Fault diagnosis method under variable working condition invisible type of multi-classifier decision-making mechanism based on pseudo-label learning

Abstract

The invention discloses a multi-classifier decision-making mechanism based on pseudo-label learning under variable working condition invisible fault types. Firstly, features of a source domain and a target domain are extracted through a feature extractor, a global classifier is utilized to generate a preliminary pseudo tag for a target domain sample, a local classifier is utilized to further verify confidence, a low-confidence sample is filtered, and an unknown category is identified. Secondly, aiming at a high-confidence sample, introducing a regularization item on the maximum average difference to realize alignment of conditional distribution and edge distribution between a source domain and a target domain, and dynamically adjusting distribution alignment intensity according to a training stage; then, a center loss function is introduced, a known category sample distribution center is drawn close in a feature space, and the distance between an unknown category and a known category is ensured; and finally, by adding unknown class output in the classifier, the detected unknown class samples are classified in a unified manner. According to the method, diagnosis of known fault types and identification of unknown fault types under variable working conditions are realized.

Description

Technical Field

[0001] The present invention provides an industrial process fault diagnosis method, in particular a fault diagnosis method under variable working condition invisible type based on a multi-classifier decision mechanism of pseudo-label learning. Background Art

[0002] Rotating machinery is a widely used and indispensable core equipment in modern industry. With the rapid development of industrial technology, rotating machinery is evolving towards high automation, complex structure and large scale. However, this trend also puts forward more stringent requirements on the operating safety and stability of the equipment. Rotating machinery components usually work under non-steady-state time-varying conditions, and the failure of its key components may lead to huge economic losses or even catastrophic accidents. Condition monitoring and fault diagnosis technology can effectively improve the reliability of equipment operation and reduce the economic losses caused by mechanical equipment failure and damage. Therefore, in modern industrial production, condition monitoring and intelligent fault diagnosis of equipment under time-varying conditions are of great significance.

[0003] In actual industrial processes, the operating state of rotating equipment is often affected by actual production needs and external influences and changes continuously. The change of working conditions will directly affect the monitoring data, such as the frequency and amplitude of the vibration signal, and thus affect the distribution of the data in the feature space. The change of the operating state of the mechanical system will directly cause the distribution difference between the training data and the test data, and this difference in data distribution will affect the generalization performance of the data-driven fault diagnosis method. Under variable working conditions, the goal of fault diagnosis is to use the fault information of the equipment under the current working condition to predict its fault type under another working condition (such as different speed or load). However, the existing variable working condition fault diagnosis methods mostly rely on the assumption that the target domain data is labeled, but this assumption is often difficult to meet in actual operation and maintenance.

[0004] Although fault detection technology has been widely used in fault diagnosis of rotating equipment, the following problems still exist: (1) Existing data-driven fault diagnosis methods, such as support vector machines, convolutional neural networks, and long short-term memory, have shown significant performance in fault diagnosis of rotating machinery. However, they are usually trained under steady-state conditions, strictly assuming that the training data and test data have the same distribution. However, in actual diagnosis scenarios, equipment is usually operated under non-steady-state variable operating conditions. (2) Since there may be operating condition changes and unknown fault types in the target domain, the initially generated pseudo-labels are often inaccurate. Incorrect pseudo-labels can cause the model to misclassify the target domain data in the early stage, thereby causing negative transfer. (3) Current fault diagnosis methods based on domain adaptation are achieved by constructing multiple classifiers and reducing the deviation between source domain features and target domain features. However, the feature distribution corresponding to the categories in different domains is rarely considered. When the working conditions change dramatically, it is difficult to learn domain-invariant knowledge. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a fault diagnosis method under variable working condition invisible type based on a multi-classifier decision mechanism of pseudo-label learning.

[0006] In order to achieve the above object, according to one aspect of the present invention, a fault diagnosis method under variable working condition invisible type based on a multi-classifier decision mechanism of pseudo-label learning is provided, and the method comprises:

[0007] 1. A fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism, characterized by comprising the following steps:

[0008] Step 1: Collect data from the industrial production process through sensors, divide the collected industrial data under different working conditions into source domain and target domain, and define the source domain training set and target domain test set as in, is the i-th sample in the source domain training set, is the label of the i-th sample in the source domain training set, is the jth sample in the target domain test set, N s and N t are the number of samples in the source domain and the target domain respectively;

[0009] Step 2: Use the data obtained in step 1 to input the feature extractor to obtain features. The constructed feature extractor is composed of three identical feature extraction units connected in series, each of which includes a one-dimensional convolutional layer, an activation layer, a pooling layer, and a batch normalization layer;

[0010] Step 3: Use the source domain features extracted by the feature extractor to train Figure 1The global classifier and local classifier shown in the figure, the global classifier predicts the probability of the fault label of the target domain sample, and the local classifier predicts its probability under each source domain type. The confidence ranking of the two is combined to generate a pseudo label. Through the global and local confidence analysis of the target domain samples, they are sorted in a sequence from high to low. The higher the ranking, the greater the possibility of the known category, and the lower the ranking, the greater the possibility of the unknown category;

[0011] Step 4: For high-confidence pseudo-labels, regularization terms are introduced in the maximum average difference using high-confidence pseudo-labels to achieve conditional distribution alignment and marginal distribution alignment between the source domain and the target domain. The strength of distribution alignment is dynamically adjusted according to the training stage to avoid negative transfer. At the same time, the center loss function is used to bring the center of known category samples closer to maintain the distance between unknown samples and known categories.

[0012] Step 5: By adding unknown class outputs in the classifier, the detected unknown samples are uniformly classified to achieve the diagnosis of known fault categories and the identification of unknown fault categories under variable working conditions.

[0013] 2. According to the method of claim 1, the fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism is characterized in that the step 2 comprises the following specific steps:

[0014] Step 2.1: First, the convolution layer extracts local convolution features by performing a dot product operation between the convolution kernel and the corresponding signal segment in the sliding window; then the extracted features are input into the activation layer, where the nonlinear activation function ReLU(·) is used to perform a nonlinear transformation operation, thereby realizing nonlinear feature extraction. The corresponding formula is as follows:

[0015]

[0016] in, is the hth output feature map of the lth layer, which serves as the input feature map of the next layer. is the g-th input feature map of the l-1th layer. The input feature map is the original signal and are the weight and bias of the hth convolution kernel in the lth layer, * is the convolution operation, ReLU(·) is the nonlinear activation function, p and q are the number of input feature maps in the l-1th layer and the number of output feature maps in the lth layer, respectively;

[0017] Step 2.2: Then, the output obtained by the above steps is sent to the Max Pooling layer to reduce the feature dimension. The formula corresponding to the pooling layer is:

[0018]

[0019] in, is the output value of the rth position in the hth pooling core of the lth layer, is the hth input feature map of the l-1th layer in the pooling area R r The value inside;

[0020] Step 2.3: Finally, the output of the pooling layer Input to the batch norm layer to complete the extraction of source domain and target domain features. The corresponding formula for batch normalization is as follows:

[0021]

[0022] in, is the output of the pooling layer, μ and σ 2 They are respectively The mean and variance of are used for normalization and standardization. ε is used to prevent the denominator from being zero and is usually set to 10 -5 , is a learnable scaling parameter used to adjust the distribution range of the normalized features, and β is a learnable bias parameter used to adjust the offset of the normalized features.

[0023] 3. According to the method of claim 1, the fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism is characterized in that the step 3 comprises the following specific steps:

[0024] Step 3.1: The global classifier performs confidence assessment on the target domain sample. Given a target domain sample x t , the global classifier will be C G Output a vector [C G (x t ,1),C G (x t ,2),...,C G (x t ,k)], each value represents the confidence that the sample belongs to the corresponding category, and the corresponding formula is:

[0025] C G (x t ,c)=G(x t )[c],c∈{1,2,...,k}

[0026] Among them, G(x t )[c] is the global classifier for sample x t In the output of , the confidence that the sample belongs to category c, k is the total number of known categories in the source domain;

[0027] Step 3.2: The local classifier consists of a binary classifier trained for each type under the source domain. Each sub-binary classifier outputs a probability for each target domain sample to measure the possibility that the target domain sample belongs to each known category. Then given the target domain sample x t , the corresponding local classifier C L The overall output is a probability vector [p1(x t ),p2(x t ),...,p k (x t )], reflecting the probability value of the sample belonging to class c, the corresponding formula is:

[0028] P L (x t )=C L (x t ,c),c∈{1,2,...,k}

[0029] Among them, P L (x t ) is the local classifier for the target domain sample x t In the output, the probability that the sample belongs to category c, C L is a local binary classifier of category c;

[0030] Step 3.3: Through global and local confidence analysis of the target domain samples, we get the target domain sample x t The comprehensive confidence value C t (x t ), the corresponding formula is:

[0031]

[0032] Among them, α and β are weight parameters, which are used to balance the contribution of the global classifier and the local classifier;

[0033] Step 3.4: Sort Rank(x) in descending order t ), the higher the ranking, the more likely it is that it is a known category, and the lower the ranking, the more likely it is that it is an unknown category. The corresponding formula is:

[0034] Rank(x t )=argsort(C t (x t ),descending)

[0035] Among them, argsort(,descending) is to sort the input values ​​according to the confidence level C t (x t ) in order from high to low and returns the sorted index.

[0036] 4. According to the method of fault diagnosis under variable working condition invisible type of multi-classifier decision mechanism based on pseudo label learning in claim 1, it is characterized in that in step 3, the relational expression of the loss function of the global classifier is expressed as:

[0037]

[0038] Among them, L cls is a source domain sample The cross entropy loss of the trained global classifier, N s is that the source domain contains N samples in total, K is the total number of fault categories, is a source domain sample The real label, is the global classifier for the sample The predicted probability of belonging to the Kth class;

[0039] The relational expression of the loss function of the local classifier is expressed as:

[0040]

[0041] Among them, L LOS is the cross entropy loss of the trained local classifier, C s is the total number of source domain categories, The category c of the source domain contains a total of N samples, is the set of source domain samples of category c, is the indicator function for binary classification, is the local classifier prediction sample The probability of belonging to class c.

[0042] 5. According to the method of fault diagnosis under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism of claim 1, it is characterized in that in the step 4, the relationship of the conditional distribution alignment loss function is as follows:

[0043]

[0044] Among them, L CMMD is the conditional distribution alignment loss function, and Represents the number of samples of category c in the source domain and target domain, respectively, i s,c and denote the features of samples of category c in the source domain and target domain respectively, k(·,·) is the kernel function, λ||θ|| 2 is the regularization term;

[0045] The relationship between the marginal distribution alignment loss function is as follows:

[0046]

[0047] Among them, L MMMD is the conditional distribution alignment loss function, n s and n t Represents the number of samples in the source domain and the target domain respectively, and f i s and denote the features of source and target domain samples respectively, k(·,·) is the kernel function, λθ 2 is the regularization term;

[0048] The relationship of the center loss function is as follows:

[0049]

[0050] Among them, L center is the center loss function, C is the total number of known categories, n r and n s are the number of samples of known category c and unknown category u, respectively. and are the features of the rth sample in the known category c, is the feature of the sth sample in the unknown category u, μ c is the center point of the known category c in the feature space, μ u is the center point of the unknown category u in the feature space, and γ is the weight coefficient.

[0051] 6. According to claim 4 or 5, a fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism is characterized in that the relationship of the total loss function is as follows:

[0052] L total =L CLC +L LOC +L CMMD +L MMMD +L center

[0053] 7. A fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism, characterized by:

[0054] The system includes an actuator, which is used to execute the fault diagnosis method under variable working condition invisible type based on a multi-classifier decision mechanism of pseudo-label learning as described in any one of claims 1-6.

[0055] According to another aspect of the present invention, a fault diagnosis method under variable operating conditions and invisible types of multi-classifier decision mechanisms based on pseudo-label learning is provided, characterized in that the method utilizes an actuator to execute the above-mentioned fault diagnosis method under variable operating conditions and invisible types of multi-classifier decision mechanisms based on pseudo-label learning.

[0056] In general, the technical solution proposed by the present invention has the following beneficial effects compared with the prior art:

[0057] 1. The present invention constructs a multi-classifier prediction pseudo-label mechanism, which includes a global classifier and a local classifier. The global classifier predicts the fault label probability of the target domain sample, and the local classifier predicts its probability under each source domain type. The confidence ranking of the two is combined to generate a high-confidence pseudo-label; the low-confidence pseudo-label is marked as an unknown fault type.

[0058] 2. The present invention obtains high-confidence pseudo-labels by constructing a multi-classifier prediction mechanism for high-confidence pseudo-labels, introduces regularization terms in the alignment of marginal distribution and conditional distribution, and dynamically adjusts the distribution strength to avoid the occurrence of negative transfer.

[0059] 3. The present invention constructs a center function and introduces a center loss function in the feature space to bring the center of the known fault category samples closer and maintain the distance between the unknown fault samples and the known fault categories. By adding the unknown class output in the classifier, the detected unknown samples are uniformly classified, realizing the diagnosis of known fault categories and the identification of unknown fault categories under variable working conditions.

[0060] 4. The present invention provides a fault diagnosis method under variable operating conditions and invisible types based on a multi-classifier decision mechanism of pseudo-label learning. A high-confidence pseudo-label is obtained through a multi-classifier prediction pseudo-label mechanism, and a regularization term is introduced to dynamically adjust the degree of feature distribution alignment. The center function is used to narrow the center of the known fault category and maintain the distance between the known fault category and the unknown fault category, thereby realizing the diagnosis of known fault categories and the identification of unknown fault categories under variable operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 . Overall framework diagram of the present invention

[0062] Figure 2 .Feature extractor diagram

[0063] Figure 3 . Flowchart of the present invention

[0064] Figure 4 . Schematic diagram of the relationship between the classification loss function and the number of iterations in the training process provided by the implementation case of the present invention

[0065] Figure 5. Schematic diagram of the relationship between the number of correctly classified samples and the number of iterations in the training process of Task 2 provided in the implementation case of the present invention

[0066] Figure 6 . Schematic diagram of the relationship between the accuracy of Task 2 and the number of iterations in the training process provided by the implementation case of the present invention

[0067] Figure 7 . Confusion matrix diagram of the diagnosis results of Task 2 in the implementation case of the present invention DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention will be further explained in detail below in conjunction with the accompanying drawings and preferred implementation examples. It should be understood that the specific examples described herein are only used to explain the present invention and are not intended to limit the present invention.

[0069] The detailed steps of the present invention are as follows:

[0070] Step 1: The implementation method of the present invention constructs labeled source domain training sets based on the data samples of the implementation case and the unlabeled target domain test set in, is the i-th sample in the source domain training set, is the label of the i-th sample in the source domain training set, is the jth sample in the target domain test set, N s and N t are the number of samples in the source domain and the target domain, respectively.

[0071] The present invention uses an open SEU reduction gearbox dataset as an implementation case for illustration. The SEU reduction gearbox dataset is collected from a transmission system dynamic simulator (DDS) and is collected under two different working conditions where the speed system load is set to 20HZ-0V or 30HZ-2V. The different fault types of the bearing are detailed in Table 1. The dataset contains 4 gear fault states and 1 health state (Health). The gear fault states are missing tooth fault, surface fault, root fault, and clipped tooth fault, and the sampling frequency is 5120HZ.

[0072] The implementation method of the present invention defines a combination of speed and load as a working condition. For each health condition of each working condition, 1000 samples with a data length of 1024 are collected for experiments. 75% of the samples are used for training and 25% of the samples are used for testing. The samples of one working condition are set as source domain data, and the samples of other working conditions are set as target domain data. Two working conditions A and B are defined, as shown in Table 1.

[0073] Table 1: Classification of different working conditions

[0074]

[0075] Fault diagnosis is performed through the vibration signal of the reduction gearbox dataset. Working condition A is selected as the source domain working condition, and working condition B is selected as different target domain working conditions. Working condition A-working condition B means that the source domain working condition is A and the target domain working condition is B. The source domain dataset A corresponds to data of four types of faults, including normal, fault 1, fault 2, and fault 3 at a speed of 1200rpm. The target domain dataset B corresponds to data of five types of gearboxes, including normal, fault 1, fault 2, fault 3, and unknown fault at a speed of 1800rpm. Among them, there are no training samples of unknown fault data types in the source domain dataset. By analogy, a total of ten cross-condition fault diagnosis tasks are designed, as shown in Table 2.

[0076] Table 2: Division of cross-operating fault diagnosis tasks

[0077]

[0078] Step 2: In the implementation of the present invention, based on Figure 2 The convolutional neural network shown in the figure extracts features, combines the global classifier and the local classifier to train the source domain features, and establishes a fault diagnosis model for invisible fault types under variable working conditions.

[0079] Step 2.1: First, the convolution layer extracts local convolution features by performing a dot product operation between the convolution kernel and the corresponding signal segment in the sliding window; then the extracted features are input into the activation layer, where the nonlinear activation function ReLU(·) is used to perform a nonlinear transformation operation, thereby realizing nonlinear feature extraction. The corresponding formula is as follows:

[0080]

[0081] in, is the hth output feature map of the lth layer, which serves as the input feature map of the next layer. is the g-th input feature map of the l-1th layer. The input feature map is the original signal and are the weight and bias of the hth convolution kernel in the lth layer, * is the convolution operation, ReLU(·) is the nonlinear activation function, p and q are the number of input feature maps in the l-1th layer and the number of output feature maps in the lth layer, respectively;

[0082] Step 2.2: Then, the output obtained by the above steps is sent to the Max Pooling layer to reduce the feature dimension. The formula corresponding to the pooling layer is:

[0083]

[0084] in, is the output value of the rth position in the hth pooling core of the lth layer, is the hth input feature map of the l-1th layer in the pooling area R r The value inside;

[0085] Step 2.3: Finally, the output of the pooling layer Input to the batch norm layer to complete the extraction of source domain and target domain features. The corresponding formula for batch normalization is as follows:

[0086]

[0087] in, is the output of the pooling layer, μ and σ 2 They are respectively The mean and variance of are used for normalization and standardization. ε is used to prevent the denominator from being zero and is usually set to 10 -5 , is a learnable scaling parameter used to adjust the distribution range of the normalized features, and β is a learnable bias parameter used to adjust the offset of the normalized features.

[0088] Step 3: Obtain the features of the source domain and target domain through the above steps and input them into Figure 1 The multi-classifier shown is composed of a global classifier and a local classifier, and the global classifier and the local classifier are trained using the source domain features extracted by the feature extractor.

[0089] Step 3.1: The global classifier performs confidence assessment on the target domain sample. Given a target domain sample x t , the global classifier will be C G Output a vector [C G (x t ,1),C G (x t ,2),...,C G (x t ,k)], each value represents the confidence that the sample belongs to the corresponding category, and the corresponding formula is:

[0090] C G (x t ,c)=G(x t )[c],c∈{1,2,...,k}

[0091] Among them, G(x t )[c] is the global classifier for sample x t In the output of , the confidence that the sample belongs to category c, k is the total number of known categories in the source domain;

[0092] Step 3.2: The local classifier consists of a binary classifier trained for each type under the source domain. Each sub-binary classifier outputs a probability for each target domain sample to measure the possibility that the target domain sample belongs to each known category. Then given the target domain sample x t , the corresponding local classifier C L The overall output is a probability vector [p1(x t ),p2(x t ),...,p k (x t )], reflecting the probability value of the sample belonging to class c, the corresponding formula is:

[0093] P L (x t )=C L (x t ,c),c∈{1,2,...,k}

[0094] Among them, P L (x t ) is the local classifier for the target domain sample x t In the output, the probability that the sample belongs to category c, C L is a local binary classifier of category c;

[0095] Step 3.3: Through global and local confidence analysis of the target domain samples, we get the target domain sample x t The comprehensive confidence value C t (x t ), the corresponding formula is:

[0096]

[0097] Among them, α and β are weight parameters, which are used to balance the contribution of the global classifier and the local classifier;

[0098] Step 3.4: Sort Rank(x) in descending order t ), the higher the ranking, the more likely it is that it is a known category, and the lower the ranking, the more likely it is that it is an unknown category. The corresponding formula is:

[0099] Rank(x t )=argsort(C t (x t ),descending)

[0100] Among them, argsort(,descending) is to sort the input values ​​according to the confidence level C t (x t ) in order from high to low and returns the sorted index.

[0101] Step 4: Obtain high-confidence pseudo labels through the above steps, and use the high-confidence pseudo labels to introduce regularization terms in the maximum average difference, so as to achieve conditional distribution alignment and marginal distribution alignment between the source domain and the target domain. Dynamically adjust the distribution alignment strength according to the training stage to avoid negative transfer. The corresponding conditional distribution alignment and marginal distribution alignment loss functions are as follows:

[0102]

[0103] Among them, L CMMD is the conditional distribution alignment loss function, and Represents the number of samples of category c in the source domain and target domain, respectively, i s,c and denote the features of samples of category c in the source domain and target domain respectively, k(·,·) is the kernel function, λ||θ|| 2 is the regularization term;

[0104] The relationship between the marginal distribution alignment loss function is as follows:

[0105]

[0106] Among them, L MMMD is the conditional distribution alignment loss function, n s and n t Represents the number of samples in the source domain and the target domain respectively, and f i s and denote the features of source and target domain samples respectively, k(·,·) is the kernel function, λ||θ|| 2 is the regularization term;

[0107] At the same time, the central loss function is introduced to pull the distribution center of known category samples closer in the feature space and maintain the distance between unknown samples and known categories. The corresponding central loss function is as follows:

[0108]

[0109] Among them, L center is the center loss function, C is the total number of known categories, n r and n s are the number of samples of known category c and unknown category u, respectively. and are the features of the rth sample in the known category c, is the feature of the sth sample in the unknown category u, μ c is the center point of the known category c in the feature space, μ u is the center point of the unknown category u in the feature space, and γ is the weight coefficient.

[0110] Finally, the detected unknown samples are uniformly classified by adding unknown class outputs in the classifier.

[0111] Step 5: Based on the proposed multi-classifier decision mechanism under pseudo-label learning, the training data set divided in the implementation case is used to train the model's diagnostic capabilities for invisible fault types and known fault types under variable working conditions. The total classification loss function of the model during the training process changes with the number of iterations as shown in the figure below: Figure 4 As shown, in this embodiment, the maximum training iteration round is set to 50. After all training iteration rounds are completed, a trained model is obtained.

[0112] In the implementation case, a variable working condition fault diagnosis problem from the source domain working condition to the target domain working condition is set. The source domain dataset A corresponds to data of four types of faults, namely normal, fault 1, fault 2, and fault 3, at a speed of 1200rpm. The target domain dataset B corresponds to five types of gearbox data, namely normal, fault 1, fault 2, fault 3, and unknown fault, at a speed of 1800rpm. Among them, the source domain dataset lacks training samples of unknown fault data types. By analogy, a total of ten cross-working condition fault diagnosis tasks are designed. Here we take Task 2 as an example. The schematic diagrams of the curves of the number of correctly classified samples and the accuracy rate with the number of iterations are shown in Figure 2. Figure 5 and Figure 6 As shown, the confusion matrix of the diagnosis results is as follows Figure 7 As shown, only 29 samples of unknown fault type were misidentified as fault type 2.

Claims

1. A fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism, characterized in that: The following steps are involved: Step 1: Collect data from the industrial production process through sensors, divide the industrial data under different working conditions into source domain and target domain, and define the source domain training set and target domain test set as in, is the i-th sample in the source domain training set, is the label of the i-th sample in the source domain training set, is the jth sample in the target domain test set, N s and N t are the number of samples in the source domain and the target domain respectively; Step 2: Use the data obtained in step 1 to input the feature extractor shown in Figure 2. The constructed feature extractor is composed of three identical feature extraction units connected in series, each of which includes a one-dimensional convolutional layer, an activation layer, a pooling layer, and a batch normalization layer; Step 3: Use the source domain features extracted by the feature extractor to train the global classifier and local classifier as shown in Figure 1. The global classifier predicts the probability of the fault label of the target domain sample, and the local classifier predicts its probability under each source domain type. Combine the confidence ranking of the two to generate a pseudo label. Through the global and local confidence analysis of the target domain samples, they are sorted in a sequence from high to low. The higher the ranking, the more likely it is a known category, and the lower the ranking, the more likely it is an unknown category. Step 4: For high-confidence pseudo-labels, regularization terms are introduced in the maximum average difference using high-confidence pseudo-labels to achieve conditional distribution alignment and marginal distribution alignment between the source domain and the target domain. The strength of distribution alignment is dynamically adjusted according to the training stage to avoid negative transfer. At the same time, the center loss function is used to bring the center of known category samples closer to maintain the distance between unknown fault categories and known fault categories. Step 5: By adding unknown class outputs in the classifier, the detected unknown samples are uniformly classified to achieve the diagnosis of known fault categories and the identification of unknown fault categories under variable working conditions.

2. According to claim 1, a fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism is characterized in that: The step 2, specifically comprises the following steps: Step 2.1: First, the convolution layer extracts local convolution features by performing a dot product operation between the convolution kernel and the corresponding signal segment in the sliding window; then the extracted features are input into the activation layer, where the nonlinear activation function ReLU(·) is used to perform a nonlinear transformation operation, thereby realizing nonlinear feature extraction. The corresponding formula is as follows: in, is the hth output feature map of the lth layer, which serves as the input feature map of the next layer. is the g-th input feature map of the l-1th layer. The input feature map is the original signal and are the weight and bias of the hth convolution kernel in the lth layer, * is the convolution operation, ReLU(·) is the nonlinear activation function, p and q are the number of input feature maps in the l-1th layer and the number of output feature maps in the lth layer, respectively; Step 2.2: Then, the output obtained by the above steps is sent to the maximum pooling layer MaxPooling to reduce the feature dimension. The formula corresponding to the pooling layer is: in, is the output value of the rth position in the hth pooling core of the lth layer, is the hth input feature map of the l-1th layer in the pooling area R r The value inside; Step 2.3: Finally, the output of the pooling layer Input to the batch norm layer to complete the extraction of source domain and target domain features. The corresponding formula for batch normalization is as follows: in, is the output of the pooling layer, μ and σ 2 They are respectively The mean and variance of are used for normalization and standardization. ε is used to prevent the denominator from being zero and is usually set to 10 -5 , is a learnable scaling parameter used to adjust the distribution range of the normalized features, and β is a learnable bias parameter used to adjust the offset of the normalized features.

3. According to claim 1, the fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism is characterized in that: The step 3 described above comprises the following specific steps: Step 3.1: The global classifier performs confidence assessment on the target domain sample. Given a target domain sample x t , the global classifier will be C G Output a vector [C G (x t ,1),C G (x t ,2),...,C G (x t ,k)], each value represents the confidence that the sample belongs to the corresponding category, and the corresponding formula is: C G (x t ,c)=G(x t )[c],c∈{1,2,...,k} Among them, G(x t )[c] is the global classifier for sample x t In the output of , the confidence that the sample belongs to category c, k is the total number of known categories in the source domain; Step 3.2: The local classifier consists of a binary classifier trained for each type under the source domain. Each sub-binary classifier outputs a probability for each target domain sample to measure the possibility that the target domain sample belongs to each known category. Then given the target domain sample x t , the corresponding local classifier C L The overall output is a probability vector [p1(x t ),p2(x t ),...,p k (x t )], reflecting the probability value of the sample belonging to class c, the corresponding formula is: P L (x t )=C L (x t ,c),c∈{1,2,...,k} Among them, P L (x t ) is the local classifier for the target domain sample x t In the output, the probability that the sample belongs to category c, C L is a local binary classifier of category c; Step 3.3: Through global and local confidence analysis of the target domain samples, we get the target domain sample x t The comprehensive confidence value C t (x t ), the corresponding formula is: Among them, α and β are weight parameters, which are used to balance the contribution of the global classifier and the local classifier; Step 3.4: Sort Rank(x) in descending order t ), the higher the ranking, the more likely it is that it is a known category, and the lower the ranking, the more likely it is that it is an unknown category. The corresponding formula is: Rank(x t )=argsort(C t (x t ),descending) Among them, argsort(,descending) is to sort the input values ​​according to the confidence level C t (x t ) in order from high to low and returns the sorted index.

4. The fault diagnosis method under variable working condition invisible type based on a multi-classifier decision mechanism of pseudo-label learning according to claim 1 is characterized in that: In step 4, the relational expression of the loss function of the global classifier is expressed as: Among them, L cls is a source domain sample The cross entropy loss of the trained global classifier, N s is that the source domain contains N samples in total, K is the total number of fault categories, is a source domain sample The real label, is the global classifier for the sample The predicted probability of belonging to the Kth class; The relational expression of the loss function of the local classifier is expressed as: Among them, L LOS is the cross entropy loss of the trained local classifier, C s is the total number of source domain categories, The category c of the source domain contains a total of N samples, is the set of source domain samples of category c, is the indicator function for binary classification, is the local classifier prediction sample The probability of belonging to class c.

5. The fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism according to claim 1 is characterized in that: In step 4, the relationship between the conditional distribution alignment loss function is as follows: Among them, L CMMD is the conditional distribution alignment loss function, and Represents the number of samples of category c in the source domain and target domain, respectively, i s,c and denote the features of samples of category c in the source domain and target domain respectively, k(·,·) is the kernel function, λ||θ|| 2 is the regularization term; The relationship between the marginal distribution alignment loss function is as follows: Among them, L MMMD is the conditional distribution alignment loss function, n s and n t Represents the number of samples in the source domain and the target domain respectively, and f i s and denote the features of source and target domain samples respectively, k(·,·) is the kernel function, λ||θ|| 2 is the regularization term; The relationship of the center loss function is as follows: Among them, L center is the center loss function, C is the total number of known categories, n r and n s are the number of samples of known category c and unknown category u, respectively. and are the features of the rth sample in the known category c, is the feature of the sth sample in the unknown category u, μ c is the center point of the known category c in the feature space, μ u is the center point of the unknown category u in the feature space, and γ is the weight coefficient.

6. The fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism according to claim 4 or 5, characterized in that: The relationship of the total loss function is as follows: L total =L CLC +L LOC +L CMMD +L MMMD +L center 7. A fault diagnosis method under variable working condition invisible type based on pseudo label learning multi-classifier decision mechanism, characterized by: The system includes an actuator, which is used to execute the fault diagnosis method under variable working condition invisible type based on a multi-classifier decision mechanism of pseudo-label learning as described in any one of claims 1-6.

Images