Model calibration based trusted passive field adaptation mechanical equipment fault diagnosis method

Abstract

The application provides a kind of model calibration-based trusted passive field adaptive mechanical equipment intelligent diagnosis method, which comprises the following steps: S1: install vibration acceleration sensor on the surface of mechanical equipment, collect vibration monitoring data of different fault types of mechanical equipment under different working conditions;S2: normalize the vibration monitoring data for pretreatment;S3: build a model calibration-based trusted passive field adaptive mechanical equipment fault diagnosis model; including establishing an initial model, generating a source model, building an adaptive fault diagnosis model and calibrating the adaptive fault diagnosis model;S4: use the adaptive fault diagnosis model to diagnose the fault type of the mechanical equipment.The application directly uses vibration monitoring data as input, realizing end-to-end cross-condition intelligent fault diagnosis;The application introduces a robust model cross-domain adaptation based on determinant-based mutual information;The application also establishes a temperature scaling theory based on target simulation data set to improve the reliability of prediction confidence.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, and in particular to a reliable passive domain-adaptive mechanical equipment fault diagnosis method based on model calibration. Background Technology

[0002] As machinery and equipment become increasingly automated and intelligent, the importance of their safe operation is becoming increasingly prominent. Especially for safety-critical components that ensure continuous equipment operation, frequent exposure to harsh conditions such as high speed, high temperature, high pressure, and load fluctuations can easily lead to degradation or even failure. Therefore, timely fault diagnosis is not only crucial for ensuring the safe operation of equipment, but also directly related to the safety of workers.

[0003] Traditional fault diagnosis methods rely primarily on existing expert diagnostic knowledge, but they often face significant challenges when dealing with fault detection in structurally complex equipment. With changing industrial demands, the need for automated diagnostic methods to shorten maintenance cycles and improve diagnostic accuracy is increasing. As artificial intelligence theory continues to develop, the fault diagnosis process is evolving towards intelligent solutions, enabling the automatic detection and identification of equipment fault types. In particular, deep learning-based fault diagnosis methods have attracted widespread attention from researchers due to their ability to adaptively extract fault-related information.

[0004] Despite significant progress in deep learning-based intelligent diagnostics research in recent years, its application still relies on the assumption that training and test data are independently and identically distributed (ISD). However, differences in equipment operating conditions and service environments inevitably lead to different data distributions. To address this challenge, Domain Adaptive Intelligent Diagnosis (DA-ID) methods have gradually gained attention. This method captures cross-domain shared diagnostic information through feature alignment mechanisms between labeled data in the source domain and unlabeled data in the target domain, thereby enabling fault diagnosis under different data distributions. Traditional DA-ID methods mainly include domain adaptation based on statistical metrics and domain adaptation based on adversarial game theory.

[0005] While DA-ID has achieved some success in cross-domain fault diagnosis, this method still faces a core challenge: the availability of source domain labeled data during the adaptation process. This requirement conflicts with engineering practice, mainly in two aspects. First, to protect commercial privacy, many industrial fields (such as aero-engine and wind turbine fault diagnosis) have inherent restrictions on access to source data. Second, the requirements of real-time diagnosis make real-time access to source domain datasets infeasible due to data storage and transmission challenges. 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 it focuses on improving diagnostic accuracy while neglecting the security of crucial information such as calibrating prediction uncertainty in unsupervised adaptive environments. This neglect makes it difficult for maintenance personnel to trust the prediction uncertainty in SFAFD applications. For example, if manual review is skipped due to excessive confidence in fault type identification, the fault may not be detected in time, leading to disastrous consequences; conversely, excessively high uncertainty in diagnostic results often triggers false alarms, leading to repeated checks and wasting resources. Therefore, calibration analysis of the SFAFD model is crucial for ensuring the safe application of neural network models in mechanical fault diagnosis, a gap that remains in existing research. Summary of the Invention

[0006] To address the aforementioned key challenges, this paper proposes a Credible Source-Free Adaptation Diagnosis with Calibration (CSFADC) method for fault diagnosis of mechanical equipment under unknown operating conditions. Based on the fault classification models of mechanical equipment corresponding to different acquired signals under known operating conditions, CSFADC obtains the fault classifications of mechanical equipment corresponding to different acquired signals under unknown operating conditions. CSFADC comprises three stages: source model generation, model adaptation, and model calibration. In the source model generation stage, the parameters of the feature encoder and classifier modules are optimized by combining a label smoothing supervised loss function and a sharpness-aware minimization optimization strategy. This process fully utilizes source domain information while maximizing the generalization ability of the diagnostic model. In the model adaptation stage, source domain classification information is preserved by fixing the classifier parameters. Then, pseudo-label learning is used to update the parameter parameters of the feature encoder module from unlabeled target data to achieve model adaptation. To mitigate the adverse effects of erroneous pseudo-labels on model performance, this paper draws inspiration from information-theoretic loss functions and employs deterministic mutual information to enhance the robustness to noisy labels during the adaptation process. Finally, in the model calibration phase, it is assumed that the diagnostic model has fully learned the target domain information. To achieve model calibration using only the unlabeled target domain dataset, a hybrid technique is employed in the model transfer phase to create a target simulation dataset. Based on the target simulation dataset, temperature scaling theory is applied to achieve model calibration in passive scenarios.

[0007] To achieve the above objectives, this invention provides a reliable passive domain-adaptive 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] A vibration acceleration sensor is installed on the surface of the mechanical equipment to collect vibration monitoring data of the mechanical equipment under different fault types in the first working condition. There are K types of fault types. Vibration monitoring data of the mechanical equipment under the second working condition is then collected.

[0010] The vibration monitoring data obtained under the first working condition is used as the source domain dataset, 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 dataset, 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 were preprocessed by normalization to obtain the corresponding normalized data.

[0013] S3: Construct a reliable passive domain-adaptive mechanical equipment fault diagnosis model based on model calibration;

[0014] Constructing a reliable, passive, domain-adaptive fault diagnosis model for mechanical equipment based on model calibration mainly includes the following sub-steps:

[0015] S31: Establish the initial model

[0016] Initial model D Base Consists of a feature encoding module E Base and a classification module C Base Composition, including feature encoding module E Base Used to extract sensitive features, the output dimension is d, and 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] Based on the normalized source domain dataset samples and their corresponding labels, the initial model is trained using a label smoothing supervised loss function and a sharpness-aware minimization optimization strategy to obtain the source model;

[0019] S33: Establish an adaptive fault diagnosis model

[0020] A prototype-based pseudo-label learning technique is used to generate normalized target domain dataset samples x. t The corresponding first pseudo-label is then used to construct a new class prototype, thereby generating the second pseudo-label;

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

[0022] S34: Adaptation and calibration of fault diagnosis model

[0023] A target simulation model calibration strategy is adopted. The target simulation temperature parameter T is obtained through the target simulation dataset. The temperature scaling strategy can be used to calibrate the prediction confidence of the adapted fault diagnosis model.

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

[0025] Vibration monitoring data of 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 based on S34, thereby diagnosing the fault type of the mechanical equipment.

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

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

[0028] The fault types refer to the fault types of rolling bearings in multi-stage transmission systems, which are divided into 7 categories: normal condition, 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 initial model in S31 has the following specific structure:

[0030] Feature encoding module E Base The classification module C is constructed by sequentially connecting convolutional layers, batch normalization layers, non-linear activation functions, and pooling layers into groups, repeating this process four times, and then sequentially connecting dropout layers, fully connected layers, batch normalization, and non-linear activation functions. Base It consists of a Dropout layer and a fully connected layer. When in use, the normalized vibration monitoring data is used as the input of the feature encoding module to extract sensitive features. Then, the extracted sensitive features are input into the classification module to realize fault type identification.

[0031] Preferably, when the initial model is trained to obtain the source model using the label smoothing supervised loss function and the sharpness-aware minimization optimization strategy in step S32, the expression of the objective function of the source model is as follows:

[0032]

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

[0034]

[0035] Among them, w d For the source model parameters, x s For the source domain dataset sample, y s These are the true labels corresponding to the samples in the source domain dataset; Indicates x s As the initial model D Base The input vector obtained, To make vector After softmax, the output is the k-th element of category K, where ε is the smoothing parameter; q k For intermediate parameters, This represents the expectation obtained by sampling from the source domain dataset.

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

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

[0038] A new class prototype was constructed using the first pseudo-label data. This generates a second pseudo-tag. The expression for this process is as follows:

[0039]

[0040] in, For the target domain dataset sample x t The first pseudo-label; χ t Represents the target domain dataset; when conditions When true, Ind(·) equals 1; otherwise, Ind(·) equals 0. t For the target domain dataset sample, Indicates x t The sensitive features obtained after being input into the feature encoding module of the source model Indicates x t The vector obtained as input to the source model Indicates the vector The softmax function outputs the k-th element of category K. represents the prototype of the k-th class; k represents the predicted class of the target domain dataset sample.

[0041] Preferably, in step S33, a target domain dataset with a second pseudo-label is used to train the source model based on 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 following objective function is used to adapt the fault diagnosis model:

[0043]

[0044] in, express The noise version, express and Joint distribution on, express The matrix format is defined by `det(·)`, which calculates the determinant of the matrix. Specifically, for... The empirical estimates are as follows:

[0045]

[0046] in, This is the expectation operation, where Pr[·] represents the probability distribution. For the noise version of k, for The noise version, 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 equation (10) into equation (9), we obtain the empirical loss function for DMI:

[0048]

[0049] Where M is composed of The matrix formed Indicates row and column subscripts,

[0050] Minimize formula (11) as the objective function of the adapted fault diagnosis model, fix the classification module parameters of the source model, update only the feature encoding module, and use the stochastic gradient descent algorithm to complete the generation of the adapted fault diagnosis model.

[0051] Preferably, step S34 employs a target simulation model calibration strategy, obtaining the target simulation temperature parameters using the target simulation dataset. Temperature scaling strategies can be used to calibrate the prediction confidence of adapted fault diagnosis models; specifically:

[0052] A target simulation dataset was introduced. And use this dataset to estimate the target simulated temperature parameters. As shown in the following formula:

[0053]

[0054] in, This represents the negative log-likelihood function, i.e. Indicates will The vector obtained as input to the fault diagnosis model, where T is the temperature parameter. Indicates the vector The result is output after passing through the softmax output layer;

[0055] The prediction confidence of the adapted fault diagnosis model is calibrated using a temperature scaling strategy. The calibrated confidence can be expressed as:

[0056]

[0057] Among them, the sample The vector obtained by inputting into the adaptive fault diagnosis model is defined as the logical z. i , To simulate the target temperature parameters, Indicates the vector The softmax function outputs the k-th element of category K.

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

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

[0060] (2) This invention is based on pseudo-label learning and introduces mutual information based on determinant to enhance the robustness of the adaptive process to noisy labels, 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 technology to create a target simulation dataset, and combines temperature scaling theory to achieve model calibration in passive scenarios, thereby improving the reliability of the prediction confidence of the passive adaptation fault diagnosis model.

[0062] (4) This invention effectively addresses the practical engineering problem of source domain datasets being inaccessible 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. Attached Figure Description

[0063] Figure 1 A flowchart for a model-calibrated, reliable, passive domain-adapted mechanical equipment fault diagnosis method;

[0064] Figure 2 A flowchart for constructing a reliable passive domain-adaptive rolling bearing fault diagnosis model based on model calibration;

[0065] Figure 3 This describes the network structure and parameter settings of the feature encoding module in this embodiment of the invention.

[0066] Figure 4 This describes the network structure and parameter settings of the classification module in this embodiment of the invention.

[0067] Figure 5This is a schematic diagram of the multi-stage acceleration transmission experimental device in this embodiment of the invention. Detailed Implementation

[0068] To better understand the technical solution of the present invention, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0069] The following example illustrates the detection of rolling bearing faults in a multi-stage transmission system. Figure 1 and Figure 2 As shown, the specific method for fault diagnosis of reliable passive domain-adaptive mechanical equipment based on model calibration of the present invention is as follows:

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

[0071] Vibration acceleration sensors were mounted on the surface of a multi-stage transmission system. Vibration monitoring data of the rolling bearings under different fault types were collected under the first operating condition. There are K types of fault types. Vibration monitoring data of the rolling bearings under the second operating condition were then collected. The vibration monitoring data obtained under the first operating condition was used as the source domain dataset, with the fault type as the label information. The vibration monitoring data obtained under the second operating condition was used as the target domain dataset, without 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 housing 6, a vibration acceleration sensor 7, and a magnetic powder brake 8, as shown in the schematic diagram below. Figure 5 As shown. Power is provided by drive motor 1, and after being accelerated through planetary gearbox 3 and fixed-axis gearbox 4, it reaches magnetic powder brake 8 via bearing test base. Magnetic powder brake 8 can apply different loads to multi-stage transmission system.

[0073] A vibration acceleration sensor is mounted on the surface of a multi-stage transmission system to collect vibration monitoring data. This invention focuses on rolling bearings in the bearing housing of a multi-stage transmission system. Different rolling bearing faults can be injected into the experimental multi-stage transmission system, resulting in seven fault types: 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. The vibration acceleration sensor is mounted on the surface of the multi-stage transmission system to collect vibration monitoring data of different rolling bearing fault types under different operating conditions. This application utilizes a sampling frequency of 10kHz. The system operating conditions are: a speed of approximately 1200 rpm, and loads of 0.7 Nm, 1.2 Nm, 2 Nm, and 5 Nm, respectively, denoted as B1, B2, B3, and B4. Vibration acceleration signals of the multi-stage transmission system under these four different operating conditions are collected. For each operating condition of the rolling bearing, 1000 vibration monitoring data samples are collected, with each sample containing 1024 data points. These conditions collectively generated 12 different cross-domain diagnostic tasks, as shown in Table 1. Here, B1→B2 indicates that B1 is the first working condition, and the vibration monitoring data samples obtained under the first working condition B1 are used as the source domain dataset, containing label information; B2 is the second working condition, and the vibration monitoring data samples obtained under the second working condition B2 are used as the target domain dataset, which is considered to not contain label information.

[0074] Table 1 Bearing Migration Diagnosis Tasks for Multistage Transmission Systems

[0075]

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

[0077] Using the vibration monitoring data collected under the first working condition as the source domain dataset and the vibration monitoring data collected under the second working condition as the target domain dataset, the vibration monitoring data under different working conditions are preprocessed by normalization using equation (1):

[0078]

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

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

[0081] S3: Construct a reliable passive domain-adaptive rolling bearing fault diagnosis model based on model calibration;

[0082] Constructing a reliable, passive, domain-adaptive rolling bearing fault diagnosis model based on model calibration mainly includes the following sub-steps:

[0083] S31: Establish the initial model

[0084] Initial model D Base Consists of a feature encoding module E Base and a classification module C Base Composition, including feature encoding module E Base Used to extract sensitive features, the output dimension is d, and the classification module C Base The number of categories is K, therefore the output is a K-dimensional vector consisting of the probabilities of each category; the specific structure is: Feature Encoding Module E Base like Figure 3 As shown, convolutional layers, batch normalization layers, non-linear activation functions, and pooling layers are connected sequentially as a group, and this process is repeated four times. Then, a Dropout layer, a fully connected layer, batch normalization, and a non-linear activation function are connected sequentially to form the classification module C. Base like Figure 4 As shown, it consists of a Dropout layer and a fully connected layer. In use, normalized vibration monitoring data is used as input to the feature encoding module to extract sensitive features. Then, the extracted sensitive features are input to the classification module to identify the type of rolling bearing fault.

[0085] S32: Source Model Generation

[0086] Based on the source domain dataset and its corresponding labels, a supervised loss function is used as the optimization objective to train the initial model and obtain the source model. This stage utilizes label smoothing techniques to mitigate the impact of mislabeled data, thereby enhancing the generalization of the initial model. The resulting label-smoothed cross-entropy loss function is described as follows:

[0087]

[0088] Among them, w d For the source model parameters, x s For the source domain dataset sample, y s These are the true labels corresponding to the samples in the source domain dataset; Indicates x s As the initial model D Base The input vector obtained, To make vector The output layer outputs the k-th element of category K after softmax (output layer), where ε is a smoothing parameter set to 1; q k For intermediate parameters, This represents the expectation obtained by sampling from the source domain dataset.

[0089] In this embodiment, the source domain dataset sample x s This refers to the vibration monitoring data after normalization under the first working condition B1, and the corresponding true label y of the source domain dataset samples. s The fault types corresponding to the vibration monitoring data are 7 in number, K.

[0090] Given that subsequent model adaptation and diagnostic procedures cannot access the source domain dataset in passive scenarios, the necessary generalization performance of the source model becomes crucial. Therefore, a Sharpness-Aware Minimization (SAM) optimization strategy is adopted. Traditional optimization techniques only aim to minimize the training loss, often resulting in poor model quality. In contrast, SAM strives to minimize both the magnitude and sharpness of the loss, aiming to identify parameter configurations within regions characterized by consistently low loss. This approach effectively improves the generalization ability of the source model. Based on the label-smoothed cross-entropy loss function of the initial model, the expression for 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 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; α is the balancing weight of the two loss functions, which is set to 1 in this embodiment.

[0093] The vibration monitoring data obtained from the first working condition B1 after normalization is used as the source domain dataset input to 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: Establish an adaptive fault diagnosis model

[0095] Since the source domain dataset is inaccessible, conventional domain adaptation methods cannot be used. To address this challenge, the classification module parameters of the source model are fixed, and only the feature encoding module is updated. This allows the adapted fault diagnosis model to retain diagnostic knowledge related to the inaccessible source data.

[0096] Furthermore, this invention employs a prototype-based pseudo-label learning technique, utilizing the target domain dataset x. t This facilitates 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] Where, x tFor the target domain dataset sample, X t Represents the target domain dataset; Indicates x t The sensitive features obtained after being input into the feature encoding module of the source model Indicates x t The vector obtained as input to the source model Indicates the vector After passing through the softmax (output layer), the k-th element of category K is output. This represents the prototype of the k-th class.

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

[0100] Based on this, the cosine distance between each unlabeled target domain dataset sample and each class prototype can be calculated, thereby assigning the sample to the nearest class to generate the first pseudo-label:

[0101]

[0102] in, For the target domain dataset sample x t The first pseudo-tag.

[0103] To address the issue of unreliable pseudo-labels during the initial training phase, a novel class prototype was constructed using the first pseudo-label data. This novel prototype was then used to refine the pseudo-label data, thereby generating more reliable second pseudo-labels. The mathematical formula for this process is as follows:

[0104]

[0105] Among them, when the condition When the condition is true, Ind(·) equals 1; otherwise, Ind(·) equals 0. k represents the predicted class of the sample in the target domain dataset. However, in the context of job condition migration, the acquisition of pseudo-labels may not be accurate enough, making the guarantee of label reliability a challenge. Traditional distance-based loss functions, such as cross-entropy loss, are easily distorted by incorrect labels, leading to overfitting of the model to noisy information. Therefore, this invention introduces the derivative of Shannon mutual information, called determinant-based mutual information (DMI). DMI ensures that even when training with noisy labels, the training process is similar to that using clean labels, except for a constant offset. Therefore, the following objective function is adopted when establishing the adaptive fault diagnosis model to enhance the robustness of the diagnostic model to noisy labels:

[0106]

[0107] in, express The noise version, express and Joint distribution on, express The matrix format is defined by `det(·)`, which calculates the determinant of the matrix. Specifically, for... The empirical estimates are as follows:

[0108]

[0109] in, This is the expectation operation, where Pr[·] represents the probability distribution. For the noise version of k, for The noise version, 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 equation (10) into equation (9), we obtain the empirical loss function for DMI:

[0111]

[0112] Where M is composed of The matrix formed Indicates row and column subscripts,

[0113] The vibration monitoring data after normalization of the second working condition B2 obtained in step S2 is used as the target domain dataset input to the source model. The minimum of formula (11) is used as the objective function of the adaptation 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 adaptation fault diagnosis model.

[0114] S34: Adaptation and calibration of fault diagnosis model

[0115] Existing intelligent diagnostic models primarily focus on the accuracy of fault identification, but this alone is insufficient to support practical decision-making in engineering. Therefore, based on this consideration, this invention introduces the concept of model calibration to improve the reliability of prediction uncertainty and ensure that the confidence level transmitted by the model accurately reflects its prediction uncertainty. Thus, the intelligent diagnostic model proposed in this invention not only pursues accurate prediction results but also prioritizes reliable calibration. Inspired by the model calibration technique Platt scaling, temperature scaling has been widely applied in fields such as computer vision and semantic segmentation. However, due to the lack of labeled target datasets in source-independent adaptive diagnostic scenarios, traditional supervised calibration methods (such as temperature scaling) face challenges in their applicability. To address this issue, this invention proposes a target-mimic-oriented model calibration (TMOMC) strategy.

[0116] First, this invention treats 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 is lowered to enhance confidence; conversely, when there are a large number of incorrect predictions, the temperature parameter is increased to mitigate prediction fluctuations. Therefore, when datasets exhibit statistical similarity in correct and incorrect predictions, similar temperature parameters are generated. To achieve this goal, this invention introduces a target simulation dataset, consisting of pseudo-labeled mixed target samples, used to simulate the correct and incorrect diagnostic statistics of target data diagnosis. Given two pseudo-labeled target domain datasets from different fault type categories: and in, and Samples and The first pseudo-label, sample and Selected from the target domain dataset. Based on this, the present invention can obtain a target simulation sample. And construct 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 this invention.

[0119] Obviously, the target sample is a pseudo-label. Dominant, and It is considered a disturbance term. If If the location is far from the classification boundary of the diagnostic model, then discrimination is possible. It will become easier to target simulated samples. It is also easier to identify. Conversely, when Approaching the classification boundary may lead to misclassification, thus... and Both are susceptible to misclassification. This indicates that... and target simulation samples The statistical characteristics of correct-incorrect diagnoses are consistent. Therefore, this invention uses the target simulated dataset. Revised to To facilitate calculations, and to use this dataset to estimate the target simulated temperature parameters. As shown in the following formula:

[0120]

[0121] in, This represents the negative log-likelihood function, i.e. Indicates will The vector obtained as input to the fault diagnosis model, where T is the temperature parameter. Indicates the vector The output result is processed through the softmax layer.

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

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

[0124]

[0125] Among them, the sample The vector obtained by inputting into the adaptive fault diagnosis model is defined as the logical z. i , To simulate the target temperature parameters, Indicates the vector The softmax function outputs the k-th element of category K.

[0126] Ultimately, based on the predicted confidence level of the calibrated and adapted fault diagnosis model, the rolling bearing fault type is identified, and the corresponding confidence level is output simultaneously, which can be used by maintenance personnel to make maintenance decisions.

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

[0128] Vibration data of the multi-stage transmission system under the second operating condition are collected and input into the adaptive fault diagnosis model. Based on the classification output by the model, rolling bearing faults are obtained, and the confidence level of the classification is obtained according to formula (14). The adaptive fault diagnosis model is a reliable passive domain adaptive mechanical fault diagnosis model based on model calibration.

[0129] To further demonstrate the effectiveness of the method proposed in this invention, the dataset collected in step S1 was subjected to the method proposed in this invention and various comparison methods, and the obtained diagnostic results are shown in Table 2. In this table, "B" i →B j "Indicates B" i As the source domain, B j As the target domain. For clarity, the best results for each task are highlighted in bold. Analysis of these findings yields the following insights: (1) Despite the lack of access to the source domain dataset during adaptation, CSFADC still achieved an impressive average diagnostic accuracy of 99.42% across 12 cross-domain fault diagnosis tasks, outperforming all comparable methods. (2) Among all comparable methods, C-ID, SHOT-ID, and SFA-ID outperformed CSFADC in some cross-domain diagnostic tasks. However, the maximum accuracy gap between CSFADC and the optimal solution did not exceed 1%. (3) Despite the lack of access to the source data during adaptation, SHOT-ID and SFA-ID performed well on most tasks, with notable exceptions: SHOT-ID showed reduced performance on tasks B4→B1 and B3→B1, while SFA-ID exhibited similar limitations on task B3→B1. In contrast, C-ID benefited from its access to the source data during adaptation and in more detailed adaptation processes, demonstrating strong competitiveness. (4) Due to the lack of adaptation mechanism, No-TL did not show any substantial disadvantages in cross-domain diagnostic tasks, resulting in a significant reduction in diagnostic accuracy compared with other methods.

[0130] Table 2. Classification accuracy and standard deviation (%) for various methods in the embodiments.

[0131]

[0132]

[0133] To verify the effectiveness of the unsupervised model calibration method TMOMC of this invention, the expected calibration error (ECE) is introduced. The ECE index is measured by N... v Model prediction results for each sample Its corresponding confidence level value and the true label of the sample Characterize the difference between confidence and accuracy. First, convert the vector... according to Divide the data into M equal-width intervals (each interval has a width of 1 / M). Then, in each interval B... m Within each range, the difference between the accuracy and the mean confidence level is calculated. Finally, the mathematical expression for ECE is as follows:

[0134]

[0135] in, For comparison purposes, this invention introduces temperature scaling with labeled target data as a benchmarking method, called target_orientd_TS.

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

[0137] Table 3. ECE values ​​(%) for various tasks in the embodiments.

[0138]

[0139]

[0140] In the verification process of the embodiments, the following intelligent diagnostic methods based solely on the source domain dataset (No-TL), cross-domain intelligent diagnostic methods based on maximum mean difference (M-ID), cross-domain intelligent diagnostic methods based on adversarial games (D-ID), cross-domain intelligent diagnostic methods based on conditional adversarial games (C-ID), passive cross-domain intelligent diagnostic methods based on source hypothesis transfer (SHOT-ID), and passive cross-domain intelligent diagnostic methods based on robust self-training and nuclear norm maximization (SFA-ID) were selected for comparison with the present invention to verify the effectiveness of the method. No-TL represents an intelligent fault diagnosis method without a transfer mechanism. It involves directly training the model on the source data and then using it to diagnose the target domain dataset. M-ID is a typical representative of domain-adaptive diagnostic methods based on higher-order moment measures, which uses the maximum mean difference constraint of a multi-Gaussian kernel to constrain the distribution difference. D-ID and C-ID are two typical representatives of domain-adaptive diagnostic methods based on adversarial games. Among them, D-ID represents a cross-domain intelligent diagnostic method based on a domain discriminator, which uses an adversarial game between the domain discriminator and the 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, thus helping to preserve class-specific information. SHOT-ID and SFA-ID are two passive cross-domain intelligent diagnostic methods. SHOT-ID represents a cross-domain intelligent diagnostic method based on source hypothesis transfer; SFA-ID represents a passive adaptive intelligent diagnostic method that combines a robust self-training mechanism and target prediction matrix constraints. The two mechanisms work together to achieve excellent model fit using only unlabeled target domain datasets. To ensure the validity of the comparison, the network architecture and parameter settings of the feature encoder and classification modules are consistent in the proposed CSFADC and the six comparative methods.

Claims

1. A model calibration-based trusted passive field adaptation mechanical equipment fault diagnosis method, characterized in that: It comprises 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 different fault types of the mechanical equipment under the first working condition, the categories of fault types have K categories, and collect vibration monitoring data of the mechanical equipment under the 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 and does not contain label information; S2: Normalization preprocessing of vibration monitoring data; The vibration monitoring data obtained under different working conditions is respectively subjected to normalization preprocessing to obtain corresponding normalized data; S3: Constructing a reliable passive field adaptive mechanical equipment fault diagnosis model based on model calibration S31: Establishing an initial model S32: Source model generation Initial model D Base A feature encoding module E Base And a classification module C Base , wherein the feature encoding module E Base is used to extract sensitive features, and the output dimension is d, and the classification module C Base has a class number K, and the output is a K-dimensional vector composed of the probability of each class. The initial model is trained to obtain a source model by using a label smoothing supervised loss function and a sharpness perception minimization optimization strategy depending on the normalized source domain data set samples and their corresponding labels; S33: Establishing an adaptive fault diagnosis model The source model is trained based on DMI using the target domain data set with the second pseudo label, the classification module parameters of the source model are fixed, only the feature encoding module is updated, and an adaptive fault diagnosis model is obtained; Adopting a prototype-based pseudo-label learning technique, normalized target domain dataset samples x are generated t A corresponding first pseudo-label is then generated, and a new type of class prototype is constructed using the first pseudo-label data, thereby generating a second pseudo-label; S34: Calibration of the adaptive fault diagnosis model The target simulation model calibration strategy is adopted, the target simulation temperature parameter T is obtained through the target simulation data set, and the temperature scaling strategy can calibrate the prediction confidence of the adaptive fault diagnosis model; S4: Diagnosing the fault type of the mechanical equipment using the adaptive fault diagnosis model 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 the calibrated prediction confidence is obtained according to S34, so that the fault type of the mechanical equipment is diagnosed. The mechanical equipment is a multi-stage transmission system.

2. The model-based calibration trusted passive field adaptation mechanical equipment fault diagnosis method according to claim 1, characterized in that: In S1, the categories of fault types have K categories, specifically:

3. The model-calibration-based trustable passive-domain adaptation mechanical equipment fault diagnosis method according to claim 2, characterized in that: The fault types are the fault types of rolling bearings in the multi-stage transmission system, and there are 7 categories, 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. The specific structure of the initial model in S31 is:

4. The model-calibration-based trustable passive-domain adaptation mechanical equipment fault diagnosis method according to claim 1, characterized in that: When the initial model is trained to obtain a source model by using a label smoothing supervised loss function and a sharpness perception minimization optimization strategy in S32, the expression of the source model objective function is as follows: A group is sequentially connected by a convolutional layer, a batch normalization layer, a nonlinear activation function, and a pooling layer, four groups are sequentially repeated, and then a Dropout layer, a fully connected layer, a batch normalization, and a nonlinear activation function are sequentially connected to form a feature encoding module E Base ; a classification module C Base A Dropout layer and a fully connected layer are used; during use, 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 model-calibration-based trustable passive-domain adaptation mechanical equipment fault diagnosis method according to claim 1, characterized in that: In S33, the source model is trained based on DMI using the target domain data set with the second pseudo label, the classification module parameters of the source model are fixed, only the feature encoding module is updated, and an adaptive fault diagnosis model is obtained; specifically: wherein wherein w d is a source model parameter, is a model parameter to be trained, and the initial value is an initial model parameter; represents a disturbance randomly generated for the model parameter, ||·||2 represents an L2 norm, and p is a hyperparameter of a neighborhood range, p>0; represents a definition, and a is a weight; where 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; represents the vector obtained by inputting x s as the initial model D Base , is the vector outputted by the softmax, and ε is a smoothing parameter; q k is an intermediate parameter, represents the expectation obtained by sampling from the source domain dataset.

6. The model-calibration-based trustable passive-domain adaptation mechanical equipment fault diagnosis method according to claim 1, characterized in that: The S33 adopts a prototype-based pseudo-label learning technology to generate normalized target domain dataset samples x t A corresponding first pseudo-label is then generated, and a new type of class prototype is constructed using the first pseudo-label data, thereby generating a second pseudo-label; specifically: A prototype-based pseudo-label learning technique is employed, where the prototype of each class is the centroid of the features of that class A cosine distance between each sample in the unlabeled target domain dataset and each class prototype is calculated, thereby assigning the sample to the nearest class to generate first pseudo labels; A new type of prototype-like class is constructed using the first pseudo-label data Thus generating a second pseudo-label The expression of this process is as follows: wherein, is a first pseudo label of a sample x t of the target domain dataset; denotes the target domain dataset; when the condition is met, Ind(·) equals 1, otherwise, Ind(·) equals 0, x t is a sample of the target domain dataset, denotes sensitive features obtained after inputting x t into a feature encoding module of the source model, and k denotes a class to be predicted by the target domain dataset sample.

7. The model-calibration-based trustable passive field-adaptive mechanical equipment fault diagnosis method according to claim 6, characterized in that: The adaptive fault diagnosis model adopts the following objective function: ​ where denotes a noisy version of denotes the vector obtained as input to the source model t ; denotes and a joint distribution over denotes a matrix format of det(·) denotes the computation of the determinant of a matrix; in particular, an empirical estimate of is given by wherein, is a desired operation, Pr[·] denotes a probability distribution, is a noisy version of k, is is a noisy version of is a second pseudo label for the i-th sample of the target domain dataset, N denotes the total number of samples of the target domain dataset. Let and substituting equation (10) into equation (9), we obtain the empirical estimated loss function of DMI: where M is a matrix composed of denotes the row and column indices, respectively,​ The formula (11) is taken as a target function of the adaptive fault diagnosis model, the classification module parameters of the source model are fixed, only the feature coding module is updated, and a random gradient descent algorithm is used to complete generation of the adaptive fault diagnosis model.

8. The model-calibration-based trustable passive-domain adaptation mechanical equipment fault diagnosis method according to claim 1, characterized in that: The target simulation type model calibration strategy is adopted in the S34, and the target simulation temperature parameter is obtained through a target simulation data set The temperature scaling strategy can calibrate the prediction confidence of the adaptive fault diagnosis model. Specifically, A target simulation data set is introduced, and the target simulation temperature parameter is estimated using this data set as follows: wherein, represents the negative log-likelihood function, i.e. represents the vector as an input to the adapted fault diagnosis model, T is a temperature parameter, represents the vector outputted by a softmax output layer; 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: wherein the input of the sample to the adapted failure diagnosis model is defined as a vector of logic z i , is a target simulation temperature parameter, denotes the vector of the input is output through softmax to the kth element in the category K.

Images