Multi-source Hybrid Transfer Diagnosis Method for Mechanical Faults Based on the AICDA Model

Abstract

The present invention relates to a mechanical fault multi-source hybrid transfer diagnosis method based on the AICDA model, belonging to the technical field of mechanical fault diagnosis. The method mainly includes constructing the AICDA model, using a one-dimensional convolutional neural network as the backbone network, including four "ConV1D" convolutional blocks, two FC convolutional layers, four DIA layers, four class-level DIA layers and an AdaSoftmax classification loss; then inputting the divided training samples into the constructed AICDA model, and iteratively optimizing and training the AICDA model by using the overall optimization objective function composed of the pseudo-label self-supervised learning function of the target domain samples and the true label learning function of multiple source domain samples. The present invention improves the accuracy and generalization ability of transfer diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault diagnosis, and relates to a mechanical fault multi-source hybrid transfer diagnosis method based on an Adaptive Intermediate Class-wise Distribution Alignment (AICDA) model. Background Art

[0002] At present, many transfer learning methods have been proposed and applied to mechanical fault intelligent diagnosis. Their loss functions generally consist of a healthy state classification loss, a distribution distance loss, and a related regularization loss. However, the trade-off parameters and internal hyperparameters among the three need to be randomly searched through personalized grids according to different diagnostic tasks. Therefore, the generalization ability of these methods in most diagnostic tasks is low. On the other hand, since the distribution alignment objectives of the vast majority of transfer diagnosis methods are in real-time dynamic changes, this will lead to problems such as large fluctuations in the optimization loss, slow convergence, and poor robustness. They only consider the diagnostic prior knowledge of a single source domain and also only serve a specific diagnostic task between the target domain and the source domain. However, in practical engineering applications, it is very difficult for those skilled in the art to predict in advance the types of healthy states of target machinery.

[0003] Therefore, there is an urgent need for a simple and general transfer diagnosis method for multi-source mechanical faults. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a mechanical fault multi-source hybrid transfer diagnosis method based on an Adaptive Intermediate Class-wise Distribution Alignment (AICDA) model, which is simple and applicable to multi-source mechanical fault transfer diagnosis.

[0005] To achieve the above object, the present invention provides the following technical solution:

[0006] A mechanical fault multi-source hybrid transfer diagnosis method based on the AICDA model. First, in the AICDA, a Dynamic Intermediate Alignment (DIA) layer is designed to adaptively reduce the distribution difference between the source domain and the target domain, and at the same time it does not require a distribution distance loss and a related regularization loss. To ensure the classification effect of the AICDA on the healthy state of the target mechanical equipment, an Adaptive Softmax (AdaSoftmax) is constructed to increase the separability, and this loss also has no internal hyperparameters. The method specifically includes the following steps:

[0007] S1: Install the acceleration sensor on the mechanical equipment to collect the original vibration signals under different working conditions. Then, use the sliding sampling technique to expand the fault sample data set for the collected original vibration signals, and divide the training sample set and the test sample set.

[0008] S2: Construct the AICDA model, which uses a one-dimensional convolutional neural network as the backbone network, including four "ConV1D" convolutional blocks, two FC convolutional layers, four DIA layers, four class-level dynamic intermediate alignment (Dynamicintermediate alignment, DIA) layers, and an AdaSoftmax classification loss.

[0009] S3: Input the divided training samples into the constructed AICDA model, and perform iterative optimization training on the AICDA model using the overall optimization objective function composed of the pseudo-label self-supervised learning function of the target domain samples and the true label learning functions of multiple source domain samples.

[0010] S4: After multiple iterative trainings, observe whether the loss curve tends to converge stably. If the curve converges, the model training is completed, and the trained AICDA model will be used for mechanical multi-source hybrid transfer fault diagnosis under different working conditions.

[0011] Further, in step S2, the expression of the dynamic intermediate alignment is:

[0012]

[0013] Among them, γ and β are training parameters that can be automatically updated by gradient backpropagation.

[0014]

[0015] Among them, X S and X T respectively represent the source domain feature and the target domain feature, and correspondingly follow the normal distribution and and respectively represent the distribution parameters of the source domain and the target domain; Z S and Z T respectively represent the source domain feature and the target domain feature after distribution alignment.

[0016] Further, in step S2, the expression of the AdaSoftmax classification loss L y is:

[0017]

[0018] Among them, F i (c) represents the feature of the i-th sample output by the last fully connected layer of the network Fi The element value corresponding to the label index position, F i (j) represents the other elements except the element at the label index position, m represents the number of samples; k is the decision margin.

[0019] Furthermore, in step S2, the expression of the decision margin k is:

[0020] k = e (Homogeneity+Completeness+V-measure) -1

[0021] where Homogeneity, Completeness, and V-measure represent homogeneity, completeness, and V-measure respectively;

[0022] Homogeneity = 1 - H(C|K) / H(C)

[0023] Completeness = 1 - H(K|C) / H(K)

[0024]

[0025] where H(C|K) and H(K|C) represent the conditional entropy given the cluster and class respectively, and H(C) and H(K) represent the information entropy given the class and cluster respectively.

[0026] Furthermore, in step S3, optimizing the training of the AICDA model specifically includes: overall optimizing the objective function L through the constructed AdaSoftmax classification loss and the target domain pseudo-labels W which is expressed as:

[0027]

[0028] where m T represents the number of samples in the target domain, represents the number of samples in the nth source domain, and respectively represent the feature vector and pseudo-label of the ith sample in the target domain, and respectively represent the feature vector and true label of the ith sample in the nth source domain, L T (·) and respectively represent the cross-entropy loss acting on the target domain samples and source domain samples;

[0029] Using the above formula, the trainable parameters of the global sub-model Θ G and the class-level sub-model Θ C are updated to:

[0030]

[0031]

[0032] Among them, ε is a hyperparameter representing the learning rate.

[0033] The beneficial effects of the present invention are as follows: The present invention improves the accuracy and generalization ability of mechanical fault migration diagnosis, which is specifically reflected in the following aspects:

[0034] (1) The present invention adopts a dynamic intermediate alignment (DIA) layer to adaptively reduce the distribution difference between the source domain and the target domain, avoiding problems such as loss oscillation, slow convergence, and weak robustness during the distribution alignment process, avoiding the occurrence of negative transfer effects, and at the same time, it does not require distribution distance loss and related regularization loss.

[0035] (2) The AdaSoftmax classification loss adopted by the present invention can directly optimize the network parameters through gradient backpropagation without other techniques and additional computational complexity, that is, while ensuring the classification effect of AICDA on the health state of the target mechanical equipment, the computational complexity of the overall network model is reduced.

[0036] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0038] Figure 1 is the structural diagram of the adaptive intermediate class-level distribution alignment (AICDA) model;

[0039] Figure 2 is the working principle diagram of the dynamic intermediate alignment (DIA) layer;

[0040] Figure 3 is the schematic diagram of the DDS test bench. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The following specific examples illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0042] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, rather than physical diagrams, and should not be construed as limiting the present invention; in order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0043] In the drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, it is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0044] Please refer to Figures 1 to 3 , the embodiments of the present invention provide a multi-source hybrid migration diagnosis method for planetary gearboxes based on the Adaptive Intermediate Class-Level Distribution Alignment (AICDA) model. The specific process is as follows:

[0045] 1. Install acceleration sensors on mechanical equipment to collect the original vibration signals under different working conditions, so as to construct subsequent migration diagnosis tasks. Then, the collected original vibration signals are used to expand the fault sample data set by using the sliding sampling technique, and the training sample set and the test sample set are divided.

[0046] 2. Build the backbone network structure of the adaptive class-level distribution alignment model through a one-dimensional convolutional neural network; then, combine the proposed AdaSoftmax classification loss and the DIA layer to construct the Adaptive Intermediate Class-Level Distribution Alignment (AICDA) model.

[0047] As Figure 1As shown in the figure, the AICDA model uses a one-dimensional convolutional neural network as the backbone network to adaptively extract health status features. The model is mainly divided into two parts: a global sub-model and a class-level sub-model. The global sub-model provides pseudo-labels for the target domain samples of the subsequent class-level sub-model. The specific network structure of this model includes five "ConV1D" convolutional blocks, two FC convolutional layers, five DIA layers, five class-level DIA layers, and an AdaSoftmax classification loss. Among them, each "ConV1D" convolutional block consists of a convolutional layer, a non-linear activation layer, and a max pooling layer. The DIA layer, AdaSoftmax classification loss, and the optimization process of AICDA in the adaptive intermediate class-level distribution alignment model will be introduced in detail below.

[0048] 1) Dynamic intermediate alignment layer (DIA)

[0049] The working principle of the DIA layer is as Figure 2 shown. Assume that the source domain feature X S and the target domain feature X T follow normal distributions and respectively. Then these two normal distributions can be normalized to a specific intermediate distribution N(0,1) through the following formula.

[0050]

[0051] Among them, and represent the distribution parameters (mean and variance) of the source domain and the target domain respectively.

[0052] Even through Equation (1), the distribution difference between the source domain and the target domain can be eliminated. However, the intermediate distribution is preferably adaptively learned from the common attributes between the source domain and the target domain. It can be seen from Figure 2 that the fixed intermediate distribution used in Equation (1) may be far from the feature distributions of the source domain and the target domain. If forced to perform normalized distribution alignment, it will inevitably lead to problems such as loss oscillation, slow convergence, weak robustness, and even the appearance of negative transfer effects. Therefore, an adaptive intermediate distribution alignment is designed to solve the above problems.

[0053]

[0054] Among them, γ and β are training parameters that can be automatically updated by gradient backpropagation. Through Equation (2), the dynamic intermediate alignment layer will automatically learn and adapt an intermediate distribution suitable for the source domain and the target domain under the prior knowledge of the normal distribution, thereby effectively accelerating network convergence and enhancing model robustness.

[0055] 2) AdaSoftmax Classification Loss

[0056] The Adaptive Intermediate-Class-Level Distribution Alignment Model is dedicated to learning domain-invariant and separable health state features. The DIA layer only ensures the learning of domain-invariant features. Therefore, a classification loss with a strict decision boundary and capable of adaptive learning is required to enhance the learning of separable features. Based on the above analysis, the present invention constructs the AdaSoftmax classification loss, whose decision margin can be adaptively learned according to the changes in the attributes of sample features during the model training process, thus avoiding the waste of computing resources caused by grid random search. The AdaSoftmax classification loss L y is shown as follows:

[0057]

[0058] where, F i (c) represents the element value corresponding to the label index position of the output of the i-th sample feature F i of the last fully-connected layer of the network, and F i (j) represents the other elements except the element at the label index position. m represents the number of samples; k is the decision margin, and its expression is as follows:

[0059] k = e (Homogeneity+Completeness+V-measure) -1 (4)

[0060] where, Homogeneity, Completeness, and V-measure represent homogeneity, completeness, and V-measure respectively.

[0061] Homogeneity = 1 - H(C|K) / H(C) (5)

[0062] Completeness = 1 - H(K|C) / H(K) (6)

[0063]

[0064] where, H(C|K) and H(K|C) represent the conditional entropy given the cluster and class respectively, and H(C) and H(K) represent the information entropy given the class and cluster respectively.

[0065] Assuming that P represents the vector output by AdaSoftmax and Y is the corresponding one-hot encoded vector, the gradient of the AdaSoftmax classification loss can be calculated by the following formula:

[0066]

[0067] where, Θ represents the learnable parameters of the network model, then it can be expressed as:

[0068]

[0069] From equations (8) and (9), it can be known that the gradients of the AdaSoftmax classification loss and the original Softmax classification loss are the same, which means that the AdaSoftmax classification loss can directly optimize the network parameters through backpropagation of gradients without other tricks and additional computational complexity.

[0070] 3. Input the divided training samples into the constructed adaptive intermediate class-level distribution alignment model, and iteratively optimize and train the transfer diagnosis model using the overall optimization objective function (Equation (10)) composed of the pseudo-label self-supervised learning function of the target domain samples and the true-label learning functions of multiple source domain samples, as shown in Equations (11) and (12).

[0071] Model optimization process:

[0072] Through the constructed AdaSoftmax classification loss and the pseudo-labels of the target domain, the entire objective function L W can be expressed as:

[0073]

[0074] where m T represents the number of samples in the target domain, represents the number of samples in the nth source domain, and represent the feature vector and pseudo-label of the ith sample in the target domain respectively, and represent the feature vector and true label of the ith sample in the nth source domain respectively, L T (·) and represent the cross-entropy losses acting on the target domain samples and source domain samples respectively. Using Equation (10), Figure 1 the trainable parameters of the global sub-model Θ G and the class-level sub-model Θ C can be updated:

[0075]

[0076]

[0077] where ε is a hyperparameter representing the learning rate.

[0078] 4. After multiple iterations of training, observe whether the loss curve tends to converge stably. If the curve converges, the diagnosis model training is completed, and the trained transfer network will be used for multi-source hybrid transfer fault diagnosis of planetary gearboxes under different working conditions.

[0079] The above are the process steps of the proposed adaptive intermediate class-level distribution alignment model for the hybrid transfer fault diagnosis of a wind turbine planetary gearbox. The following experimental results have demonstrated the effectiveness of the method.

[0080] Verification experiment: The original vibration signals of the gears collected in this experiment come from a DDS test bench, as Figure 3 shown. The DDS test bench mainly consists of five parts: a motor, a planetary gearbox, a parallel gearbox, and a magnetic particle brake. Different working condition signals can be obtained by loading and simulating through the magnetic particle brake. The planetary gearbox has five health conditions: normal (NC), surface wear (SW), root crack (RC), tooth defect (CT), and tooth breakage (MT). Accelerometers A and B are installed on the planetary gearbox to collect the original vibration signals. By controlling the current magnitude of the magnetic particle brake, four working conditions are simulated, including 0 N·m (L1), 1.4 N·m (L2), 2.8 N·m (L3), and 25.2 N·m (L4). As shown in Table 1, combining the above five health state information and four working conditions, a multi-source hybrid transfer diagnosis task is constructed, which mainly includes three sub-class transfer diagnosis tasks: multi-source closed-set transfer, multi-source partial transfer, and multi-source open-set transfer.

[0081] Table 1 Details of the multi-source hybrid transfer diagnosis task

[0082]

[0083] Comparative experiment:

[0084] To prove the superiority of the multi-source hybrid transfer diagnosis method based on the adaptive intermediate class-level distribution alignment model (AICDA) of the present invention, it is compared with the current typical single-source marginal distribution alignment network models (DDC, DCORAL, DANN, MCD), the classical multi-source marginal distribution alignment (MSDDC, MSCORAL), and the well-known multi-source joint distribution alignment (MSJDA). To ensure the reliability of the experimental results, each multi-source hybrid transfer diagnosis task is repeated five times, and the experimental results (including accuracy and standard deviation) of the 12 multi-source hybrid transfer diagnosis tasks are shown in Tables 2 and 3 respectively.

[0085] Table 2 Experimental results

[0086]

[0087] Table 3 Experimental results

[0088]

[0089] As can be seen from Table 2 and Table 3, the adaptive intermediate class-level distribution alignment model proposed by the present invention has higher transfer diagnosis accuracy and stronger generalization ability (the bold indicates the highest diagnosis accuracy rate under each transfer diagnosis task). The average diagnosis accuracy rate of AICDA on 12 multi-source hybrid transfer diagnosis tasks has reached over 88.87%, showing an obvious improvement compared with other comparison methods in most diagnosis tasks; in particular, for the multi-source partial transfer diagnosis sub-class tasks, the average diagnosis accuracy rate of AICDA has reached 91.86%, which is at least 14.07% higher than other classical transfer diagnosis tasks.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and all of them should be covered by the scope of the claims of the present invention.

Claims

1. A mechanical fault multi-source hybrid migration diagnosis method based on the AICDA model, characterized in that The method specifically includes the following steps: S1: Install an acceleration sensor on the mechanical equipment to collect the original vibration signals under different working conditions; then, use the sliding sampling technique to expand the fault sample data set for the collected original vibration signals, and divide the training sample set and the test sample set; S2: Construct an AICDA model, which uses a one-dimensional convolutional neural network as the backbone network, including four "ConV1D" convolutional blocks, two FC convolutional layers, four DIA layers, four class-level DIA layers, and an AdaSoftmax classification loss; among them, AICDA represents Adaptive Intermediate Class-level Distribution Alignment; DIA represents Dynamic Intermediate Alignment, and the expression is: Among them, γ and β are training parameters that can be automatically updated by gradient backpropagation; Among them, X S and X T respectively represent the source domain feature and the target domain feature, and respectively follow a normal distribution and and respectively represent the distribution parameters of the source domain and the target domain; Z S and Z T respectively represent the source domain feature and the target domain feature after distribution alignment; The AdaSoftmax classification loss represents the adaptive classification loss, and the AdaSoftmax classification loss L y has the following expression: Among them, F i (c) represents the element value corresponding to the label index position of the output of the i-th sample feature F of the last fully connected layer of the network i of, and F i (j) represents the other elements except the element at the label index position, m represents the number of samples; k is the decision margin; S3: Input the divided training samples into the constructed AICDA model, and use the overall optimization objective function composed of the pseudo-label self-supervised learning function of the target domain samples and the true label learning function of multiple source domain samples to perform iterative optimization training on the AICDA model; S4: After multiple iterative trainings, observe whether the loss curve tends to converge stably. If the curve converges, the model training is completed, and the trained AICDA model will be used for mechanical multi-source hybrid transfer fault diagnosis under different working conditions.

2. The mechanical fault multi-source hybrid migration diagnosis method according to claim 1, wherein In step S2, the expression of the decision margin k is: k = e (Homogeneity+Completeness+V-measure) -1 Among them, Homogeneity, Completeness, and V-measure represent homogeneity, completeness, and V-measure respectively; Homogeneity = 1 - H(C|K) / H(C) Completeness = 1 - H(K|C) / H(K) Among them, H(C|K) and H(K|C) represent the conditional entropy under the given cluster and class respectively, and H(C) and H(K) represent the information entropy under the class and cluster respectively.

3. The mechanical fault multi-source hybrid migration diagnosis method according to claim 1, characterized in that In step S3, the AICDA model is optimized, specifically including: overall optimizing the objective function L through the constructed AdaSoftmax classification loss and target domain pseudo-labels W which is expressed as: Among them, m T represents the number of samples in the target domain, represents the number of samples in the n-th source domain, and respectively represent the feature vector and the pseudo-label of the i-th sample in the target domain, and respectively represent the feature vector and the true label of the i-th sample in the n-th source domain, L T (·) and respectively represent the cross-entropy losses acting on the target domain samples and the source domain samples; Using the above global sub-model Θ G and the trainable parameters of the class-level sub-model Θ C are updated to: Among them, ε is a hyperparameter representing the learning rate.

Images