CNN-MLB-AB-based sliding bearing water pollution identification method and system

Abstract

The invention discloses a CNN-MLB-AB-based sliding bearing water pollution identification method and system, and the method comprises the steps: obtaining a sliding bearing vibration signal, and sequentially carrying out the signal decomposition and signal reconstruction processing through a variational mode decomposition algorithm, and obtaining a reconstructed sliding bearing vibration signal; discrete wavelet decomposition and normalization processing are sequentially carried out on the reconstructed sliding bearing vibration signals, wavelet coefficient time domain features are calculated, and a sliding bearing vibration signal feature matrix is obtained; introducing a multi-layer bidirectional long and short term layer, and constructing a CNN-MLB-AB sliding bearing water pollution identification neural network model; and carrying out fault identification on the sliding bearing vibration signal characteristic matrix to obtain a sliding bearing water pollution identification result. According to the invention, the transient characteristics of the vibration signal of the sliding bearing can be captured on different frequency bandwidths, so that the accuracy of water pollution identification of the sliding bearing is improved. The sliding bearing water pollution identification method and system based on CNN-MLB-AB can be widely applied to the technical field of sliding bearing water pollution identification.

Description

Technical Field

[0001] The invention relates to the technical field of water contamination identification of sliding bearings, and in particular to a method and system for identifying water contamination of sliding bearings based on CNN-MLB-AB. Background Art

[0002] Since most sliding bearings work in extreme and harsh environments for a long time, it is difficult to avoid contamination of lubricating oil, such as water contamination. The working performance of sliding bearings depends on the state of lubricating oil film. Lubricating oil contamination is the core cause of oil film failure or rupture. Condition monitoring of lubricating oil contamination is the key to sliding bearing fault identification and health maintenance. Once a sliding bearing fails, it will cause sudden damage to rotating equipment, forcing large equipment to shut down and causing huge economic losses. Vibration analysis is the most widely used in rotating machinery fault diagnosis due to its non-invasive monitoring advantages. However, due to the complexity and nonlinear characteristics of mechanical equipment fault signals, a single signal processing method is difficult to effectively capture the complex fault characteristics during equipment operation. Traditional frequency domain or time domain analysis methods cannot accurately extract implicit features in fault signals, resulting in low accuracy of fault detection. For the extraction of fault features of vibration signals, variational mode decomposition (VMD) is an adaptive decomposition method that can decompose complex signals into multiple intrinsic mode functions (IMFs) with different frequency bandwidths. However, the selection of components is directly related to the quality of extracted features, and the fault signals of sliding bearings are often transient, and related methods find it difficult to identify and obtain the key features therein. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a sliding bearing water contamination identification method and system based on CNN-MLB-AB, which can capture the transient characteristics of the sliding bearing vibration signal at different frequency bandwidths, thereby improving the accuracy of sliding bearing water contamination identification.

[0004] The first technical solution adopted by the present invention is: a method for identifying water contamination of sliding bearings based on CNN-MLB-AB, comprising the following steps:

[0005] Acquire the sliding bearing vibration signal and perform signal decomposition and signal reconstruction processing in sequence through the variational mode decomposition algorithm to obtain a reconstructed sliding bearing vibration signal, wherein the sliding bearing vibration signal includes a sliding bearing water contamination vibration signal and a sliding bearing fault-free vibration signal;

[0006] The reconstructed sliding bearing vibration signal is sequentially subjected to discrete wavelet decomposition and normalization processing, and the time domain characteristics of the wavelet coefficients are calculated to obtain the sliding bearing vibration signal characteristic matrix;

[0007] Multi-layer bidirectional long-short term layers were introduced to construct the CNN-MLB-AB neural network model for water pollution identification of sliding bearings.

[0008] Based on the CNN-MLB-AB sliding bearing water contamination identification neural network model, the sliding bearing vibration signal feature matrix is ​​used for fault identification, and the sliding bearing water contamination identification result is obtained.

[0009] Further, the step of acquiring the sliding bearing vibration signal and sequentially performing signal decomposition and signal reconstruction processing by using a variational mode decomposition algorithm to obtain a reconstructed sliding bearing vibration signal specifically includes:

[0010] Collect the vibration signal of sliding bearing;

[0011] The vibration signal of the sliding bearing is decomposed by using the variational mode decomposition algorithm to obtain several decomposed IMF components of the vibration signal of the sliding bearing;

[0012] Obtain variance contribution rates corresponding to several decomposed IMF components of sliding bearing vibration signals;

[0013] According to the variance contribution rate, the first n IMF components of the decomposed sliding bearing vibration signal are selected for signal reconstruction processing to obtain the reconstructed sliding bearing vibration signal.

[0014] Furthermore, the step of sequentially performing discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal, and calculating the time domain characteristics of the wavelet coefficients to obtain the sliding bearing vibration signal characteristic matrix specifically includes:

[0015] Decomposing the reconstructed sliding bearing vibration signal by discrete wavelet decomposition algorithm to obtain a sliding bearing vibration signal decomposition coefficient vector, wherein the sliding bearing vibration signal decomposition coefficient vector includes an approximation coefficient and four groups of detail coefficients;

[0016] Extract features from the decomposition coefficient vector of the sliding bearing vibration signal to obtain a feature vector of the sliding bearing vibration signal;

[0017] The characteristic vectors of the sliding bearing vibration signal are normalized and aggregated to obtain the characteristic matrix of the sliding bearing vibration signal.

[0018] Furthermore, the sliding bearing vibration signal feature vector body includes wavelet coefficient variance, wavelet coefficient kurtosis, wavelet shape factor, wavelet peak factor, wavelet scale centroid, Shannon entropy, logarithmic energy entropy, wavelet maximum value, wavelet minimum value, wavelet margin factor, third-order central moment, fourth-order central moment and wavelet energy entropy.

[0019] Furthermore, the CNN-MLB-AB sliding bearing water pollution identification neural network model specifically includes an input layer, a sequence folding layer, a convolution layer, a batch normalization layer, an activation layer, a pooling layer, a sequence expansion layer, a flattening layer, a multi-layer bidirectional long-short term layer, a fully connected layer, a Softmax layer and a classification layer, wherein the multi-layer bidirectional long-short term layer includes a first BiLSTM layer and a second BiLSTM layer, and the multi-layer bidirectional long-short term layer adopts an AdaBoost ensemble learning method.

[0020] Further, the step of performing fault identification on the sliding bearing vibration signal feature matrix based on the CNN-MLB-AB sliding bearing water contamination identification neural network model to obtain the sliding bearing water contamination identification result specifically includes:

[0021] The characteristic matrix of sliding bearing vibration signal is input into the sliding bearing water pollution identification neural network model based on CNN-MLB-AB;

[0022] Based on the input layer of the CNN-MLB-AB sliding bearing water contamination identification neural network model, the characteristic matrix of the sliding bearing vibration signal is obtained;

[0023] Based on the sequence folding layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the sliding bearing vibration signal feature matrix is ​​integrated and processed to obtain the two-dimensional feature data of the sliding bearing vibration signal suitable for convolution operation.

[0024] Based on the convolution layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the two-dimensional feature data of the sliding bearing vibration signal suitable for convolution operation is convoluted to extract the local time series features and obtain the convolution feature data of the sliding bearing vibration signal;

[0025] Based on the batch normalization layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the convolution feature data of the sliding bearing vibration signal is normalized to obtain the normalized convolution feature data of the sliding bearing vibration signal;

[0026] Based on the activation layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the normalized sliding bearing vibration signal convolution feature data is processed nonlinearly to obtain the nonlinear sliding bearing vibration signal convolution feature data;

[0027] Based on the pooling layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the convolution feature data of the nonlinear sliding bearing vibration signal is pooled to reduce the feature dimension and obtain the convolution feature data of the sliding bearing vibration signal after dimensionality reduction;

[0028] Based on the sequence expansion layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the convolution feature data of the sliding bearing vibration signal after dimensionality reduction is expanded and restored to the time series feature data of the sliding bearing vibration signal;

[0029] Based on the flat layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the sliding bearing vibration signal time series feature data is flattened to obtain one-dimensional sliding bearing vibration signal time series feature data;

[0030] Based on the multi-layer bidirectional long-short term layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the time series characteristic data of the one-dimensional sliding bearing vibration signal is analyzed to obtain the bidirectional time series characteristics of the sliding bearing vibration signal;

[0031] Based on the fully connected layer of the CNN-MLB-AB sliding bearing water contamination recognition neural network model, feature mapping and fusion are performed on the bidirectional time series features of the sliding bearing vibration signal, and the high-dimensional features are compressed into the output feature space of the corresponding number of categories.

[0032] Based on the Softmax layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the high-dimensional features are compressed to the output feature space of the corresponding number of categories and mapped into the probability distribution of each category to obtain the probability value of the sliding bearing water pollution prediction result;

[0033] Based on the classification layer of the CNN-MLB-AB sliding bearing water contamination recognition neural network model, the category with the largest probability is selected as the final classification result according to the probability value of the sliding bearing water contamination prediction result, and the sliding bearing water contamination recognition result is obtained.

[0034] Further, the multi-layer bidirectional long-short term layer based on the CNN-MLB-AB sliding bearing water pollution identification neural network model performs time series analysis on the one-dimensional sliding bearing vibration signal time series feature data to obtain the bidirectional time series features of the sliding bearing vibration signal. This step specifically includes:

[0035] Input the one-dimensional sliding bearing vibration signal time series characteristic data into the multi-layer bidirectional long-term and short-term layers;

[0036] Based on the first BiLSTM layer of the multi-layer bidirectional long-term and short-term layers, the time feature extraction is performed on the one-dimensional sliding bearing vibration signal time series feature data to obtain the underlying time series features of the sliding bearing vibration signal;

[0037] The second BiLSTM layer based on the multi-layer bidirectional long-term and short-term layers performs time dependency analysis on the underlying time series features of the sliding bearing vibration signal to obtain the high-level time series features of the sliding bearing vibration signal.

[0038] Based on the AdaBoost framework, the prediction results of multiple CNN-BiLSTM weak classifiers are weighted and combined, and the output results of the multi-layer bidirectional long-term and short-term layers are integrated and calculated using the weight optimization strategy to obtain the bidirectional time series characteristics of the sliding bearing vibration signal.

[0039] The second technical solution adopted by the present invention is: a sliding bearing water pollution identification system based on CNN-MLB-AB, comprising:

[0040] The first module is used to obtain the vibration signal of the sliding bearing and perform signal decomposition and signal reconstruction processing in sequence through the variational mode decomposition algorithm to obtain the reconstructed sliding bearing vibration signal, wherein the sliding bearing vibration signal includes the sliding bearing water contamination vibration signal and the sliding bearing fault-free vibration signal;

[0041] The second module is used to perform discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal in sequence, and calculate the time domain characteristics of the wavelet coefficients to obtain the sliding bearing vibration signal characteristic matrix;

[0042] The third module is used to introduce multi-layer bidirectional long-short term layers and construct the CNN-MLB-AB sliding bearing water pollution recognition neural network model;

[0043] The fourth module is used to perform fault identification on the sliding bearing vibration signal feature matrix based on the CNN-MLB-AB sliding bearing water contamination identification neural network model to obtain the sliding bearing water contamination identification result.

[0044] The beneficial effects of the method and system of the present invention are as follows: the present invention obtains the sliding bearing vibration signal and sequentially performs signal decomposition and signal reconstruction processing through a variational mode decomposition algorithm, and further sequentially performs discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal, so as to capture the transient characteristics of the sliding bearing vibration signal at different frequency bandwidths, thereby comprehensively reflecting the different frequency band information of the sliding bearing vibration signal, adapting to the non-stationarity of the signal, and then introducing a multi-layer bidirectional long-term and short-term layer, constructing a CNN-MLB-AB sliding bearing water pollution recognition neural network model, and finally performing fault recognition on the sliding bearing vibration signal feature matrix, which can fully utilize the ability of a convolutional neural network (CNN) to extract local features and the ability of a multi-layer BiLSTM to capture time series dependencies, thereby enhancing the overall classification performance of the model, making it pay more attention to samples that are difficult to classify, thereby improving the accuracy of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of the steps of the sliding bearing water contamination identification method based on CNN-MLB-AB of the present invention;

[0046] Figure 2 It is a structural block diagram of the sliding bearing water pollution identification system based on CNN-MLB-AB of the present invention;

[0047] Figure 3 is a schematic diagram of VMD decomposition results provided by a specific embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram of variance contribution rate of each component provided by a specific embodiment of the present invention;

[0049] Figure 5 is a schematic diagram comparing the reconstructed vibration signal and the original vibration signal provided by a specific embodiment of the present invention;

[0050] Figure 6 It is a schematic diagram of detail coefficients and approximation coefficients obtained by discrete wavelet decomposition according to a specific embodiment of the present invention;

[0051] Figure 7 It is a CNN-MLB-AB neural network structure diagram provided by a specific embodiment of the present invention;

[0052] Figure 8 is a schematic diagram of CNN-MLB-AB fault diagnosis results based on raw data provided by a specific embodiment of the present invention;

[0053] Fig. 9 It is a schematic diagram of a confusion matrix based on CNN fault diagnosis provided by a specific embodiment of the present invention;

[0054] Fig.10 It is a schematic diagram of a confusion matrix for fault diagnosis based on CNN-BiLSTM provided by a specific embodiment of the present invention;

[0055] Fig.11 It is a schematic diagram of fault diagnosis results based on CNN-MLB-AB provided by a specific embodiment of the present invention;

[0056] Fig.12 It is a schematic diagram of a confusion matrix for fault diagnosis based on CNN-MLB-AB provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0058] Reference Figure 1 The present invention provides a method for identifying water contamination of sliding bearings based on CNN-MLB-AB, the method comprising the following steps:

[0059] S100, acquiring a sliding bearing vibration signal, and sequentially performing signal decomposition and signal reconstruction processing using a variational mode decomposition algorithm to obtain a reconstructed sliding bearing vibration signal, wherein the sliding bearing vibration signal includes a sliding bearing water contamination vibration signal and a sliding bearing fault-free vibration signal;

[0060] S110, collecting vibration signals of sliding bearings;

[0061] In this embodiment, vibration signals of the sliding bearing are collected under the conditions of no fault and water contamination of 0%, 3%, and 5% and rotation speed of 900 r / min, 1200 r / min, and 1500 r / min.

[0062] S120, performing signal decomposition processing on the sliding bearing vibration signal by using a variational mode decomposition algorithm to obtain a plurality of decomposed sliding bearing vibration signal IMF components;

[0063] S130, obtaining variance contribution rates corresponding to a number of decomposed IMF components of sliding bearing vibration signals;

[0064] S140, selecting the first n decomposed IMF components of the sliding bearing vibration signal according to the variance contribution rate to perform signal reconstruction processing to obtain a reconstructed sliding bearing vibration signal.

[0065] In this embodiment, VMD is used to decompose the vibration signal of the sliding bearing, the decomposition level is 5, the penalty factor is 2500, and the decomposition result is as follows: Figure 3 After using VMD to decompose the vibration signal of the sliding bearing, as shown in Figure 4 The variance contribution rate of each component is shown. According to the variance contribution rate, the top three IMF components with the highest variance contribution rate are selected for signal reconstruction. The schematic diagram of the comparison between the reconstructed signal and the original signal is shown in Figure 5 As shown, the variance contribution rate is calculated by the following formula:

[0066]

[0067] In the above formula, is the variance of the kth modal component, indicating the energy in this component, T is the total duration of the signal, u k (t) represents the variation of the kth modal component with time t.

[0068] S200, performing discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal in sequence, and calculating the time domain characteristics of the wavelet coefficients to obtain a sliding bearing vibration signal characteristic matrix;

[0069] S210, decomposing the reconstructed sliding bearing vibration signal by a discrete wavelet decomposition algorithm to obtain a sliding bearing vibration signal decomposition coefficient vector, wherein the sliding bearing vibration signal decomposition coefficient vector includes an approximation coefficient and four groups of detail coefficients;

[0070] S220, extracting features from the decomposition coefficient vector of the sliding bearing vibration signal to obtain a feature vector of the sliding bearing vibration signal;

[0071] In this embodiment, the reconstructed signal is decomposed by discrete wavelet, using Daubechies 4 wavelet basis functions, with 4 decomposition layers, where the calculation formula for the i-th layer detail coefficient is as follows:

[0072]

[0073] In the above formula, D i represents the detail coefficient of the i-th layer, representing the high-frequency component of the signal, x rec (n) represents the discrete sampling value of the reconstructed signal, ψ(t) is the wavelet basis function Daubechies 4, t is the time variable, 2 i Indicates different decomposition scales.

[0074] The calculation formula of the approximation coefficient is as follows:

[0075]

[0076] In the above formula, A4 represents the fourth-layer approximation coefficient, representing the low-frequency component of the signal, and φ(t) represents the approximation function.

[0077] The coefficients generated by wavelet decomposition include approximation coefficient A4 and detail coefficients D1, D2, D3, D4, and 13 features of the decomposition coefficients are calculated; among them, Figure 6 As shown, the reconstructed signal is decomposed by discrete wavelet to obtain a decomposition coefficient vector, and then the decomposition coefficient vector containing four layers of detail coefficients and approximation coefficients is used to calculate various features of the decomposition coefficient vector, including:

[0078] The decomposition coefficient obtained is a long vector, which stores the approximation coefficients and detail coefficients of different decomposition levels in sequence. For 4-layer wavelet decomposition, the vector structure is as follows:

[0079] First layer detail coefficient (high frequency component, D1);

[0080] Second layer detail coefficient (D2);

[0081] The third layer detail coefficient (D3);

[0082] The fourth layer detail coefficient (D4);

[0083] The fourth layer approximation coefficient (low-frequency component, A4);

[0084] The coefficients of these five parts are stored in a vector, and the coefficients are stored in order: first-layer detail coefficient (D1), second-layer detail coefficient (D2), third-layer detail coefficient (D3), fourth-layer detail coefficient (D4), and fourth-layer approximation coefficient (A4). The detail coefficients (D1-D4) represent the high-frequency components of the signal, and the frequency gradually decreases from rapid changes to smooth changes, capturing different levels of detail information. The approximation coefficient (A4) represents the low-frequency components of the signal, reflecting the overall trend and main characteristics of the signal. The original signal is divided into A1 (low-frequency part) and D1 (high-frequency part), and A1 is further decomposed into A2 and D2. By analogy, the signal S can be expressed as:

[0085] S=A4+D4+D3+D2+D1

[0086] Calculate the characteristics of the decomposition coefficients: wavelet coefficient variance, wavelet coefficient kurtosis, wavelet shape factor, wavelet peak factor, wavelet scale centroid, Shannon entropy, logarithmic energy entropy, wavelet maximum, wavelet minimum, wavelet margin factor, third-order central moment, fourth-order central moment, wavelet energy entropy, a total of 13 characteristics.

[0087] S230, performing feature normalization and aggregation processing on the characteristic vectors of the sliding bearing vibration signal to obtain a characteristic matrix of the sliding bearing vibration signal.

[0088] In this embodiment, based on the features calculated by the decomposition coefficients, the features are normalized, the mean of each feature is adjusted to 0, and the standard deviation is adjusted to 1. After each feature is normalized, the normalized features are aggregated into a feature matrix.

[0089] S300, introduce multi-layer bidirectional long-short term layers to build the CNN-MLB-AB sliding bearing water pollution identification neural network model;

[0090] Specifically, the CNN-MLB-AB sliding bearing water pollution identification neural network model specifically includes an input layer, a sequence folding layer, a convolution layer, a batch normalization layer, an activation layer, a pooling layer, a sequence expansion layer, a flattening layer, a multi-layer bidirectional long-short term layer, a fully connected layer, a Softmax layer and a classification layer, wherein the multi-layer bidirectional long-short term layer includes a first BiLSTM layer and a second BiLSTM layer, and the multi-layer bidirectional long-short term layer adopts an AdaBoost ensemble learning method.

[0091] S400, based on the CNN-MLB-AB sliding bearing water contamination identification neural network model, fault identification is performed on the sliding bearing vibration signal feature matrix to obtain the sliding bearing water contamination identification result.

[0092] First of all, it should be noted that the embodiment of the present invention performs discrete wavelet transform on the original signal, extracts features and normalizes them, and then inputs them into the constructed CNN-MLB-AB model to obtain fault diagnosis results. The original signal is then decomposed by VMD, the first three components are selected for reconstruction, and discrete wavelet decomposition is performed again. After extracting features, they are input into the same model for diagnosis. 75% of the data set is used for training and 25% for testing. In addition, CNN and CNN-BiLSTM sliding bearing water contamination recognition neural networks are constructed, and the fault recognition performance is compared with CNN-MLB-AB. Figure 7 As shown in the figure, the hierarchical structure of the CNN-MLB-AB sliding bearing water pollution recognition neural network mainly includes input layer, sequence folding layer, convolution layer, batch normalization layer, activation layer, pooling layer, sequence expansion layer, flattening layer, multi-layer bidirectional long-short term layer, fully connected layer, Softmax layer and classification layer.

[0093] The input layer receives the processed feature data and then extracts the local features of the input feature data through the convolution layer. The convolution kernel is [6, 1] and the number of convolution kernels is 8. The batch normalization layer normalizes the data. After the nonlinear processing of the activation layer, the pooling layer takes the local maximum value of the region to retain important features. The multi-layer bidirectional long-term and short-term layer processes the forward and backward time series data simultaneously. The first layer of BiLSTM is set to 120 hidden units.

[0094] Specifically, the sliding bearing vibration signal feature matrix is ​​input into the sliding bearing water pollution recognition neural network model based on CNN-MLB-AB; based on the input layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the sliding bearing vibration signal feature matrix is ​​obtained; based on the sequence folding layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the sliding bearing vibration signal feature matrix is ​​integrated and processed to obtain the two-dimensional feature data of the sliding bearing vibration signal suitable for convolution operation; based on the convolution layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the two-dimensional feature data of the sliding bearing vibration signal suitable for convolution operation is convolved to extract the local time series features to obtain the sliding bearing water pollution recognition neural network model. Convolution feature data of bearing vibration signal; based on the batch normalization layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the convolution feature data of the sliding bearing vibration signal is normalized to obtain the normalized convolution feature data of the sliding bearing vibration signal; based on the activation layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the normalized convolution feature data of the sliding bearing vibration signal is subjected to nonlinear activation processing to obtain the nonlinear convolution feature data of the sliding bearing vibration signal; based on the pooling layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the nonlinear sliding bearing vibration signal convolution feature data is subjected to pooling processing to reduce the feature dimension to obtain the sliding bearing vibration after dimensionality reduction. Signal convolution feature data; based on the sequence expansion layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the convolution feature data of the sliding bearing vibration signal after dimensionality reduction is expanded to restore it to the time series feature data of the sliding bearing vibration signal; based on the flattening layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the time series feature data of the sliding bearing vibration signal is flattened to obtain one-dimensional time series feature data of the sliding bearing vibration signal; based on the multi-layer bidirectional long-short term layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the one-dimensional time series feature data of the sliding bearing vibration signal is analyzed in time series to obtain the bidirectional time series features of the sliding bearing vibration signal ; Based on the fully connected layer of the CNN-MLB-AB sliding bearing water contamination identification neural network model, the bidirectional time series features of the sliding bearing vibration signal are feature mapped and fused, and the high-dimensional features are compressed to the output feature space of the corresponding number of categories; Based on the Softmax layer of the CNN-MLB-AB sliding bearing water contamination identification neural network model, the high-dimensional features are compressed to the output feature space of the corresponding number of categories and mapped into the probability distribution of each category to obtain the probability value of the sliding bearing water contamination prediction result; Based on the classification layer of the CNN-MLB-AB sliding bearing water contamination identification neural network model, according to the probability value of the sliding bearing water contamination prediction result, the category with the largest probability is selected as the final classification result to obtain the sliding bearing water contamination identification result.

[0095] Among them, the one-dimensional sliding bearing vibration signal time series feature data is input into the multi-layer bidirectional long-term and short-term layer; based on the first BiLSTM layer of the multi-layer bidirectional long-term and short-term layer, the time feature of the one-dimensional sliding bearing vibration signal time series feature data is extracted to obtain the underlying time series features of the sliding bearing vibration signal; based on the second BiLSTM layer of the multi-layer bidirectional long-term and short-term layer, the time dependence analysis of the underlying time series features of the sliding bearing vibration signal is performed to obtain the high-level time series features of the sliding bearing vibration signal; based on the AdaBoost framework, the prediction results of multiple CNN-BiLSTM weak classifiers are weighted combined, and the output results of the multi-layer bidirectional long-term and short-term layer are integrated and calculated using the weight optimization strategy to obtain the bidirectional time series features of the sliding bearing vibration signal.

[0096] In this embodiment, a two-layer BiLSTM structure is used to capture the complex time dependency in the sequence. The first layer gradually extracts the time series features, and the second layer integrates these features to obtain the overall information of the sequence for subsequent classification tasks.

[0097] Specifically, the first layer of BiLSTM is processed as follows:

[0098] Forward LSTM: read the sequence X = {x1, x2, ..., xT} from time step t = 1 to t = T, and calculate the hidden state

[0099] The forward LSTM expression is as follows:

[0100]

[0101] In the above formula, xt represents the characteristics of the input sequence at time step t, T represents the length of the sequence, is the memory state of the forward LSTM unit, is the hidden state of the forward LSTM at time step t+1.

[0102] Reverse LSTM: read the sequence backwards from time step t=T to t=1 and calculate the hidden state

[0103] The reverse LSTM expression is as follows:

[0104]

[0105] In the above formula, xt represents the characteristics of the input sequence at time step t, T represents the length of the sequence, is the memory state of the reverse LSTM unit, is the hidden state of the reverse LSTM at time step t+1.

[0106] The output expression of the first layer BiLSTM is as follows:

[0107]

[0108] In the above formula, Is the positive and negative hidden state as the input of the second layer.

[0109] The second layer of BiLSTM is set to 100 hidden units and is set to only output the state of the last time step, thus summarizing the information of the entire sequence.

[0110] The second layer forward and reverse LSTM calculation expressions are as follows:

[0111]

[0112] In the above formula, is the hidden state of the second layer forward LSTM at time step t+1, is the hidden state of the second layer of reverse LSTM at time step t+1, is the memory state of the second layer forward LSTM unit, is the memory state of the second layer of reverse LSTM unit. The remaining characters are the same as above.

[0113] Print the last time step:

[0114]

[0115] In the above formula, H (2) is the hidden state at the last time step of the sequence.

[0116] Adaboost combines multiple independently trained BiLSTM weak classifiers to form a strong classifier. By gradually adjusting the sample weights, it pays more attention to difficult-to-classify samples during training and improves the overall classification performance.

[0117] Specifically, the Adaboost processing flow includes:

[0118] Initialize sample weights, train weak classifiers, combine weak classifiers for final prediction, and independently predict each BiLSTM weak classifier to obtain the final classification result. When initializing sample weights, the weight of each sample is equal to ensure that all samples are considered equally in the first round of training. The expression of the initialization sample weight is as follows:

[0119]

[0120] In the above formula, represents the initial weight of the i-th sample in the first round of training, N is the total number of training samples, and i is the sample index.

[0121] Training weak classifiers Adaboost gradually generates multiple weak classifiers, each of which is an independent multi-layer BiLSTM model. Assume that there are T weak classifiers in total. The training steps for the tth weak classifier include:

[0122] According to the current sample weight After sampling the training data and training the t-th BiLSTM model ht(x), the classification error rate is calculated. The classification error rate expression is as follows:

[0123]

[0124] In the above formula, ∈t represents the classification error rate of the tth weak classifier, represents the weight of the i-th sample in the t-th round, ht(xi) represents the predicted output of the t-th weak classifier for the i-th sample, yi represents the true label of the i-th sample, and I(ht(xi)≠yi) represents the indicator function.

[0125] After calculating the classification error rate, the classifier weight is calculated according to the classification error rate. The classifier weight expression is as follows:

[0126]

[0127] In the above formula, αt represents the weight of the t-th weak classifier, and ∈t represents the classification error rate of the t-th classifier.

[0128] After calculating the classifier weights, update the sample weights. The update expression is as follows:

[0129]

[0130] In the above formula, represents the weight of the i-th sample in the t+1th round, represents the weight of the i-th sample in the t-th round, αt represents the weight of the t-th weak classifier, yi represents the true label of the i-th sample, and ht(xi) represents the prediction of the t-th weak classifier for the i-th sample.

[0131] Finally, the weak classifiers are combined for the final prediction. After completing T rounds of training, Adaboost combines the outputs of all weak classifiers to generate the final strong classifier. The expression for combining weak classifiers for the final prediction is as follows:

[0132]

[0133] In the above formula, H(x) represents the prediction output of the final strong classifier for the input x, T represents the total number of weak classifiers trained, αt represents the weight of the tth weak classifier, and ht(x) represents the prediction output of the tth weak classifier for the input x.

[0134] After the combined classifier makes the final prediction, each BiLSTM weak classifier is independently predicted to obtain the final classification result. In the test phase, Adaboost uses the trained BiLSTM weak classifier to make weighted predictions on new test samples and finally generates the classification result. Among them:

[0135] The expressions for predicting classification according to each BiLSTM weak classifier are as follows:

[0136]

[0137] In the above formula, H(xtest) represents the prediction result of the final strong classifier for the test sample xtest, αt represents the weight of the tth weak classifier, and ht(xtest) represents the prediction output of the tth weak classifier for the test sample.

[0138] The original signal and the reconstructed signal are decomposed by discrete wavelet, and the characteristics of the decomposition coefficients are calculated respectively. The decomposition coefficients are input into the CNN-MLB-AB sliding bearing water pollution recognition neural network to obtain the diagnosis results.

[0139] Specifically, the original signal is decomposed by discrete wavelet, the decomposition coefficient characteristics are calculated, and the feature data set is obtained, which is then input into the CNN-MLB-AB sliding bearing water pollution recognition neural network. The first 75% of the data set is used for training, and the last 25% is used for testing. The fault diagnosis schematic diagram is as follows: Figure 8 As shown, the diagnostic accuracy is only 76.33%.

[0140] The reconstructed signal is decomposed by wavelet discrete, the characteristics of the decomposition coefficients are calculated, and the feature data set is obtained and input into the CNN and BiLSTM sliding bearing water pollution recognition neural network. The first 75% of the data set is used for training and the last 25% is used for testing. Fig. 9 and Fig.10 As shown in Figure 2, the diagnostic accuracy of the CNN model and the CNN-BiLSTM model are 80.68% and 91.3%, respectively.

[0141] The reconstructed signal is decomposed by wavelet discrete, the characteristics of the decomposition coefficients are calculated, and the characteristic data set is obtained and input into the CNN-MLB-AB sliding bearing water pollution recognition neural network. The first 75% of the data set is used for training and the last 25% is used for testing. Fig.11 and Fig.12 As shown, the diagnostic accuracy rate is as high as 99.03%.

[0142] In summary, by reconstructing the signal using the variance contribution rate after VMD decomposition, and then performing discrete wavelet decomposition and extracting the decomposition coefficient features, noise and irrelevant information can be further removed, and the characteristics of the signal in different frequency ranges can be effectively captured, retaining the key features related to the fault. This feature extraction method greatly enhances the ability to distinguish between different fault states. Compared with the traditional CNN and CNN-BiLSTM models, the present invention further improves the overall classification performance of the model by combining CNN, Multi-Layer-BiLSTM and Adaboost technology. Multi-layer BiLSTM performs well in capturing temporal dependencies, while the Adaboost method improves the model's attention and accuracy to complex signals through a weighted combination of multiple weak classifiers. Experimental results show that the diagnostic accuracy of the CNN-MLB-AB model is as high as 99.03%, which is significantly better than the traditional CNN model (80.68%) and CNN-BiLSTM model (91.3%). This deep learning method combining multi-level feature extraction with ensemble learning not only improves the diagnostic accuracy, but also enhances the robustness and generalization ability of the model to complex signals, especially in the identification of complex water pollution. The sliding bearing water contamination identification method of the present invention can identify faults more accurately, greatly improving the accuracy and reliability of fault detection.

[0143] Reference Figure 2 , the sliding bearing water contamination identification system based on CNN-MLB-AB includes:

[0144] The first module 201 is used to obtain a sliding bearing vibration signal and perform signal decomposition and signal reconstruction processing in sequence through a variational mode decomposition algorithm to obtain a reconstructed sliding bearing vibration signal, wherein the sliding bearing vibration signal includes a sliding bearing water contamination vibration signal and a sliding bearing fault-free vibration signal;

[0145] The second module 202 is used to perform discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal in sequence, and calculate the time domain characteristics of the wavelet coefficients to obtain the sliding bearing vibration signal characteristic matrix;

[0146] The third module 203 is used to introduce a multi-layer bidirectional long-term and short-term layer to construct a CNN-MLB-AB sliding bearing water pollution recognition neural network model;

[0147] The fourth module 204 is used to perform fault identification on the sliding bearing vibration signal feature matrix based on the CNN-MLB-AB sliding bearing water contamination identification neural network model to obtain the sliding bearing water contamination identification result.

[0148] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0149] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for identifying water contamination in sliding bearings based on CNN-MLB-AB, characterized in that: The following steps are involved: Acquire the sliding bearing vibration signal and perform signal decomposition and signal reconstruction processing in sequence through the variational mode decomposition algorithm to obtain a reconstructed sliding bearing vibration signal, wherein the sliding bearing vibration signal includes a sliding bearing water contamination vibration signal and a sliding bearing fault-free vibration signal; The reconstructed sliding bearing vibration signal is sequentially subjected to discrete wavelet decomposition and normalization processing, and the time domain characteristics of the wavelet coefficients are calculated to obtain the sliding bearing vibration signal characteristic matrix; Multi-layer bidirectional long-short term layers were introduced to construct the CNN-MLB-AB neural network model for water pollution identification of sliding bearings. Based on the CNN-MLB-AB sliding bearing water contamination identification neural network model, the sliding bearing vibration signal feature matrix is ​​used for fault identification, and the sliding bearing water contamination identification result is obtained.

2. The method for identifying water contamination in sliding bearings based on CNN-MLB-AB according to claim 1 is characterized in that: The step of obtaining the sliding bearing vibration signal and sequentially performing signal decomposition and signal reconstruction processing by using a variational mode decomposition algorithm to obtain a reconstructed sliding bearing vibration signal specifically includes: Collect the vibration signal of sliding bearing; The vibration signal of the sliding bearing is decomposed by using the variational mode decomposition algorithm to obtain several decomposed IMF components of the vibration signal of the sliding bearing; Obtain variance contribution rates corresponding to several decomposed IMF components of sliding bearing vibration signals; According to the variance contribution rate, the first n IMF components of the decomposed sliding bearing vibration signal are selected for signal reconstruction processing to obtain the reconstructed sliding bearing vibration signal.

3. The method for identifying water contamination in sliding bearings based on CNN-MLB-AB according to claim 2 is characterized in that: The step of sequentially performing discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal, and calculating the time domain characteristics of the wavelet coefficients to obtain the sliding bearing vibration signal characteristic matrix specifically includes: Decomposing the reconstructed sliding bearing vibration signal by discrete wavelet decomposition algorithm to obtain a sliding bearing vibration signal decomposition coefficient vector, wherein the sliding bearing vibration signal decomposition coefficient vector includes an approximation coefficient and four groups of detail coefficients; Extract features from the decomposition coefficient vector of the sliding bearing vibration signal to obtain a feature vector of the sliding bearing vibration signal; The characteristic vectors of the sliding bearing vibration signal are normalized and aggregated to obtain the characteristic matrix of the sliding bearing vibration signal.

4. The method for identifying water contamination in sliding bearings based on CNN-MLB-AB according to claim 3 is characterized in that: The sliding bearing vibration signal feature vector body includes wavelet coefficient variance, wavelet coefficient kurtosis, wavelet shape factor, wavelet peak factor, wavelet scale centroid, Shannon entropy, logarithmic energy entropy, wavelet maximum value, wavelet minimum value, wavelet margin factor, third-order central moment, fourth-order central moment and wavelet energy entropy.

5. The method for identifying water contamination in sliding bearings based on CNN-MLB-AB according to claim 4 is characterized in that: The CNN-MLB-AB sliding bearing water pollution identification neural network model specifically includes an input layer, a sequence folding layer, a convolution layer, a batch normalization layer, an activation layer, a pooling layer, a sequence expansion layer, a flattening layer, a multi-layer bidirectional long-short term layer, a fully connected layer, a Softmax layer and a classification layer, wherein the multi-layer bidirectional long-short term layer includes a first BiLSTM layer and a second BiLSTM layer, and the multi-layer bidirectional long-short term layer adopts an AdaBoost ensemble learning method.

6. The method for identifying water contamination in sliding bearings based on CNN-MLB-AB according to claim 5 is characterized in that: The step of performing fault identification on the sliding bearing vibration signal feature matrix based on the CNN-MLB-AB sliding bearing water contamination identification neural network model to obtain the sliding bearing water contamination identification result specifically includes: The characteristic matrix of sliding bearing vibration signal is input into the sliding bearing water pollution identification neural network model based on CNN-MLB-AB; Based on the input layer of the CNN-MLB-AB sliding bearing water contamination identification neural network model, the characteristic matrix of the sliding bearing vibration signal is obtained; Based on the sequence folding layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the sliding bearing vibration signal feature matrix is ​​integrated and processed to obtain the two-dimensional feature data of the sliding bearing vibration signal suitable for convolution operation. Based on the convolution layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the two-dimensional feature data of the sliding bearing vibration signal suitable for convolution operation is convoluted to extract the local time series features and obtain the convolution feature data of the sliding bearing vibration signal; Based on the batch normalization layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the convolution feature data of the sliding bearing vibration signal is normalized to obtain the normalized convolution feature data of the sliding bearing vibration signal; Based on the activation layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the normalized sliding bearing vibration signal convolution feature data is processed nonlinearly to obtain the nonlinear sliding bearing vibration signal convolution feature data; Based on the pooling layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the convolution feature data of the nonlinear sliding bearing vibration signal is pooled to reduce the feature dimension and obtain the convolution feature data of the sliding bearing vibration signal after dimensionality reduction; Based on the sequence expansion layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the convolution feature data of the sliding bearing vibration signal after dimensionality reduction is expanded and restored to the time series feature data of the sliding bearing vibration signal; Based on the flat layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the sliding bearing vibration signal time series feature data is flattened to obtain one-dimensional sliding bearing vibration signal time series feature data; Based on the multi-layer bidirectional long-short term layer of the CNN-MLB-AB sliding bearing water pollution identification neural network model, the time series characteristic data of the one-dimensional sliding bearing vibration signal is analyzed to obtain the bidirectional time series characteristics of the sliding bearing vibration signal; Based on the fully connected layer of the CNN-MLB-AB sliding bearing water contamination recognition neural network model, feature mapping and fusion are performed on the bidirectional time series features of the sliding bearing vibration signal, and the high-dimensional features are compressed into the output feature space of the corresponding number of categories. Based on the Softmax layer of the CNN-MLB-AB sliding bearing water pollution recognition neural network model, the high-dimensional features are compressed to the output feature space of the corresponding number of categories and mapped into the probability distribution of each category to obtain the probability value of the sliding bearing water pollution prediction result; Based on the classification layer of the CNN-MLB-AB sliding bearing water contamination recognition neural network model, the category with the largest probability is selected as the final classification result according to the probability value of the sliding bearing water contamination prediction result, and the sliding bearing water contamination recognition result is obtained.

7. The method for identifying water contamination in sliding bearings based on CNN-MLB-AB according to claim 6 is characterized in that: The multi-layer bidirectional long-short term layer based on the CNN-MLB-AB sliding bearing water pollution identification neural network model performs time series analysis on the one-dimensional sliding bearing vibration signal time series feature data to obtain the bidirectional time series feature of the sliding bearing vibration signal. This step specifically includes: Input the one-dimensional sliding bearing vibration signal time series characteristic data into the multi-layer bidirectional long-term and short-term layers; Based on the first BiLSTM layer of the multi-layer bidirectional long-term and short-term layers, the time feature extraction is performed on the one-dimensional sliding bearing vibration signal time series feature data to obtain the underlying time series features of the sliding bearing vibration signal; The second BiLSTM layer based on the multi-layer bidirectional long-term and short-term layers performs time dependency analysis on the underlying time series features of the sliding bearing vibration signal to obtain the high-level time series features of the sliding bearing vibration signal. Based on the AdaBoost framework, the prediction results of multiple CNN-BiLSTM weak classifiers are weighted and combined, and the output results of the multi-layer bidirectional long-term and short-term layers are integrated and calculated using the weight optimization strategy to obtain the bidirectional time series characteristics of the sliding bearing vibration signal.

8. The sliding bearing water pollution identification system based on CNN-MLB-AB is characterized by: Includes the following modules: The first module is used to obtain the vibration signal of the sliding bearing and perform signal decomposition and signal reconstruction processing in sequence through the variational mode decomposition algorithm to obtain the reconstructed sliding bearing vibration signal, wherein the sliding bearing vibration signal includes the sliding bearing water contamination vibration signal and the sliding bearing fault-free vibration signal; The second module is used to perform discrete wavelet decomposition and normalization processing on the reconstructed sliding bearing vibration signal in sequence, and calculate the time domain characteristics of the wavelet coefficients to obtain the sliding bearing vibration signal characteristic matrix; The third module is used to introduce multi-layer bidirectional long-short term layers and construct the CNN-MLB-AB sliding bearing water pollution recognition neural network model; The fourth module is used to perform fault identification on the sliding bearing vibration signal feature matrix based on the CNN-MLB-AB sliding bearing water contamination identification neural network model to obtain the sliding bearing water contamination identification result.

Images