Intelligent diagnosis method and system for mechanical equipment failure in open set

Abstract

The application discloses an intelligent diagnosis method and system for mechanical equipment faults in an open set, which comprises the following steps: a new FCformer encoder based on Fourier and convolution embedding is used to construct long-distance dependency in the frequency domain and extract more representative local domain invariant fault features; a dynamic sub-domain adaptive module based on DLMMD is designed to align the conditional feature distribution of known classes by masking potential unknown classes; in order to reduce the influence of experience threshold setting on unknown class detection, an adaptive threshold learning method is used to establish the decision boundary of known classes and unknown classes. The shared feature extractor and fault classifier in the model are trained by a min-max entropy game training strategy to suppress the overconfidence problem of the model when the entropy value is minimized. The application can solve the problems of cross-domain knowledge transfer and intelligent detection of unknown faults of mechanical equipment in an open set, and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of mechanical equipment fault diagnosis, in particular to an intelligent diagnosis method and system for mechanical equipment faults in an open set. BACKGROUND

[0002] With the development of modern industrial technology, mechanical equipment is gradually developing towards precision, complexity and intelligence, which puts forward higher requirements for the safety and reliability of mechanical equipment. Rolling bearings in mechanical equipment may cause serious economic losses and safety accidents due to pitting, fracture and other faults during operation. Therefore, health monitoring of mechanical equipment is of great significance to ensure mechanical safety, reduce maintenance costs and protect personal safety.

[0003] In recent years, deep learning technology has become an effective solution in the field of mechanical equipment fault diagnosis due to its powerful nonlinear feature extraction capability. The excellent performance of these methods depends on the independent and identically distributed of training data and test data. However, in actual industrial scenarios, mechanical equipment often operates under complex and variable working conditions, and the fault feature frequency and amplitude range of the collected vibration signals show obvious differences. Direct application of domain knowledge learned from training data to test data will lead to a decline in diagnostic performance. Therefore, domain adaptation technology, as a branch of transfer learning, can align the features from the source domain and the target domain, thereby transferring the knowledge of the source domain to the target domain. The success of domain adaptation technology is due to the same label space of the source domain and the target domain. However, in actual industrial scenarios, the working conditions of mechanical equipment show instability and uncertainty, and new unknown fault types may occur at any time, making the fault diagnosis method based on domain adaptation no longer applicable. In this case, the label space of the target domain is greater than that of the source domain, which is an open set fault diagnosis scenario. Although domestic and foreign scholars have carried out extensive research on open set mechanical equipment fault diagnosis, there are still two problems:

[0004] (1) Most open set fault diagnosis methods mainly focus on using domain adversarial training strategies to align the global feature distribution of the source domain and the target domain, while ignoring the specific conditional distribution bias and discriminative features of the target domain samples. At the same time, due to the existence of unknown classes, traditional domain adaptation methods will incorrectly classify unknown classes as known classes, and simply using LMMD to align the conditional distribution of the two domains will lead to negative transfer of features.

[0005] (2) In the aspect of unknown class detection, the traditional open set fault diagnosis method usually uses a fixed threshold based on prior expert knowledge as the evaluation index of unknown class, which has poor generalization and stability and is difficult to apply to actual industrial scenes. In addition, some open set fault diagnosis methods based on adaptive threshold usually only use confidence or distance metric as the detection threshold, lacking the correlation information between samples. SUMMARY

[0006] The purpose of the present application is to provide an intelligent diagnosis method and system for mechanical equipment faults under an open set, which can overcome the interference of unknown classes on feature alignment and automatically generate a more discriminative unknown class detection threshold, realizing accurate fault diagnosis under different open set scenes.

[0007] The technical solution for achieving the purpose of the present application is:

[0008] An intelligent diagnosis method for mechanical equipment faults under an open set, comprising:

[0009] S1, using a linear embedding layer to serialize a one-dimensional vibration signal, converting it into a one-dimensional token embedding sequence that meets the requirements of a Transformer encoder;

[0010] S2, constructing an FCformer encoder with Fourier and convolution embedding layers, and inputting the one-dimensional token embedding sequence into the FCformer encoder for time-frequency feature extraction;

[0011] S3, using a global average pooling layer to obtain shared features, combining a linear embedding layer and T FCformer encoders to form a shared feature extractor;

[0012] S4, designing a fault classifier based on a convolutional neural network to identify fault patterns from shared features;

[0013] S5, building an open set dynamic sub-domain adaptive module to capture fine-grained information of each known class, realizing dynamic sub-domain feature alignment under an open set;

[0014] S6, designing an adaptive unknown class detection threshold learning method, using a sample evaluation index function to automatically update the unknown class detection threshold in each iteration cycle, and obtaining the final unknown class detection threshold after iterative training;

[0015] S7, for source domain sample data, designing a source domain loss function containing cross-entropy loss and center loss to distinguish fault features;

[0016] S8, for the target domain sample data, an entropy loss is introduced to improve the class distinction ability of the target domain sample, and a DLMMD loss function is used for dynamic conditional feature distribution alignment;

[0017] S9, a minimum-maximum entropy game training strategy is designed, and the shared feature extractor and the fault classifier are trained through the source domain loss function, the entropy loss function and the DLMMD loss function;

[0018] S10, the evaluation index function is used to generate the evaluation index of the target domain test sample, and the evaluation index is compared with the unknown class detection threshold, so that the intelligent detection of the unknown class and the cross-domain identification of the known class fault mode are completed.

[0019] An intelligent diagnosis system for mechanical equipment faults in an open set, comprising:

[0020] A data acquisition and dataset creation module: the vibration signals of mechanical equipment in various health states are collected under different working conditions, and the vibration signal data is divided into samples and normalized for pretreatment; part of the source domain data of the health state is randomly selected as the known class sample, and the target domain data contains all health states, that is, the target domain contains both known classes and unknown classes; the samples in the source domain and the target domain are selected in proportion to create a training dataset, and quantitative target domain data is selected as a test dataset;

[0021] A model building and training module: a dynamic subdomain adaptive network model architecture is constructed, including a shared feature extractor, a fault classifier, an open set dynamic subdomain adaptive module and an adaptive unknown class detection threshold learning method, and the model structure parameters are initialized; the model is trained by using a minimum-maximum entropy game strategy, the parameters of the model are updated by optimization, and the training of the dynamic subdomain adaptive network model is completed;

[0022] A model performance verification module: the trained dynamic subdomain adaptive network model is verified by using the test dataset, the test results are analyzed, and the optimal model and the corresponding unknown class detection threshold are output;

[0023] A fault diagnosis module: in different open set domain adaptive fault diagnosis scenes, the optimal dynamic subdomain adaptive network model is used for fault diagnosis of the real-time collected vibration signals, so as to realize the cross-domain identification of the known fault mode and the intelligent detection of the unknown fault mode.

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

[0025] 1. The present application proposes a dynamic subdomain adaptive network model for intelligent diagnosis of mechanical equipment faults in an open set, which realizes the cross-domain knowledge transfer of known classes and the intelligent detection of unknown classes.

[0026] 2. The application constructs an improved Transformer encoder based on Fourier and convolution embedding, provides a lightweight Transformer encoder structure and a long-distance modeling scheme in the frequency domain.

[0027] 3. Considering the sub-domain space difference and sub-domain sample imbalance caused by unknown classes, the application proposes a new open set dynamic sub-domain adaptive module based on DLMMD, which can fine-align the sub-domain features of known classes by dynamically assigning specific class weights.

[0028] 4. The application uses the correlation information between samples to design an adaptive threshold learning method to automatically generate unknown class detection thresholds, which can reduce the dependence on prior knowledge when setting thresholds and improve the generalization ability of the model for different open set diagnosis tasks.

[0029] 5. The application uses a min-max entropy game training strategy between the shared feature extractor and the fault classifier to suppress the problem of excessive confidence of the model for unknown classes caused by blindly minimizing the entropy value. Finally, the fault data set of cylindrical roller bearings in offshore wind power hoisting equipment is used to verify the stable and accurate fault diagnosis performance of the model proposed in the application in the open set domain adaptation scene. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The specific implementation flowchart of the intelligent diagnosis method and system for mechanical equipment faults in an open set.

[0031] Figure 2 is a dynamic sub-domain adaptive network model structure diagram of the application.

[0032] Figure 3 is an FCformer encoder structure diagram of the application.

[0033] Figure 4 is an open set dynamic sub-domain adaptive module structure diagram of the application.

[0034] Figure 5 is an appearance diagram of the cylindrical roller bearing test bench in the embodiment.

[0035] Figure 6 is an average H-Score comparison diagram of the model proposed in the application and four open set domain adaptation models in the embodiment.

[0036] Figure 7 is a confusion matrix diagram of the model proposed in the application and different benchmark models in the T9 task under the first test in the embodiment. DETAILED DESCRIPTION

[0037] The specific embodiments of the present application are described below with reference to the accompanying drawings so that those skilled in the art can better understand the present application. It should be particularly noted that in the following description, when detailed descriptions of known functions and designs may obscure the main content of the present application, these descriptions will be omitted here.

[0038] Figure 1 is a specific implementation flowchart of the intelligent diagnosis method and system for mechanical equipment failure in an open set of the present application. As shown in Figure 2 the dynamic sub-domain self-adaptive network structure of the present application mainly consists of a shared feature extractor G and a fault classifier C. Among them, the shared feature extractor mainly includes a linear embedding layer, T FCformer encoders and a global average pooling layer, and the fault classifier is used to identify the fault mode of the shared features. The specific implementation steps of the present application are as follows:

[0039] Step 1: using a linear embedding layer to perform serialization processing on the collected one-dimensional vibration signal , so as to convert it into a one-dimensional token embedding sequence z0 that meets the input requirements of the traditional Transformer encoder. The specific steps are as follows:

[0040] Assuming that the size of the sequence block is fixed as L, the collected one-dimensional vibration signal is converted into a new sequence block: where W is the length of the input data, and Q=W / L is the number of sequence blocks. Then, the sequence block X is projected into a token embedding sequence with a dimension of N using a linear embedding function ξ(·), and the token embedding sequence z0 for inputting the Transformer encoder can be obtained as:

[0041]

[0042] In the formula, x q represents the qth sequence block.

[0043] Step 2: construct a new Transformer framework with Fourier and convolutional embedding (FCE) layer-FCformer encoder, and input the one-dimensional token embedding sequence z0 obtained in step 1 into the FCformer encoder. As shown in Figure 3 , the FCformer encoder uses a lightweight FCE layer to replace the multi-head self-attention layer of the traditional Transformer. The FCformer encoder includes an FCE layer, a layer normalization (LN), a feedforward layer, and a layer normalization, wherein the FCE layer includes a 1D-DFT layer, a one-dimensional convolutional layer, a batch normalization (BN) layer, and a rectified linear unit (ReLU) layer; the feedforward layer includes a one-dimensional convolutional layer, a BN layer and a RE LU layer, a dropout layer and a one-dimensional convolutional layer.

[0044] Specifically, the 1D-DFT layer is given a vibration signal sequence S n , where n∈[0,...,N-1], is processed by a discrete Fourier transform (DFT) as:

[0045]

[0046] In the formula, m is a frequency index, F(m) is a frequency domain feature representation, is a rotation factor, and j is an imaginary unit.

[0047] Further, by Euler's formula, the discrete Fourier transform is decomposed into a real part and an imaginary part:

[0048] F(m) = F real (m) + jF imag (m) (3)

[0049] In the formula, F real (m) and F imag (m) are the real part and the imaginary part of F(m), respectively.

[0050] Using the frequency feature learning capability of the 1D-DFT layer, an FCE layer is designed in the FCformer encoder to perform DFT transformation on the length dimension of the token embedding sequence. It should be noted that the FCE layer only retains the real part after the 1D-DFT layer transformation. Therefore, for the input token embedding sequence z t-1 of the tthFCformer encoder, the feature map z′ t processed by the 1D-DFT layer is:

[0051]

[0052] In the formula, is a function that retains the real part after DFT transformation along the length dimension.

[0053] Subsequently, a one-dimensional convolution layer is introduced into the FCformer encoder, and a BN layer and a RELU layer are connected, to obtain a feature map z t ″:

[0054]

[0055] In the formula, σ(·) represents the processing function of the BN layer and the RELU layer, and f(·) is a one-dimensional convolution operation.

[0056] Using residual connection, the feature maps z t ′ and z t-1 are superimposed to obtain a feature map z″′ tFor:

[0057] z t ′′=LN(z t ′ t ′)+z t-1 (6)

[0058] where LN(·) is a layer normalization function.

[0059] Finally, a feed-forward layer is connected to provide a non-linear transformation and access to the LN layer and residual connection. At this time, the output token embedding sequence z t of the t-th FCformer encoder can be represented as:

[0060]

[0061] where FF(·) denotes the processing function of the feed-forward layer.

[0062] The token embedding sequence z T output after time-frequency feature extraction by T FCformer encoders is input into a global average pooling layer (GAP), and the shared feature u1 can be obtained as:

[0063]

[0064] where R denotes the kernel size of the GAP layer, and GAP(·) denotes the GAP operation.

[0065] Combining the linear embedding layer in step 1 and the T FCformer encoders in step 2, a shared feature extractor G is formed.

[0066] Step 4: Design a fault classifier C based on a convolutional neural network to perform fault pattern recognition on the shared feature u1 in step 3. The fault classifier includes a convolutional extraction layer and two fully connected layers (FC). The convolutional extraction layer includes a one-dimensional convolutional layer, a BN layer, a ReLU layer, and a max pooling layer (MP). The convolutional extraction layer and the fully connected layers are flattened. Therefore, the output feature u2 of the first FC layer and the final output probability vector y of the second FC layer are respectively represented as:

[0067]

[0068] where F C (·) denotes the operation of the convolutional extraction layer, ψ(·) is a flattening function, f FC1 (·) and f FC2 (·) denote the operations of the first and second FC layers, respectively. D is the output dimension of the first FC layer, and K is the number of known classes.

[0069] Step 5: Build an open set dynamic subfield adaptive module to realize dynamic subfield feature alignment under the open set by capturing fine-grained information of each known class. The core idea of the open set dynamic subfield adaptive module is to dynamically align the subfield distribution of known classes by shielding potential unknown classes. This module overcomes the risk of feature negative transfer due to unknown classes in the traditional LMMD method. Figure 4 The structure diagram of the open set dynamic subfield adaptive module is shown, and the specific implementation steps of the module are described as follows:

[0070] 1. Quantitatively evaluate the probability of whether each target domain sample belongs to an unknown class to ensure that the model accurately suppresses the alignment of unknown classes during subfield adaptation. Since unknown class samples belong to any known class with a low probability, their confidence sensitivity for uncertain prediction is low. On the contrary, the entropy value for representing the uncertainty of the class distribution is high. Therefore, the open set dynamic subfield adaptive module in the present application utilizes the complementary advantages of entropy and confidence to construct a sample evaluation feature to comprehensively represent each sample. For a sample with a probability vector y, its evaluation feature function fea(·) can be represented as:

[0071] fea(y) = [ε1, 1-ε2] (12)

[0072] In the formula, is the entropy value of the sample probability vector, where y(k) represents the kth element in the probability vector y; ε2 = max(y(k)), k = 1, 2, …, K is the confidence of the sample probability vector (consistent with the number of known classes).

[0073] 2. Cluster analysis is performed on the evaluation features of the target domain samples using the K-means algorithm to provide pseudo-labels of known classes and unknown classes for subsequent training of the model. Since the cluster number k' in this example is 2, i.e., k' = 2, the present application does not need to consider the impact of cluster number setting on sample clustering performance. It should be noted that the cluster samples closer to zero in the clustering results are known classes. Therefore, the pseudo-label group of known classes or unknown classes in the target domain samples can be represented as:

[0074]

[0075] In the formula, R(·) is a marking operation, Kmeans(·) represents the K-means algorithm operation, n t is the number of target domain samples, represents the probability of the ath sample in the target domain. represents the pseudo-label value of the ath target domain sample. When , the target domain sample is a known class; , the target domain sample is an unknown class.

[0076] 3. Based on the pseudo-label h of the target domain sample t The probability vector for the a-th target domain sample By masking the unknown class samples, the optimized probability vector is obtained as follows:

[0077]

[0078] 4. Considering that potentially unknown class samples do not participate in feature alignment, this will lead to an imbalance in the distribution of subdomain samples between the source and target domains. Traditional LMMD assigns the same weights to different subdomains, resulting in inaccurate feature alignment within the subdomain space. Therefore, to more precisely align subdomain features of known classes, a Dynamic Local Maximum Mean Difference (DLMMD) method is designed in the open set dynamic subdomain adaptation module. This method can assign different weights to specific subdomains based on the number of subdomain samples, thereby more accurately aligning subdomain features of known classes. Therefore, for the source domain feature distribution P... s and target domain feature distribution P T The result obtained by performing DLMMD calculation is as follows:

[0079]

[0080] In the formula, For DLMMD functions, n s The number of samples in the target domain. and They represent the i-th source domain sample respectively. and the a-th target domain sample The weights belonging to class k, H is the reproducing kernel Hilbert space (RKHS), and φ(·) represents the nonlinear mapping function from the feature space to the RKHS, ||·|| 2 This represents the L2 norm. For samples in the source domain... Convert the real labels into one-hot vectors. To calculate weights For unlabeled target domain samples The optimized prediction vector Weights are calculated as one-hot vectors. This represents the number of samples in the source domain that belong to class k. This represents the number of samples in the target domain that belong to the unknown class. This represents the number of target domain samples whose data belongs to a known class and class k. Its specific definition is as follows:

[0081]

[0082] In the formula, G(·) is the indicator function, that is: if the target domain sample is known class and belongs to category k, then return 1, otherwise return 0. represents the a-th target domain sample whose prediction result is category k.

[0083] Step 6: An adaptive unknown class detection threshold learning method is designed to realize automatic updating of the unknown class detection threshold in each iteration period.

[0084] 1. The average value of all elements in the sample evaluation feature is taken as the evaluation index of the sample, and the maximum evaluation index of the source domain sample in each training batch is calculated:

[0085]

[0086] In the formula, B is the batch size, P=n s / B is the number of batches in each iteration period, is an operation function for taking the average of all elements in the evaluation feature, represents the probability vector of the b-th source domain sample in the p-th batch.

[0087] 2. The maximum evaluation index of each batch obtained in the e-th iteration period is combined into a vector:

[0088]

[0089] In the formula, represents the maximum evaluation index of the p-th batch obtained in the e-th iteration period, p∈[1,P].

[0090] 3. In order to retain the feature information of the previous iteration period and reduce the cumulative error, the vector of the current iteration period is fused with the vector of the previous iteration period, so that the unknown class detection threshold is:

[0091]

[0092] In the formula, κ is a balance factor for balancing and The relative importance between them. It is worth noting that when the iteration period E is obtained by training, the final unknown class detection threshold

[0093] 4、Since the feature distribution of known classes in source and target domains will gradually converge as the model training proceeds, the evaluation indicators will continuously converge. Therefore, in order to accurately measure the strength of the linear relationship between the evaluation indicators in the convergence process, the Pearson correlation coefficient ρ(·,·) is used to analyze the correlation between vectors and , which is further used to adaptively update the balance factor in the unknown class detection threshold function, that is:

[0094]

[0095] When vectors and are not correlated, that is, the model is still in the convergence state, κ is calculated as 0.5 at this time, that is, the unknown detection threshold will be the average of the maximum value of and .

[0096] 5、After the E-th iteration training is completed, the final unknown class detection threshold

[0097] Step 7: For source domain sample data, two optimization objectives are designed to promote the model to learn more discriminative fault features. First, the cross-entropy loss is used to increase the class separability by taking the distance between source samples as a penalty:

[0098]

[0099] In the formula, denotes the probability vector of the b-th source domain sample in the current batch, and are the k-th and l-th elements of the Softmax parameter in the source domain sample, and are the k-th and l-th elements of the probability vector of the b-th source domain sample in the current batch. 1{·} is an indicator function, that is, it returns 1 if the condition is true, and 0 otherwise.

[0100] In addition, a center loss is introduced to improve the intra-class compactness of the source domain sample, that is, by calculating the distance between the feature and its feature center as a penalty, features with the same label are close to each other, and the loss function is represented as:

[0101]

[0102] In the formula, c yb denotes the feature center of y b class, and u 2,b denotes the first FC layer output feature of the b-th sample in the current batch.

[0103] Combining formula (22) and formula (23), the loss function of the source domain sample is obtained as follows:

[0104]

[0105] In the formula, μ is a penalty factor.

[0106] Step 8: For the target domain sample data, an entropy loss function is introduced to improve the class discrimination ability of the target domain sample. As an evaluation component of the feature, entropy minimization can provide more reliable detection basis for the DLMMD method and adaptive threshold. Therefore, the entropy loss function of the target domain sample is expressed as follows:

[0107]

[0108] In the formula, represents the kth element in the bth probability vector in the current batch in the target domain.

[0109] Subsequently, the DLMMD method described in step 5 is used to dynamically align the conditional feature distribution of the shared feature u1 described in step 3 and the first FC layer feature u2 described in step 4, and the loss function is expressed as follows:

[0110]

[0111] In the formula, λ is a weight factor, and represent the shared features in the source domain and the target domain, respectively, and represent the output features of the first FC layer in the source domain and the target domain, respectively.

[0112] Step 9: Design a min-max entropy game training strategy to train the shared feature extractor G described in step 3 and the fault classifier C described in step 4 through the source domain loss function L s described in step 7 and the entropy loss function and the DLMMD loss function L DA described in step 8 to suppress the overconfidence problem that may be caused by minimizing the entropy value. The principle of the min-max entropy game strategy is to maximize the entropy of the target domain sample to increase the gap between the known classes and the unknown classes in the evaluation index, while minimizing the entropy of the target domain sample to improve the discrimination of the known classes. The core idea of this strategy is to encourage the model to generate more discriminative decision boundaries to identify known classes and separate unknown classes through adversarial learning of entropy values. Therefore, the model training steps based on the min-max entropy game strategy are described as follows:

[0113] 1. Update the shared feature extractor parameters θ​​G and fault classifier parameter C The structure parameters of the shared feature extractor and the fault classifier are preliminarily trained by using the training samples of the source domain and the target domain, and the loss function of the model is:

[0114]

[0115] In the formula, a and b are weighting coefficients.

[0116] 2. Update the shared feature extractor parameter G and fix the fault classifier parameter C The target domain sample is optimized by maximum entropy value while minimizing the intra-class and inter-class distances of the source domain sample, and the loss function of the model is:

[0117]

[0118] 3. Repeat the training steps 1-2 above, constantly optimize the class boundary and align the cross-domain feature distribution of the known class, and finally generate a more accurate and clear open set fault decision boundary.

[0119] Step 10: For the sample of the probability vector y test , the evaluation feature fea(y test ) of the test sample is generated by using the evaluation feature function described in step 5, and the evaluation index is obtained by averaging all elements in the evaluation feature, and compared with the final unknown class detection threshold described in step 6, and finally the detection of unknown class and the identification of fault mode in known class are completed. Specifically, if the evaluation index is less than the threshold , the sample is a known class; otherwise, the sample is an unknown class. Thus, the final output result of the network model proposed in the present application is:

[0120]

[0121] In the formula, is the final diagnosis output result of the test sample.

[0122] Based on the above method, the present embodiment further proposes an intelligent diagnosis system for mechanical equipment faults under an open set, and the specific content includes:

[0123] Data acquisition and dataset creation module: collect vibration signals of mechanical equipment in different working conditions and under various health states, and perform sample division and normalization preprocessing on the vibration signal data; according to the open set task requirements, randomly select part of the source domain data of the health state as known class samples, and the target domain data contains all health states, that is, the target domain contains both known classes and unknown classes; select samples in the source domain and the target domain in proportion to create a training dataset, and select quantitative target domain data as a test dataset.

[0124] Model building and training module: build a dynamic sub-domain adaptive network model architecture, including a shared feature extractor, a fault classifier, an open set dynamic sub-domain adaptive module, and an adaptive unknown class detection threshold learning method, and initialize the model structure parameters; train the model using the minimax entropy game strategy, and constantly update the model parameters by optimizing the target to complete the training of the dynamic sub-domain adaptive network model.

[0125] Model performance verification module: verify the diagnostic performance of the trained dynamic sub-domain adaptive network model using the test dataset, analyze the test results, and output the optimal model and the corresponding unknown class detection threshold.

[0126] Fault diagnosis module: in different open set domain adaptive fault diagnosis scenarios, use the optimal dynamic sub-domain adaptive network model to diagnose the real-time collected vibration signals, thereby realizing cross-domain recognition of known fault modes and intelligent detection of unknown fault modes.

[0127] 1Example 1

[0128] To better illustrate the technical effects of the present application, a specific example is used to experimentally verify the present application. The specific example is as follows:

[0129] Step 1: Data acquisition and dataset creation. The fault dataset of the cylindrical roller bearing test bench of the hoisting transmission system of the offshore wind power hoisting equipment is used for experimental verification. As shown in Figure 5 , the test bench mainly includes an upper computer, a motor, a servo control system, a data acquisition system, a magnetic brake, a tension controller, an acceleration sensor, etc. In this example, the running speed of the motor is set to 900 rpm, the load is set to two working conditions of 10 N / m and 20 N / m, and the vibration signal of the cylindrical roller bearing is acquired at a sampling frequency of 10240 Hz. In terms of fault setting, the cylindrical roller bearing with faults is installed in the bearing seat close to the magnetic brake, and the test bearing is set to four fault states (corresponding to labels 0-4) of rolling element fault (RF), inner ring fault (IF), outer ring fault (OF) and composite fault (CF, rolling element and outer ring fault). Among them, the damage diameter width of the fault is 0.42 mm, and the depth is 0.8 mm.

[0130] Table 1 Domain adaptation tasks

[0131]

[0132] Based on the above fault data description, this example adopts five different health states of vibration signals, each health state contains 300 samples, and each sample consists of 1024 data points. Therefore, according to the number of source domain health states in different domain adaptation tasks, 180 (60%) samples are selected from each source domain health state, and 18x n (where n is the number of source domain health states) samples are selected from each target domain health state to form the training set. Finally, 60 (20%) remaining target domain samples are randomly selected from each health state as the test data set. It should be noted that the labels of the source domain samples are randomly selected, while the target domain samples contain all health states. Thus, the domain adaptation tasks of the 10 cylindrical roller bearing fault data sets shown in Table 1 are obtained.

[0133] Step 2: Model structure and parameter setting. Table 2 gives the specific structure parameter table of the dynamic sub-domain adaptation network model in this case. Among them, each one-dimensional convolutional layer adopts zero padding to ensure the consistency of the feature dimension.

[0134] Table 2 Dynamic sub-domain adaptation network model structure parameters

[0135]

[0136]

[0137] In addition, the main training hyperparameters in the model training process of this case are shown in Table 3. It should be noted that the learning rate is reduced by 90% after every 10 iteration periods, and each model training and testing is independently run 10 times, and the average of the 10 experimental test results is taken as the final diagnostic performance of the model to reduce the influence of randomness.

[0138] Table 3 Model training hyperparameters

[0139]

[0140] Step 3: Comparison of benchmark model settings. In order to prove the superiority of the model proposed in this invention, this example uses existing advanced and relevant methods to compare with this invention. The specific information of these benchmark models is described as follows:

[0141] (1) DANN: A domain adversarial neural network model composed of a feature extractor, a task classifier, and a domain classifier (for details, see: Ganin Y, Ustinova E, Ajakan H, et al. Domain-Adversarial Training of Neural Networks. J Mach Learn Res. 2016, 17, 2096-30.).

[0142] (2) OSBP: An open set domain adaptation model based on backpropagation, which identifies unknown classes by adding additional prediction probability elements to the output layer (for details, see: Saito K, Yamamoto S, Ushiku Y, Harada T. Open Set Domain Adaptation by Backpropagation. Computer Vision-Eccv 2018, Pt V. 2018, 11209, 156-71.).

[0143] (3) UDA: A classic universal domain adaptation method that can achieve feature alignment of known classes through a weighted strategy based on sample domain similarity (for details, see: You KC, Long MS, Cao ZJ, Wang JM, Jordan MI. Universal Domain Adaptation. Proc Cvpr Ieee 2019. p. 2715-24.).

[0144] (4) OSDAM: An advanced open set domain adaptation method for mechanical fault diagnosis (for details, see: Zhang W, Li X, Ma H, Luo Z, Li X. Open-Set Domain Adaptation in Machinery Fault Diagnostics Using Instance-Level Weighted Adversarial Learning. Ieee T Ind Inform. 2021, 17, 7445-55.).

[0145] (5) DANet: An open set fault diagnosis method, the model structure of which includes an auxiliary domain discriminator, an extended classifier and a feature generator, and the cross-domain diagnosis of known classes and the separation of unknown classes are realized through dual adversarial learning (for details, see: Zhao C, Shen WM. Dual adversarial network for cross-domain open set fault diagnosis. Reliab Eng Syst Safe. 2022, 221, 108358.).

[0146] Step 4: Analysis of experimental results. In order to more finely verify the effectiveness and superiority of the model in the present application, four performance indicators are used for quantitative comparison: the accuracy of all test samples OS, the accuracy of known classes OS*, the accuracy of unknown classes UK and the harmonic mean H-Score of OS and UK. Table 4 lists the open set diagnosis test results of all comparison models. From the table, it can be seen that DANN has the best diagnosis performance for closed set domain adaptation tasks T1 and T2, while for all other open set domain adaptation tasks, the average OS of the model proposed in the present application is the highest. Overall, the average OS of the model proposed in the present application is the highest 98.60% in all test tasks, while the average OS of DANN, OSBP, UDA, OSDAM and DANet in all test tasks is only 67.84%, 74.07%, 75.11%, 69.75% and 81.41% respectively, which is much lower than the model proposed in the present application. Therefore, the model proposed in the present application has better diagnosis performance and robustness, and shows great potential in open set mechanical equipment fault diagnosis.

[0147] Table 4 Open set diagnosis test results

[0148]

[0149] In addition, Figure 6 The average H-Score comparison chart of the model proposed in the present application and the four open set domain adaptation models is given, and from the chart it can be seen that the average H-Score of the model proposed in the present application in each open set task is better than that of other open set domain adaptation models, and the minimum average H-Score is also greater than 94%. In addition, the average H-Score of the model proposed in the present application in the eight open set domain adaptation tasks is the highest 98.35%. Overall, the model proposed in the present application has better accuracy and stability in simultaneously processing cross-domain diagnosis and unknown class detection.

[0150] In order to intuitively understand the accuracy of the model proposed in the present application for each health state, confusion matrix is used to visually analyze representative DANN, OSBP and DANet. Figure 7The confusion matrix diagram of the first test of the model of the application and different benchmark models in the T9 task is shown. It can be seen from the diagram that in the detection of unknown categories, DANN cannot detect unknown categories, although OSBP has the highest accuracy in diagnosing unknown categories, but 6.67% of the second category is identified as unknown categories. In addition, the diagnostic performance of DANet for known categories is the worst. On the contrary, the diagnostic accuracy of the unknown categories of the model of the application is 93.33%, and the diagnostic accuracy of all known categories reaches 100%, which shows that the model of the application has better comprehensive diagnostic performance under the open set.

[0151] Although the above describes the specific embodiments of the application in detail, so that those skilled in the art can understand the application, it should be clear that the application is not limited to the scope of the specific embodiments, and for those skilled in the art, as long as various changes are within the spirit and scope of the application defined and determined by the appended claims, all the application and creation using the concept of the application are protected.

Claims

1. An intelligent diagnostic method for mechanical equipment faults under open sets, characterized in that, include: S1. Use a linear embedding layer to serialize the one-dimensional vibration signal and convert it into a one-dimensional token embedding sequence that meets the requirements of the Transformer encoder. S2. Construct an FCformer encoder with Fourier and convolutional embedding layers, and input the one-dimensional token embedding sequence into the FCformer encoder for time-frequency feature extraction; The FCformer encoder comprises an FCE layer, a layer normalization layer, a residual connection layer, a feedforward layer, a layer normalization layer, and a residual connection layer in sequence. The FCE layer includes a 1D-DFT layer, a one-dimensional convolutional layer, a batch normalization (BN) layer, and a rectified linear unit (ReLU) layer. The 1D-DFT layer retains the real part of the feature map after DFT transformation, and the residual connection superimposes the feature maps before and after normalization of the corresponding layer; the feedforward layer includes a one-dimensional convolutional layer, a BN layer and a ReLU layer, a dropout layer and a one-dimensional convolutional layer. S3. Utilize a global average pooling layer to obtain shared features, combined with a linear embedding layer and A set of FCformer encoders form a shared feature extractor; S4. Design a fault classifier based on convolutional neural networks to identify fault modes based on shared features; S5. Build an open set dynamic subdomain adaptive module to achieve dynamic subdomain feature alignment under the open set by capturing fine-grained information of each known class. Specifically, it includes: 5.1 Constructing sample evaluation features using entropy and confidence levels; 5.

2. Use the K-means algorithm to perform cluster analysis on the evaluation features of the target domain samples to provide pseudo-labels for known and unknown classes for subsequent training; 5.

3. Based on the pseudo-labels of the target domain samples , for the The probability vector of each target domain sample Unknown samples are masked to obtain an optimized probability vector. 5.4 Assign different weights to specific subdomains based on the number of subdomain samples to align subdomain features of known classes; S6. Design an adaptive unknown class detection threshold learning method. Use the sample evaluation index function to automatically update the unknown class detection threshold in each iteration cycle. After the iteration training is completed, the final unknown class detection threshold is obtained. S7. For source domain sample data, design a source domain loss function that includes cross-entropy loss and center loss to identify fault features; S8. For target domain sample data, entropy loss is introduced to improve the class discrimination ability of target domain samples, and the DLMMD loss function is used to align the dynamic conditional feature distribution. S9. Design a minimum-maximum entropy game training strategy to train the shared feature extractor and fault classifier using the source domain loss function, entropy loss function and DLMMD loss function. S10. Use the evaluation index function to generate the evaluation index of the target domain test sample, and compare it with the detection threshold of the unknown class to complete the intelligent detection of the unknown class and the cross-domain identification of fault modes in the known class.

2. The intelligent diagnostic method according to claim 1, characterized in that, The output of the FCformer encoder for time-frequency feature extraction is: In the formula, The output token embedding sequence of the FCformer encoder. For layer normalization function, This represents the processing function of the feedforward layer. The feature map is obtained from the second residual connection, where Q is the number of sequence blocks and N is the dimension of the token embedding sequence.

3. The intelligent diagnostic method according to claim 1, characterized in that, The fault classifier includes a convolutional extraction layer and two fully connected layers. The convolutional extraction layer includes a one-dimensional convolutional layer, a batch normalization (BN) layer, a ReLU layer, and a max pooling layer. The convolutional extraction layer and the fully connected layers are flattened.

4. The intelligent diagnostic method according to claim 3, characterized in that, Output features of the first FC layer and the final output probability vector of the second FC layer They are represented as follows: In the formula, This indicates the operation of the convolution extraction layer. This is a function for flattening. and These represent the operations of the first and second FC layers, respectively. The output dimension of the first FC layer. The number of known classes.

5. The intelligent diagnostic method according to claim 1, characterized in that, The formula for calculating the detection threshold for unknown classes is: in, In the formula, It is a balancing factor used to balance the vector in the current iteration cycle. The vector of the previous iteration period The relative importance between them The Pearson correlation coefficient; Indicates the first The result of the iteration cycle is the first The maximum evaluation index for each batch.

6. The intelligent diagnostic method according to claim 1, characterized in that, The source domain sample loss function is: Where cross-entropy loss is: Central loss: In the formula, As a penalty factor, For batch size, Given the number of classes, Indicates the first in the current batch The probability vector of each source domain sample. and The first parameter of the Softmax parameter in the source domain sample is the 6th parameter. The and the first One element, and These are the first in the current batch The i-th source domain sample probability vector The and the first One element; It is an indicator function, meaning it returns 1 if the condition is true, and 0 otherwise. express The feature center of the class Indicates the first in the current batch The first FC layer outputs features for each sample.

7. The intelligent diagnostic method according to claim 1, characterized in that, The DLMMD loss function is: In the formula, For DLMMD functions, and These are the feature distributions of the source domain and the target domain, respectively. The number of samples in the target domain. The number of samples in the target domain. and They represent the first Source domain samples and the Target domain samples Category The weights, where K is the number of known classes. For the regenerative nucleus Hilbert space (RKHS). This represents a nonlinear mapping function from the feature space to RKHS. Represents the L2 norm; Indicates that the source domain belongs to the category The number of samples, This indicates the number of samples in the target domain that belong to the unknown class; For source domain samples Convert real labels into one-hot vectors To calculate weights For unlabeled target domain samples The optimized prediction vector Weights are calculated as one-hot vectors. ; This indicates that the data in the target domain is of a known class and belongs to the category. The number of samples in the target domain is defined as follows: In the formula, It is an indicator function, that is: if the target domain sample It is a known class and belongs to the category. If the condition is met, return 1; otherwise, return 0. Indicates the prediction result is a category. The One target domain sample; Indicates the first The pseudo-label values ​​of each target domain sample. This indicates that the target domain sample is of a known class.

8. An intelligent diagnostic system for mechanical equipment faults under an open set, used to execute the intelligent diagnostic method according to any one of claims 1-7, characterized in that, include: Data acquisition and dataset creation module: Vibration signals of mechanical equipment in various health states under different working conditions are collected, and the vibration signal data are preprocessed by sample division and normalization; a portion of the source domain data in health states is randomly selected as known class samples, while the target domain data contains all health states, that is, the target domain contains both known and unknown classes; samples from the source and target domains are selected proportionally to create a training dataset, and a quantitative amount of target domain data is selected as a test dataset; Model building and training module: Construct a dynamic subdomain adaptive network model architecture, including a shared feature extractor, a fault classifier, an open set dynamic subdomain adaptive module, and an adaptive unknown class detection threshold learning method, and initialize the model structure parameters; use a minimum maximum entropy game strategy to train the model, and update the model parameters by optimizing the target to complete the training of the dynamic subdomain adaptive network model; Model performance validation module: Validates the diagnostic performance of the trained dynamic subdomain adaptive network model using the test dataset, analyzes the test results, and outputs the optimal model and the corresponding unknown class detection threshold; Fault diagnosis module: Under different open domain adaptive fault diagnosis scenarios, the optimal dynamic subdomain adaptive network model is used to perform fault diagnosis on the real-time acquired vibration signals, thereby realizing cross-domain identification of known fault modes and intelligent detection of unknown fault modes.

Images