Multi-information fusion based adaptive transfer fault diagnosis method for rotating machinery

Abstract

The application discloses a kind of multi-element information fusion's rotating machinery adaptive migration fault diagnosis method, to solve the problems of low classification precision and poor generalization ability of prior art, its implementation scheme is: obtaining old working condition multi-sensor dataset and new working condition multi-sensor dataset;Multi-element information interaction fusion network is constructed;Old working condition multi-sensor dataset and new working condition multi-sensor dataset are input into multi-element information interaction fusion network, the network is trained, to minimize its total loss value, update network parameters, until the number of iterations reaches the maximum set value, obtain the trained multi-element information interaction fusion network;New multi-sensor data collected under new working condition are input into the trained fusion network, and the rotating machinery fault diagnosis result is obtained.The application can adaptively adjust the contribution of different characteristics, improve the fault diagnosis precision under the condition that the distribution difference of multi-sensor data is larger, and can be used for intelligent detection of rotating machinery under variable working conditions.

Description

Technical Field

[0001] This invention belongs to the field of mechanical technology, and further relates to an adaptive migration fault diagnosis method for rotating machinery, which can be used for fault detection of rotating machinery. Background Technology

[0002] Rotating machinery is widely used in mechanical engineering due to its versatility. However, these machines operate in harsh environments for extended periods, causing wear and tear on components such as gears and bearings over time. Accurate and automated fault diagnosis of rotating machinery is crucial for timely detection and resolution of equipment problems, preventing further economic losses and personnel casualties. In complex mechanical systems, relying on a single sensor signal can lead to the loss of critical information. Therefore, the comprehensive use of multi-sensor data for effective fault diagnosis and health monitoring is essential. However, in practical applications, the operating conditions of rotating machinery are complex and variable. As the distribution of multi-sensor data changes, the performance of fault diagnosis deteriorates.

[0003] To address the problem of multi-sensor collaborative fault diagnosis under different operating conditions, scholars have drawn on ideas such as multi-sensor information fusion and transfer learning to solve this problem. The main idea is to fuse multi-sensor data and utilize transfer learning to align the distribution of multi-sensor data under different operating conditions, thereby achieving fault diagnosis across operating conditions.

[0004] In their paper "A novel fusion diagnosis method for rotor system fault based on deep learning and multi-sourced heterogeneous monitoring data" (Measurement Science and Technology, 2018, doi:10.1088 / 1361-6501 / aadfb3), Zhuang Yuan et al. proposed a rotor system fault diagnosis method based on deep learning and multi-sourced heterogeneous monitoring data fusion. The method's implementation steps are as follows: First, multi-source heterogeneous monitoring data of the rotor system is collected; second, a multi-mode convolutional neural network is constructed to automatically learn fault-sensitive features from the raw multi-sensor data composed of vibration signals and infrared images; then, t-distributed random neighbor embedding is introduced to fuse deep features, further improving the quality of the learned features; finally, the fused features are used for fault classification. While this method can diagnose rotor system faults, it fails to consider the distribution alignment of multi-sensor data under new operating conditions, thus hindering effective fault diagnosis under these conditions. In other words, the diagnostic accuracy is poor when dealing with cross-condition scenarios.

[0005] In patent application number 202211104454.4, Shanchen Pang et al. proposed a transferable multi-scale rotating machinery fault diagnosis method. The method first acquires a training dataset; second, it constructs a multi-scale feature extraction network to extract and fuse multi-scale features from the input samples, obtaining fused features; then, it constructs a fully connected neural network classifier to classify the input fused features, obtaining fault category prediction results; finally, it constructs a deconvolution-based feature alignment network to identify the domain labels of the fault category prediction results, distinguishing whether the fault category prediction results originate from the target domain or the source domain. While this method can diagnose faults in rotating machinery systems, it does not consider the contribution of different features to the diagnostic results, cannot adaptively adjust the contribution of different features according to changes in operating conditions, and only reduces the domain distribution differences under new operating conditions, without reducing the feature distribution differences. Therefore, when multi-sensor data is more complex and the differences between the source and target domains are greater, the diagnostic accuracy is low. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by proposing a multi-source information fusion adaptive migration fault diagnosis method for rotating machinery, which adaptively adjusts the contribution of different features to improve fault diagnosis accuracy when multi-sensor data is more complex and the source and target domains have greater distribution differences.

[0007] The technical approach to achieving the objective of this invention is as follows: by utilizing feature interaction and attention mechanisms to fuse multi-sensor data, the contribution of different features can be adaptively adjusted; by utilizing transfer learning methods to reduce the distribution differences between new and old operating conditions, the degree of fusion of multi-sensor data and the accuracy of fault diagnosis under cross-operating conditions can be improved.

[0008] Based on the above ideas, the technical solution of the present invention includes the following steps:

[0009] (1) Obtain the old working condition multi-sensor dataset D S and the new working condition multi-sensor dataset D T ;

[0010] (2) Constructing a multi-dimensional information interaction and integration network:

[0011] (2a) Construct feature extractors G for each of the multiple sensors. i Extract the corresponding sensor features h i ;

[0012] (2b) Using the Kronecker product to measure h i Interactive features k are obtained through feature interaction. i ;

[0013] (2c) Calculate interaction features k using an attention mechanism i The feature weight values, and the interaction feature k i Multiplying it by its feature weight value yields the multi-sensor fusion feature f. i To complete the construction of a multi-dimensional information interaction and integration network;

[0014] (3) Training a multi-dimensional information interaction and fusion network:

[0015] (3a) Set the model iteration count Epoch, and the current model iteration count e = 1;

[0016] (3b) Transfer the old working condition multi-sensor dataset D S The old working condition fusion feature f is obtained from the input multi-source information interaction and fusion network. i S The new working condition multi-sensor dataset D T The new working condition fusion feature f is obtained by inputting it into a multi-source information interaction and fusion network. i T ;

[0017] (3c) Integrate old operating conditions with features f i S The input is fed into the existing classifier C to obtain the classification result F for the old working condition. i S Calculate the fusion characteristics f of the old working conditions respectively. i S Integration features with new operating conditions f i T The covariance matrix C S and C T ;

[0018] (3d) Based on the covariance matrix C S C T The old working condition fusion feature f is calculated using the square root correlation alignment formula. i S Integration features with new operating conditions f i T Domain adaptation loss l SC ;

[0019] (3e) Based on the classification results of the old working conditions F i S Calculate the classification loss of the old working condition based on the label of the old working condition. cls And based on the domain adaptation loss of the fusion characteristics of the new and old working conditions l SC Classification of losses under old operating conditions cls Calculate the total loss of the multi-source information interaction and fusion network l total Utilizing domain adaptation loss SC And classification loss lcls Update network parameters;

[0020] (3f) Determine whether e = Epoch is true.

[0021] If so, obtain the trained multi-source information interaction and fusion network, and proceed to step (4).

[0022] Otherwise, let e = e + 1 and return to (3b);

[0023] (4) Input the new multi-sensor data collected under the new working conditions into the trained multi-source information interaction and fusion network to obtain the fault diagnosis results of rotating machinery.

[0024] Compared with the prior art, the present invention has the following advantages:

[0025] First, when this invention integrates features from multiple sensors, it utilizes a feature interaction method to perform fine-grained interaction of features, which makes the network able to extract richer features. Furthermore, by determining the contribution of different features through an attention mechanism and adaptively adjusting the contribution of different sensors according to changes in operating conditions, the network can extract more useful features.

[0026] Second, in the fault diagnosis task under new operating conditions, this invention adopts a fusion feature difference metric and reduces the distribution difference of fusion features by reducing domain adaptation loss and classification loss, so that the extracted features have stronger generalization ability and improve the diagnostic ability of the fault diagnosis model under new operating conditions.

[0027] Experimental results show that, compared with existing transferable multi-scale rotating machinery fault diagnosis methods, the present invention has higher fault diagnosis accuracy in multi-sensor cross-condition collaborative fault diagnosis tasks. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0029] Figure 2 This is a diagram of the experimental platform for acquiring multi-sensor datasets in this invention;

[0030] Figure 3 This is a time-domain signal diagram of data collected by different sensors in this invention;

[0031] Figure 4 This is a schematic diagram of the multi-source information interaction and fusion network structure constructed in this invention;

[0032] Figure 5 This is a diagram showing the results of fault diagnosis based on multi-sensor data collected under new operating conditions according to the present invention. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0034] Reference Figure 1 The implementation steps for this example are as follows:

[0035] Step 1, Obtain the old working condition multi-sensor dataset D S and the new working condition multi-sensor dataset D T And migration tasks.

[0036] The new and old operating condition datasets for this example come from the Paderborn University Bearing Dataset. This bearing dataset consists of vibration and current sensors, featuring high resolution and a high sampling rate. It collects operational data for 26 damaged bearing states and 6 undamaged (healthy) bearing states under four operating conditions. It also measures data such as load force, load torque, speed, and temperature under operating conditions.

[0037] The Paderborn University bearing dataset is from, for example... Figure 2 The test platform shown consists of several modules, from left to right: motor, torque measurement shaft, rolling bearing test module, flywheel, and load motor. Experimental data is generated by installing ball bearings with different damage types in the bearing test module. All tested bearings are 6203 rolling bearings. Failed bearings can be categorized as either man-made damage or actual damage. Man-made damage, such as cracks, spalling, and pitting, is caused by electrical discharge machining (EDM), drilling, or electro-engraving. Actual bearing damage is obtained on the bearing accelerated life test bench. The motor current sampling frequency is 64 kHz. The vibration signal sampling frequency is 64 kHz. The mechanical parameters, namely the loading force, loading torque, and speed, are all sampled at a frequency of 4 kHz. The temperature sampling frequency is 1 Hz.

[0038] In this example, only multi-sensor data from 10 bearings operating under 4 different conditions were selected. The multi-sensor data from any one of these conditions was chosen as the old operating condition multi-sensor dataset D. S Select the multi-sensor data from another working condition as the new working condition multi-sensor dataset D. T .

[0039] Specific parameters are shown in Table 1 and Table 2.

[0040] Table 1 Four operating conditions

[0041]

[0042] Table 2 Different Fault Types

[0043]

[0044] The Paderborn University bearing dataset contains frequency domain signals of vibration sensor data, force sensor data, and load sensor data. Time domain plots of the different sensor data are shown below. Figure 3 As shown, where Figure 3 The Vibration plot in the middle represents the time-domain data from the vibration sensor. Figure 3 The Force field represents the time-domain plot of the force sensor data. Figure 3 Torque is a time-domain plot of the load sensor data.

[0045] This example demonstrates 12 multi-sensor fusion transfer tasks based on domain adaptation and cross-condition multi-sensor fusion across four operating conditions. The number of samples from different sensor datasets is equal for both the old and new operating conditions. The dimensions of the vibration, force, and torque data are 2000, 400, and 400, respectively, and the dimensions of the input networks via Fast Fourier Transform are 1000, 200, and 200, respectively. Twelve transfer tasks are established: A→B, A→C, A→D, B→A, B→C, B→D, C→A, C→B, C→D, D→A, D→B, D→C, D→C, D→A, D→B, D→C, D→C, D→C, D→A, D→B, D→C, D→C, D→C.

[0046] Step 2: Construct a multi-source information fusion network.

[0047] 2.1) Establish a feature extractor corresponding to multi-sensor data:

[0048] This example uses three types of sensors: vibration sensor data corresponds to feature extractor 1, force sensor data corresponds to feature extractor 2, and torque sensor data corresponds to feature extractor 3. Each feature extractor is composed of BatchNorm1d layer, Linear layer, BatchNorm1d layer, ReLU layer, Linear layer, BatchNorm1d layer, ReLU layer, and Linear layer connected sequentially.

[0049] 2.2) Set up an attention mechanism layer;

[0050] 2.3) Set the classifier C

[0051] 2.4) Set the parameters for each feature extractor, attention mechanism layer, and classifier C, as shown in Table 3;

[0052] 2.5) Place the three feature extractors in parallel, and then cascade them with the attention mechanism layer and classifier C respectively. This completes the construction of the multi-source information interaction and fusion network.

[0053] Table 3 Network Structure Settings

[0054]

[0055] Step 3: Train the multi-source information interaction and fusion network.

[0056] Reference Figure 4 The implementation of this step includes the following:

[0057] 3.1) Set the network iteration count Epoch to 50, the learning rate to 0.001, the batch size to 50, the penalty coefficient λ to 1, and the current model iteration count e to 1.

[0058] 3.2) Transfer the tagged multi-sensor data from the old operating conditions to D S And new working conditions without tag-free multi-sensor data D T Input the multi-source information fusion network and extract the corresponding sensor features h1, h2, h3 through the feature extractor;

[0059] 3.3) The Kronecker product is used to perform feature interaction on the three sensor features h1, h2, and h3 to obtain several interactive features k. 11 k 12 k 13 k 22 k 23 k 33 , with k 11 For example:

[0060]

[0061] Among them, h 1,i Let i represent the i-th feature of the first sensor, where i ranges from 1 to D, and D is the total number of features of the first sensor.

[0062] 3.4) Calculate the weight of each interaction feature using an attention mechanism:

[0063]

[0064] in, Let PWConv2 represent a 2D 1×1 dotted convolution, δ represent the ReLU activation function, PWConv1 represent a 1D 1×1 dotted convolution, and GAP represent global average pooling operation. ij ) represents feature k ij The weight value.

[0065] 3.5) Multiply the interaction feature by its weight value to obtain the multi-sensor fusion feature f. i The formula is as follows:

[0066]

[0067] in, Representing feature k ij Multiply by the weight value;

[0068] 3.6) Calculate the multi-sensor fusion features f separately. i Includes old working condition fusion features f i S Integration features with new operating conditions f i T The covariance matrix C S and C T :

[0069]

[0070]

[0071] Where, n S and n T f represents the total number of training samples in the source and target domains, respectively. i S and f i T These represent the old working condition fusion features and the new working condition fusion features, respectively. 1 indicates a column vector with all elements being 1, and the superscript T indicates transpose.

[0072] 3.7) Based on the covariance matrix C S C T The old working condition fusion feature f is calculated using the square root correlation alignment formula. i S Integration features with new operating conditions f i T Domain adaptation loss l SC :

[0073]

[0074] in Represents the F-norm, C S and C T f represents the fusion feature of the old working condition, respectively. i S Integration features with new operating conditions f i T The covariance matrix.

[0075] 3.8) Integrate features from old operating conditions f i S The input is fed into classifier C to obtain the classification result of the old working condition fusion feature, and the old working condition fusion feature f is calculated by using the classification result and the label of the old working condition fusion feature. i S Classification loss lcls :

[0076]

[0077] Where n S F represents the number of samples for the old operating condition. j Let y be the j-th probability vector output by the softmax layer. i Let l represent the label of the i-th sample. cls This represents the cross-entropy loss between the predicted label and the actual label.

[0078] 3.9) Classification loss l cls Domain Adaptation Loss SC Calculate the total loss l total The formula is as follows:

[0079] l total =l cls +λl SC

[0080] Where λ is the domain adaptation penalty term;

[0081] 3.7) The Adam optimization method is adopted, and all parameters are updated using domain adaptation loss, classification loss, and backpropagation algorithm, as shown in the following formula:

[0082]

[0083] Where α is the learning rate, and θ is the old parameters of the model. This is the updated result of θ;

[0084] 3.8) Determine if e = Epoch is true:

[0085] If so, a well-trained multi-dimensional information interaction and fusion network is obtained, and step (4) is executed;

[0086] Otherwise, let e = e + 1 and return to step (3.2).

[0087] Step 4) Obtain the fault diagnosis results of rotating machinery.

[0088] New multi-sensor data collected under new operating conditions are input into a trained multi-source information interaction and fusion network to obtain fault diagnosis results, such as... Figure 5 As shown in the figure, the horizontal axis represents the fault category corresponding to the predicted label, the vertical axis represents the fault category corresponding to the true label, and the numbers represent the accuracy of the classification results.

[0089] from Figure 5It can be seen that the diagnostic accuracy rate for the second type of fault is 0.96, the diagnostic accuracy rate for the third type of fault is 0.99, and the diagnostic accuracy rate for the other eight fault types is 1.

[0090] The numbers used in the above steps are only for the purpose of clearly describing the embodiments of the present invention, and their order is not limited.

[0091] The technical effects of the present invention will be described in detail below with reference to specific experiments:

[0092] I. Experimental conditions and contents:

[0093] On a system with an Intel(R) Core(TM) i5-10400 CPU @ 2.90GHz, RAM = 16.0GB, and Windows 10 operating system, the fault diagnosis results were tested using Python 3.7 software.

[0094] 2. Experimental content and results:

[0095] The fault categories in Table 2 were classified using the present invention and six comparative methods, and the classification accuracy (Acc) for each method was calculated.

[0096]

[0097] In the formula, Let y be the predicted label for the j-th target domain test sample. j This represents the actual label of the j-th target domain test sample.

[0098] The accuracy calculations yielded a comparison of the fault diagnosis accuracy of this invention and six comparative methods, as shown in Table 5.

[0099] Table 5 Fault diagnosis results of different methods

[0100]

[0101] As can be seen from Table 5, the classification accuracy of the multi-source information fusion method for diagnosing rotating machinery migration faults proposed in this invention is approximately 99%, which is significantly higher than the other six diagnostic methods.

[0102] In summary, this invention enables fault diagnosis under various operating conditions, thereby improving the accuracy of fault diagnosis.

[0103] The above description is merely a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, those skilled in the art, after understanding the content and principles of the present invention, may make various modifications and changes in form and details without departing from the principles and structure of the present invention. However, these modifications and changes based on the ideas of the present invention are still within the scope of protection of the claims of the present invention.

Claims

1. A multi-source information fusion-based adaptive migration fault diagnosis method for rotating machinery, characterized in that, Includes the following steps: (1) Obtain the old working condition multi-sensor dataset D S and the new working condition multi-sensor dataset D T ; (2) Constructing a multi-dimensional information interaction and integration network: (2a) Construct feature extractors G for each of the multiple sensors. i Extract the corresponding sensor features h i ; (2b) Using the Kronecker product to measure h i Interactive features k are obtained through feature interaction. i ; (2c) Calculate interaction features k using an attention mechanism i The feature weight values, and the interaction feature k i Multiplying it by its feature weight value yields the multi-sensor fusion feature f. i To complete the construction of a multi-dimensional information interaction and integration network; (3) Training a multi-dimensional information interaction and fusion network: (3a) Set the model iteration count Epoch, and the current model iteration count e = 1; (3b) Transfer the old working condition multi-sensor dataset D S The old working condition fusion feature f is obtained from the input multi-source information interaction and fusion network. i S The new working condition multi-sensor dataset D T The new working condition fusion feature f is obtained by inputting it into a multi-source information interaction and fusion network. i T ; (3c) Integrate features from old operating conditions i S The input is fed into the existing classifier C to obtain the classification result F for the old working condition. i S Calculate the fusion characteristics f of the old working conditions respectively. i S Integration features with new operating conditions f i T The covariance matrix C S and C T ; (3d) Based on the covariance matrix C S C T The old working condition fusion feature f is calculated using the square root correlation alignment formula. i S Integration features with new operating conditions f i T Domain adaptation loss (3e) Based on the classification results of the old working conditions F i S Calculate the classification loss of the old working condition based on the label of the old working condition. And based on the domain adaptation loss of the fusion characteristics of the new and old operating conditions Classification of losses under old operating conditions Calculate the total loss of a multi-source information interaction and fusion network. Utilizing domain adaptation loss and classification loss Update network parameters; (3f) Determine whether e = Epoch is true. If so, obtain the trained multi-source information interaction and fusion network, and proceed to step (4). Otherwise, let e = e + 1 and return to (3b); (4) Input the new multi-sensor data collected under the new working conditions into the trained multi-source information interaction and fusion network to obtain the fault diagnosis results of rotating machinery.

2. The method according to claim 1, characterized in that, In step (2a), feature extractors G corresponding to multiple sensors are constructed respectively. i Each feature extractor is formed by sequentially connecting BatchNorm1d layer, Linear layer, BatchNorm1d layer, ReLU layer, Linear layer, BatchNorm1d layer, ReLU layer, and Linear layer. The number of channels in each layer is set according to the feature dimension of the input data from the corresponding sensor.

3. The method according to claim 1, characterized in that, Step (2b) uses the Kronecker product to convert h i Interactive features k are obtained through feature interaction. i The formula is as follows: Among them, h i,i Let i represent the i-th feature of the i-th sensor, where i ranges from 1 to D, and D is the total number of features of the i-th sensor.

4. The method according to claim 1, characterized in that, In step (2c), the interaction feature k is calculated using an attention mechanism. i The feature weights are calculated using the following formula: in, Let PWConv2 represent a 2D 1×1 dotted convolution, δ represent the ReLU activation function, PWConv1 represent a 1D 1×1 dotted convolution, and GAP represent global average pooling operation. i ) represents feature k i The weight value.

5. The method according to claim 1, characterized in that, In step (3c), the old working condition fusion feature f is calculated respectively. i S Integration features with new operating conditions f i T The covariance matrix C S and C T The formula is as follows: Where, n S and n T f represents the total number of training samples in the source and target domains, respectively. i S and f i T These represent the fusion features of the old operating conditions and the fusion features of the new operating conditions, respectively. 1 indicates a column vector with all elements being 1, and the superscript T indicates transpose.

6. The method according to claim 1, characterized in that, In step (3d), the old working condition fusion feature f is calculated. i S Integration features with new operating conditions f i T Domain adaptation loss The formula is as follows: in Represents the F-norm, C S and C T f represents the fusion feature of the old working condition, respectively. i S Integration features with new operating conditions f i T The covariance matrix.

7. The method according to claim 1, characterized in that, In step (3e), based on the classification results F of the old working conditions i S Calculate the classification loss of the old working condition based on the label of the old working condition. The formula is as follows: Where n S F represents the number of samples for the old operating condition. j Let y be the j-th probability vector output by the softmax layer. i Let represent the label of the i-th sample. This represents the cross-entropy loss between the predicted label and the actual label.

8. The method according to claim 1, characterized in that, In step (3e), the domain adaptation loss is calculated based on the fusion characteristics of the new and old operating conditions. Classification of losses under old operating conditions Calculate the total loss of a multi-source information interaction and fusion network. The formula is as follows: Where λ is the domain adaptation penalty term.

9. The method according to claim 1, characterized in that, In step (3e), the domain adaptation loss is utilized. and classification loss The formula for updating network parameters is as follows: Where α is the learning rate, and θ is the old parameters of the model. This is the updated result of θ.

Images