenced by noise, and the detection of the degradation starting point depends on the signal amplitude and neglects the dynamic behavior of the signal are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and relates to a method for detecting the starting point of bearing degradation based on an autoencoder and AOMM. Background Art

[0002] As a key component in a mechanical system, the operating state of a bearing is directly related to the reliability and safety of the entire system. In large mechanical equipment such as wind power generation equipment, aeroengines, and high-end CNC machine tools, the service life of the bearing affects the normal operation of the equipment. Accurately predicting the remaining useful life (RUL) of a bearing is of great significance for preventing equipment failures, reducing maintenance costs, and improving production efficiency.

[0003] In the process of bearing life prediction, the identification of the degradation starting point is a crucial basic link. Accurately identifying the degradation starting point can further confirm the degradation state of the bearing and provide a key basis for RUL prediction. The identification of the degradation starting point is not only the starting point for predicting the performance change of the bearing, but also the basis for establishing the subsequent life prediction model. Only by accurately identifying the degradation starting point can the prediction deviation caused by misjudgment be effectively avoided, thereby providing a reliable time window for equipment maintenance and fault prevention.

[0004] However, existing methods for detecting the degradation starting point mostly rely on manual feature extraction in feature extraction. This method is difficult to comprehensively and accurately capture the key features reflecting the degradation starting point. In addition, the vibration signal of a bearing usually contains a large amount of noise, which will mask the true degradation features and lead to a decrease in detection accuracy. Another problem with existing methods for detecting the degradation starting point is that they mostly rely on signal amplitude to detect anomalies and ignore the dynamic behavior of the signal. This method that relies solely on signal amplitude cannot effectively detect the degradation starting point in some cases, especially when the signal change is relatively complex or the noise interference is large. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for detecting the starting point of bearing degradation based on an autoencoder and AOMM (Adaptive Order Markov Model), which can more accurately identify the degradation starting point, and solves the problems of incomplete feature extraction of the degradation starting point, being greatly affected by noise, and the detection of the degradation starting point mostly relying on signal amplitude and ignoring the dynamic behavior of the signal.

[0006] To solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0007] In a first aspect, the present invention proposes a method for detecting the starting point of bearing degradation based on an autoencoder and AOMM, including:

[0008] Collect the real-time vibration signal of the bearing;

[0009] Preprocess the collected real-time vibration signal to obtain the preprocessed real-time vibration signal;

[0010] Use an autoencoder to extract features from the preprocessed real-time vibration signal, obtain the feature sequence of the real-time vibration signal, remove invalid features, and extract a multi-dimensional feature vector;

[0011] Globally symbolize the feature sequence of the real-time vibration signal according to the pre-trained optimal number of symbols to obtain the symbolized feature sequence of the real-time vibration signal;

[0012] Segment the symbolized feature sequence of the real-time vibration signal according to the pre-trained optimal window size to obtain a number of sliding windows;

[0013] Apply the AOMM method to adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and concatenate all the state stationary probability vectors into a state stationary probability matrix;

[0014] Input the state stationary probability matrix and the label data into the pre-trained adaptive boosting classifier for prediction, and output the predicted degradation starting point.

[0015] Combined with the first aspect, further, the application of the AOMM method to adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and concatenate all the state stationary probability vectors into a state stationary probability matrix includes:

[0016] Calculate the symbol distribution entropy of each window, compare it with the symbol distribution entropy of the previous window, and obtain the change of the symbol distribution entropy;

[0017] Predict the change value of the symbol distribution entropy based on the change of the symbol distribution entropy and the polynomial regression model, and dynamically adjust the Markov order;

[0018] Based on the dynamically adjusted Markov order, calculate the state stationary probability vector of each sliding window, and concatenate all the state stationary probability vectors into a state stationary probability matrix.

[0019] Combined with the first aspect, further, the method for obtaining the optimal number of symbols and the optimal window size is:

[0020] Step S1, collect the original vibration signal of the whole life cycle of the bearing;

[0021] Step S2: Preprocess the original vibration signal to obtain the preprocessed original vibration signal;

[0022] Step S3: Use an autoencoder to extract features from the preprocessed original vibration signal to obtain the feature sequence of the original vibration signal;

[0023] Step S4: Globally symbolize the feature sequence of the original vibration signal to obtain the symbolized feature sequence of the original vibration signal;

[0024] Step S5: Segment the symbolized feature sequence of the original vibration signal to obtain a number of sliding windows;

[0025] Step S6: Apply the AOMM method to adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and concatenate all the state stationary probability vectors into a state stationary probability matrix;

[0026] Step S7: Input the state stationary probability matrix and the label data as training samples to train the adaptive boosting classifier, and obtain the probability that each training sample predicted by the adaptive boosting classifier belongs to the abnormal class;

[0027] Step S8: Based on the probability that each training sample predicted by the adaptive boosting classifier belongs to the abnormal class, use the regularized cross-entropy loss function to optimize the sliding window size and the number of symbols, and feedback the optimized number of symbols to Step S4 and the optimized sliding window size to Step S5, and finally obtain the optimal number of symbols and the optimal window size.

[0028] In combination with the first aspect, further, the step of inputting the state stationary probability matrix and the label data as training samples to train the adaptive boosting classifier to obtain the probability that each training sample predicted by the adaptive boosting classifier belongs to the abnormal class includes:

[0029] Initialize the weights of the training samples;

[0030] Based on the initialized weights of the training samples, train the weak classifier and calculate the classification weighted error rate under all training sample distributions;

[0031] Calculate the weak classifier weights according to the classification weighted error rate to obtain the classification results of each training sample;

[0032] Update the weights of the training samples according to the classification results of each training sample;

[0033] Normalize all the updated weights of the training samples to obtain the normalized weights of each training sample;

[0034] Determine whether the preset number of iterations is reached. If the number of iterations is not reached, return to the step of training the weak classifier. If the number of iterations is reached, linearly combine all the weak classifiers according to the weights after normalization to obtain the final strong classifier.

[0035] Combined with the first aspect, further, the use of the regularized cross-entropy loss function to optimize the sliding window size and the number of symbols includes:

[0036] Design the loss function :

[0037] ;

[0038] Among them, represents the cross-entropy loss function; represents the regularization function;

[0039] Cross-entropy loss function The mathematical expression of is:

[0040] ;

[0041] Among them, is the number of training samples, is the true label of the th training sample, taking values of 0 or 1, is the probability that the th training sample predicted by the adaptive boosting classifier belongs to the abnormal class;

[0042] Regularization function The mathematical expression of is:

[0043] ;

[0044] Among them, is the number of weights of the adaptive boosting classifier, is the th weight, is the regularization coefficient, controlling the strength of the regularization term;

[0045] Calculate the loss according to the loss function and output the loss value. The loss value is used to measure the matching degree between the prediction result of the adaptive boosting classifier and the true label. The smaller the loss value, the better the prediction performance of the adaptive boosting classifier;

[0046] Keep the current number of symbols unchanged. In each loop, adjust the window size sequentially and train using an adaptive boosting classifier. By comparing the loss values under different window sizes, determine whether the change in the loss value is less than the set loss value tolerance. If the change in the loss value is less than the set loss value tolerance, determine the optimal window size and the corresponding optimal loss value under the current number of symbols, stop further adjustment of the window size simultaneously, and update the globally symbolized number of symbols.

[0047] Under the updated globally symbolized number of symbols, repeat the previous step to continue optimizing the window size until the change in the optimal loss value is less than the set optimal loss value tolerance. At this time, stop the entire optimization process to obtain the optimal sliding window size, and the corresponding updated globally symbolized number of symbols is the optimal number of symbols.

[0048] Combined with the first aspect, further, the use of an autoencoder to extract features from the preprocessed original vibration signal to obtain a feature sequence of the original vibration signal includes:

[0049] Let be the input data set, that is, the set of preprocessed original vibration signals, and define , denote a feature vector in the input data set , is the index of the feature vector in the input data set , , denote the input data set this set contains feature vectors, that is, the size of the input data set is , denote dimensional real space, denote the dimension of the feature vector; the encoder transmits the high-dimensional input data set from the input layer to the hidden layer, and in the hidden layer, the input data set is compressed into a low-dimensional space , and define , denote a low-dimensional feature vector in the low-dimensional space , which is the compressed feature representation, is the index of the special vector, ; denote the low-dimensional space this set contains feature vectors, denote dimensional real space, Represents the dimension of the compressed feature vector; subsequently, the decoder remaps the compressed feature vector to the input data at the output layer , as the output result; the encoder and the decoder The basic mathematical expressions are:

[0050] ;

[0051] ;

[0052] The encoder converts the input into a low-dimensional representation , and subsequently, the decoder converts the low-dimensional representation into the output to reconstruct the input .

[0053] Combined with the first aspect, further, globally symbolizing the feature sequence of the original vibration signal to obtain the symbolized feature sequence of the original vibration signal includes: globally symbolizing the feature sequence of the original vibration signal using a uniform partitioning method;

[0054] For the feature sequence , is the sequence length, is the feature value; initialize the parameter , represents how many symbols the feature sequence is to be divided into, that is, the number of symbols;

[0055] Calculate the minimum value and the maximum value in the feature sequence , as the boundaries of the partitioning interval;

[0056] Evenly divide the value range of the feature sequence into intervals; the width of each interval is ;

[0057] Map each feature value in the feature sequence to the corresponding symbol to obtain the symbolized feature sequence.

[0058] Combined with the first aspect, further, collect the real-time vibration signals of the bearing during operation, and the directions of the vibration signals include the horizontal direction and the vertical direction. The full life cycle refers to the entire process from when the bearing starts to be used until it is scrapped and no longer has the ability to operate.

[0059] In combination with the first aspect, further, preprocess the collected vibration signals to obtain preprocessed vibration signals, including: perform normalization processing on the collected vibration signals to obtain normalized vibration signals, which can eliminate the influence of dimensions between different signals and unify the data scale.

[0060] In combination with the first aspect, still further, the performing normalization processing on the collected vibration signals to obtain normalized vibration signals includes: adopting max - min normalization processing:

[0061] ;

[0062] wherein, is the minimum value in the vibration signal, is the maximum value in the vibration signal, is each value in the vibration signal, is the normalized value, such that all data points fall within the interval [0, 1].

[0063] It should be noted that: the vibration signals in the preprocessing of the collected vibration signals to obtain preprocessed vibration signals include real - time vibration signals and original vibration signals, and the preprocessing methods for real - time vibration signals and original vibration signals are the same.

[0064] In the second aspect, the present invention proposes a bearing degradation starting point detection system based on an auto - encoder and AOMM, including:

[0065] A data acquisition module, configured to collect real - time vibration signals of a bearing;

[0066] A data preprocessing module, configured to preprocess the collected real - time vibration signals to obtain preprocessed real - time vibration signals;

[0067] An auto - encoder module, configured to use an auto - encoder to extract features from the preprocessed real - time vibration signals to obtain a feature sequence of the real - time vibration signals;

[0068] A symbolic feature sequence module, configured to globally symbolize the feature sequence of the real - time vibration signals according to the pre - trained optimal number of symbols to obtain a symbolic feature sequence of the real - time vibration signals;

[0069] A sliding window segmentation module, configured to segment the symbolic feature sequence of the real - time vibration signals according to the pre - trained optimal window size to obtain a number of sliding windows;

[0070] The AOMM method module is configured to apply the AOMM method, adaptively adjust the Markov order based on the symbolic distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and splice all the state stationary probability vectors into a state stationary probability matrix;

[0071] The adaptive boosting classifier module is configured to input the state stationary probability matrix and the label data into a pre-trained adaptive boosting classifier for prediction, and output the predicted degradation starting point.

[0072] In a third aspect, the present invention proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned bearing degradation starting point detection method based on the autoencoder and AOMM are implemented.

[0073] In a fourth aspect, the present invention proposes a computer device, including:

[0074] A memory for storing a computer program;

[0075] A processor for executing the computer program to implement the steps of the above-mentioned bearing degradation starting point detection method based on the autoencoder and AOMM.

[0076] In a fifth aspect, the present invention proposes a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the above-mentioned bearing degradation starting point detection method based on the autoencoder and AOMM are implemented.

[0077] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0078] (1) The present invention can more accurately identify the degradation starting point, solves the problems of incomplete feature extraction of the degradation starting point, being greatly affected by noise, and the detection of the degradation starting point mostly relying on the signal amplitude and ignoring the dynamic behavior of the signal.

[0079] (2) The present invention uses an autoencoder to automatically learn the features in the data, without the need for manual design of feature extraction rules, and has a certain robustness to noise, can effectively remove noise interference, and improves the efficiency and accuracy of feature extraction.

[0080] (3) The present invention combines the symbolic feature sequence with the AOMM method, models the symbolic sequence, and adaptively adjusts the Markov order to accurately capture the dynamic change law of the signal, rather than simply relying on the amplitude information of the signal.

[0081] (4) The present invention dynamically adjusts the window size and the number of symbols, searches for the optimal window and the optimal number of symbols, ensures that each window contains enough information to capture the dynamic behavior, and balances the model complexity and the detection performance.

[0082] (5) The present invention combines an adaptive boosting classifier, which can quickly adapt to signal changes, accurately identify the starting point of bearing degradation, and effectively reduce detection latency, providing a reliable time window for equipment maintenance and fault prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 is a schematic flow chart of the method for detecting the starting point of bearing degradation of the present invention;

[0084] Figure 2 is a schematic flow chart of symbol quantity optimization, window size optimization, and adaptive boosting classifier training in the present invention;

[0085] Figure 3 is a schematic structural diagram of the autoencoder (AE) in the present invention;

[0086] Figure 4 is a schematic flow chart of the adaptive boosting classifier in the present invention;

[0087] Figure 5 is a schematic diagram of the starting point of degradation of bearing 1 identified by the detection method of the present invention;

[0088] Figure 6 is a schematic diagram of the starting point of degradation of bearing 2 identified by the detection method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0089] The technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other.

[0090] The term "and / or" is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " generally represents an "or" relationship between the associated objects before and after.

[0091] Embodiment 1

[0092] As Figure 1 shown, the steps of the method for detecting the starting point of bearing degradation based on the autoencoder and AOMM in this embodiment are as follows:

[0093] Collect the real-time vibration signal of the bearing;

[0094] Preprocess the collected real-time vibration signal to obtain the preprocessed real-time vibration signal;

[0095] Use an autoencoder to extract features from the preprocessed real-time vibration signal, obtain the feature sequence of the real-time vibration signal, remove invalid features, and extract a multi-dimensional feature vector;

[0096] Globally symbolize the feature sequence of the real-time vibration signal according to the optimal number of symbols obtained by pre-training to obtain the symbolized feature sequence of the real-time vibration signal;

[0097] Segment the symbolized feature sequence of the real-time vibration signal according to the optimal window size obtained by pre-training to obtain a number of sliding windows;

[0098] Apply the AOMM method, adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and concatenate all the state stationary probability vectors into a state stationary probability matrix;

[0099] Input the state stationary probability matrix and the label data into the pre-trained adaptive boosting classifier for prediction, and output the predicted degradation starting point.

[0100] In a specific implementation manner in this embodiment, as Figure 2 shown, the method for training the optimal number of symbols, the optimal window size, and the adaptive boosting classifier includes the following steps:

[0101] Step S1, collect the original vibration signal of the bearing's full life cycle.

[0102] The full life cycle refers to the entire process from the start of use of the bearing to scrapping and no longer having the ability to operate. The directions of the vibration signal include the horizontal direction and the vertical direction.

[0103] It should be noted that: in this embodiment, the bearing is specifically a rolling bearing.

[0104] Step S2, perform normalization preprocessing on the original vibration signal of the full life cycle to eliminate the influence of dimensions between different signals, unify the data scale, obtain the preprocessed original vibration signal, and further divide the preprocessed original vibration signal to obtain a training set and a test set.

[0105] Among them, the normalization preprocessing adopts the maximum-minimum normalization processing:

[0106] ;

[0107] Among them, is the minimum value in the original vibration signal of the full life cycle, is the maximum value in the original vibration signal of the full life cycle, is each value in the original vibration signal of the entire life cycle, is the normalized value such that all data points fall within the interval [0, 1].

[0108] Step S3: Use an autoencoder (AE) to extract features from the preprocessed original vibration signal, remove zero features and invalid features, extract multi-dimensional features, and obtain the feature sequence of the original vibration signal.

[0109] Among them, as Figure 3 shown, let be the input data set, that is, the set of preprocessed original vibration signals, and define , represents a feature vector in the input data set , is the index of the feature vector in the input data set , , represents the input data set This set contains feature vectors, that is, the size of the input data set is , represents dimensional real number space, represents the dimension of the feature vector; the encoder transmits the high-dimensional input data set from the input layer to the hidden layer, and in the hidden layer, the input data set is compressed into a low-dimensional space , and define , represents a low-dimensional feature vector in the low-dimensional space , which is the compressed feature representation, is the index of the feature vector, ; represents the low-dimensional space This set contains feature vectors, represents dimensional real number space, represents the dimension of the compressed feature vector. Subsequently, the decoder remaps the compressed feature vector to the input data in the output layer, is the output result. The basic mathematical expressions of the encoder and the decoder are:

[0110] ;

[0111] ;

[0112] Encoder converts the input into a low-dimensional representation . Subsequently, the decoder converts the low-dimensional representation into an output to reconstruct the input .

[0113] Initialize the weights and biases of the encoder and decoder respectively, and select the Sigmod function as the activation function . Train the autoencoder as follows:

[0114] ;

[0115] ;

[0116] wherein represents an element in the input data set ; represents an element in the low-dimensional space ; represents an element in the output result ; represents the activation function of the encoder represents the activation function of the decoder represents the weight of the encoder represents the weight of the decoder represents the bias of the encoder represents the bias of the decoder

[0117] Use the mean square error loss function , and its expression is:

[0118] ;

[0119] When training the autoencoder, use the Adam optimizer to adaptively adjust the learning rate. After extracting the features, remove the invalid features. Calculate the variance of each dimension of the features and set a threshold (such as 0.0001) to remove the features with variance lower than the threshold.

[0120] Step S4: Use the uniform partitioning method to globally symbolize the feature sequence obtained in step S3 to obtain a symbolized feature sequence.

[0121] For the feature sequence , is the sequence length, is the eigenvalue. Initialize the parameter , indicating the number of symbols into which the feature sequence is to be divided, i.e., the number of symbols. Calculate the minimum value in the feature sequence and the maximum value , and use them as the boundaries of the division interval. Divide the value range of the feature sequence evenly into intervals. The width of each interval is . Map each eigenvalue in the feature sequence to the corresponding symbol. The symbols adopt an integer sequence starting from the simple numeric character "0". Finally, convert the feature sequence into a symbolized feature sequence , where , is the th symbol, , and is a symbol table composed of a finite number of symbols.

[0122] Step S5. Use the sliding window method to segment the symbolized feature sequence obtained in Step S4 to generate multiple windows.

[0123] Step S51. Initialize the window size (window_size) and the step size (step_size). The window size and the step size are two key parameters of the sliding window method. The window size determines the number of symbols contained in each window, and the step size determines the step distance at which the window slides on the symbolized feature sequence.

[0124] Step S52. Slide the window in a loop. In the loop, the window starts from the starting position of the symbolized feature sequence and slides one step size each time until the end position of the window exceeds the sequence length of the symbolized feature sequence.

[0125] Step S6. Apply the AOMM method to each window, adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window obtained, and splice all the state stationary probability vectors into a state stationary probability matrix.

[0126] The present invention proposes the AOMM method, i.e., the adaptive order Markov model, which can dynamically adjust the order of the Markov model according to the distribution of symbols , it can more precisely characterize the long-range dependence and complex dynamic behavior of the sequence, reduce the dependence on computing resources, and at the same time achieve a better effect of symbolic sequence feature extraction.

[0127] Step S61, generate the state sequence of the Markov model for the symbolic feature sequence of each window. The probability that the th state in the state set transfers to the th state is:

[0128] ;

[0129] where is the state transition probability, is the state set, is the th state in the state set, is the th state in the state set, is the probability function, which is used to calculate the probability of an event occurring.

[0130] For a Markov model, the state transition probability matrix is the transition probability between all states and is a non-degenerate matrix. It describes the transition probability between all possible states.

[0131] Step S62, calculate the symbol distribution entropy of the current window. The symbol distribution entropy is calculated by counting the number of times each symbol appears in the sequence in the window and converting it into a probability distribution. Then the Shannon entropy formula is used to measure the degree of dispersion of these probabilities. The mathematical formula is:

[0132] ;

[0133] where represents the symbol distribution entropy, is the total number of symbols, is the probability that the th symbol appears in the sequence.

[0134] Step S63, dynamically adjust the order of the Markov model in each window. First, calculate the change in symbol distribution entropy between the current window and the previous window. Add the current symbol distribution entropy change value to the historical symbol distribution entropy list. If the symbol distribution entropy change exceeds the set threshold, it indicates an increase in data complexity and the order At this time, if the historical symbol distribution entropy data is sufficient (at least 3 data points), the historical symbol distribution entropy change value is used as the training data. Then, a polynomial regression model is used to fit the trend of the historical symbol distribution entropy change and predict the symbol distribution entropy change of the next window. In this application, a fourth-degree polynomial regression model is selected, which can capture more complex curve relationships and is applicable to situations where the data has a certain degree of smoothness and trend. The mathematical expression of the fourth-degree polynomial regression model is as follows:

[0135] ;

[0136] where, represents the symbol distribution entropy change value, represents the window index, is the intercept, representing the predicted value of the polynomial regression model at ; are the coefficients of the polynomial regression model, corresponding to the weights of the linear term, quadratic term, cubic term, and quartic term respectively, is the error term.

[0137] When predicting the symbol distribution entropy change of the next window, this application uses the fitted polynomial regression model, substitutes the index of the next window into the polynomial regression model to obtain the predicted symbol distribution entropy change value, and adjusts the order of the Markov model according to the prediction result . If the predicted symbol distribution entropy change still exceeds the preset symbol distribution entropy threshold, the order of the Markov model is increased significantly , that is, incremented by 2; otherwise, the order of the Markov model is increased slightly , that is, incremented by 1. If the historical symbol distribution entropy data is insufficient, the order of the Markov model is directly increased slightly , that is, incremented by 1. If the symbol distribution entropy change is less than or equal to the preset symbol distribution entropy threshold, the order of the current Markov model remains unchanged. The adjustment strategy of the order is as follows:

[0138] ;

[0139] where, represents the order of the Markov model corresponding to the th window; represents the entropy change value between the th window and the previous window; represents the preset symbol distribution entropy change threshold; represents the number of historical symbol distribution entropy data points; represents the The predicted value of the change in the distribution entropy of window symbols.

[0140] Step S64. For the symbolized feature sequence , slide the window to the right by the order of the length, and count the number of symbol strings with lengths of and , which are respectively counted as and , represents the count, represents a symbol string with a length of , represents a symbol string with a length of , where is the sequence composed of the first symbols, is the symbol immediately following, represents the -th symbol in the symbol sequence, and the subscript is used to identify the index in the symbol sequence. If , then the probability of the state occurring is 0. If , then calculate the transition probability of a certain state as:

[0141] ;

[0142] where the corresponding states are represented as and .

[0143] The left eigenvector corresponding to the unit eigenvalue of the state transition probability matrix is the state stationary probability vector, which represents the state distribution of the bearing operating state represented by the symbolized feature sequence under stable conditions. The state stationary probability vector, as the feature vector of each window, will change significantly when the bearing operating state changes from the normal state to the abnormal state. Finally, concatenate the state stationary probability vectors of each window into a matrix, that is, the state stationary probability matrix. Use the state stationary probability matrix as the input feature of the adaptive boosting classifier, and at the same time make label data for the training set, marking the time ranges of the normal state and the abnormal state. The feature sequence window before the bearing degradation occurs is marked as normal (label 0), and the feature sequence window after the bearing degradation occurs is marked as abnormal (label 1). Use these labeled feature data as training samples to train the adaptive boosting classifier, so as to realize the abnormal detection of the real-time vibration signal during the bearing operation.

[0144] Step S7. Use the state stationary probability matrix and the label data as the input for the model training of the adaptive boosting classifier.​​​​​​​​​​​​​

[0145] Among them, as Figure 4 shown, the adaptive boosting classifier is an ensemble learning algorithm with adaptive capabilities. Its core idea is to construct a strong classifier by combining multiple weak classifiers, thereby giving full play to the advantages of each base learner and making up for the deficiencies of a single base learner. The adaptive boosting classifier adopts the forward stagewise algorithm, which can quickly reduce the loss function and gradually approach the optimal solution, thus simplifying the complexity of ensemble learning.

[0146] First, initialize the weights of the training samples. Suppose there are training samples, and initially each training sample weight is assigned the same value:

[0147] ;

[0148] Among them, represents the th training sample; represents the weight of the th training sample at the beginning.

[0149] In each round of iteration, train a weak classifier , that is, a decision stump, representing the weak classifier trained in the th round of iteration.

[0150] Calculate the weighted classification error rate of the weak classifier under the current training sample distribution (the weights of the current training samples):

[0151] ;

[0152] Among them, represents the weighted error rate of the th weak classifier, represents the weight of the th th training sample, is an indicator function that takes the value of 1 when the condition holds and 0 otherwise.

[0153] Calculate the weight of the weak classifier according to the weighted classification error rate, representing the weight of the th weak classifier. The lower the weighted classification error rate, the higher the weight of the weak classifier:

[0154] ;

[0155] For each training sample, update its weight according to the classification result.

[0156] ;

[0157] Among them, represents the weight of the th training sample in the th round of the next iteration. is the true label of the training sample. The weights of misclassified training samples will increase, and the weights of correctly classified training samples will decrease.

[0158] To keep the sum of the weights of the training samples equal to 1, normalize the weights of all training samples:

[0159] ;

[0160] Among them, represents the temporary index variable when summing the weights of all training samples in the denominator part.

[0161] Linearly combine the obtained weak classifiers according to their weights to get the final strong classifier:

[0162] ;

[0163] Among them, is the output of the final strong classifier, is the total number of weak classifiers, represents the weight of the th weak classifier, the th output of the weak classifier, represents the training sample data input to the weak classifier. =1 indicates that the strong classifier predicts that the input feature sequence window is in an abnormal state, =0 indicates that the strong classifier predicts that the input feature sequence window is in a normal state.

[0164] Step S8: Use the regularized cross-entropy loss function to optimize the sliding window size and the number of symbols to obtain the optimal number of symbols and the optimal window size.

[0165] Step S81: Design the loss function . Cross-Entropy Loss is an index that measures the difference between the predicted probability distribution of the model and the true label distribution. For a binary classification problem, the mathematical expression of the cross-entropy loss function is:

[0166]

[0167] Among them, is the number of training samples, is the true label (0 or 1) of the th training sample, and is the probability that the

[0168] th training sample predicted by the adaptive boosting classifier belongs to the abnormal class. Regularization is a technique to prevent overfitting, which constrains the complexity of the adaptive boosting classifier weights by adding a regularization term to the cross-entropy loss function. The mathematical expression of the regularization function

[0169]

[0170] is as follows: where is the number of weights of the adaptive boosting classifier, is the th weight, and

[0171] is the regularization coefficient, which controls the strength of the regularization term. The regularized cross-entropy loss function combines the cross-entropy loss and the regularization term to form the final loss function

[0172]

[0173] Step S82: Calculate the loss according to the obtained final loss function. The regularized cross-entropy loss function calculates the difference between the input predicted probability distribution and the true label distribution, and outputs a specific loss value. This loss value measures the matching degree between the prediction result of the adaptive boosting classifier and the true label. The smaller the loss value, the better the prediction performance of the adaptive boosting classifier.

[0174] Step S83: Keep the current number of symbols unchanged. In each loop, adjust the size of the sliding window in turn and use the adaptive boosting classifier for training. By comparing the loss values under different sliding window sizes, determine whether the change in the loss value is less than the set loss value tolerance (such as 0.001). If the change in the loss value is less than the set loss value tolerance, determine the optimal sliding window size and the corresponding optimal loss value under the current number of symbols, and at the same time stop further adjusting the size of the sliding window and update the globally symbolized number of symbols.

[0175] Step S84: Under the number of globally symbolized symbols updated in step S83, repeat step S83 to continue optimizing the sliding window size. This process continues until the change in the optimal loss value is less than the set tolerance for the optimal loss value (e.g., 0.0001). At this point, stop the entire optimization process to obtain the optimal sliding window size, and the number of globally symbolized symbols updated in step S83 is the optimal number of symbols.

[0176] Through steps S81 to S84, the optimal number of symbols and the optimal sliding window size can be determined, thus providing the most ideal parameter configuration for subsequent detection of the bearing degradation starting point.

[0177] Step S9: Apply the trained adaptive boosting classifier to predict the test set data to identify the starting point of the degradation process.

[0178] Combined with the optimal number of symbols and the optimal window size, apply the trained adaptive boosting classifier to predict the test set data to obtain the probability that each window belongs to the abnormal class.

[0179] The degradation process is usually a gradual process. The degradation starting point is often the first significant change point where the bearing operating state begins to deviate from the normal state. As described above, when the bearing operating state changes from the normal state to the abnormal state, the state stationary probability vector changes significantly, so that the corresponding window is recognized as abnormal with a significantly increased probability in the adaptive boosting classifier. Therefore, when the probability that a window belongs to the abnormal class is the highest, this window can effectively capture the degradation starting point.

[0180] In the present invention, by selecting the window with the highest probability and taking its starting position as the degradation starting point, the experimental results of bearing 1 and bearing 2 are as Figure 5 and Figure 6 shown, where the feature sequence extracted by the autoencoder is represented by a gray solid line. Figure 5 and Figure 6 A statistical method based on the normal distribution threshold is also introduced as a comparison method, the threshold is marked by a gray dashed line, and the area outside this range is regarded as abnormal. The light gray area indicates the time period when the data enters the degradation stage, and the degradation starting points predicted by the method proposed in the present invention are respectively at Figure 5 1358 in Figure 6 and 3663 in

[0181] Example 2

[0182] Based on the same inventive concept as in Embodiment 1, this embodiment introduces a bearing degradation starting point detection system based on an autoencoder and AOMM, including:

[0183] A data acquisition module configured to collect real-time vibration signals of a bearing;

[0184] A data preprocessing module configured to preprocess the collected real-time vibration signals to obtain preprocessed real-time vibration signals;

[0185] An autoencoder module configured to use an autoencoder to extract features from the preprocessed real-time vibration signals to obtain a feature sequence of the real-time vibration signals;

[0186] A symbolic feature sequence module configured to globally symbolize the feature sequence of the real-time vibration signals according to the optimal number of symbols obtained by pre-training to obtain a symbolized feature sequence of the real-time vibration signals;

[0187] A sliding window segmentation module configured to segment the symbolized feature sequence of the real-time vibration signals according to the optimal window size obtained by pre-training to obtain a plurality of sliding windows;

[0188] An AOMM method module configured to apply the AOMM method, adaptively adjust the Markov order based on the symbol distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and splice all the state stationary probability vectors into a state stationary probability matrix;

[0189] An adaptive boosting classifier module configured to input the state stationary probability matrix and the label data into a pre-trained adaptive boosting classifier for prediction and output the predicted degradation starting point.

[0190] Embodiment 3

[0191] Based on the same inventive concept as other embodiments, this embodiment introduces a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned bearing degradation starting point detection based on an autoencoder and AOMM are implemented.

[0192] Embodiment 4

[0193] Based on the same inventive concept as other embodiments, this embodiment introduces a computer device, including: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the above-mentioned bearing degradation starting point detection method based on an autoencoder and AOMM.

[0194] Embodiment 5

[0195] Based on the same inventive concept as other embodiments, this embodiment introduces a computer program product, including a computer program which, when executed by a processor, implements the steps of the above-mentioned bearing degradation starting point detection method based on an autoencoder and AOMM.

[0196] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0198] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0199] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0200] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative rather than restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms, and these all fall within the protection scope of the present invention.

Claims

1. A bearing degradation starting point detection method based on autoencoder and AOMM, characterized in that: include: Collect real-time vibration signals of bearings; Preprocessing the collected real-time vibration signal to obtain a preprocessed real-time vibration signal; The autoencoder is used to extract features of the preprocessed real-time vibration signal to obtain a feature sequence of the real-time vibration signal; The feature sequence of the real-time vibration signal is globally symbolized according to the optimal number of symbols obtained by pre-training to obtain a symbolized feature sequence of the real-time vibration signal; The symbolic feature sequence of the real-time vibration signal is segmented according to the optimal window size obtained by pre-training to obtain a number of sliding windows; The AOMM method is applied to adaptively adjust the Markov order based on the symbolic distribution entropy and the polynomial regression model, and the state stationary probability vector of each sliding window is calculated, and all the state stationary probability vectors are spliced ​​into a state stationary probability matrix; The state stability probability matrix and label data are input into the pre-trained adaptive boosting classifier for prediction, and the predicted degradation starting point is output.

2. The bearing degradation starting point detection method based on autoencoder and AOMM according to claim 1 is characterized in that: The method for obtaining the optimal number of symbols and the optimal window size is: Step S1, collecting the original vibration signal of the bearing throughout its life cycle; Step S2, preprocessing the original vibration signal to obtain a preprocessed original vibration signal; Step S3, using an autoencoder to extract features from the preprocessed original vibration signal to obtain a feature sequence of the original vibration signal; Step S4, globally symbolizing the feature sequence of the original vibration signal to obtain a symbolized feature sequence of the original vibration signal; Step S5, segmenting the symbolic feature sequence of the original vibration signal to obtain a plurality of sliding windows; Step S6, applying the AOMM method, adaptively adjusting the Markov order based on the symbolic distribution entropy and the polynomial regression model, calculating the state stationary probability vector of each sliding window, and splicing all the state stationary probability vectors into a state stationary probability matrix; Step S7, inputting the state stable probability matrix and the label data as training samples, training the adaptive enhanced classifier, and obtaining the probability that each training sample predicted by the adaptive enhanced classifier belongs to an abnormal category; Step S8, based on the probability that each training sample predicted by the adaptive enhanced classifier belongs to the abnormal category, the regularized cross entropy loss function is used to optimize the sliding window size and the number of symbols, and the optimized number of symbols is fed back to the step S4, and the optimized sliding window size is fed back to the step S5, and finally the optimal number of symbols and the optimal window size are obtained.

3. The bearing degradation starting point detection method based on autoencoder and AOMM according to claim 2 is characterized in that: The state stationary probability matrix and label data are used as training sample inputs to train the adaptive enhanced classifier to obtain the probability that each training sample predicted by the adaptive enhanced classifier belongs to an abnormal category, including: Initialize the weights of the training samples; Based on the weights of the initialized training samples, the weak classifier is trained to calculate the weighted classification error rate under the distribution of all training samples; Calculating the weight of the weak classifier according to the classification weighted error rate to obtain the classification result of each training sample; Update the weight of each training sample according to the classification result of the training sample; Normalize all updated training sample weights to obtain the normalized weight of each training sample; Determine whether the preset number of iterations has been reached. If not, return to the step of training the weak classifiers. If the number of iterations has been reached, linearly combine all weak classifiers according to the normalized weights to obtain the final strong classifier.

4. The bearing degradation starting point detection method based on autoencoder and AOMM according to claim 2 is characterized in that: The use of the regularized cross entropy loss function to optimize the sliding window size and the number of symbols includes: Design loss function : ； in, represents the cross entropy loss function; represents the regularization function; Cross Entropy Loss Function The mathematical expression is: ； in, is the number of training samples, It is The true label of the training sample is 0 or 1. is the first prediction of the adaptive boosting classifier The probability that a training sample belongs to an abnormal category; Regularization function The mathematical expression is: ； in, is the number of adaptive boosting classifier weights, It is weights, is the regularization coefficient, which controls the strength of the regularization term; According to the loss function Calculate the loss and output the loss value, which is used to measure the matching degree between the prediction result of the adaptive boosting classifier and the true label. The smaller the loss value, the better the prediction performance of the adaptive boosting classifier. Keep the current number of symbols unchanged. In each cycle, adjust the window size in turn and use the adaptive boosting classifier for training. By comparing the loss values ​​under different window sizes, determine whether the change in the loss value is less than the set loss value tolerance. If the change in the loss value is less than the set loss value tolerance, determine the optimal window size and the corresponding optimal loss value under the current number of symbols, stop further adjusting the window size, and update the number of globally symbolized symbols. Under the updated number of globally symbolized symbols, repeat the previous step and continue to optimize the window size until the change in the optimal loss value is less than the set optimal loss value tolerance. At this time, stop the entire optimization process and obtain the optimal sliding window size. The corresponding updated number of globally symbolized symbols is the optimal number of symbols.

5. The bearing degradation starting point detection method based on autoencoder and AOMM according to claim 4 is characterized in that: The method of extracting features from the preprocessed original vibration signal using the autoencoder to obtain a feature sequence of the original vibration signal includes: make For the input data set, i.e. the set of preprocessed raw vibration signals, define , Represents the input dataset A feature vector in is the input dataset The index of the eigenvector in , , Represents the input dataset This collection contains feature vectors, that is, the size of the input data set is , express dimensional real number space, Represents the dimension of the feature vector; encoder The high-dimensional input dataset Pass from the input layer to the hidden layer, input the data set in the hidden layer Compressed into low-dimensional space In the definition , Represents a low-dimensional space A low-dimensional feature vector in is a compressed feature representation. is the index of the particular vector, ; Represents a low-dimensional space This collection contains feature vectors, express dimensional real number space, Represents the dimension of the compressed feature vector; the decoder The output layer remaps the compressed feature vector to the input data , To output the result; encoder and decoder The basic mathematical expression is: ； ； Encoder Enter Convert to low-dimensional representation , then the decoder The low-dimensional representation Convert to output , to reconstruct the input .

6. The bearing degradation starting point detection method based on autoencoder and AOMM according to claim 2 is characterized in that: The globally symbolizing the feature sequence of the original vibration signal to obtain the symbolized feature sequence of the original vibration signal includes: globally symbolizing the feature sequence of the original vibration signal using a uniform partitioning method; For feature sequences , is the sequence length, is the characteristic value; initialization parameter , Indicates that the feature sequence How many symbols are divided into, that is, the number of symbols; Calculate feature sequence The minimum value in and maximum value , as the boundary of the partition interval; The feature sequence The value range is evenly divided into intervals; the width of each interval is ; The feature sequence Each eigenvalue in is mapped to the corresponding symbol to obtain a symbolic feature sequence.

7. A bearing degradation starting point detection system based on autoencoder and AOMM, characterized in that: include: A data acquisition module is configured to acquire real-time vibration signals of the bearing; A data preprocessing module is configured to preprocess the collected real-time vibration signal to obtain a preprocessed real-time vibration signal; The autoencoder module is configured to extract features from the preprocessed real-time vibration signal using the autoencoder to obtain a feature sequence of the real-time vibration signal; A symbolized feature sequence module is configured to globally symbolize the feature sequence of the real-time vibration signal according to the optimal number of symbols obtained by pre-training to obtain a symbolized feature sequence of the real-time vibration signal; A sliding window segmentation module is configured to segment the symbolic feature sequence of the real-time vibration signal according to an optimal window size obtained through pre-training to obtain a plurality of sliding windows; An AOMM method module is configured to apply the AOMM method, adaptively adjust the Markov order based on the symbolic distribution entropy and the polynomial regression model, calculate the state stationary probability vector of each sliding window, and splice all the state stationary probability vectors into a state stationary probability matrix; The adaptive boosting classifier module is configured to input the state stationary probability matrix and label data into a pre-trained adaptive boosting classifier for prediction, and output a predicted degradation starting point.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for detecting the starting point of bearing degradation based on an autoencoder and AOMM described in any one of claims 1 to 6 are implemented.

9. A computer device, characterized in that: include: Memory for storing computer programs; A processor is used to execute the computer program to implement the steps of the bearing degradation starting point detection method based on autoencoder and AOMM according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the bearing degradation starting point detection method based on an autoencoder and AOMM described in any one of claims 1 to 6 are implemented.

Images