Carrier roller fault diagnosis method based on aliasing period fusion features

Abstract

The invention discloses a carrier roller fault diagnosis method based on aliasing period fusion features, and particularly relates to the field of roller fault diagnosis. According to the scheme, the operation data of the carrier roller of the conveyor are collected through the sensor, and the collected original sensing data are preprocessed; the method comprises the following steps: dividing carrier roller sensing data into samples based on sliding window sampling, and dividing into a training set, a verification set and a test set according to 6: 3: 1; a carrier roller fault diagnosis model based on the Times-SCConv network is constructed, and parameters are initialized; training and verifying the Times-SCConv network based on the training set and the verification set, optimizing network parameters by using an optimization algorithm, and storing an optimal weight; and inputting the test set into the optimal Times-SCConv network model to carry out carrier roller fault diagnosis, and finally obtaining a carrier roller fault classification result and test accuracy. According to the technical scheme, the problem that fault feature extraction is difficult when sensing data signals of the carrier roller are subjected to aliasing coupling is solved, and modeling extraction of carrier roller data periodic change features is achieved.

Description

Technical Field

[0001] This invention relates to the field of roller fault diagnosis, and in particular to a method for diagnosing idler roller faults based on the characteristics of mixed-cycle fusion. Background Technology

[0002] Belt conveyors offer advantages such as continuous long-distance transport and large capacity, making them an indispensable part of modern process industries. They are widely used in factory assembly lines, logistics sorting, and many other fields. Idler rollers are key components of belt conveyors, primarily serving to support and guide the conveyor belt during material transport. Their number increases rapidly with the transport distance, accounting for a significant portion of the conveyor's overall weight and value. Therefore, idler roller failures can severely impact production efficiency and even lead to safety accidents. Because the entire conveyor comprises numerous components, and multiple components are involved in the transport process, the collected idler roller sensor data signals are aliased signals resulting from the coupling of various state signals. The nonlinear relationships between these different states make it difficult to accurately extract fault characteristics from the complex data relationships, posing significant challenges to idler roller fault diagnosis.

[0003] When faults such as idler bearing failure, roller wear, and seal failure occur, the rotational motion of the idler generates periodic signals at a specific frequency. These periodic components can be disturbed by other non-stationary factors, such as uneven changes in conveyor belt load and wear variations in the idler itself, making it difficult to model and extract the periodic variation characteristics of the idler data. Summary of the Invention

[0004] The present invention aims to provide a fault diagnosis method for idler rollers based on the fusion characteristics of aliasing periodicity, which solves the problem of difficulty in extracting fault features when sensor data signals of idler rollers are aliased and coupled.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: a method for diagnosing idler roller faults based on the characteristics of cascading periodic fusion, comprising the following steps:

[0006] S1. Collect sensor data of the conveyor rollers during operation and preprocess the collected sensor data.

[0007] S2. Based on sliding window sampling, the idler roller sensing data is divided into samples, and the training set, validation set and test set are divided in a ratio of 6:3:1.

[0008] S3. Construct a roller fault diagnosis model based on the Times-SCConv network and initialize the parameters;

[0009] S4. Train and validate the Times-SCConv network based on the training set and validation set. Update and optimize the network parameters using the optimization algorithm. If the network achieves the optimal effect, proceed to step S5; otherwise, proceed to step S3, modify the initial parameters, continue training, and save the optimal weights.

[0010] S5. Input the test set into the optimal Times-SCConv network model from step S4 to perform idler roller fault diagnosis, and finally obtain the idler roller fault classification results and test accuracy.

[0011] Furthermore, the sensor data preprocessing method in step S1 is as follows: interpolation is used to fill in the missing values ​​in the sensor data, and normalization is used to adjust the sensor data to the same scale. The specific calculation formula is as follows:

[0012]

[0013] In the formula, s * The data is normalized, s is the sensor data, max represents the maximum value, and min represents the minimum value.

[0014] Furthermore, the preprocessed idler roller sensing sequence data is segmented using a sliding window. A suitable window size and sliding step size are selected, and the specific calculation formula is as follows:

[0015]

[0016] Wherein, the total signal length is H, the unit sample length is h, the window shift length is m, the number of divisible samples is N, and [·] indicates rounding down to the nearest integer.

[0017] Furthermore, the idler roller fault diagnosis model in step S3 includes a data cycle reshaping module, a spatial reshaping convolution module, a channel reshaping convolution module, a feature adaptive fusion module, a normalization layer, a global pooling layer, and a fully connected layer.

[0018] Furthermore, the period reshaping module calculates the absolute value of the amplitude of the energy intensity of the frequency component in the data sequence through Fast Fourier Transform (FFT) and performs dimensional averaging.

[0019] A = Avg(FFT(X) 1D ));

[0020] In the formula, A is the averaged sequence amplitude vector, Avg averages the sequence amplitude intensity along its dimension, FFT represents the Fast Fourier Transform, and X... 1D This is a sequence of data from idler roller sensor monitoring.

[0021] The k components of amplitude intensity within half a frequency period are selected to obtain their corresponding frequencies. The length of the first k periods is obtained by dividing the sample sequence length by the component frequencies. Based on the period length, the idler roller sensing data sequence X is divided and reshaped into a corresponding two-dimensional tensor. The specific calculation formula is as follows:

[0022]

[0023] In the formula, For the reshaped 2D tensor, Reshape represents the change in data dimension, p i f represents the length of the i-th period. i A represents i The corresponding frequency, Padding indicates that the part that is not long enough to be filled with 0.

[0024] Furthermore, the spatial reshaping module uses group normalization to process tensor features, and the specific calculation formula is as follows:

[0025]

[0026] In the formula, Xout represents the normalized feature, X represents the input feature, GN represents the group normalization operation, μ and σ are the mean and standard deviation of X, ε is a constant to prevent division by zero, γ and β are the scaling factor and offset factor, respectively, and the parameter γ measures the degree of change in feature space information. After normalization processing, the normalized weights of different channels are obtained. The specific calculation formula is as follows:

[0027]

[0028] In the formula, w i This represents the relative importance of the feature maps from different channels, where C is the number of channels;

[0029] The output of the group normalization layer is weighted by normalized weights, and the channel separation weights are obtained by combining the Sigmoid function with threshold gating. The channel separation weights are then used to separate redundant features from the input features. The specific calculation formula is as follows:

[0030] W = Gate(Sigmoid(W) γ (X out )));

[0031]

[0032] In the formula, W is the channel separation weight, Gate is the gate threshold function, Sigmoid is the Sigmoid function, Xw1 is the feature containing more information, and Xw2 is the redundant feature with less information.

[0033] The features are further segmented and cross-fused, and the specific calculation formula is as follows:

[0034]

[0035] In the formula, X w11 For the first part of Xw1, X w12 For the latter part of Xw1, X w21 For the first part of Xw2, X w22 The latter part is Xw2, where Xw is the fusion feature.

[0036] Furthermore, the channel reshaping convolution module separates and channels the features output by the spatial reshaping convolution module, and the specific calculation formula is as follows:

[0037]

[0038] In the formula, Conv 1×1 For convolution, α is the channel compression ratio, and C is the number of channels;

[0039] X1 is used for high-level feature extraction using group convolution and point convolution, while X2 is used for shallow-level feature extraction. The specific calculation formulas are as follows:

[0040] Y1 = GConv 3×3 (X1)+PConv 1×1 (X1),Y2=PConv 1×1 (X2)UX2;

[0041] In the formula, Conv 3×3 For group convolution, PConv 1×1 For point convolution, Y1 and Y2 are the extracted features;

[0042] The extracted features Y1 and Y2 are adaptively fused using global average pooling and Softmax, respectively. The specific calculation formula is as follows:

[0043]

[0044] In the formula, Y represents the fused feature, and Gap represents the global average pooling operation.

[0045] Furthermore, the feature adaptive fusion module performs weighted fusion of local features within and between extraction cycles based on the different intensities of different frequency components to obtain the final aggregated features. Finally, the roller fault diagnosis results are output through normalization, global pooling, and a fully connected layer.

[0046] Furthermore, step S4 optimizes the network parameters by backpropagating the error between the predicted and true sample labels using a loss function, selecting cross-entropy as the loss function:

[0047]

[0048] In the formula, L is the final calculated loss value, N is the total number of samples, C is the total number of categories, and y n,i One-hot encoding for the real label. The probability that the model predicts the nth sample as belonging to the i-th class;

[0049] The AdamW optimizer is used to train and optimize the network parameters during backpropagation.

[0050] Compared with existing technologies, the beneficial effects of this solution are:

[0051] 1. This invention segments and reshapes the mixed multi-cycle frequency components in the original data of the idler roller, extracts local features within and between cycles of different frequency components, and performs adaptive fusion based on the intensity of different component signals to obtain fusion features that characterize data changes, thereby further realizing the effective identification of idler roller faults.

[0052] 2. This invention collects sensor data during the idler roller's operation, uses interpolation to handle missing values, and employs normalization to convert sensor data of different magnitudes to the same scale, accelerating the model training process and improving model performance. Sliding window sampling moves across the original data with a specific step size and window size, focusing on local features of the original long sequence while preserving the temporal order and contextual information of the data. Furthermore, since fault features may repeat in multiple windows, and noise is difficult to maintain stably across different windows due to its randomness, the invention reduces the impact of random noise to a certain extent.

[0053] 3. This invention employs the Times-SCConv network for training, validation, and diagnostic testing based on the processed sample dataset. The Times-SCConv network reshapes periodic data into a two-dimensional tensor based on the frequencies of different components within the data. To address the feature extraction redundancy problem easily caused by convolutional network models, SCConv is used to spatially and channel-wise reshape the features, extracting local features within and between periods of the idler roller data. Furthermore, the extracted local features are adaptively fused according to the different intensities of each frequency component. The extracted fused features are then normalized, globally pooled, and processed through fully connected layers to obtain the final fault diagnosis result. Therefore, this invention can accurately and effectively classify and identify faults occurring during idler roller operation. Attached Figure Description

[0054] Figure 1 This is a Times-SCConv network diagram in a roller fault diagnosis method based on cascading periodic fusion characteristics according to the present invention;

[0055] Figure 2This is a spatially reshaped convolutional network diagram in a roller fault diagnosis method based on aliasing periodic fusion features according to the present invention;

[0056] Figure 3 This is a channel reshaping convolutional network diagram in a roller fault diagnosis method based on aliasing periodic fusion features according to the present invention. Detailed Implementation

[0057] The present invention will be further described in detail below through specific embodiments:

[0058] Example

[0059] like Figures 1 to 3 As shown, a method for diagnosing idler roller faults based on the fusion characteristics of cascading cycles includes the following steps:

[0060] S1. Collect sensor data of the conveyor idlers during operation and preprocess the collected sensor data. The preprocessing method is as follows: interpolation is used to fill in missing values ​​in the sensor data of the raw collected idler data, and normalization is used to adjust the sensor data to the same scale. This extracts more complete features while avoiding differences in feature extraction caused by different units of measurement, thus eliminating the influence of different data units. The specific calculation formula is as follows:

[0061]

[0062] In the formula, s * The data is normalized, s is the sensor data, max represents the maximum value, and min represents the minimum value.

[0063] S2. The preprocessed idler roller sensing data from step S1 is segmented using sliding window sampling. A suitable window size and sliding step are selected. Sliding window sampling effectively captures local patterns or trends in the time series and, to some extent, increases the number of samples used for training. This achieves the goal of retaining complete feature information while reducing sample length. Furthermore, the increased sample size reduces the risk of overfitting during subsequent model training. The training, validation, and test sets are divided in a 6:3:1 ratio. The specific calculation formula is as follows:

[0064]

[0065] Wherein, the total signal length is H, the unit sample length is h, the window shift length is m, the number of divisible samples is N, and [·] indicates rounding down to the nearest integer.

[0066] S3. Construct a roller fault diagnosis model based on the Times-SCConv network and initialize the parameters. The roller fault diagnosis model includes a data periodic reshaping module, a spatial reshaping convolution module, a channel reshaping convolution module, a feature adaptive fusion module, a normalization layer, a global pooling layer, and a fully connected layer.

[0067] The period reshaping module calculates the absolute value of the energy intensity of the frequency component in the data sequence using Fast Fourier Transform (FFT) and performs dimensional averaging to eliminate the influence of different dimensions among multiple variables and highlight common periodic information. The specific calculation method is as follows:

[0068] A = Avg(FFT(X) 1D ));

[0069] In the formula, A is the averaged sequence amplitude vector, Avg averages the sequence amplitude intensity along its dimension, FFT represents the Fast Fourier Transform, and X... 1D This is a sequence of data from idler roller sensor monitoring.

[0070] The k components of amplitude intensity within half a frequency period are selected to obtain their corresponding frequencies. The length of the first k periods is obtained by dividing the sample sequence length by the component frequencies. Based on the period length, the idler roller sensing data sequence X is divided and reshaped into a corresponding two-dimensional tensor. The specific calculation formula is as follows:

[0071]

[0072] In the formula, For the reshaped 2D tensor, Reshape represents the change in data dimension, p i f represents the length of the i-th period. i A represents i The corresponding frequency, Padding indicates that the part that is not long enough to be filled with 0.

[0073] The spatial reshaping module uses group normalization to process tensor features. The specific calculation formula is as follows:

[0074]

[0075] In the formula, Xout represents the normalized feature, X represents the input feature, GN represents the group normalization operation, μ and σ are the mean and standard deviation of X, ε is a constant to prevent division by zero, γ and β are the scaling factor and offset factor, respectively, and the parameter γ measures the degree of change in feature space information. After normalization processing, the normalized weights of different channels are obtained. The specific calculation formula is as follows:

[0076]

[0077] In the formula, wi This represents the relative importance of the feature maps from different channels, where C is the number of channels;

[0078] The output of the group normalization layer is weighted by normalized weights, and the channel separation weights are obtained by combining the Sigmoid function with threshold gating. The channel separation weights are then used to separate redundant features from the input features. The specific calculation formula is as follows:

[0079] W = Gate(Sigmoid(W) γ (X out )));

[0080]

[0081] In the formula, W is the channel separation weight, Gate is the gate threshold function, Sigmoid is the Sigmoid function, Xw1 is the feature containing more information, and Xw2 is the redundant feature with less information.

[0082] The features are further segmented and cross-fused to achieve complementary feature information. The specific calculation formula is as follows:

[0083]

[0084] In the formula, X w11 For the first part of Xw1, X w12 For the latter part of Xw1, X w21 For the first part of Xw2, X w22 The latter part is Xw2, where Xw represents the fused feature. The spatial reshaping convolutional module separates effective and redundant features, performs cross-reconstruction on the separated features, and finally outputs the fused feature, thus suppressing spatial redundancy.

[0085] The channel reshaping convolution module separates and compresses the features output by the spatial reshaping convolution module. The specific calculation formula is as follows:

[0086]

[0087] In the formula, Conv 1×1 For convolution, α is the channel compression ratio, and C is the number of channels;

[0088] X1 uses group convolution and point convolution for high-level feature extraction, while X2 uses point convolution for shallow feature extraction, thus enriching feature diversity. The specific calculation formula is as follows:

[0089] Y1 = GConv 3×3 (X1)+PConv 1×1 (X1),Y2=PConv 1×1 (X2)UX2;

[0090] In the formula, Conv 3×3 For group convolution, PConv 1×1 For point convolution, Y1 and Y2 are the extracted features;

[0091] The extracted features Y1 and Y2 are adaptively fused using global average pooling and Softmax, respectively. The specific calculation formula is as follows:

[0092]

[0093] In the formula, Y represents the fused feature, and Gap represents the global average pooling operation. The channel reshaping convolution module further reduces the channel redundancy caused by the repeated use of standard convolution. The features output by the spatial reshaping convolution module are then processed sequentially through segmentation, feature transformation, and adaptive fusion to suppress channel redundancy.

[0094] The feature adaptive fusion module performs weighted fusion of local features within and between extraction cycles based on the different intensities of different frequency components to obtain the final aggregated features. Finally, the roller fault diagnosis results are output through normalization, global pooling, and a fully connected layer.

[0095] S4. Train and validate the Times-SCConv network using the training and validation sets. Update and optimize the network parameters using an optimization algorithm. If the network reaches its optimal performance, proceed to step S5; otherwise, proceed to step S3, modify the initial parameters, continue training, and save the optimal weights. Backpropagate the error between the predicted and true sample labels using a loss function to optimize the network parameters. Cross-entropy is chosen as the loss function.

[0096]

[0097] In the formula, L is the final calculated loss value, N is the total number of samples, C is the total number of categories, and y n,i One-hot encoding for the real label. The probability that the model predicts the nth sample as belonging to the i-th class;

[0098] Backpropagation uses the AdamW optimizer to train and optimize network parameters. Unlike the Adam optimizer, where weight decay occurs before gradient calculation, AdamW applies weight decay only after gradient calculation. Therefore, using AdamW typically results in faster convergence and better generalization.

[0099] S5. Input the test set into the optimal Times-SCConv network model from step S4 to perform idler roller fault diagnosis, and finally obtain the idler roller fault classification results and test accuracy.

[0100] The above are merely embodiments of the present invention, and common knowledge such as specific structures and / or characteristics in the solutions are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for diagnosing idler roller faults based on the characteristics of cascading cycle fusion, characterized in that, Includes the following steps: S1. Collect sensor data of the conveyor rollers during operation and preprocess the collected sensor data. S2. Based on sliding window sampling, the idler roller sensing data is divided into samples, and the training set, validation set and test set are divided in a ratio of 6:3:

1. S3. Construct a roller fault diagnosis model based on the Times-SCConv network and initialize the parameters; S4. Train and validate the Times-SCConv network based on the training set and validation set. Update and optimize the network parameters using the optimization algorithm. If the network achieves the optimal effect, proceed to step S5; otherwise, proceed to step S3, modify the initial parameters, continue training, and save the optimal weights. S5. Input the test set into the optimal Times-SCConv network model from step S4 to perform idler roller fault diagnosis, and finally obtain the idler roller fault classification results and test accuracy.

2. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 1, characterized in that: The sensor data preprocessing method in step S1 is as follows: interpolation is used to fill in the missing values ​​in the sensor data, and normalization is used to adjust the sensor data to the same scale. The specific calculation formula is as follows: In the formula, s * The data is normalized, s is the sensor data, max represents the maximum value, and min represents the minimum value.

3. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 2, characterized in that: The preprocessed idler roller sensor sequence data is segmented using a sliding window. The appropriate window size and sliding step size are selected, and the specific calculation formula is as follows: Wherein, the total signal length is H, the unit sample length is h, the window shift length is m, the number of divisible samples is N, and [·] indicates rounding down to the nearest integer.

4. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 1, characterized in that: The idler roller fault diagnosis model in step S3 includes a data cycle reshaping module, a spatial reshaping convolution module, a channel reshaping convolution module, a feature adaptive fusion module, a normalization layer, a global pooling layer, and a fully connected layer.

5. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 4, characterized in that: The period reshaping module calculates the absolute value of the amplitude of the energy intensity of the frequency component in the data sequence through Fast Fourier Transform (FFT) and performs dimensional averaging. A=Avg(FFT(X 1D )); In the formula, A is the averaged sequence amplitude vector, Avg averages the sequence amplitude intensity along its dimension, FFT represents the Fast Fourier Transform, and X... 1D This is a sequence of data from idler roller sensor monitoring. The k components of amplitude intensity within half a frequency period are selected to obtain their corresponding frequencies. The length of the first k periods is obtained by dividing the sample sequence length by the component frequencies. Based on the period length, the idler roller sensing data sequence X is divided and reshaped into a corresponding two-dimensional tensor. The specific calculation formula is as follows: In the formula, For the reshaped 2D tensor, Reshape represents the change in data dimension, p i f represents the length of the i-th period. i A represents i The corresponding frequency, Padding indicates that the part that is not long enough to be filled with 0.

6. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 4, characterized in that: The spatial reshaping module uses group normalization to process tensor features, and the specific calculation formula is as follows: In the formula, Xout represents the normalized feature, X represents the input feature, GN represents the group normalization operation, μ and σ are the mean and standard deviation of X, ε is a constant to prevent division by zero, γ and β are the scaling factor and offset factor, respectively, and the parameter γ measures the degree of change in feature space information. After normalization processing, the normalized weights of different channels are obtained. The specific calculation formula is as follows: In the formula, w i This represents the relative importance of the feature maps from different channels, where C is the number of channels; The output of the group normalization layer is weighted by normalized weights, and the channel separation weights are obtained by combining the Sigmoid function with threshold gating. The channel separation weights are then used to separate redundant features from the input features. The specific calculation formula is as follows: W=Gate(Sigmoid(W γ (X out ))); In the formula, W is the channel separation weight, Gate is the gate threshold function, Sigmoid is the Sigmoid function, and X... w1 For features that contain more information, X w2 Redundant features with limited information; The features are further segmented and cross-fused, and the specific calculation formula is as follows: In the formula, X w11 For the first part of Xw1, X w12 For the latter part of Xw1, X w21 For the first part of Xw2, X w22 The latter part is Xw2, where Xw is the fusion feature.

7. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 6, characterized in that: The channel reshaping convolution module separates and compresses the features output by the spatial reshaping convolution module, and the specific calculation formula is as follows: In the formula, Conv 1×1 For convolution, α is the channel compression ratio, and C is the number of channels; X1 is used for high-level feature extraction using group convolution and point convolution, while X2 is used for shallow-level feature extraction. The specific calculation formulas are as follows: Y1=GConv 3×3 (X1)+PConv 1×1 (X1),Y2=PConv 1×1 (X2)U X2; In the formula, Conv 3×3 For group convolution, PConv 1×1 For point convolution, Y1 and Y2 are the extracted features; The extracted features Y1 and Y2 are adaptively fused using global average pooling and Softmax, respectively. The specific calculation formula is as follows: In the formula, Y represents the fused feature, and Gap represents the global average pooling operation.

8. The method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to claim 4, characterized in that: The feature adaptive fusion module performs weighted fusion of local features within and between extraction cycles based on the different intensities of different frequency components to obtain the final aggregated features. Finally, the roller fault diagnosis results are output through normalization, global pooling, and a fully connected layer.

9. A method for diagnosing idler roller faults based on the fusion characteristics of overlapping cycles according to any one of claims 1-8, characterized in that: Step S4 optimizes the network parameters by backpropagating the error between the predicted and true sample labels using a loss function, choosing cross-entropy as the loss function: In the formula, L is the final calculated loss value, N is the total number of samples, C is the total number of categories, and y n,i One-hot encoding for the real label. The probability that the model predicts the nth sample as belonging to the i-th class; The AdamW optimizer is used to train and optimize the network parameters during backpropagation.

Images