Trusted passive field adaptive mechanical equipment fault diagnosis method based on model calibration

Abstract

The invention provides a credible passive field adaptive mechanical equipment intelligent diagnosis method based on model calibration, and the method comprises the following steps: S1, installing a vibration acceleration sensor on the surface of mechanical equipment, and collecting the vibration monitoring data of different fault types of the mechanical equipment under different working conditions; s2, performing normalization preprocessing on the vibration monitoring data; S3, constructing a credible passive field adaptive mechanical equipment fault diagnosis model based on model calibration; comprising the steps of establishing an initial model, generating a source model, establishing an adaptive fault diagnosis model and calibrating the adaptive fault diagnosis model. And S4, using the adaptive fault diagnosis model to diagnose the fault type of the mechanical equipment. According to the invention, vibration monitoring data is directly used as input, and end-to-end cross-working-condition intelligent fault diagnosis is realized; according to the method, determinant-based mutual information is introduced to realize robust model cross-domain adaptation; the method also establishes a temperature scaling theory based on a target simulation data set, and improves the reliability of the prediction confidence.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to a trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration. Background Art

[0002] As mechanical equipment develops towards higher automation and intelligence, the importance of its safe operation becomes increasingly prominent. Especially for those safety-critical components that ensure the continuous operation of the equipment, they are prone to degradation or even failure due to frequent exposure to harsh working conditions such as high speed, high temperature, high pressure and load fluctuations. Therefore, timely fault diagnosis is not only crucial to ensure the safe operation of the equipment, but also directly related to the life safety of the staff.

[0003] Traditional fault diagnosis methods mainly rely on existing expert diagnostic knowledge, but often face great challenges when dealing with fault detection of complex equipment. With the changes in industrial needs, there is an increasing demand for automated diagnostic methods to shorten maintenance cycles and improve diagnostic accuracy. With the continuous development of artificial intelligence theory, the fault diagnosis process is evolving towards intelligent solutions to achieve automatic detection and identification of equipment fault types. In particular, fault diagnosis methods based on deep learning have attracted widespread attention from scholars because of their ability to adaptively extract information related to faults.

[0004] Although the research on intelligent diagnosis based on deep learning has made significant progress in recent years, its application still relies on the training data and test data satisfying the independent and identically distributed assumption. However, differences in equipment operating conditions and service environments will inevitably lead to different data distributions. To address this challenge, the domain adaptation-based intelligent diagnosis (DA-ID) method has gradually attracted attention. This method captures cross-domain shared diagnostic information through the feature alignment mechanism between the labeled data in the source domain and the unlabeled data in the target domain, thereby realizing fault diagnosis under different data distributions. Traditional DA-ID methods mainly include domain adaptation based on statistical metrics and domain adaptation based on adversarial games.

[0005] Although DA-ID has achieved certain results in cross-domain fault diagnosis, the method still faces a core challenge: the availability of source domain annotated data is required during the adaptation process. This requirement is in conflict with engineering practice, mainly in two aspects. First, in order to protect commercial privacy, many industrial fields (such as aircraft engine and wind turbine fault diagnosis) have natural restrictions on access to source data. Second, the requirements of real-time diagnosis make the challenges of data storage and transmission make real-time access to source domain datasets infeasible. Therefore, cross-domain fault diagnosis in passive scenarios has gradually attracted widespread attention from researchers. However, existing research on source-independent adaptive fault diagnosis (SFAFD) is still relatively limited, and most of them focus on improving diagnostic accuracy, while ignoring the security issue of calibrating prediction uncertainty, a key information in unsupervised adaptation environments. Ignoring this issue makes it difficult for maintenance personnel to trust the prediction uncertainty in SFDAD applications. For example, if manual review is skipped due to excessive confidence in fault type identification, the fault may not be discovered in time, resulting in catastrophic consequences; conversely, too high uncertainty diagnosis results often trigger false alarms, leading to repeated inspections, thereby wasting resources. Therefore, the calibration analysis of the SFAFD model is crucial to ensure the safe application of neural network models in mechanical fault diagnosis, which is still a gap in existing research. Summary of the invention

[0006] In view of the above key challenges, this paper proposes a credible source-free adaptation diagnosis method for mechanical equipment with calibration (CSFADC). According to the fault classification model of mechanical equipment corresponding to different acquired signals under known working conditions, the fault classification of mechanical equipment corresponding to different acquired signals under unknown working conditions is obtained. CSFADC consists of three stages: source model generation, model adaptation, and model calibration. In the source model generation stage, the parameters of feature encoder and classifier modules are optimized by combining label smoothing supervision loss function and sharpness-aware minimization optimization strategy. This process maximizes the generalization ability of the diagnosis model while making full use of source domain information. In the model adaptation stage, the source domain classification information is retained by fixing the classifier parameters. Then, pseudo-label learning is used to update the parameters of feature encoder module from unlabeled target data to achieve model adaptation. In order to mitigate the adverse effects of erroneous pseudo-labels on model performance, this paper draws inspiration from information theory loss function and adopts deterministic mutual information to enhance the robustness to noisy labels in the adaptation process. Finally, in the model calibration phase, it is assumed that the diagnostic model has fully learned the target domain information. In order to achieve model calibration using only the unlabeled target domain dataset, a hybrid technique is used to create a target simulation dataset in the model transfer phase. Based on the target simulation dataset, the temperature scaling theory is applied to achieve model calibration in the passive scenario.

[0007] In order to achieve the above object, the present invention provides a trusted passive domain adaptation mechanical fault diagnosis method based on model calibration, comprising the following steps:

[0008] S1: Collect vibration monitoring data of different fault types of mechanical equipment under different working conditions;

[0009] Install a vibration acceleration sensor on the surface of the mechanical equipment, collect vibration monitoring data of the mechanical equipment under different fault types under a first working condition, where there are K types of fault types, and collect vibration monitoring data of the mechanical equipment under a second working condition;

[0010] The vibration monitoring data obtained under the first working condition is used as the source domain data set, and the label information is the category of the fault type; the vibration monitoring data obtained under the second working condition is used as the target domain data set, which does not contain label information;

[0011] S2: normalize and preprocess the vibration monitoring data;

[0012] The vibration monitoring data obtained under different working conditions are respectively normalized and preprocessed to obtain corresponding normalized data;

[0013] S3: Construct a trusted passive domain adaptation mechanical equipment fault diagnosis model based on model calibration;

[0014] Constructing a trusted passive domain adaptation mechanical equipment fault diagnosis model based on model calibration mainly includes the following sub-steps:

[0015] S31: Build initial model

[0016] Initial Model D Base A feature encoding module E Base and a classification module C Base The feature encoding module E Base Used to extract sensitive features, the output dimension is d, the classification module C Base The number of categories is K, and the output is a K-dimensional vector consisting of the probability of each category;

[0017] S32: Source model generation

[0018] Relying on the normalized source domain dataset samples and their corresponding labels, the label smoothing supervision loss function and sharpness perception minimization optimization strategy are used to train the initial model to obtain the source model;

[0019] S33: Establishing an adaptive fault diagnosis model

[0020] Use prototype-based pseudo-label learning technology to generate normalized target domain dataset samples x t The corresponding first pseudo label, and then constructing a new type of class prototype using the first pseudo label data to generate a second pseudo label;

[0021] Using the target domain dataset with the second pseudo-label, the source model is trained based on DMI, the classification module parameters of the source model are fixed, and only the feature encoding module is updated to obtain an adapted fault diagnosis model;

[0022] S34: Adaptive Fault Diagnosis Model Calibration

[0023] The target simulation model calibration strategy is adopted. The target simulation temperature parameter T is obtained through the target simulation data set. The prediction confidence of the adaptive fault diagnosis model can be calibrated by using the temperature scaling strategy.

[0024] S4: Use the adapted fault diagnosis model to diagnose the fault type of mechanical equipment;

[0025] The vibration monitoring data of the mechanical equipment under the second working condition is collected and input into the adaptive fault diagnosis model. The adaptive fault diagnosis model outputs the probability of each fault type and obtains the calibrated prediction confidence according to S34, so as to diagnose the fault type of the mechanical equipment.

[0026] Preferably, the mechanical equipment is a multi-stage transmission system.

[0027] Preferably, in S1, there are K types of fault types, specifically:

[0028] The fault types are the fault types of rolling bearings in a multi-stage transmission system, and there are 7 types in total, namely: normal state, inner ring crack fault, inner ring wear fault, inner and outer ring crack fault, outer ring crack fault, outer ring wear fault and cage crack fault.

[0029] Preferably, the specific structure of the initial model in S31 is:

[0030] Feature encoding module E Base The convolution layer, batch normalization layer, nonlinear activation function, and pooling layer are connected in sequence as a group. After four groups are repeated in sequence, the Dropout layer, fully connected layer, batch normalization, and nonlinear activation function are connected in sequence. The classification module C Base It consists of a Dropout layer and a fully connected layer. When used, the normalized vibration monitoring data is used as the input of the feature encoding module to extract sensitive features, and then the extracted sensitive features are input into the classification module to realize fault type identification.

[0031] Preferably, when the label smoothing supervision loss function and the sharpness perception minimization optimization strategy are used to train the initial model to obtain the source model in S32, the expression of the source model objective function is as follows:

[0032]

[0033] Among them, w d is the source model parameter, which is the model parameter to be trained, and the initial value is the initial model parameter; ζ represents the perturbation randomly generated for the model parameter, ||·||2 represents the L2 norm, ρ is the hyperparameter of the neighborhood range, ρ≥0; represents the definition, α is the weight;

[0034]

[0035] Among them, w d is the source model parameter, x s is the source domain dataset sample, y s is the corresponding true label of the source domain dataset sample; Indicates that x s As the initial model D Base The vector obtained by input is To vector After softmax, the kth element in the output category K is obtained. ε is the smoothing parameter. q k is the intermediate parameter, represents the expectation obtained by sampling from the source domain dataset.

[0036] Preferably, in S33, a prototype-based pseudo-label learning technique is used to generate a normalized target domain data set sample x t The corresponding first pseudo label is then used to construct a new class prototype using the first pseudo label data, thereby generating a second pseudo label; specifically:

[0037] A prototype-based pseudo-label learning technique is used, where the prototype of each class is the centroid of the features of that class. Calculate the cosine distance between each sample and each category prototype in the unlabeled target domain dataset, so as to assign the sample to the nearest category to generate the first pseudo label;

[0038] A new class prototype was constructed using the first pseudo-labeled data Thus generating the second pseudo label The expression of this process is as follows:

[0039]

[0040] in, is the target domain dataset sample x t The first pseudo label of t represents the target domain dataset; when the condition When it is established, Ind(·) is equal to 1, otherwise, Ind(·) is equal to 0, x t is the target domain dataset sample, Indicates that x t The sensitive features obtained after being input into the feature encoding module of the source model, Indicates that x t The vector obtained as input to the source model, Represents the vector After softmax output, the kth element in category K is Represents the prototype of the kth class; k represents the predicted category of the target domain dataset sample.

[0041] Preferably, in S33, the target domain dataset with the second pseudo label is used to train the source model based on the DMI, the classification module parameters of the source model are fixed, and only the feature encoding module is updated to obtain the adapted fault diagnosis model; specifically:

[0042] The adaptive fault diagnosis model adopts the following objective function:

[0043]

[0044] in, express The noisy version of express and The joint distribution of express The matrix format of , det(·) represents the calculation of the matrix determinant. Specifically, The empirical estimates are as follows:

[0045]

[0046] in, is the expectation operation, Pr[·] represents the probability distribution, is the noise version of k, for The noisy version of is the second pseudo label of the i-th sample in the target domain dataset, and N represents the total number of samples in the target domain dataset;

[0047] set up Substituting formula (10) into formula (9), we can obtain the empirical estimation loss function of DMI:

[0048]

[0049] Among them, M is composed of The matrix composed of Indicates row and column subscripts,

[0050] The minimum value of formula (11) is used as the objective function of the adaptive fault diagnosis model, the classification module parameters of the source model are fixed, only the feature encoding module is updated, and the generation of the adaptive fault diagnosis model is completed using the stochastic gradient descent algorithm.

[0051] Preferably, in S34, a target simulation model calibration strategy is adopted to obtain a target simulation temperature parameter through a target simulation data set. The temperature scaling strategy can be used to calibrate the prediction confidence of the adaptive fault diagnosis model; specifically:

[0052] A target simulation dataset is introduced. The target simulation temperature parameters are estimated using this data set As follows:

[0053]

[0054] in, represents the negative log-likelihood function, that is Indicates that The vector obtained as the input of the adaptive fault diagnosis model, T is the temperature parameter, Represents the vector Output the result through the softmax output layer;

[0055] The temperature scaling strategy is used to calibrate the prediction confidence of the adaptive fault diagnosis model. The calibrated confidence can be expressed as:

[0056]

[0057] Among them, the sample The vector obtained by inputting the adapted fault diagnosis model is defined as the logical z i , Simulate temperature parameters for the target, Represents the vector After softmax, the kth element in category K is output.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] (1) The present invention can directly use vibration monitoring data as input, realizing end-to-end cross-operating condition intelligent fault diagnosis;

[0060] (2) Based on pseudo-label learning, the present invention introduces determinant-based mutual information to enhance the robustness to noisy labels in the adaptive process, thereby reducing the adverse effects of erroneous pseudo-labels on model performance and achieving robust cross-domain model adaptation;

[0061] (3) The present invention designs a hybrid technique to create a target simulation data set and combines the temperature scaling theory to achieve model calibration in passive scenarios, thereby improving the reliability of the prediction confidence of the passive adaptive fault diagnosis model;

[0062] (4) The present invention effectively addresses the practical engineering problem of inaccessibility of source domain datasets due to data privacy issues or the need to reduce the burden of data storage and transmission, and achieves high-precision passive domain adaptive mechanical fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Flowchart of the mechanical equipment fault diagnosis method for trusted passive domain adaptation based on model calibration;

[0064] Figure 2 Flowchart for building a trusted passive domain adaptation rolling bearing fault diagnosis model based on model calibration;

[0065] Figure 3 The network structure and parameter settings of the feature encoding module in the embodiment of the present invention;

[0066] Figure 4 The network structure and parameter settings of the classification module in the embodiment of the present invention;

[0067] Figure 5Schematic diagram of the multi-stage acceleration transmission experimental device in this embodiment of the invention. DETAILED DESCRIPTION

[0068] In order to better understand the technical solution of the present invention, the specific implementation of the present invention is further described in detail below in conjunction with the drawings and embodiments.

[0069] The following takes the rolling bearing fault detection in a multi-stage transmission system as an example. Figure 1 and Figure 2 As shown, the trusted passive field adaptation mechanical equipment fault diagnosis method based on model calibration of the present invention is specifically as follows:

[0070] S1: Collect vibration monitoring data of rolling bearings with different fault types under different working conditions;

[0071] The vibration acceleration sensor is installed on the surface of the multi-stage transmission system, and the vibration monitoring data of the rolling bearing of the multi-stage transmission system under different fault types under the first working condition are collected. There are K types of fault types in total. The vibration monitoring data of the rolling bearing of the multi-stage transmission system under the second working condition are collected. The vibration monitoring data obtained under the first working condition is used as the source domain data set, and the label information is the fault type; the vibration monitoring data obtained under the second working condition is used as the target domain data set, which does not contain label information.

[0072] The multi-stage transmission system includes a drive motor 1, a coupling 2, a planetary gearbox 3, a fixed axis gearbox 4, a tachometer 5, a bearing seat 6, a vibration acceleration sensor 7 and a magnetic powder brake 8, as shown in the schematic diagram. Figure 5 The power is provided by the driving motor 1, and the power is increased through the planetary gearbox 3 and the fixed axis gearbox 4, and then reaches the magnetic powder brake 8 through the bearing test base. The magnetic powder brake 8 can apply different loads to the multi-stage transmission system.

[0073] The vibration acceleration sensor is installed on the surface of the multi-stage transmission system to collect vibration monitoring data. The present invention takes the rolling bearing in the bearing seat of the multi-stage transmission system as the research object. The multi-stage transmission system used for the experiment can be injected with different rolling bearing faults, so that its fault types include normal state, inner ring crack fault, inner ring wear fault, inner and outer ring crack fault, outer ring crack fault, outer ring wear fault and cage crack fault, a total of 7 types. The vibration acceleration sensor is installed on the surface of the multi-stage transmission system to collect vibration monitoring data of different fault types of rolling bearings under different operating conditions of the multi-stage transmission system. In this application, a sampling frequency of 10kHz is used, and the system conditions are: the speed is about 1200rpm, and the loads of the conditions are 0.7 Newton meters (Nm), 1.2Nm, 2Nm, and 5Nm, respectively. The four operating conditions are represented as B1, B2, B3 and B4 respectively. The vibration acceleration signals of the multi-stage transmission system operating under four different working conditions are collected. For each operating condition of the rolling bearing, 1000 vibration monitoring data samples are collected, and each sample contains 1024 data points. These conditions together generate 12 different cross-domain diagnostic tasks, as shown in Table 1. Among them, B1→B2 means that B1 is taken as the first working condition, and the vibration monitoring data samples obtained under the first working condition B1 are taken as the source domain data set, which contains label information; B2 is taken as the second working condition, and the vibration monitoring data samples obtained under the second working condition B2 are taken as the target domain data set, and the target domain data set is considered to contain no label information.

[0074] Table 1 Bearing migration diagnosis tasks for multi-stage transmission system

[0075]

[0076] S2: normalize and preprocess the vibration monitoring data;

[0077] The vibration monitoring data collected under the first working condition is used as the source domain data set, and the vibration monitoring data collected under the second working condition is used as the target domain data set. The vibration monitoring data under different working conditions are normalized and preprocessed using formula (1):

[0078]

[0079] Wherein, x represents the vibration monitoring data, μ represents the mean value, σ represents the standard deviation, and x′ represents the normalized vibration monitoring data.

[0080] After normalization, the source domain data set consists of the normalized vibration monitoring data collected under the first working condition, and the target domain data set consists of the normalized vibration monitoring data collected under the second working condition.

[0081] S3: Construct a trusted passive domain adaptation rolling bearing fault diagnosis model based on model calibration;

[0082] The construction of a trusted passive domain adaptation rolling bearing fault diagnosis model based on model calibration mainly includes the following sub-steps:

[0083] S31: Build initial model

[0084] Initial Model D Base A feature encoding module E Base and a classification module C Base The feature encoding module E Base Used to extract sensitive features, the output dimension is d, the classification module C Base The number of categories is K, so the output is a K-dimensional vector consisting of the probability of each category; the specific structure is: feature encoding module E Base like Figure 3 As shown in the figure, the convolution layer, batch normalization layer, nonlinear activation function, and pooling layer are sequentially connected as a group. After four groups are repeated in sequence, the Dropout layer, fully connected layer, batch normalization, and nonlinear activation function are sequentially connected to form the classification module C. Base like Figure 4 As shown in FIG, it is composed of a Dropout layer and a fully connected layer. When used, the normalized vibration monitoring data is used as the input of the feature encoding module to extract sensitive features, and then the extracted sensitive features are input into the classification module to realize the identification of rolling bearing fault types.

[0085] S32: Source model generation

[0086] Relying on the source domain dataset and its corresponding labels, the supervised loss function is used as the optimization target to train the initial model to obtain the source model. In this stage, the label smoothing technique is used to reduce the impact of incorrect labels, thereby enhancing the generalization of the initial model. The resulting label smoothing cross entropy loss function is described as follows:

[0087]

[0088] Among them, w d is the source model parameter, x s is the source domain dataset sample, y s is the corresponding true label of the source domain dataset sample; Indicates that x s As the initial model D Base The vector obtained by input is To vector After the softmax (output layer), the kth element in the output category K is output. ε is the smoothing parameter, which is set to 1. k is the intermediate parameter, represents the expectation obtained by sampling from the source domain dataset.

[0089] In this embodiment, the source domain dataset sample x s It is the normalized vibration monitoring data of the first working condition B1, and the corresponding true label y of the source domain dataset sample s is the fault type corresponding to the vibration monitoring data, and there are 7 types of fault types K.

[0090] Given that in passive scenarios, the subsequent model adaptation and diagnostic procedures cannot access the source domain dataset, the necessary generalization performance of the source model becomes crucial. In view of this necessity, the sharpness-aware minimization (SAM) optimization strategy is adopted. Traditional optimization techniques only minimize the training loss, which often leads to poor model quality. In contrast, SAM strives to minimize both the loss magnitude and its sharpness, aiming to identify parameter configurations in regions characterized by consistently low losses. This approach effectively improves the generalization ability of the source model. According to the label-smoothed cross-entropy loss function of the initial model, the expression of the objective function in the source model generation stage is obtained as follows.

[0091]

[0092] Among them, w d is the source model parameter, which is the model parameter that needs to be trained, and the initial value is the initial model parameter; ζ represents the disturbance randomly generated for the model parameter, ||·||2 represents the L2 norm, ρ is the hyperparameter of the neighborhood range, ρ≥0; α is the balance weight of the two loss functions, which is set to 1 in this embodiment.

[0093] The normalized vibration monitoring data of the first working condition B1 obtained by S2 is used as the source domain data set to input the initial model of S31. Formula (4) is used as the objective function of the source model generation stage, and the stochastic gradient descent algorithm is used to complete the generation of the source model.

[0094] S33: Establishing an adaptive fault diagnosis model

[0095] The source domain dataset is inaccessible, so conventional domain adaptation methods cannot be used. To address the above challenges, the classification module parameters of the source model are fixed and only the feature encoding module is updated, so that the adapted fault diagnosis model retains the diagnostic knowledge related to the inaccessible source data.

[0096] In addition, the present invention adopts a prototype-based pseudo-label learning technique and uses the target domain dataset x t to facilitate the self-training process. The prototype of each class is the centroid of the features of that class Its mathematical expression is as follows:

[0097]

[0098] Among them, x tis the target domain dataset sample, X t represents the target domain dataset; Indicates that x t The sensitive features obtained after being input into the feature encoding module of the source model, Indicates that x t The vector obtained as input to the source model, Represents the vector After the softmax (output layer), the kth element in the output category K is output. Represents the prototype of the kth class.

[0099] In this embodiment, the target domain dataset sample x t It is the normalized vibration monitoring data of the second working condition B2 obtained in step S2.

[0100] On this basis, the cosine distance between each unlabeled target domain dataset sample and each category prototype can be calculated, so that the sample is assigned to the nearest category to generate the first pseudo label:

[0101]

[0102] in, is the target domain dataset sample x t The first pseudo label of .

[0103] In order to deal with the problem of unreliable pseudo-labels in the initial training stage, a new class prototype is constructed using the first pseudo-label data. This new class prototype is then used to refine the pseudo-labeled data, thereby generating a more reliable second pseudo-label The mathematical formula for this process is as follows:

[0104]

[0105] Among them, when the condition When it holds, Ind(·) is equal to 1, otherwise, Ind(·) is equal to 0. k represents the predicted category of the target domain dataset sample. However, in the context of working condition migration, the acquisition of pseudo labels may not be accurate enough, which makes the guarantee of label reliability a challenge. Traditional distance-based loss functions, such as cross entropy loss, are easily distorted by incorrect labels, which in turn causes the model to overfit on noisy information. Therefore, the present invention introduces the derivative of Shannon mutual information, called determinant-based mutual information (DMI). DMI ensures that even when training under noisy labels, the training process is similar to the training process using clean labels, except that there is a constant offset. Therefore, the following objective function is used when establishing an adaptive fault diagnosis model to enhance the robustness of the diagnosis model to noisy labels:

[0106]

[0107] in, express The noisy version of express and The joint distribution of express The matrix format of , det(·) represents the calculation of the matrix determinant. Specifically, The empirical estimates are as follows:

[0108]

[0109] in, is the expectation operation, Pr[·] represents the probability distribution, is the noise version of k, for The noisy version of is the second pseudo label of the i-th sample in the target domain dataset, and N represents the total number of samples in the target domain dataset.

[0110] set up Substituting formula (10) into formula (9), we can obtain the empirical estimation loss function of DMI:

[0111]

[0112] Among them, M is composed of The matrix composed of Indicates row and column subscripts,

[0113] The normalized vibration monitoring data of the second working condition B2 obtained in step S2 is used as the target domain data set to input into the source model. The minimum of formula (11) is used as the objective function of the adaptive fault diagnosis model generation stage. The classification module parameters of the source model are fixed, and only the feature encoding module is updated. The stochastic gradient descent algorithm is used to complete the generation of the adaptive fault diagnosis model.

[0114] S34: Adaptive Fault Diagnosis Model Calibration

[0115] Existing intelligent diagnosis models mainly focus on the accuracy of fault identification, but this alone is not enough to support actual decision-making in engineering. Therefore, based on this consideration, the present invention introduces the concept of model calibration to improve the credibility of prediction uncertainty and ensure that the confidence conveyed by the model can accurately reflect the uncertainty of its prediction. Therefore, the intelligent diagnosis model proposed in the present invention not only pursues accurate prediction results, but also gives priority to achieving reliable calibration. Inspired by Platt Scaling, a model calibration technology, temperature scaling has been widely used in fields such as computer vision and semantic segmentation. However, due to the inability to obtain labeled target data sets in source-independent adaptive diagnosis scenarios, traditional supervised calibration methods (such as temperature scaling) face challenges in their applicability. In order to solve this problem, the present invention proposes a target-mimic-oriented model calibration (TMOMC) strategy.

[0116] First, the present invention regards this problem as an unsupervised target-specific model calibration problem. It is well known that when there are a large number of correct predictions, the temperature parameter will be lowered to enhance confidence; while when there are a large number of incorrect predictions, the temperature parameter will be increased to slow down the prediction fluctuation. Therefore, when the dataset shows statistical similarity in correct and incorrect predictions, similar temperature parameters will be generated. To achieve this goal, the present invention introduces a target simulation dataset, which consists of pseudo-labeled mixed target samples to simulate the correct and incorrect diagnosis statistics of target data diagnosis. Given two pseudo-labeled target domain datasets from different fault type categories: and in, and The samples are and The first pseudo label of the sample and Select from the target domain data set. On this basis, the present invention can obtain a target simulation sample And build the target simulation dataset As shown in formula (12).

[0117]

[0118] The mixing ratio λ belongs to [0, 1], and is designed to be 0.75 in the present invention.

[0119] Obviously, the pseudo-labeled target samples Dominant, while is considered as a disturbance term. If If it is far from the classification boundary of the diagnostic model, It will become easier to target simulation samples On the contrary, when When it is close to the classification boundary, it may cause misjudgment, which will make and are susceptible to misclassification. This suggests that and target simulation samples The statistical characteristics of correct and incorrect diagnosis are consistent. Therefore, the present invention uses the target simulation data set Corrected to To facilitate calculation, the target simulation temperature parameters are estimated using this data set. As follows:

[0120]

[0121] in, represents the negative log-likelihood function, that is Indicates that The vector obtained as the input of the adaptive fault diagnosis model, T is the temperature parameter, Represents the vector The result is output through softmax (output layer).

[0122] According to formula (13), the value of T is

[0123] Given a sample The vector obtained by inputting the adapted fault diagnosis model is defined as the logical z i , the temperature scaling strategy is used to calibrate the prediction confidence of the adaptive fault diagnosis model, and the calibrated confidence can be expressed as:

[0124]

[0125] Among them, the sample The vector obtained by inputting the adapted fault diagnosis model is defined as the logical z i , Simulate temperature parameters for the target, Represents the vector After softmax, the kth element in category K is output.

[0126] Finally, the prediction confidence of the calibrated adaptive fault diagnosis model can realize the identification of the rolling bearing fault type and output the corresponding confidence at the same time, which the operation and maintenance personnel can use as a basis for making operation and maintenance decisions.

[0127] S4: Use the adapted fault diagnosis model to diagnose the fault type of mechanical equipment;

[0128] The vibration data of the multi-stage transmission system under the second working condition is collected and input into the adaptive fault diagnosis model. According to the classification of the model output, the rolling bearing fault is obtained, and the confidence of the classification is obtained according to formula (14). The adaptive fault diagnosis model is a trusted passive field adaptive mechanical fault diagnosis model based on model calibration.

[0129] To further demonstrate the effectiveness of the method proposed in the present invention, the data set collected in step S1 is subjected to the method proposed in the present invention and various comparison methods, and the obtained diagnostic results are shown in Table 2. i →B j " means B i As the source domain, B j as the target domain. For clarity, the best result for each task is highlighted in bold. Analyzing these findings, the following insights emerge. (1) Although CSFADC does not have access to the source domain dataset during the adaptation process, it still achieves an impressive average diagnostic accuracy of 99.42% on 12 cross-domain fault diagnosis tasks, surpassing all the compared methods. (2) Among all the compared methods, C-ID, SHOT-ID, and SFA-ID have higher accuracy than CSFADC in some cross-domain diagnosis tasks. However, the maximum accuracy gap between CSFADC and the optimal solution does not exceed 1%. (3) Despite the lack of access to the source data during the adaptation process, SHOT-ID and SFA-ID perform well on most tasks, but there are notable exceptions: SHOT-ID's performance degrades on tasks B4→B1 and B3→B1, while SFA-ID shows similar limitations on task B3→B1. In contrast, C-ID benefits from its access to the source data during the adaptation process and in a more exhaustive adaptation process, showing strong competitiveness. (4) Due to the lack of adaptation mechanism, No-TL does not show substantial disadvantages in cross-domain diagnosis tasks, resulting in significantly lower diagnostic accuracy compared with other methods.

[0130] Table 2 Classification accuracy and standard deviation of various methods in the embodiment (%)

[0131]

[0132]

[0133] In order to verify the effectiveness of the unsupervised model calibration method TMOMC of the present invention, the expected calibration error (ECE) is introduced. The ECE index is calculated by N v The model prediction results of samples Its corresponding confidence value and the true label of the sample Characterize the difference between confidence and accuracy. First, the vector according to Divide into M equal-width interval groups (the width of each interval is 1 / M). Then, in each interval B m The difference between the accuracy and confidence mean is calculated. Finally, the mathematical expression of ECE is as follows:

[0134]

[0135] in, For comparison, the present invention introduces temperature scaling with labeled target data as a benchmark method, called target_orientd_TS.

[0136] Table 3 shows the calibration effect of TMOMC based on the ECE results. Based on these findings, it is clear that Target_orientd_TS exhibits superior model calibration capabilities in all tasks, which is due to the use of labeled target domain datasets as a validation set for supervised estimated temperature parameters in the temperature scaling strategy. In contrast, the proposed TMOMC model calibration method completes model calibration in a passive scenario without accessing target domain labels or processing source data. Although TMOMC shows a slight disadvantage compared to Target_orientd_TS, it significantly improves the calibration performance compared to the uncalibrated scenario.

[0137] Table 3 ECE values ​​of various tasks in the embodiment (%)

[0138]

[0139]

[0140] During the verification of the embodiment, an intelligent diagnosis method based only on the source domain data set (No-TL), a cross-domain intelligent diagnosis method based on the maximum mean difference (M-ID), a cross-domain intelligent diagnosis method based on adversarial game (D-ID), a cross-domain intelligent diagnosis method based on conditional adversarial game (C-ID), a passive cross-domain intelligent diagnosis method based on source hypothesis migration (SHOT-ID), and a passive cross-domain intelligent diagnosis method based on robust self-training and nuclear norm maximization (SFA-ID) are selected for method comparison with the present invention to verify the effectiveness of the method of the present invention. Among them, No-TL represents an intelligent fault diagnosis method without a migration mechanism. It involves directly training a model only on the source data, which is then used to diagnose the target domain data set. M-ID is a typical representative of a domain adaptation diagnosis method based on a high-order moment measure, which uses the maximum mean difference constraint distribution difference of multiple Gaussian kernels. D-ID and C-ID are two typical representatives of domain adaptation diagnosis methods based on adversarial games. Among them, D-ID represents a cross-domain intelligent diagnosis method based on domain discriminator, which adopts adversarial game between domain discriminator and feature encoder to extract shared diagnostic features under different operating conditions; C-ID extends the concept of adversarial training by incorporating conditional information into the adversarial learning framework, which is beneficial to retain class-specific information. SHOT-ID and SFA-ID are two passive cross-domain intelligent diagnosis methods. Among them, SHOT-ID represents a cross-domain intelligent diagnosis method based on source hypothesis transfer as the backbone; SFA-ID represents an intelligent diagnosis method based on passive adaptation, which combines a robust self-training mechanism and target prediction matrix constraints. The two mechanisms work together to achieve excellent model adaptation using only unlabeled target domain datasets. To ensure the effectiveness of the comparison, the network architecture and parameter settings of the feature encoder module and the classification module are kept consistent in the proposed CSFADC and the six comparison methods.

Claims

1. A trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration, characterized by: It includes the following steps: S1: Collect vibration monitoring data of different fault types of mechanical equipment under different working conditions; Install a vibration acceleration sensor on the surface of the mechanical equipment, collect vibration monitoring data of the mechanical equipment under different fault types under a first working condition, where there are K types of fault types, and collect vibration monitoring data of the mechanical equipment under a second working condition; The vibration monitoring data obtained under the first working condition is used as the source domain data set, and the label information is the category of the fault type; the vibration monitoring data obtained under the second working condition is used as the target domain data set, which does not contain label information; S2: normalize and preprocess the vibration monitoring data; The vibration monitoring data obtained under different working conditions are respectively normalized and preprocessed to obtain corresponding normalized data; S3: Construct a trusted passive domain adaptation mechanical equipment fault diagnosis model based on model calibration; Constructing a trusted passive domain adaptation mechanical equipment fault diagnosis model based on model calibration mainly includes the following sub-steps: S31: Building the initial model Initial Model D Base A feature encoding module E Base and a classification module C Base The feature encoding module E Base Used to extract sensitive features, the output dimension is d, the classification module C Base The number of categories is K, and the output is a K-dimensional vector consisting of the probability of each category; S32: Source model generation Relying on the normalized source domain dataset samples and their corresponding labels, the label smoothing supervision loss function and sharpness perception minimization optimization strategy are used to train the initial model to obtain the source model; S33: Establishing an adaptive fault diagnosis model Use prototype-based pseudo-label learning technology to generate normalized target domain dataset samples x t The corresponding first pseudo label, and then constructing a new type of class prototype using the first pseudo label data to generate a second pseudo label; Using the target domain dataset with the second pseudo-label, the source model is trained based on DMI, the classification module parameters of the source model are fixed, and only the feature encoding module is updated to obtain an adapted fault diagnosis model; S34: Adaptive Fault Diagnosis Model Calibration The target simulation model calibration strategy is adopted to obtain the target simulation temperature parameters through the target simulation data set. The prediction confidence of the adaptive fault diagnosis model can be calibrated by using the temperature scaling strategy; S4: Use the adapted fault diagnosis model to diagnose the fault type of mechanical equipment; The vibration monitoring data of the mechanical equipment under the second working condition is collected and input into the adaptive fault diagnosis model. The adaptive fault diagnosis model outputs the probability of each fault type and obtains the calibrated prediction confidence according to S34, so as to diagnose the fault type of the mechanical equipment.

2. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 1 is characterized in that: The mechanical device is a multi-stage transmission system.

3. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 2 is characterized in that: In S1, there are K types of fault types, specifically: The fault types are the fault types of rolling bearings in a multi-stage transmission system, and there are 7 types in total, namely: normal state, inner ring crack fault, inner ring wear fault, inner and outer ring crack fault, outer ring crack fault, outer ring wear fault and cage crack fault.

4. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 1 is characterized in that: The specific structure of the initial model in S31 is: Feature encoding module E Base The convolution layer, batch normalization layer, nonlinear activation function, and pooling layer are connected in sequence as a group. After four groups are repeated in sequence, the Dropout layer, fully connected layer, batch normalization, and nonlinear activation function are connected in sequence. The classification module C Base It consists of a Dropout layer and a fully connected layer. When used, the normalized vibration monitoring data is used as the input of the feature encoding module to extract sensitive features, and then the extracted sensitive features are input into the classification module to realize fault type identification.

5. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 1 is characterized in that: When the label smoothing supervision loss function and the sharpness perception minimization optimization strategy are used to train the initial model to obtain the source model in S32, the expression of the source model objective function is as follows: Among them, w d is the source model parameter, which is the model parameter that needs to be trained, and the initial value is the initial model parameter; represents the randomly generated disturbance of the model parameters, ||·||2 represents the L2 norm, ρ is the hyperparameter of the neighborhood, ρ≥0; represents the definition, α is the weight; Among them, w d is the source model parameter, x s is the source domain dataset sample, y s is the corresponding true label of the source domain dataset sample; Indicates that x s As the initial model D Base The vector obtained by input is To vector After softmax, the kth element in the output category K is obtained. ε is the smoothing parameter. q k is the intermediate parameter, represents the expectation obtained by sampling from the source domain dataset.

6. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 1 is characterized in that: In S33, a prototype-based pseudo-label learning technique is used to generate a normalized target domain dataset sample x t The corresponding first pseudo label is then used to construct a new class prototype using the first pseudo label data, thereby generating a second pseudo label; specifically: A prototype-based pseudo-label learning technique is used, where the prototype of each class is the centroid of the features of that class. , calculate the cosine distance between each sample and each category prototype in the unlabeled target domain dataset, so as to assign the sample to the nearest category to generate the first pseudo label; A new class prototype was constructed using the first pseudo-labeled data , thereby generating the second pseudo label , the expression of this process is as follows: in, is the target domain dataset sample x t The first pseudo label of represents the target domain dataset; when the condition When it is established, Ind(·) is equal to 1, otherwise, Ind(·) is equal to 0, x t is the target domain dataset sample, Indicates that x t The sensitive features obtained after being input into the feature encoding module of the source model, Indicates that x t The vector obtained as input to the source model, Represents the vector After softmax output, the kth element in category K is Represents the prototype of the kth class; k represents the predicted category of the target domain dataset sample.

7. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 6 is characterized in that: In S33, the target domain dataset with the second pseudo label is used to train the source model based on the DMI, the classification module parameters of the source model are fixed, and only the feature encoding module is updated to obtain an adapted fault diagnosis model; specifically: The adaptive fault diagnosis model adopts the following objective function: in, express The noisy version of express and The joint distribution of express The matrix format of det(·) represents the calculation of the matrix determinant; specifically, The empirical estimates are as follows: in, is the expectation operation, Pr[·] represents the probability distribution, is the noise version of k, for The noisy version of is the second pseudo label of the i-th sample in the target domain dataset, and N represents the total number of samples in the target domain dataset; set up Substituting formula (10) into formula (9), we can obtain the empirical estimation loss function of DMI: Among them, M is composed of The matrix composed of Indicates row and column subscripts, The minimum value of formula (11) is used as the objective function of the adaptive fault diagnosis model, the classification module parameters of the source model are fixed, only the feature encoding module is updated, and the stochastic gradient descent algorithm is used to complete the generation of the adaptive fault diagnosis model.

8. The trusted passive domain adaptation mechanical equipment fault diagnosis method based on model calibration according to claim 1 is characterized in that: In the step S34, a target simulation model calibration strategy is adopted to obtain the target simulation temperature parameter through the target simulation data set. , the temperature scaling strategy can be used to calibrate the prediction confidence of the adaptive fault diagnosis model; specifically: A target simulation dataset is introduced. The target simulation temperature parameters are estimated using this data set , as follows: in, represents the negative log-likelihood function, that is Indicates that The vector obtained as the input of the adaptive fault diagnosis model, T is the temperature parameter, Represents the vector Output the result through the softmax output layer; The temperature scaling strategy is used to calibrate the prediction confidence of the adaptive fault diagnosis model. The calibrated confidence can be expressed as: Among them, the sample The vector obtained by inputting the adapted fault diagnosis model is defined as the logical z i , Simulate temperature parameters for the target, Represents the vector After softmax, the kth element in category K is output.

Images