Rotating machinery fault migration diagnosis method and system

Abstract

This invention provides a method and system for fault transfer diagnosis of rotating machinery, comprising the following steps: acquiring labeled and unlabeled monitoring datasets of rotating components under different operating conditions; performing the same data preprocessing on both the source and target domain datasets; selecting a deep transfer learning algorithm and constructing a deep neural network model; inputting labeled source domain data to obtain the output predicted label; inputting source and target domain data and calculating the loss between the features from the two domains; calculating the deviation between the predicted and actual results; performing gradient derivative and optimization on the above loss; repeating the above steps iteratively to obtain a trained neural network model; and obtaining the predicted label for the new monitoring data after the data preprocessing in step 2. This invention achieves the integration of adaptive methods from different domains at the algorithm level by constructing an adaptive ensemble framework.

Description

Technical Field

[0001] This invention relates to the field of mechanical condition monitoring and fault diagnosis technology, specifically to a method and system for diagnosing fault migration in rotating machinery, and more particularly to a method and system for diagnosing fault migration in rotating machinery based on dynamic domain adaptive integration. Background Technology

[0002] Rotating machinery, as one of the core pieces of equipment in manufacturing, is widely used. Conducting research on rotating machinery fault diagnosis and ensuring its safe and reliable operation is of great practical significance for improving enterprise production efficiency and safeguarding national economic security. Traditional rotating machinery fault diagnosis methods often rely on time-frequency domain analysis techniques to extract typical fault features. These methods are mostly cumbersome to operate and require a high level of expertise from testing personnel. In the context of today's massive amounts of monitoring data, they are no longer adequate for the demands of rapid, real-time, and efficient diagnostic analysis. Since the advent of the artificial intelligence era, intelligent diagnostic methods based on deep learning and other AI technologies have been widely applied. These intelligent diagnostic methods, with their powerful feature extraction and fitting capabilities, continue to make new research progress in the field of fault diagnosis.

[0003] While these deep intelligent diagnostic methods offer many advantages, two factors limit their application in complex real-world scenarios. First, these deep learning models require training and testing data to follow the same data distribution, but in practical applications, the collected monitoring data often covers different operating conditions and even different mechanical equipment. Furthermore, monitoring signal data lacks fault labels, and manual labeling is extremely expensive, resulting in a severe shortage of labeled fault data. With the continuous advancement of related technologies, transfer learning promises to alleviate the data acquisition requirements, offering a possible solution to the aforementioned challenges. The goal of transfer learning is to find a way to combine a source domain rich in information with a target domain lacking information, utilizing knowledge learned from the source domain to apply to relevant tasks in the target domain.

[0004] Patent document CN113076834B discloses a method, system, terminal, and medium for processing fault information of rotating machinery. It constructs a neural network model including a deep feature extractor, a domain classifier, and a state predictor. The neural network model uses the deep feature extractor to automatically extract migration fault features from laboratory simulation data and monitoring data of rotating components in actual engineering equipment. The domain classifier is used to reduce the difference between the two data distributions. The state predictor, with the introduction of domain adaptation constraints, forms a fault diagnosis model based on a deep domain adaptive adversarial network, and this model is used to achieve intelligent fault diagnosis of rotating machinery. However, this patent document still has the drawback of conforming to the same data distribution and lacking fault labels on the monitoring signal data. Summary of the Invention

[0005] In view of the deficiencies in the prior art, the purpose of this invention is to provide a method for diagnosing fault migration in rotating machinery.

[0006] A method for diagnosing fault migration in rotating machinery according to the present invention includes the following steps:

[0007] Step 1: Obtain labeled and unlabeled monitoring datasets of rotating parts under different operating conditions to form source domain dataset and target domain dataset;

[0008] Step 2: Perform the same data preprocessing on the source domain dataset and the target domain dataset to obtain two sample sets that can be recognized by the ensemble network algorithm;

[0009] Step 3: Select different deep transfer learning algorithms and construct deep neural network models according to the ensemble transfer learning framework;

[0010] Step 4: Input labeled source domain data, obtain the output predicted label through forward propagation in the constructed network, and compare the predicted label with the true label to obtain the loss function;

[0011] Step 5: Input source domain data and target domain data, and use the selected deep transfer learning algorithm to calculate the loss between the features from the two domain data;

[0012] Step 6: Calculate the deviation between the prediction results and the actual results of different transfer learning methods, and use this as a basis to calculate the corresponding optimization adaptive factor of different deep transfer learning methods in the ensemble framework;

[0013] Step 7: Calculate and optimize the gradient of the above loss;

[0014] Step 8: Repeat steps 4 to 7 iteratively until the convergence condition is met, then stop training and obtain the trained neural network model.

[0015] Step 9: After the new monitoring data has undergone data preprocessing in Step 2, it is input into the trained deep neural network model to obtain the predicted label of the data sample.

[0016] Preferably, in step 1, the source domain dataset is represented as The target domain dataset is represented as follows: in, These are the monitoring data of the i-th sample in the source domain dataset and its corresponding health status label, respectively. Let be the i-th sample in the target domain dataset, and n be the number of training samples in the minimum batch.

[0017] Preferably, in step 6, the deviation between the predicted results and the actual results of different migration methods is calculated using A-distance.

[0018] Preferably, in step 7, the gradient derivative of the above loss is calculated and optimized using the backpropagation algorithm.

[0019] Preferably, step 2 specifically includes the following steps:

[0020] Step 2.1: Remove outliers from the original vibration signal of the rotating component using Laida's method, as shown in the following formula:

[0021]

[0022] in, x represents the average value of a signal segment. i Let be the value of the i-th sample point in the signal sample; Where n is the standard deviation and n is the total number of sampling points in the signal segment;

[0023] Step 2.2: Slice the time-series vibration signal after outlier removal. Each slice contains 4096 vibration signal sampling points. Perform continuous wavelet transform to obtain wavelet time-frequency graph samples, which are used as input to the network model. The continuous wavelet transform formula is as follows:

[0024]

[0025] Where Φ(t) is the wavelet mother function, τ is the time shift coefficient, a is the scaling coefficient, and a≠0.

[0026] Preferably, in step 3, the deep neural network model is a neural network DDAENN based on DDAE;

[0027] The DDAE-based neural network DDAENN includes a feature extractor G. f Domain Discriminator G d and label predictor G l ;

[0028] The DDAE-based neural network DDANN employs domain adversarial techniques and MMD-based domain adaptive techniques.

[0029] Preferably, in step 4, the classification accuracy of the deep neural network for the labeled data in the source domain is trained using cross-entropy loss, with the following loss function:

[0030]

[0031] Among them, W y and b y It is the label predictor Gl The matrix-vector pair obtained after linear transformation, f represents the matrix-vector pair obtained by the feature extractor G. f The extracted features are: F(·) denotes the softmax function, I{·} denotes the indicator function, and k is the i-th sample in the source domain. Corresponding real sample labels;

[0032] In step 5, the loss function for domain adversarial techniques is expressed as:

[0033]

[0034] Where φ represents the sigmoid function, W d and b d It is a linear transformation matrix-vector pair used in the domain classifier, d i f represents the domain label of the i-th training sample. s and f t These represent the features extracted by feature extractor G. f Extracted source and target domain features, where m is the total number of samples in a training batch;

[0035] For the domain adaptation technique based on MMD, its loss function is expressed as:

[0036]

[0037] Where, n s and n t Let represent the number of training samples in the batch from the source domain and the target domain, respectively. and This represents the neighborhood-invariant features of the two domains output by the deep feature extractor. It is a regenerated kernel Hilbert space with k characteristic kernels.

[0038] Preferably, step 6 specifically includes the following steps:

[0039] Step 6.1: Calculate the A-distance after using different migration techniques. The calculation formula is as follows:

[0040]

[0041]

[0042] Where, x j Let G represent the j-th sample. c (x j ) is the classifier G c For sample x j The output, This indicates an indicator function, where n′ represents two different domain datasets D.s and D t The size of the sample, err i (G c ) represents the data from two different domains in the i-th domain adaptation method in classifier G. c Error on; To use the i-th domain adaptation method for the next two domains D s and D t A-distance between;

[0043] Step 6.2: For the i-th domain adaptation method used, calculate its corresponding optimization adaptation factor α. i The calculation formula is as follows:

[0044]

[0045]

[0046] in, It is the A-distance parameter of the i-th neighborhood adaptation method. Sum of all A-distance parameters The ratio is used to represent the weight of the migration technique. The weight assigned to the adaptive method for the j-th domain;

[0047] Step 6.3: Calculate the obtained optimization adaptive factor α i The loss function of the domain adaptation method selected in step 5 Combining these, we obtain the total loss function of the transfer method, expressed as:

[0048]

[0049] Preferably, step 9 specifically includes the following steps:

[0050] Step 9.1: Collect online monitoring data of the actual rotating machinery through sensors, and normalize the monitoring data into suitable wavelet time-frequency diagram samples after data preprocessing in Step 2;

[0051] Step 9.2: Input the obtained time-frequency map samples into the trained neural network model obtained in Step 8 in sequence to obtain the predicted labels of the corresponding samples, and evaluate the current operating status of the device based on the prediction results of the labels.

[0052] The present invention also provides a rotating machinery fault migration diagnosis system, comprising the following steps:

[0053] Module M1: Acquire labeled and unlabeled monitoring datasets of rotating parts under different operating conditions to form source domain dataset and target domain dataset, where and are the monitoring data of the i-th sample in the source domain dataset and its corresponding health status label, respectively, is the i-th sample in the target domain dataset, and n is the number of training samples in the minimum batch.

[0054] Module M2: Performs the same data preprocessing on the source domain dataset and the target domain dataset to obtain two sample sets that can be recognized by the ensemble network algorithm;

[0055] Module M3: Select different deep transfer learning algorithms and construct deep neural network models according to the ensemble transfer learning framework;

[0056] Module M4: Input labeled source domain data, obtain the output predicted label through forward propagation in the constructed network, and obtain the loss function by comparing the predicted label with the true label;

[0057] Module M5: Input source domain data and target domain data, and use the selected deep transfer learning algorithm to calculate the loss between the features from the two domain data;

[0058] Module M6: Calculates the deviation between the prediction results and the actual results of different transfer learning methods using A-distance, and calculates the corresponding optimization adaptive factors of different deep transfer learning methods in the ensemble framework based on this.

[0059] Module M7: Utilizes the backpropagation algorithm to perform gradient differentiation and optimization on the above loss;

[0060] Module M8: Repeatedly trigger modules M4 to M7 to execute iterations until the convergence condition is met, then stop training and obtain the trained neural network model;

[0061] Module M9: After the new monitoring data has been preprocessed by Module 2, it is input into the trained deep neural network model to obtain the predicted label of the data sample.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. This invention integrates the ideas of transfer learning and ensemble learning, and achieves the integration of adaptive methods from different fields at the algorithm level by constructing an adaptive ensemble framework;

[0064] 2. This invention integrates adaptive methods from different fields at the algorithm level by introducing an optimal adaptation factor α. Based on the proposed ensemble learning framework, a feasible DDAE-based neural network (DDAENN) model and a DDAENN-based fault diagnosis process are designed.

[0065] 3. The DDAENN of the present invention includes three specific modules: a feature extractor, a domain classifier, and a label adaptive predictor. It employs two domain adaptive techniques, namely domain adversarial techniques and distribution adaptive techniques based on the maximum mean difference, to achieve integrated diagnosis of the two domain adaptive techniques. Attached Figure Description

[0066] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0067] Figure 1 This is a flowchart of the steps in the rotating machinery fault migration diagnosis method of the present invention;

[0068] Figure 2 This is a structural diagram of the DDAE frame of the present invention;

[0069] Figure 3 This is a schematic diagram of the process of the present invention;

[0070] Figure 4 A schematic diagram comparing the original vibration signals under different working conditions;

[0071] Figure 5 T-SNE distribution plot of output features;

[0072] Figure 6 This is a schematic diagram of the confusion matrix of the model's prediction results. Detailed Implementation

[0073] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0074] Example 1:

[0075] like Figures 1-6 As shown, this embodiment provides a method for diagnosing fault migration in rotating machinery, including the following steps:

[0076] Step 1: Obtain labeled and unlabeled monitoring datasets of rotating components under different operating conditions, forming a source domain dataset and a target domain dataset; the source domain dataset is represented as... The target domain dataset is represented as follows: in, These are the monitoring data of the i-th sample in the source domain dataset and its corresponding health status label, respectively. Let be the i-th sample in the target domain dataset, and n be the number of training samples in the minimum batch.

[0077] Step 2: Perform the same data preprocessing on both the source domain dataset and the target domain dataset to obtain two sample sets that can be recognized by the ensemble network algorithm; Step 2 specifically includes the following steps:

[0078] Step 2.1: Remove outliers from the original vibration signal of the rotating component using Laida's method, as shown in the following formula:

[0079]

[0080] in, x represents the average value of a signal segment. i Let be the value of the i-th sample point in the signal sample; Where n is the standard deviation and n is the total number of sampling points in the signal segment;

[0081] Step 2.2: Slice the time-series vibration signal after outlier removal. Each slice contains 4096 vibration signal sampling points. Perform continuous wavelet transform to obtain wavelet time-frequency graph samples, which are used as input to the network model. The continuous wavelet transform formula is as follows:

[0082]

[0083] Where Φ(t) is the wavelet mother function, τ is the time shift coefficient, a is the scaling coefficient, and a≠0.

[0084] Step 3: Select different deep transfer learning algorithms and construct a deep neural network model according to the ensemble transfer learning framework; the deep neural network model is a neural network DDAENN based on DDAE.

[0085] The DDAE-based neural network DDAENN includes a feature extractor G. f Domain Discriminator G d and label predictor G l ;

[0086] The DDAE-based neural network DDANN employs domain adversarial techniques and MMD-based domain adaptive techniques.

[0087] Step 4: Input labeled source domain data, obtain the predicted output labels through forward propagation in the constructed network, compare the predicted labels with the true labels to obtain the loss function; train the deep neural network to improve the classification accuracy of the labeled source domain data, using cross-entropy loss, the loss function is as follows:

[0088]

[0089] Among them, W y and b y It is the label predictor G l The matrix-vector pair obtained after linear transformation, f represents the matrix-vector pair obtained by the feature extractor G. f The extracted features are: F(·) denotes the softmax function, I{·} denotes the indicator function, and k is the i-th sample in the source domain. The corresponding real sample labels.

[0090] Step 5: Input source domain data and target domain data, and calculate the loss between the features from the two domain data using the selected deep transfer learning algorithm; for domain adversarial techniques, the loss function is expressed as:

[0091]

[0092] Where φ represents the sigmoid function, W d and b d It is a linear transformation matrix-vector pair used in the domain classifier, d i f represents the domain label of the i-th training sample. s and f t These represent the features extracted by feature extractor G. f Extracted source and target domain features, where m is the total number of samples in a training batch;

[0093] For the domain adaptation technique based on MMD, its loss function is expressed as:

[0094]

[0095] Where, n s and n t Let represent the number of training samples in the batch from the source domain and the target domain, respectively. and This represents the neighborhood-invariant features of the two domains output by the deep feature extractor. It is a regenerated kernel Hilbert space with k characteristic kernels.

[0096] Step 6: Calculate the deviation between the predicted results and the actual results of different transfer learning methods, and based on this, calculate the corresponding optimization adaptation factor of different deep transfer learning methods in the ensemble framework; use A-distance to calculate the deviation between the predicted results and the actual results of different transfer learning methods; Step 6 specifically includes the following steps:

[0097] Step 6.1: Calculate the A-distance after using different migration techniques. The calculation formula is as follows:

[0098]

[0099]

[0100] Where, x j Let G represent the j-th sample. c (x j ) is the classifier G c For sample x j The output, This indicates an indicator function, where n′ represents two different domain datasets D. s and D t The size of the sample, err i (G c ) represents the data from two different domains in the i-th domain adaptation method in classifier G. c Error on; To use the i-th domain adaptation method for the next two domains D s and D t A-distance between;

[0101] Step 6.2: For the i-th domain adaptation method used, calculate its corresponding optimization adaptation factor α. i The calculation formula is as follows:

[0102]

[0103]

[0104] in, It is the A-distance parameter of the i-th neighborhood adaptation method. Sum of all A-distance parameters The ratio is used to represent the weight of the migration technique. The weight assigned to the adaptive method for the j-th domain;

[0105] Step 6.3: Calculate the obtained optimization adaptive factor α i The loss function of the domain adaptation method selected in step 5 Combining these, we obtain the total loss function of the transfer method, expressed as:

[0106]

[0107] Here That is, the one that appears in step 5 and

[0108] Step 7: Calculate and optimize the gradient of the above loss; use the backpropagation algorithm to calculate and optimize the gradient of the above loss.

[0109] Step 8: Repeat steps 4 to 7 iteratively until the convergence condition is met, then stop training and obtain the trained neural network model.

[0110] Step 9: After the new monitoring data has undergone data preprocessing in Step 2, it is input into the trained deep neural network model to obtain the predicted label of the data sample; Step 9 specifically includes the following steps:

[0111] Step 9.1: Collect online monitoring data of the actual rotating machinery through sensors, and normalize the monitoring data into suitable wavelet time-frequency diagram samples after data preprocessing in Step 2;

[0112] Step 9.2: Input the obtained time-frequency map samples into the trained neural network model obtained in Step 8 in sequence to obtain the predicted labels of the corresponding samples, and evaluate the current operating status of the device based on the prediction results of the labels.

[0113] Example 2:

[0114] This embodiment provides a rotating machinery fault migration diagnosis system, including the following modules:

[0115] Module M1: Acquire labeled and unlabeled monitoring datasets of rotating parts under different operating conditions to form source domain dataset and target domain dataset, where and are the monitoring data of the i-th sample in the source domain dataset and its corresponding health status label, respectively, is the i-th sample in the target domain dataset, and n is the number of training samples in the minimum batch.

[0116] Module M2: Performs the same data preprocessing on the source domain dataset and the target domain dataset to obtain two sample sets that can be recognized by the ensemble network algorithm;

[0117] Module M3: Select different deep transfer learning algorithms and construct deep neural network models according to the ensemble transfer learning framework;

[0118] Module M4: Input labeled source domain data, obtain the output predicted label through forward propagation in the constructed network, and obtain the loss function by comparing the predicted label with the true label;

[0119] Module M5: Input source domain data and target domain data, and use the selected deep transfer learning algorithm to calculate the loss between the features from the two domain data;

[0120] Module M6: Calculates the deviation between the prediction results and the actual results of different transfer learning methods using A-distance, and calculates the corresponding optimization adaptive factors of different deep transfer learning methods in the ensemble framework based on this.

[0121] Module M7: Utilizes the backpropagation algorithm to perform gradient differentiation and optimization on the above loss;

[0122] Module M8: Repeat modules 4 to 7 iteratively until the convergence condition is met, then stop training and obtain the trained neural network model;

[0123] Module M9: After the new monitoring data has been preprocessed by Module 2, it is input into the trained deep neural network model to obtain the predicted label of the data sample.

[0124] Example 3:

[0125] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0126] This embodiment provides a rotating machinery fault migration diagnosis method based on dynamic domain adaptive integration, belonging to the field of mechanical condition monitoring and fault diagnosis technology, and more specifically, relating to a deep migration integration diagnosis method for rotating machinery component faults.

[0127] To address the existing technical difficulties and challenges, this embodiment provides a rotating machinery fault transfer diagnosis method based on dynamic domain adaptive integration. This method integrates the ideas of transfer learning and ensemble learning, and achieves the integration of adaptive methods from different domains at the algorithm level by constructing an adaptive integration framework.

[0128] The technical solution of this embodiment is: a method for diagnosing fault migration in rotating machinery based on dynamic domain adaptive integration, comprising the following steps:

[0129] Step 1: Obtain labeled monitoring datasets of rotating components under different operating conditions. and unlabeled monitoring dataset The dataset consists of a source domain and a target domain, where These are the monitoring data of the i-th sample in the source domain dataset and its corresponding health status label, respectively. Let be the i-th sample in the target domain dataset, and n be the number of training samples in the minimum batch.

[0130] Step 2: The two datasets obtained will undergo the same data preprocessing to obtain two sample sets that can be recognized by the ensemble network algorithm;

[0131] Step 3: Select different deep transfer algorithms and construct deep neural network models according to the ensemble transfer framework;

[0132] Step 4: Input labeled source domain data, obtain the output predicted label through forward propagation in the constructed network, and compare the predicted label with the true label to obtain the loss function;

[0133] Step 5: Input source domain data and target domain data, and use the selected transfer learning algorithm to calculate the loss between the features from the two domain data.

[0134] Step 6: Calculate the deviation between the prediction results and the actual results of different transfer methods using A-distance, and calculate the corresponding optimization adaptive factor of different transfer methods in the integration framework based on this.

[0135] Step 7: Use the backpropagation algorithm to calculate the gradient derivative of the above loss and optimize it;

[0136] Step 8: Repeat steps 4, 5, 6, and 7 iteratively until the convergence condition is met, then stop training. The trained neural network model is obtained.

[0137] Step 9: After the new monitoring data has undergone data preprocessing in Step 2, input it into the trained neural network model to obtain the predicted label of the data sample.

[0138] Because rotating machinery monitoring data often lacks sufficient fault data and related state labels, it is difficult to train intelligent diagnostic models with good generalization and robustness. This embodiment proposes a novel transfer learning ensemble framework (Dynamic Domain Adaptive Ensemble, DDAE). By introducing an optimal adaptation factor α, it achieves the integration of different domain adaptive methods (DAMs) at the algorithm level. Based on the proposed ensemble learning framework, a feasible DDAE-based neural network (DDAENN) model and a fault diagnosis process based on DDAENN are designed. DDAENN includes three specific modules: a feature extractor, a domain classifier, and a label adaptive predictor. Two domain adaptive techniques are employed: domain adversarial technique and distribution adaptive technique based on maximum mean difference, achieving integrated diagnosis using these two domain adaptive techniques. Case studies demonstrate that this method has high experimental accuracy, and the experimental results meet expectations.

[0139] Example 4:

[0140] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0141] This embodiment provides a method for diagnosing fault migration in rotating machinery based on dynamic domain adaptive integration, including the following steps:

[0142] Step 1: As Figure 3As shown, labeled monitoring datasets of rotating components under different operating conditions were obtained. and unlabeled monitoring dataset The dataset consists of a source domain and a target domain, where These are the monitoring data of the i-th sample in the source domain dataset and its corresponding health status label, respectively. Let be the i-th sample in the target domain dataset, and n be the number of training samples in the minimum batch.

[0143] Step 2: As Figure 3 As shown, the two datasets obtained will undergo the same data preprocessing to obtain two sample sets that can be recognized by the ensemble network algorithm;

[0144] Furthermore, this step includes:

[0145] Step 2-1: Remove outliers from the original vibration signal of the rotating component. Here, Laida's method is used to remove outliers, as shown in the following formula:

[0146]

[0147] in, x represents the average value of a signal segment. i Let be the value of the i-th sample point in the signal sample. Where n is the standard deviation and n is the total number of sampling points in the signal segment;

[0148] Step 2-2: The time-series vibration signal after outlier removal will be sliced, with each slice containing 4096 vibration signal sampling points. Then, a continuous wavelet transform will be performed to obtain wavelet time-frequency graph samples, which will be used as input to the network model. The continuous wavelet transform formula is as follows:

[0149]

[0150] Where Φ(t) is the wavelet mother function, τ is the time shift coefficient, a is the scaling coefficient, and a≠0.

[0151] Step 3: Select different deep transfer algorithms, according to... Figure 2 The integrated transfer framework shown constructs a deep neural network model; as... Figure 3 As shown, we constructed a neural network DDAENN based on DDAE, which consists of three modules: a feature extractor G... f Domain Discriminator G d and label predictor G l The network employs two commonly used domain adaptation methods: domain adversarial techniques and MMD-based domain adaptation techniques. The model uses a CNN backbone network to learn useful feature representations and performs domain adaptation training within the DDAE framework.

[0152] Detailed parameter configurations for the network structure will be provided in specific examples;

[0153] Step 4: Input labeled source domain data, and obtain the predicted output labels through forward propagation in the constructed network. Compare the predicted labels with the true labels to obtain the loss function. This step is mainly used to train the network's classification accuracy for labeled source domain data, using cross-entropy loss. The loss function is as follows:

[0154]

[0155] Among them, W y and b y It is the label predictor G l The matrix-vector pair obtained after linear transformation, f represents the matrix-vector pair obtained by the feature extractor G. f The extracted features are: F(·) denotes the softmax function, I{·} denotes the indicator function, and k is the i-th sample in the source domain. Corresponding real sample labels;

[0156] Step 5: Input source domain data and target domain data, and use the selected transfer learning algorithm to calculate the loss between the features of the data from the two domains;

[0157] For domain adversarial techniques, the loss function can be expressed as:

[0158]

[0159] Where φ represents the sigmoid function, W d and b d It is a linear transformation matrix-vector pair used in the domain classifier, d i f represents the domain label of the i-th training sample. s and f t These represent the features extracted by feature extractor G. f Extracted source and target domain features, where m is the total number of samples in a training batch;

[0160] For domain adaptation techniques based on MMD, the loss function can be expressed as:

[0161]

[0162] Where n s and n t Let represent the number of training samples in the batch from the source domain and the target domain, respectively. and This represents the neighborhood-invariant features of the two domains output by the deep feature extractor, while It is a regenerated kernel Hilbert space with k feature kernels, and the test accuracy of the method can be improved by using different kernel functions for the feature kernels k;

[0163] Step 6: Calculate the deviation between the prediction results and the actual results of different transfer methods using A-distance, and calculate the corresponding optimization adaptive factor of different transfer methods in the integration framework based on this.

[0164] Furthermore, this step includes:

[0165] Step 6-1: Calculate the A-distance after using different migration techniques. The calculation formula is as follows:

[0166]

[0167]

[0168] Where, x j Let G represent the j-th sample. c (x j ) is the classifier G c For sample x j The output, This indicates an indicator function, where n′ represents two different domain datasets D. s and D t The size of the sample, err i (G c ) represents the data from two different domains in the i-th domain adaptation method in classifier G. c Error on That is, after using the i-th domain adaptation method, the two domains D s and D t A-distance between;

[0169] Step 6.2: For different transfer techniques, calculate the corresponding optimization adaptation factor α. i The calculation formula is as follows:

[0170]

[0171]

[0172] in, It is the A-distance parameter of the i-th neighborhood adaptation method. Sum of all A-distance parameters The ratio is used to represent the weight of the migration technique. The weight assigned to the adaptive method for the j-th domain;

[0173] Step 6.3: Calculate the obtained optimization adaptive factor α i The loss function of the domain adaptation method in step 5 Combining these, we obtain the total loss function of the transfer method, expressed as:

[0174]

[0175] Step 7: Use the backpropagation algorithm to calculate the gradient derivative of the above loss and optimize it; the total loss function can be expressed as:

[0176]

[0177] Where λ is the regularization parameter, L y It is a classification loss function that labels training samples in a training batch. The process of backpropagation and differentiation can be expressed as:

[0178]

[0179] Where Θ c This represents the network parameters that need to be optimized in the model;

[0180] Step 8: Repeat steps 4, 5, 6, and 7 until the convergence condition is met, the total loss function decreases, and training stops. The trained neural network model is obtained.

[0181] Step 9: The trained network model can be directly used for online diagnostics of rotating machinery. Further, this step includes:

[0182] Step 9-1: Collect online monitoring data of the actual rotating machinery through sensors, and normalize the monitoring data into suitable wavelet time-frequency diagram samples after data preprocessing in Step 2;

[0183] Step 9-2: Input the obtained time-frequency map samples sequentially into the trained neural network model obtained in Step 8 to obtain the predicted labels for the corresponding samples. The current operating status of the device can be evaluated based on the predicted label results.

[0184] Example 5:

[0185] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1, Embodiment 2, and Embodiment 4.

[0186] In this embodiment, to address the problems existing in the fault diagnosis of rotating equipment, a rotating machinery fault migration diagnosis method based on dynamic domain adaptive integration is proposed. By migrating and training labeled experimental data under a single operating condition load to unlabeled gearbox data under multiple operating conditions loads, intelligent fault diagnosis of gearbox under different operating conditions and loads is finally realized.

[0187] The dataset used in this implementation case is derived from laboratory-simulated gearbox fault data. The gearbox test bench mainly consists of an AC motor, flywheel, variable speed gearbox, and a load brake applying the load. The maximum speed of the test bench is 3600 rpm, and the maximum output torque is 50 N. Vibration signals are acquired using a three-channel accelerometer. A total of four health states were simulated on the test bench, including a healthy mode and three fault modes. Faults included pitting, wear, and fracture. All faults were artificially induced. Six operating conditions were tested, including variable speed and variable load. Operating conditions (OC) 1 to 4 were constant speed loads, and OC 5 to 6 were variable speed loads. The sampling frequency was 10000 Hz. Data acquisition in each OC lasted for 30 seconds. For variable speed OCs, data was collected as the speed gradually increased from 1200 rpm to 2400 rpm over 30 seconds. Approximately 15 seconds of signal data was collected during the speed increase phase. The steady-state vibration signals for the four health states are shown below. Figure 4 As shown.

[0188] The method of this embodiment for fault diagnosis of the target domain includes the following steps:

[0189] Step 1: Perform data preprocessing on the original vibration signal. First, cut each data point into segments using a fixed-size sliding time window. After outlier removal, use formula (2) to obtain wavelet time-frequency map samples. Select four health status data under working condition 1 as the source domain and the other five working condition data as the target domain to establish source and target domain datasets.

[0190] Step 2: According to Figure 3 As shown, a neural network DDAENN based on DDAE is constructed. As mentioned earlier, this network consists of three modules, namely the feature extractor G. f Domain Discriminator G d and label predictor G l The specific configurations of each network structure are shown in Tables 1, 2, and 3, respectively:

[0191] Table 1 Feature Extractor G f

[0192] number of floors name operate Specific parameters 1 enter Input Sample 64×64×3 2 Convolutional layer 1 convolution / 3 Batch standardization layer 1 Batch standardization / 4 Pooling layer 1 Pooling 4 5 Activation layer 1 ReLU function activation 0.2 6 Convolutional layer 2 convolution / 7 Batch standardization layer 2 Batch standardization / 8 Pooling layer 2 Pooling 4 9 Activation layer 2 ReLU function activation 0.2 10 Convolutional layer 3 convolution / 11 Batch standardization layer 3 Batch standardization / 12 Pooling layer 3 Pooling 16 13 Activation layer 3 ReLU function activation 0.2 14 Fully connected layer 3 Tensor flattening /

[0193] Table 2 Tag Predictor G l

[0194] number of floors name operate Specific parameters 1 Fully connected layer 1 Linear transformation 100 2 Batch standardization layer 1 Batch standardization / 3 Activation layer 1 ReLU function activation 0.2 4 Dropout 1 Dropout 0.3 5 Fully connected layer 2 Linear transformation 100 6 Batch standardization layer 2 Batch standardization / 7 Activation layer 2 ReLU function activation 0.2 8 Fully connected layer 3 Classification Total number of categories

[0195] Table 3 Domain Discriminator G d

[0196] number of floors name operate Specific parameters 1 Fully connected layer 1 Linear transformation 100 2 Batch standardization layer 1 Batch standardization / 3 Activation layer 1 ReLU function activation 0.2 4 Fully connected layer 2 Linear transformation 100 5 Batch standardization layer 2 Batch standardization / 6 Activation layer 2 ReLU function activation 0.2 7 Fully connected layer 3 Classification 2

[0197] Step 3: Train the network model on the training sets in the source and target domains using the aforementioned training method. During training, the source domain training set consists of wavelet time-frequency image samples and their corresponding labels, while the target domain training set only contains wavelet time-frequency image samples without labels. The resulting trained network model is then obtained. The example was implemented on a Win10 system using Anaconda, Python 3.6, and PyTorch.

[0198] Step 4: After data preprocessing, the test signal is input into the trained network model and passes through the feature extractor G once via forward propagation. f and label predictor G l This yields the predicted equipment operating status output.

[0199] Figure 5 The original input data are displayed respectively. Figure 5 (a) and the output features of the trained model ( Figure 5 (b) T-SNE distribution results. The results show that in the original data, only one type of sample (pitting fault) can be separated, while other healthy state samples are mixed together and difficult to distinguish. However, the output data features of these samples processed by the trained DDAENN model can be well mapped one-to-one and form four different clusters. Each cluster consists of functions with the same health state label in the source and target domains. The results show that the proposed DDAENN-based fault diagnosis method can successfully extract common similar features from the source and target domain samples and map data features with the same type of health state together.

[0200] Figure 6 The confusion matrix of the prediction results is shown. The left figure shows the specific prediction results for 1200 test samples, and the right figure shows the normalized results. The results show that the method used has a prediction accuracy of up to 99.75% for unlabeled gearbox monitoring data. Among the total 1200 monitoring data samples, only 3 wear fault samples were predicted incorrectly. This indicates that the method can successfully achieve fault diagnosis of gearbox under varying operating loads under the condition of lacking fault labels. The model trained by this method has high diagnostic accuracy and strong generalization performance.

[0201] This invention achieves the integration of adaptive methods from different fields at the algorithm level by constructing an adaptive integration framework.

[0202] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0203] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for diagnosing fault migration in rotating machinery, characterized in that, Includes the following steps: Step 1: Obtain labeled and unlabeled monitoring datasets of rotating parts under different operating conditions to form source domain dataset and target domain dataset; Step 2: Perform the same data preprocessing on the source domain dataset and the target domain dataset to obtain two sample sets that can be recognized by the ensemble network algorithm; Step 3: Select different deep transfer learning algorithms and construct deep neural network models according to the ensemble transfer learning framework; Step 4: Input labeled source domain data, obtain the output predicted label through forward propagation in the constructed network, and compare the predicted label with the true label to obtain the loss function; Step 5: Input source domain data and target domain data, and use the selected deep transfer learning algorithm to calculate the loss between the features from the two domain data; Step 6: Calculate the deviation between the prediction results and the actual results of different transfer learning methods, and use this as a basis to calculate the corresponding optimization adaptive factor of different deep transfer learning methods in the ensemble framework; Step 7: Calculate the gradient derivative of the above loss and optimize it; Step 8: Repeat steps 4 to 7 iteratively until the convergence condition is met, then stop training and obtain the trained neural network model; Step 9: After the new monitoring data has undergone data preprocessing in Step 2, it is input into the trained deep neural network model to obtain the predicted label of the data sample; Step 6 specifically includes the following steps: Step 6.1: Calculate the A-distance after using different migration techniques. The calculation formula is as follows: in, Indicates the first j One sample, It is a classifier For the sample The output, Indicates an indicator function, For two different domain datasets and The size of the sample. Indicates the first i In domain-adaptive methods, data from two different domains are used in the classifier. Error on; For use of the i The two domains are adaptive methods. and A-distance between; Step 6.2: For the first... i For each domain-specific adaptive method, calculate its corresponding optimization adaptive factor. The calculation formula is as follows: in, It is the first i A-distance parameter of domain-specific adaptive methods Sum of all A-distance parameters The ratio is used to represent the weight of the migration technique. For the first j The weight of adaptive methods in each domain; Step 6.3: Obtain the optimized adaptive factor The loss function of the domain adaptation method selected in step 5 Combining these, we obtain the total loss function of the transfer method, expressed as: 。 2. The method for diagnosing fault migration in rotating machinery according to claim 1, characterized in that, In step 1, the source domain dataset is represented as follows: The target domain dataset is represented as ,in, , These are the first two numbers in the source domain dataset. i Monitoring data for each sample and its corresponding health status markers, For the first in the target domain dataset i One sample, n This represents the minimum number of training samples in a batch.

3. The method for diagnosing fault migration in rotating machinery according to claim 1, characterized in that, In step 6, the deviation between the predicted results and the actual results of different migration methods is calculated using A-distance.

4. The method for diagnosing fault migration in rotating machinery according to claim 3, characterized in that, In step 7, the gradient derivative of the above loss is calculated and optimized using the backpropagation algorithm.

5. The method for diagnosing fault migration in rotating machinery according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Remove outliers from the original vibration signal of the rotating component using Laida's method, as shown in the following formula: in, , representing the average value of a signal segment. For the signal sample, the first i The values ​​of each sample point; , is the standard deviation. n It is the total number of sampling points in the signal segment; Step 2.2: Slice the time-series vibration signal after outlier removal. Each slice contains 4096 vibration signal sampling points. Perform continuous wavelet transform to obtain wavelet time-frequency graph samples, which are used as input to the network model. The continuous wavelet transform formula is as follows: in, It is the wavelet mother function. It is the time shift coefficient. It is a scaling factor, and .

6. The method for diagnosing fault migration in rotating machinery according to claim 5, characterized in that, In step 3, the deep neural network model is a neural network DDAENN based on DDAE. The DDAE-based neural network DDAENN includes a feature extractor. G f Domain Discriminator G d and label predictor G l ; The DDAE-based neural network DDANN employs domain adversarial techniques and MMD-based domain adaptive techniques.

7. The method for diagnosing fault migration in rotating machinery according to claim 6, characterized in that, In step 4, the classification accuracy of the deep neural network for the source domain labeled data is trained using cross-entropy loss, and the loss function is as follows: in, and It is a label predictor G l The matrix-vector pair obtained after linear transformation f Indicated by feature extractor G f Extracted features This represents the softmax function. Indicates an indicator function, k For the source domain i Sample Corresponding real sample labels; In step 5, the loss function for domain adversarial techniques is expressed as: in, This refers to the sigmod function. and These are linear transformation matrix-vector pairs used in domain classifiers. The domain label representing the i-th training sample. and These represent the features extracted by the feature extractor. G f Extracted source and target domain features, m The total number of samples in a training batch; For the domain adaptation technique based on MMD, its loss function is expressed as: in, and Let represent the number of training samples in the batch from the source domain and the target domain, respectively. and This represents the neighborhood-invariant features of the two domains output by the deep feature extractor. It is a k The regenerated kernel Hilbert space of 1 characteristic kernel.

8. The method for diagnosing fault migration in rotating machinery according to claim 1, characterized in that, Step 9 specifically includes the following steps: Step 9.1: Collect online monitoring data of the actual rotating machinery through sensors, and normalize the monitoring data into suitable wavelet time-frequency diagram samples after data preprocessing in Step 2; Step 9.2: Input the obtained time-frequency map samples into the trained neural network model obtained in Step 8 in sequence to obtain the predicted labels of the corresponding samples, and evaluate the current operating status of the device based on the prediction results of the labels.

9. A fault migration diagnosis system for rotating machinery, characterized in that, The rotating machinery fault migration diagnosis method according to any one of claims 1 to 8 includes the following modules: Module M1: Acquires labeled and unlabeled monitoring datasets of rotating components under different operating conditions, forming source domain datasets and target domain datasets, where and are the in the source domain dataset, respectively. i The monitoring data of the sample and its corresponding health status label are the first sample in the target domain dataset. i One sample, n This represents the minimum number of training samples in a batch. Module M2: Performs the same data preprocessing on the source domain dataset and the target domain dataset to obtain two sample sets that can be recognized by the ensemble network algorithm; Module M3: Select different deep transfer learning algorithms and construct deep neural network models according to the ensemble transfer learning framework; Module M4: Input labeled source domain data, obtain the output predicted label through forward propagation in the constructed network, and obtain the loss function by comparing the predicted label with the true label; Module M5: Input source domain data and target domain data, and use the selected deep transfer learning algorithm to calculate the loss between the features from the two domain data; Module M6: Calculates the deviation between the prediction results and the actual results of different transfer learning methods using A-distance, and calculates the corresponding optimization adaptive factors of different deep transfer learning methods in the ensemble framework based on this. Module M7: Utilizes the backpropagation algorithm to perform gradient differentiation and optimization on the above loss; Module M8: Repeatedly trigger modules M4~M7 to execute iterations until the convergence condition is met, then stop training and obtain the trained neural network model; Module M9: After the new monitoring data has been preprocessed by Module 2, it is input into the trained deep neural network model to obtain the predicted label of the data sample.

Images