An adaptive semi-supervised imbalanced fault diagnosis method and system

Abstract

The application discloses a self-adaptive semi-supervised unbalanced fault diagnosis method and system, and belongs to the technical field of intelligent fault diagnosis and industrial equipment state monitoring. The method comprises the following steps: constructing a teacher fault diagnosis network, a student fault diagnosis network and a labeled reference sample set, generating weak enhanced samples and strong enhanced samples for unlabeled fault samples; forming a candidate admission order and a tentative batch according to a category prediction probability, a prediction uncertainty, a category admission shortage and a representation distance of an intermediate representation relative to a pseudo-label category prototype; and realizing formal admission and formal training through shadow student network branch tentative update, admission contribution evaluation, parameter difference backwriting and rule linkage update. The application can improve the fault diagnosis precision, training stability, minority class recognition ability and robustness under the conditions of limited labeling, category imbalance and noise disturbance.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis and industrial equipment condition monitoring technology. Specifically, it relates to a semi-supervised imbalance fault diagnosis method and system for industrial equipment such as rotating machinery under the conditions of limited labeled samples, sufficient unlabeled samples, and imbalanced fault category distribution. The system is based on teacher-student networks, tentative acceptance of candidate samples, evaluation of acceptance contribution, parameter differential write-back, and rule linkage update. Background Technology

[0002] Rotating machinery is widely used in industrial scenarios such as power, rail transportation, aerospace, metallurgy, and intelligent manufacturing. Its operating status directly affects the safety, continuity, and economy of the production system. Because key components such as bearings, gears, rotors, and couplings are susceptible to wear, impact, uneven loading, abnormal lubrication, and complex operating conditions during long-term operation, equipment is prone to various types of failures. Therefore, how to accurately identify the operating status and fault types of equipment using monitoring signals has always been a key technical issue in industrial health management and predictive maintenance.

[0003] With the development of deep learning technology, data-driven fault diagnosis methods can automatically extract fault features from vibration signals, current signals, acoustic emission signals, and multi-source sensor fusion signals. Compared with traditional methods that rely on human experience to construct features, these methods have stronger representation and adaptability. However, in real-world industrial scenarios, high-quality labeled samples usually rely on manual judgment, shutdown inspection, or expert annotation, which is costly, time-consuming, and limited in quantity. Meanwhile, the large amount of monitoring data accumulated during equipment operation often lacks reliable labels, making it difficult for fully supervised fault diagnosis methods to fully utilize unlabeled data.

[0004] To alleviate the problem of scarce labeled samples, semi-supervised learning has been gradually introduced into the field of fault diagnosis. Existing semi-supervised fault diagnosis methods typically utilize unlabeled samples through pseudo-label screening, consistency constraints, knowledge distillation, or uncertainty assessment, but they still have the following shortcomings: First, existing schemes usually decide whether to include unlabeled samples in formal training based on the results of a single screening, lacking a pre-validation mechanism for the actual training contribution of candidate samples. This makes it easy for incorrect pseudo-labels or low-yield samples to be directly accepted, leading to error accumulation and model drift. Second, existing schemes mostly deal with class imbalance problems by static reweighting, fixed quotas, or simple priority adjustments, making it difficult to balance the compensation needs of minority class samples with acceptance risk control. Third, even when introducing uncertainty assessment and teacher-student distillation, existing schemes mostly focus on pseudo-label reliability judgment or knowledge transfer itself, without feeding back the training benefits after candidate sample trial updates into the sample acceptance and rule correction process. Fourth, existing schemes generally lack a closed-loop mechanism that links the calculation rules of candidate acceptance ranking scores, acceptance thresholds in acceptance conditions, class compensation relationships, and model parameter updates, making it difficult to dynamically correct the acceptance strategy based on the actual contribution of the samples.

[0005] Therefore, there is an urgent need for a semi-supervised imbalance fault diagnosis method and system that can first verify the true training value of candidate samples in scenarios where there are limited labeled samples, a large number of unlabeled samples, and imbalanced class distributions, and then decide whether to formally accept them, and update the model parameters and acceptance rules in conjunction with the verification results, so as to improve fault diagnosis accuracy, training stability, minority class fault identification capability, and engineering deployment reliability. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive semi-supervised imbalance fault diagnosis method and system to solve the problems in the prior art, such as the direct acceptance of unlabeled samples based on static or single-time judgment results, high risk of mis-acceptance, insufficient compensation for minority class samples, difficulty in dynamically modifying the acceptance rules based on real training benefits, and insufficient coupling between teacher and student knowledge transfer and the sample acceptance process. This invention achieves pre-validation acceptance of candidate unlabeled samples, controlled write-back of beneficial updates, and closed-loop linkage adjustment of the acceptance rules, thereby improving the accuracy, robustness, and minority class friendliness of imbalance fault diagnosis under complex working conditions.

[0007] To achieve the above objectives, this invention provides a semi-supervised imbalance fault diagnosis method, comprising: constructing a teacher fault diagnosis network, a student fault diagnosis network, and a labeled reference sample set based on a labeled fault sample set; constructing weakly enhanced samples and strongly enhanced samples for unlabeled fault samples; using the teacher fault diagnosis network to output pseudo-labels, class prediction probabilities, intermediate representations, and prediction uncertainties for the weakly enhanced samples; calculating candidate acceptance ranking scores based on class prediction probabilities, prediction uncertainties, class acceptance gaps, and the representation distance of intermediate representations relative to pseudo-label class prototypes, and combining these scores with a screening threshold to filter and rank candidate acceptance samples to form a trial batch; replicating the student fault diagnosis network to obtain a shadow student network branch; and, under the condition that the teacher fault diagnosis network is fixed or slowly updated, using the trial batch to perform supervision with pseudo-labels, weak-strong enhancement consistency constraints, and teacher-student relationship management. The process involves trial updates of the student network and the teacher-student network. Based on the increase in accuracy and decrease in loss on the labeled reference sample set before and after the trial update, the improvement in predictive consistency of weak and strong enhancements in the trial batch, and the decrease in category acceptance gaps, an acceptance contribution score is calculated. When the acceptance contribution score meets the acceptance criteria, the unlabeled fault samples corresponding to the trial batch are written into the formal acceptance sample set. The parameter difference between the updated shadow student network branch parameters and the original student fault diagnosis network parameters is weighted according to the acceptance contribution score and written back to the student fault diagnosis network. Otherwise, the update of the shadow student network branch is discarded. The calculation rules, screening thresholds, acceptance criteria, and category acceptance gaps for the candidate acceptance ranking score are updated based on the acceptance contribution score. The student fault diagnosis network is then formally trained based on the labeled fault sample set and the formal acceptance sample set, and the fault diagnosis results are output.

[0008] To achieve the above objectives, this invention also provides a semi-supervised imbalanced fault diagnosis system, comprising a teacher fault diagnosis network, a student fault diagnosis network, a sample construction module, a candidate acceptance ranking module, a shadow trial learning module, an acceptance contribution evaluation module, an acceptance write-back module, a rule update module, and a formal training module. The sample construction module is used to construct a labeled reference sample set and weakly and strongly augmented samples of unlabeled fault samples based on the labeled fault sample set. The teacher fault diagnosis network outputs pseudo-labels, class prediction probabilities, intermediate representations, and prediction uncertainties for the weakly augmented samples. The candidate acceptance ranking module calculates candidate acceptance ranking scores based on class prediction probabilities, prediction uncertainties, class acceptance gaps, and the representation distance of intermediate representations relative to the pseudo-label class prototypes, and combines these scores with a screening threshold to form trial batches. The shadow trial learning module replicates the student fault diagnosis network to obtain a shadow student network branch, and utilizes... The trial batch performs trial updates with pseudo-label supervision, weak-strong reinforcement consistency constraints, and teacher-student distillation. The acceptance contribution evaluation module calculates the acceptance contribution score based on the accuracy increment and loss reduction on the labeled reference sample set before and after the trial update, the improvement in weak-strong reinforcement prediction consistency, and the reduction in category acceptance gaps in the trial batch. The acceptance write-back module writes the unlabeled fault samples corresponding to the trial batch into the formal acceptance sample set when the acceptance contribution score meets the acceptance conditions, and writes back the parameter difference to the student fault diagnosis network weighted by the acceptance contribution score; otherwise, it discards the update of the shadow student network branch. The rule update module updates the calculation rules, screening thresholds, acceptance conditions, and category acceptance gaps of the candidate acceptance ranking score based on the acceptance contribution score. The formal training module formally trains the student fault diagnosis network based on the labeled fault sample set and the formal acceptance sample set, and outputs the fault diagnosis results.

[0009] IV. Beneficial Effects Compared with the prior art, the present invention has at least the following beneficial effects:

[0010] Firstly, this invention transforms the process of utilizing candidate samples from "direct acceptance after meeting the conditions" to a closed-loop mechanism of "candidate screening - shadow probing update - contribution evaluation - formal acceptance - parameter differential write-back". This allows unlabeled samples to undergo real training benefit verification before entering formal training, thereby effectively reducing the risk of mis-acceptance caused by false labels and low-yield samples directly entering formal training, and improving the controllability and stability of the semi-supervised training process.

[0011] Secondly, this invention introduces the category acceptance gap and the representation distance of intermediate representations relative to pseudo-labeled category prototypes, and uses them together with category prediction probability and prediction uncertainty for candidate acceptance sample screening and ranking. This not only prioritizes the retention of highly reliable unlabeled samples, but also enhances the participation opportunities of minority class samples and boundary samples in the acceptance process, thereby alleviating the dominant role of majority class samples in the training process and improving the minority class fault identification capability in scenarios with imbalanced class distribution.

[0012] Third, this invention constructs a shadow student network branch to perform a preset number of trial updates on candidate samples, and calculates the acceptance contribution score based on the changes in accuracy, loss, weak-strong enhancement prediction consistency, and the decrease in class acceptance gap on the labeled reference sample set before and after the trial update. This makes the sample acceptance judgment no longer rely on a single confidence level or static threshold, but is based on a quantitative evaluation of the candidate sample's contribution to the current model's actual training, thereby improving the accuracy of sample acceptance decisions.

[0013] Fourth, this invention only writes back the parameter differences generated by the shadow student network branch to the current student fault diagnosis network in a weighted manner according to the acceptance contribution score when the acceptance contribution score meets the acceptance condition. This establishes a controlled coupling relationship between the model parameter update and the real benefits of the candidate samples, avoids low-quality trial updates from polluting the main student network, and thus improves the robustness, convergence stability and engineering deployability of the model training.

[0014] Fifth, this invention further uses the acceptance contribution score to update the calculation rules for candidate acceptance ranking scores, the acceptance threshold in the acceptance conditions, and the category acceptance gap, so that the sample acceptance mechanism can adaptively correct itself based on the actual training effect of the previous round of candidate samples, forming a closed loop of linkage between sample screening rules, acceptance conditions, and model updates, thereby improving the adaptability of unlabeled sample utilization strategies at different training stages.

[0015] Sixth, in the formal training phase, the present invention combines labeled fault samples and formally accepted samples to jointly train the student fault diagnosis network, and maintains a stable knowledge transfer relationship between the teacher fault diagnosis network and the student fault diagnosis network. Under the condition of limited labeled samples, it can further improve the recognition accuracy, generalization ability and anti-interference ability of the fault diagnosis model for complex working condition disturbances, weak fault features and class imbalance scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the adaptive semi-supervised imbalance fault diagnosis method provided in the first embodiment of the present invention.

[0017] Figure 2 This is a block diagram of the adaptive semi-supervised unbalanced fault diagnosis system provided in the second embodiment of the present invention. Detailed Implementation

[0018] First Embodiment

[0019] In the first embodiment, as Figure 1 As shown, the adaptive semi-supervised unbalanced fault diagnosis method of the present invention includes the following steps.

[0020] S1. Construction of labeled fault sample set, unlabeled fault sample set, labeled reference sample set, and weakly and strongly enhanced samples.

[0021] First, the raw monitoring signals of the device under different operating states are acquired. These raw monitoring signals are then truncated, segmented, normalized, and detrended, and samples are formed according to a preset window length and step size. This constructs a labeled fault sample set and an unlabeled fault sample set, denoted as follows: , in, This indicates that a set of faulty samples has been labeled. Indicates the first One labeled faulty sample, Indicates the first The actual fault category label corresponding to each labeled fault sample This indicates the total number of marked faulty samples; This represents the set of unlabeled fault samples. Indicates the first One unlabeled fault sample, This represents the total number of unlabeled fault samples.

[0022] A labeled reference sample set is constructed based on the labeled fault sample set. Preferably, a stratified sampling method is used to extract a portion of samples from the labeled fault sample set according to the category ratio to form a fixed labeled reference sample set. , in, Indicates the labeled reference sample set, Indicates the first One labeled reference sample, Indicates the first Each marker refers to the actual fault category label corresponding to the sample. This represents the total number of labeled reference samples. Preferably, the labeled reference sample set remains unchanged throughout the semi-supervised training process to ensure that the benchmark for comparing accuracy and loss is consistent before and after the trial update.

[0023] For each unlabeled faulty sample, weak enhancement and strong enhancement are performed separately to obtain weakly enhanced samples and strongly enhanced samples: , in, Indicates the first Weakly enhanced samples corresponding to unlabeled fault samples Indicates the first Strongly augmented samples corresponding to unlabeled fault samples This represents a weak enhancement function. This represents a strong enhancement function. Preferably, weak enhancement includes one or more of slight time shift, slight amplitude scaling, slight noise perturbation, and local smoothing; strong enhancement includes two or more of large time shift, frequency band perturbation, random occlusion, random noise superposition, and random deactivation.

[0024] During training initialization, the initial formally accepted sample set is set to an empty set, and the initial unlabeled fault sample pool is set to the unlabeled fault sample set, that is: , set up For time steps The official acceptance sample set, Set time step The pool of unlabeled faulty samples.

[0025] The purpose of this step is to establish the data foundation used by subsequent teacher fault diagnosis networks, student fault diagnosis networks, candidate acceptance ranking, trial updates, acceptance contribution evaluation, and formal training. Its beneficial effect is that it enables subsequent steps to process data around the same data chain, thereby improving the overall consistency and controllability of the method.

[0026] S2. Construction and initialization training of teacher fault diagnosis network and student fault diagnosis network.

[0027] A teacher fault diagnosis network and a student fault diagnosis network are constructed based on the labeled fault sample set. Preferably, both the teacher and student fault diagnosis networks include a feature extraction backbone network and a classification head. The feature extraction backbone network is used to extract intermediate representations, and the classification head is used to output the unnormalized classification output vector logits for each fault category and the category prediction probability. The parameters of the teacher fault diagnosis network are denoted as... The parameters of the student fault diagnosis network are denoted as follows: .

[0028] In one implementation, the teacher fault diagnosis network is first pre-trained using a labeled fault sample set, and then the student fault diagnosis network is pre-trained using the same labeled fault sample set. Preferably, the pre-training uses the Adam optimizer, and the supervised loss uses cross-entropy loss. For the student fault diagnosis network, at time step... The supervised loss for labeled faulty samples can be expressed as: , in, Indicates time step The labeled faulty sample training batch, Indicates time step The number of samples in the labeled faulty sample training batch. This indicates that the student fault diagnosis network is at time step For the Each labeled fault sample outputs its true fault category. The predicted probability. The supervised training method for the teacher fault diagnosis network is the same as that for the student fault diagnosis network.

[0029] The purpose of this step is to enable the teacher fault diagnosis network and the student fault diagnosis network to acquire initial fault identification capabilities. Its beneficial effect is to provide stable initial model parameters for subsequent pseudo-label output, category prediction probability calculation, intermediate representation extraction, prediction uncertainty estimation, and formal training.

[0030] S3. The teacher fault diagnosis network outputs pseudo-labels, category prediction probabilities, intermediate representations, and prediction uncertainties for weakly enhanced samples.

[0031] At time step , will the Weakly enhanced samples of unlabeled fault samples Inputting the teacher fault diagnosis network yields its intermediate representation, unnormalized classification output vector, and class prediction probability vector, denoted as follows: , and Among them, the teacher fault diagnosis network has a time step Judgment No. The unlabeled fault sample belongs to category The predicted probability is denoted as .

[0032] No. The pseudo-labels of the unlabeled fault samples satisfy: , in, Indicates the first A pseudo-label for an unlabeled fault sample. This indicates the total number of fault categories.

[0033] In order to obtain the Prediction uncertainty of unlabeled fault samples Preferably, multiple random forward propagations with Dropout are performed on the same weakly enhanced sample. Let the number of random forward propagations be... , No. The categories obtained from the second random forward propagation The predicted probability is denoted as Then the average class prediction probability satisfies: , Predicted entropy satisfies: , Mutual information satisfies: , Therefore, the prediction uncertainty satisfies: , in, Indicates the first An unlabeled fault sample at time step Category The average predicted probability, Indicates the first An unlabeled fault sample at time step Predicted entropy, Indicates the first An unlabeled fault sample at time step mutual information, It represents a small constant to prevent logarithmic singularities.

[0034] The purpose of this step is to provide pseudo-labels, category prediction probabilities, intermediate representations, and prediction uncertainties for subsequent calculations of category acceptance gaps, construction of pseudo-label category prototypes, calculation of representation distance, and calculation of candidate acceptance ranking scores. Its beneficial effect is that the acceptance of unlabeled fault samples no longer depends solely on a single confidence level, but can comprehensively consider the multidimensional information output by the teacher fault diagnosis network.

[0035] S4. Calculation of Category Acceptance Shortfall

[0036] At time step Based on the number of samples in each real label category in the labeled fault sample set and the time step The number of pseudo-labels accepted for each category of previously officially accepted samples is used to calculate the pseudo-label categories. The target acceptance rate and category acceptance gap. First, pseudo-label categories. The target acceptance rate is met: , Pseudo-label categories At time step Category acceptance gaps satisfy: , in, Indicates pseudo-label category The target acceptance rate Indicates the category of pseudo-labels One-to-one corresponding real tag category index, This indicates the set of labeled faulty samples and the pseudo-label category. The corresponding number of samples for the real label categories, This indicates that the true label category in the set of labeled faulty samples is The number of samples, Indicates time step pseudo-tag categories The category accepts vacancies. Indicates time step The pseudo-label category in the previously officially accepted historical samples was: The number of samples, Indicates time step The pseudo-label category in the previously officially accepted historical samples was: The number of samples.

[0037] The purpose of this step is to quantify the compensation needs of each pseudo-label category during the formal acceptance process; its beneficial effect is that it enables the subsequent candidate acceptance ranking and screening threshold generation processes to explicitly perceive the class imbalance, thereby increasing the utilization opportunities of minority class samples.

[0038] S5, Calculation of pseudo-label category prototype, representation distance, and candidate acceptance ranking score.

[0039] At time step Construct pseudo-label categories The pseudo-label category prototype. Let the time step be... Used to construct pseudo-label categories The prototype's sample set is The set From the real label category The labeled fault samples and at time step Previously, the pseudo-tag category was The historical formally accepted samples together constitute the pseudo-label category. The pseudo-label category prototype satisfies: , If there is no pseudo-label category before the current time step. The historical formally accepted samples, then the set Only by real label category It consists of labeled fault samples.

[0040] No. An unlabeled fault sample at time step The distance between the intermediate representation and its pseudo-label category prototype satisfies: , Furthermore, the first An unlabeled fault sample at time step The candidate acceptance ranking score satisfies: , in, Indicates time step pseudo-tag categories The pseudo-label category prototype, Represents a set The number of samples in This indicates that the teacher fault diagnosis network is effective for the sample. Intermediate representation of the output, Indicates the first An unlabeled fault sample at time step The distance between the intermediate representation and its pseudo-label category prototype. Indicates pseudo-tags At time step The corresponding pseudo-label category prototype, Represents the L2 norm, Indicates the first An unlabeled fault sample at time step The candidate acceptance ranking score, , , and This represents the weighting coefficient.

[0041] The purpose of this step is to evaluate the value of unlabeled faulty samples by utilizing class acceptance gaps, pseudo-label class prototypes, and representation distance; its beneficial effect is that it enables the candidate acceptance ranking score to simultaneously take into account sample reliability, class compensation needs, and the value of representation boundary expansion.

[0042] S6. Generation of screening thresholds, screening of candidate acceptance samples, and formation of trial batches.

[0043] At time step Based on the current student fault diagnosis network's recognition results on the labeled reference sample set, the average diagnostic accuracy of the labeled reference sample set is calculated: , in, This indicates that the student fault diagnosis network is for the first... Each labeled reference sample belongs to the category. The predicted probability, Indicates an indicator function.

[0044] Furthermore, time step pseudo-tag categories The category confidence threshold and uncertainty threshold satisfy the following conditions: , Therefore, time step The candidate acceptance sample set satisfies: , And ranked according to the candidate acceptance score Select from largest to smallest Each candidate acceptance sample constitutes a time step Trial batch: , in, Indicates time step pseudo-tag categories Category confidence threshold, This represents the interval cutoff function. This represents the adjustment coefficient for the acceptance deficit. and These represent the lower and upper limits of the confidence threshold, respectively. Indicates time step Uncertainty threshold, The quantile parameter is quantile function, Indicates time step quantile parameters, Indicates time step The candidate acceptance sample set Indicates time step The number of trial batch samples, Indicates time step The trial batch.

[0045] when At this time step, no trial update is performed; instead, the command is executed directly. , and And then proceed to the next time step.

[0046] The purpose of this step is to combine the candidate acceptance ranking score with the screening threshold to screen unlabeled faulty samples for candidate acceptance. Its beneficial effect is that the samples entering the trial update meet the reliability requirements while also taking into account the category compensation needs and the value of characterizing differences.

[0047] S7. Replication of shadow student network branches, distillation weight calculation, and trial update.

[0048] At time step The student fault diagnosis network is replicated to obtain a shadow student network branch. Preferably, the parameters of the shadow student network branch before the trial update are the same as those of the current student fault diagnosis network, denoted as... The parameters of the shadow student network branch after the trial update are denoted as follows: Under conditions where the teacher fault diagnosis network is fixed or updates slowly, time steps are utilized. trial batch Implement trial updates that include pseudo-label supervision, weak-strong reinforcement consistency constraints, and teacher-student distillation.

[0049] No. A trial sample at time step The distillation weights satisfy: , Time step The pseudo-label supervision loss, weak-strong reinforcement consistency loss, distillation loss, and trial update total loss satisfy the following: , , in, Indicates the first A trial sample at time step Distillation weight, Indicates time step The loss of supervision due to false labels This indicates that the student fault diagnosis network is for the first... The strongly augmented sample of the trial sample outputs its pseudo-label. The predicted probability, Indicates time step The loss of consistency due to weak and strong enhancement. and These represent the student fault diagnosis network for the first... The class prediction probability vectors output by the weakly and strongly enhanced samples of the trial samples. Indicates time step Distillation loss, Indicates the Kullback-Leibler divergence. This represents the softmax function. This indicates that the teacher fault diagnosis network is for the first... The unnormalized classification output vector of the weakly augmented sample output of the trial sample. This indicates that the student fault diagnosis network is for the first... The unnormalized classification output vector of the strongly augmented sample output of the trial sample. Indicates the distillation temperature coefficient. Indicates time step The total loss from the trial update, and This represents the loss weighting coefficient.

[0050] In one implementation, the shadow student network branch uses the Adam optimizer to perform backpropagation updates. The number of trial update steps is preferably one or more, and the trial update learning rate is less than or equal to the formal training learning rate. During the trial update process, the parameters of the teacher fault diagnosis network remain fixed; or a slow update method with a speed much slower than that of the student fault diagnosis network is used.

[0051] The purpose of this step is to pre-validate the training value of candidate acceptance samples without directly polluting the parameters of the master student fault diagnosis network. Its beneficial effect is that only samples that can bring positive contributions after trial updates will enter the formal acceptance and parameter differential write-back process.

[0052] S8. Calculation of Acceptance Contribution Score

[0053] At time step Based on the changes in accuracy, loss reduction, improvement in predictive consistency of weak and strong enhancements in the trial batch, and reduction in class acceptance gaps on the labeled reference sample set before and after the shadow student network branch trial update, the acceptance contribution score is calculated.

[0054] Time step The accuracy increment satisfies: ,in, and Representing time steps The average diagnostic accuracy of the labeled reference sample set before and after the trial update.

[0055] Time step The average loss of the labeled reference sample set can be defined using the average cross-entropy. The average loss before and after the trial update are denoted as follows: and Then the decrease in loss satisfies: ,in, In the formula, the superscript express or , This indicates that the student fault diagnosis network, before or after the trial update, affects the first... Each labeled reference sample outputs its true category. The predicted probability.

[0056] Time step The improvement in consistency of weak and strong enhancement predictions satisfies: .

[0057] Preferably, the consistency can be measured by the mean cosine similarity between the predicted probability vectors of weakly enhanced samples and strongly enhanced samples, i.e.: , In the formula, superscript express or , and These represent the student fault diagnosis network before or after the trial update, respectively, for the first... The class prediction probability vectors output by weakly and strongly enhanced samples of the trial samples.

[0058] Time step The category acceptance deficit decrease meets the following requirements: , in, Indicates time step Exploring pseudo-tag categories before update The category accepts vacancies. Indicates time step The trial batch is considered as a proposed formal acceptance of the time-fake label category. The category accepts vacancies. Preferably, By testing batches Chinese pseudo-label category is The sample is considered as the sample to be formally accepted and the category acceptance gap is recalculated.

[0059] Therefore, time step The acceptance contribution score satisfies: , in, Indicates time step Acceptance contribution score, , , and This represents the weighting coefficient.

[0060] The purpose of this step is to quantify the combined benefits of the trial batch on the main task's performance, consistency improvement, and class balance improvement into a unified acceptance contribution score; its beneficial effect is that it provides a clear quantitative basis for subsequent formal acceptance and parameter differential write-back.

[0061] S9. Admission condition judgment, parameter differential write-back, and formal admission sample set writing.

[0062] At time step Based on the contribution score for acceptance and the acceptance threshold in the acceptance conditions Calculate the parameter differential write-back coefficients: , when and At that time, it is believed that time steps If the trial batch meets the acceptance criteria, the unlabeled fault samples corresponding to the trial batch will be written into the formal acceptance sample set. At the same time, the parameter differences will be weighted according to the acceptance contribution score and written back to the student fault diagnosis network. , when or At that time, it is believed that time steps The trial batch did not meet the acceptance criteria; therefore, the formal acceptance sample set and student fault diagnosis network parameters remain unchanged. , in, Indicates time step The parameter difference write-back coefficients, Indicates time step The acceptance threshold in the acceptance conditions. This indicates that the shadow student network branch is at time step Complete the trial and update of network parameters.

[0063] Preferably, when the trial batch Upon formal acceptance, it will be removed from the current pool of unlabeled faulty samples. Remove from the pool of unlabeled faulty samples for the next time step. When testing batches Maintain .

[0064] The purpose of this step is to make a formal decision to accept or discard the trial batch based on the acceptance contribution score and acceptance conditions. Its beneficial effect is that the parameter update of the main student fault diagnosis network only absorbs the trial update results with positive training contributions, thereby reducing the adverse effects of falsely accepted samples on the main model.

[0065] S10. Updates to screening thresholds, acceptance criteria, and category-based acceptance gaps.

[0066] Complete the time step After the acceptance judgment, based on the acceptance contribution score Update the screening thresholds and acceptance criteria. Preferably, first based on time steps. The official acceptance sample set Recalculate pseudo-label categories Category Acceptance Shortfall Update the time step again Category confidence threshold and uncertainty threshold.

[0067] Time step The acceptance threshold in the acceptance conditions satisfies: , Time step pseudo-tag categories The category confidence threshold satisfies: , Time step The uncertainty threshold satisfies: , in, Indicates an indicator function, , , , , and This indicates the update step size.

[0068] It should be noted that, due to the candidate acceptance ranking score The calculation explicitly includes category acceptance gaps. Therefore, after the category acceptance gap is updated, the calculation rules for the candidate acceptance ranking score are also updated accordingly; at the same time, since the candidate sample selection also depends on... and Therefore, after the screening threshold is updated, the rules for forming candidate acceptance samples in subsequent time steps also change synchronously. Thus, the time step... Acceptance contribution score It can provide feedback to drive the linked updates of the calculation rules for candidate acceptance ranking scores, screening thresholds, acceptance conditions, and category acceptance gaps.

[0069] The purpose of this step is to enable various gating rules and compensation quantities to be dynamically adjusted based on the actual gains from the previous time step; its beneficial effect is to improve the adaptability of the strategy for utilizing unlabeled fault samples at different training stages.

[0070] S11. Formal training and fault diagnosis result output based on the labeled fault sample set and the formally accepted sample set.

[0071] At time step Based on the labeled fault sample set and the formally accepted sample set Formal training was provided to students on fault diagnosis networks. (Time step) The officially accepted sample pseudo-label supervision loss satisfies: , in, This indicates the category acceptance deficit amplification factor.

[0072] Time step The relationship maintains that distillation loss satisfies: , in, This represents the inner product operation. and These represent the teacher fault diagnosis network and the student fault diagnosis network, respectively, for the first... Intermediate representation of the output of a formally accepted sample and These represent the teacher fault diagnosis network and the student fault diagnosis network, respectively, for the first... Intermediate representation of the output of a formally accepted sample.

[0073] Time step The total training loss satisfies: .

[0074] In one implementation, the student fault diagnosis network uses the Adam optimizer based on... Perform backpropagation updates. The network parameters for teacher fault diagnosis are updated using an exponential moving average method. , in, Indicates time step The supervised loss of the labeled faulty samples, and This represents the loss weighting coefficient. This represents the momentum coefficient.

[0075] Once the set number of training rounds is reached or the convergence condition is met, the trained student fault diagnosis network will be used as the final fault diagnosis model. For the samples to be diagnosed... The data is then preprocessed in the same way as during the training phase and input into the final student fault diagnosis network to obtain the output category prediction probability vector. The fault diagnosis results meet the following requirements: , in, Indicates the sample to be diagnosed The fault diagnosis results Indicates the sample to be diagnosed Category The predicted probability.

[0076] The purpose of this step is to formally train the student fault diagnosis network using the labeled fault sample set and the formally accepted sample set, and to maintain stable teacher guidance by updating the teacher fault diagnosis network through the exponential moving average. Its beneficial effects are to improve fault diagnosis accuracy, training stability and minority class recognition ability under conditions of limited labeling, class imbalance and noise perturbation.

[0077] Second Embodiment

[0078] The second embodiment of the present invention only describes the contents that are different from those of the first embodiment; the contents that are the same will not be described again.

[0079] like Figure 2 As shown, the present invention also provides an adaptive semi-supervised imbalance fault diagnosis system. The system includes a teacher fault diagnosis network, a student fault diagnosis network, a sample construction module, a candidate acceptance ranking module, a shadow trial learning module, an acceptance contribution evaluation module, an acceptance write-back module, a rule update module, and a formal training module.

[0080] The system comprises several modules: a sample construction module for building a labeled reference sample set and weakly and strongly augmented samples of unlabeled fault samples based on the labeled fault sample set, forming an initial formal acceptance sample set and an unlabeled fault sample pool; a teacher fault diagnosis network for outputting pseudo-labels, class prediction probabilities, intermediate representations, and prediction uncertainties for weakly augmented samples; a candidate acceptance ranking module for calculating candidate acceptance ranking scores based on class prediction probabilities, prediction uncertainties, class acceptance gaps, and the representation distance of intermediate representations relative to pseudo-labeled class prototypes, and forming a trial batch based on a selection threshold; a shadow trial learning module for replicating the student fault diagnosis network to obtain a shadow student network branch, and using the trial batch to perform trial updates with pseudo-label supervision, weak-strong augmentation consistency constraints, and teacher-student distillation; and an acceptance contribution evaluation module. The system is used to calculate the acceptance contribution score based on the accuracy increment and loss reduction on the labeled reference sample set before and after the trial update, the improvement in prediction consistency of weak and strong enhancements in the trial batch, and the reduction in category acceptance gaps. The acceptance write-back module is used to write the unlabeled fault samples corresponding to the trial batch into the formal acceptance sample set when the acceptance contribution score meets the acceptance conditions, and write back the parameter difference to the student fault diagnosis network in a weighted manner according to the acceptance contribution score; otherwise, the update of the shadow student network branch is discarded. The rule update module is used to update the calculation rules, screening thresholds, acceptance conditions, and category acceptance gaps of the candidate acceptance ranking score based on the acceptance contribution score. The formal training module is used to formally train the student fault diagnosis network based on the labeled fault sample set and the formal acceptance sample set, and output the fault diagnosis results in combination with the momentum update of the teacher fault diagnosis network.

[0081] In one embodiment, the above modules can be implemented by a processor calling program instructions from memory, or by dedicated hardware circuits, programmable logic devices, or a combination of hardware and software. The system implementation and the method implementation described above correspond to each other in terms of data objects, execution order, parameter meanings, and technical effects.

[0082] experiment

[0083] To verify the effectiveness of the proposed adaptive semi-supervised unbalanced fault diagnosis method based on shadow trial update, acceptance contribution evaluation, parameter differential write-back, and rule-linked update, an experimental verification was conducted using a vibration fault dataset of a rotating machinery bearing. This dataset contains five states: normal state, inner race fault, outer race fault, rolling element fault, and combined fault, totaling 6000 samples. The dataset is divided into training, validation, and test sets in a 6:2:2 ratio, with the training set containing 3600 samples. To simulate the application conditions of a real-world industrial scenario—limited labeled samples, sufficient unlabeled samples, and an imbalanced class distribution—only 10% of the samples in the training set are retained as labeled fault samples, while the remaining 90% are treated as unlabeled fault samples. Furthermore, the number of training samples for the five states is set to 1080, 900, 720, 540, and 360, respectively, to construct an unbalanced training scenario. In the experiment, the teacher fault diagnosis network and the student fault diagnosis network used the same one-dimensional convolutional network skeleton. The optimizer used was the Adam (adaptive moment estimator), with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. All experimental results were the average of 5 independent and repeated experiments.

[0084] Evaluation metrics include overall accuracy (Acc), macro-F1 score, minority class recall, formally accepted precision, and false acceptance rate (FAR). Minority class recall represents the average recall of the two least represented fault classes in the test set; formally accepted precision represents the proportion of correctly labeled false samples among the officially accepted unlabeled samples; and false acceptance rate (FAR) represents the proportion of incorrectly labeled false samples among the officially accepted unlabeled samples.

[0085] 1. Comparison Method Settings

[0086] To fully verify the effectiveness of the technical solution defined in the current claims of this invention, the following comparison method is set up: (1) Fully supervised method: The student fault diagnosis network is trained using only labeled fault samples, without using unlabeled fault samples; (2) Direct Admission Semi-Supervised Method: Based on pseudo-labels and consistency constraints, it utilizes unlabeled fault samples, but does not employ candidate admission ranking, shadow trial update, admission contribution evaluation, parameter differential write-back, and rule linkage update. (3) Candidate ranking semi-supervised method: Based on the direct admission semi-supervised method, a candidate admission ranking based on prediction probability, prediction uncertainty and category admission gap is introduced, but shadow trial update, admission contribution evaluation and parameter differential write-back are not adopted. (4) No write-back linkage method: Candidate admission sorting, shadow trial update and admission contribution evaluation are adopted, but parameter differential write-back and rule linkage update driven by admission contribution score are not adopted; (5) The method of the present invention adopts a complete scheme of candidate acceptance sorting, shadow probing update, acceptance contribution evaluation, parameter differential write-back and rule linkage update.

[0087] 2. Overall Comparative Experimental Results

[0088] Table 1 presents the overall performance comparison results of different methods on the test set.

[0089] Table 1. Overall Comparison Results of Different Methods / %

[0090] Full supervision method 86.7 81.9 70.8 — — Directly adopt semi-supervised methods 90.4 86.2 75.6 87.2 12.8 Candidate ranking semi-supervised method 92.1 88.4 79.3 89.9 10.1 No write-back linkage method 93.5 90.7 83.8 91.4 8.6 Method of the present invention 95.8 93.9 89.7 94.9 5.1

[0091] As shown in Table 1, the method of this invention achieves optimal results in terms of overall accuracy, macro-average F1 score, minority class average recall, formal acceptance sample precision, and false acceptance rate. Compared with the fully supervised method, the method of this invention improves Acc by 9.1 percentage points, Macro-F1 by 12.0 percentage points, and minority class average recall by 18.9 percentage points, indicating that this invention can effectively utilize a large number of unlabeled samples under limited labeled sample conditions. Compared with the direct acceptance semi-supervised method, the formal acceptance sample precision of this invention increases from 87.2% to 94.9%, and the false acceptance rate decreases from 12.8% to 5.1%, demonstrating that this invention, through candidate acceptance ranking, shadow trial update, and acceptance contribution evaluation, can verify the true training value of unlabeled samples before they enter formal training. Compared with the write-back-linked method, the method of this invention further improves Acc by 2.3 percentage points, Macro-F1 by 3.2 percentage points, the average recall rate of minority classes by 5.9 percentage points, and the false acceptance rate by 3.5 percentage points, indicating that parameter differential write-back and rule linkage update have a significant effect on improving sample utilization quality and training stability.

[0092] 3. Ablation Experiment Results

[0093] To further verify the function of each component technical feature of the current claim, the following ablation scheme is proposed:

[0094] Comparative Example 1: Only pseudo-label supervision and weak-strong reinforcement consistency constraints are used; Comparative Example 2: On the basis of Comparative Example 1, candidate acceptance ranking is added; Comparative Example 3: On the basis of Comparative Example 2, representation distance constraints between the category prototype and intermediate representation and the pseudo-label category prototype are added; Comparative Example 4: On the basis of Comparative Example 3, shadow probing update and acceptance contribution evaluation are added, but parameter differential write-back and rule linkage update are not performed; The method of this invention: On the basis of Comparative Example 4, parameter differential write-back and rule linkage update are further added.

[0095] Table 2 presents the ablation experiment results. The numerical values ​​are the same as those in the original manuscript, but the experimental group names have been rewritten according to the current claim structure.

[0096] Table 2 Ablation Experiment Results / %

[0097] Comparative Example 1 90.8 86.6 76.1 87.6 12.4 Comparative Example 2 92.0 88.1 79.0 89.3 10.7 Comparative Example 3 93.1 89.8 81.6 91.0 9.0 Comparative Example 4 94.0 91.4 85.2 92.2 7.8 Method of the present invention 95.8 93.9 89.7 94.9 5.1

[0098] As shown in Table 2, candidate acceptance ranking, representation distance between category prototype and intermediate representation and pseudo-label category prototype, shadow trial update and acceptance contribution evaluation, parameter differential write-back and rule linkage update can all contribute positively to diagnostic performance. Specifically, after adding candidate acceptance ranking to Comparative Example 1 to form Comparative Example 2, Acc increased by 1.2 percentage points and Macro-F1 increased by 1.5 percentage points, indicating that prioritizing unlabeled samples helps to select high-value candidate samples. After adding the representation distance between the class prototype and intermediate representation relative to the pseudo-label class prototype to form Comparative Example 3, the minority class average recall increased from 79.0% to 81.6%, indicating that the representation distance constraint is beneficial to enhancing the utilization value of boundary samples and minority class samples. After adding shadow trial update and acceptance contribution evaluation to Comparative Example 3 to form Comparative Example 4, Acc increased by 0.9 percentage points and FAR decreased from 9.0% to 7.8%, indicating that pre-validating the real training benefits of candidate samples can further reduce false acceptances. After further adding parameter differential write-back and rule linkage update to Comparative Example 4 to form the method of this invention, Acc increased by 1.8 percentage points, Macro-F1 increased by 2.5 percentage points, minority class average recall increased by 4.5 percentage points, and FAR increased. The decrease from 7.8% to 5.1% indicates that the parameter differential write-back driven by the contribution score and the rule linkage update are the key technical aspects of this invention to achieve further performance improvement.

[0099] 4. Noise Resistance Comparison Experiment Results

[0100] To verify the robustness of the present invention under complex working conditions and noise interference, Gaussian noise with a signal-to-noise ratio of 6 dB and 0 dB was superimposed on the test set, and the performance changes of different methods were compared.

[0101] Table 3. Comparison results under different noise conditions / %

[0102] Directly adopt semi-supervised methods 90.4 84.8 79.3 86.2 80.1 74.6 No write-back linkage method 93.5 89.2 84.7 90.7 85.4 80.8 Method of the present invention 95.8 92.6 88.9 93.9 89.7 85.8

[0103] As shown in Table 3, the performance of all methods decreased after noise was added, but the decrease was the smallest for the method of this invention. Under 6 dB noise conditions, the Acc of the method of this invention still reached 92.6%, which is 3.4 percentage points higher than the method without write-back linkage; under 0 dB strong noise conditions, the Acc of the method of this invention still reached 88.9%, which is 4.2 percentage points higher than the method without write-back linkage; Macro-F1 also showed the same trend. The above results indicate that the present invention, through the collaborative design of candidate admission ranking, shadow trial update, admission contribution evaluation, parameter differential write-back, and rule linkage update, can improve the model's adaptability to noise disturbances and complex operating conditions, thereby enhancing the robustness and generalization ability of the fault diagnosis model.

[0104] 5. Experimental Conclusions

[0105] The experimental results summarized in Tables 1 to 3 show that the method of this invention outperforms the comparative methods in terms of overall accuracy, macro-average F1 score, minority class average recall, formally accepted sample precision, and false acceptance rate. Especially in scenarios with imbalanced class distribution and noise interference, the improvement in the ability to identify minority class fault samples is more significant. Experimental results demonstrate that this invention, by organically combining candidate acceptance ranking, shadow probing update, acceptance contribution evaluation, parameter differential write-back, rule-linked update, and formal training, can effectively improve the recognition accuracy, training stability, minority class friendliness, and noise robustness of the fault diagnosis model under limited labeling conditions, thus verifying the significant technical effectiveness of the present invention.

[0106] The above embodiments, in accordance with the technical features defined in the claims, have provided a detailed description of the technical solution of the present invention. It should be understood that the above embodiments are merely preferred embodiments of the present invention and not limitations on the scope of protection of the present invention. For those skilled in the art, any equivalent substitutions, obvious modifications, and improvements made to the sample augmentation methods, network skeleton, specific implementation methods of the loss function, details of evaluation index calculation, parameter update strategies, and module partitioning methods without departing from the concept and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. An adaptive semi-supervised unbalanced fault diagnosis method, characterized in that, include: The original monitoring signals of the device under different operating states are acquired, and the original monitoring signals are intercepted, segmented, normalized and detrended. Samples are formed according to the preset window length and step size, thereby constructing a labeled fault sample set and an unlabeled fault sample set. The different operating states are normal state, inner ring fault, outer ring fault, rolling element fault and compound fault. Based on the labeled fault sample set, a teacher fault diagnosis network, a student fault diagnosis network, and a labeled reference sample set are constructed. Weak enhancement samples and strong enhancement samples are constructed for the unlabeled fault samples in the unlabeled fault sample set. The teacher fault diagnosis network is used to output pseudo-labels, category prediction probabilities, intermediate representations, and prediction uncertainties for weakly enhanced samples. Based on the predicted probability of the category, the prediction uncertainty, the category acceptance gap, and the representation distance of the intermediate representation relative to the pseudo-label category prototype, the candidate acceptance ranking score is calculated, and the candidate acceptance samples are selected and ranked in combination with the screening threshold to form a trial batch. The shadow student network branch is obtained by replicating the student fault diagnosis network. Under the condition that the teacher fault diagnosis network is fixed or slowly updated, the trial update with pseudo-label supervision, weak-strong reinforcement consistency constraints and teacher-student distillation is performed by using trial batches. Based on the accuracy increase and loss decrease on the labeled reference sample set before and after the trial update, the improvement in prediction consistency of the weak and strong enhancements in the trial batch, and the decrease in category acceptance gap, the acceptance contribution score is calculated, wherein the acceptance contribution score satisfies: ， ， ， ， ， in, Indicates time step The increase in the average diagnostic accuracy of the reference sample set before and after the shadow student network branch trial update. and Representing time steps The average diagnostic accuracy of the labeled reference sample set before and after the trial update; Indicates time step The average loss reduction of the reference sample set before and after the shadow student network branch trial update. and Representing time steps The average loss of the labeled reference sample set before and after the trial update; Indicates time step The improvement in prediction consistency between weakly enhanced samples and strongly enhanced samples in the trial batch. and Representing time steps The predictive consistency between weakly enhanced samples and strongly enhanced samples in the trial batch before and after the trial update. Indicates time step The trial batch is considered as the decrease in the category acceptance shortfall when formal acceptance is planned. This represents the pseudo-label category index. Indicates time step Exploring pseudo-tag categories before update The category accepts vacancies. Indicates time step The trial batch is considered as a proposed formal acceptance of the time-fake label category. Category acceptance of vacancies; Indicates time step Acceptance contribution score, , , and Indicates the weighting coefficient; When the acceptance contribution score meets the acceptance conditions, the unmarked fault samples corresponding to the trial batch are written into the formal acceptance sample set, and the parameter difference between the shadow student network branch parameters after the trial update and the student fault diagnosis network parameters before the trial update is weighted according to the acceptance contribution score and written back to the student fault diagnosis network; otherwise, the update of the shadow student network branch is discarded. The calculation rules, screening thresholds, admission conditions, and category admission gaps for candidate admission ranking scores are updated based on the admission contribution scores. The student fault diagnosis network is formally trained based on the labeled fault sample set and the formal admission sample set, and the fault diagnosis results are output.

2. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 1, characterized in that, The category acceptance shortfall meets the following requirements: ， ， in, Indicates pseudo-label category The target acceptance rate; Indicates the pseudo-label category index; Indicates the actual label category index; This represents the pseudo-label category index in the summation of the denominator; Indicates the category of pseudo-labels One-to-one corresponding index of real label categories; This indicates the set of labeled faulty samples and the pseudo-label category. The number of samples corresponding to the real label categories; This indicates that the true label category in the set of labeled faulty samples is The number of samples; Indicates the total number of fault categories; Indicates time step pseudo-tag categories Category acceptance of vacancies; Indicates time step The pseudo-label category in the previously officially accepted historical samples was: The number of samples; Indicates time step The pseudo-label category in the previously officially accepted historical samples was: The number of samples; This represents a small constant to prevent the denominator from being zero.

3. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 2, characterized in that, The pseudo-label category prototype, the representation distance of the intermediate representation relative to the pseudo-label category prototype, and the candidate acceptance ranking score respectively satisfy: ， ， ， in, Indicates time step pseudo-tag categories The pseudo-label category prototype; Indicates time step When the real label category is The labeled fault samples and at time step Previously, the pseudo-tag category was The set of historically formally accepted samples; Represents a set The number of samples in; This indicates that the teacher fault diagnosis network is effective for the sample. Intermediate representations of the output; Indicates the first An unlabeled fault sample at time step The distance between the intermediate representation and its pseudo-label category prototype; This indicates that the teacher fault diagnosis network is for the first... Intermediate representations of the output of unlabeled fault samples; Indicates the first A pseudo-label for an unlabeled faulty sample; Represents the L2 norm; Indicates the first An unlabeled fault sample at time step The candidate acceptance ranking score; This indicates the teacher's fault diagnosis network determination. The unlabeled fault sample belongs to its pseudo-label. The predicted probability; Indicates the first Prediction uncertainty for an unlabeled fault sample; , , and This represents the weighting coefficient.

4. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 3, characterized in that, The screening threshold and trial batch satisfy the following: ， ， ， And according to the candidate acceptance ranking score Select from largest to smallest The trial batch consists of 10 candidate acceptance samples. ; in, Indicates time step pseudo-tag categories Category confidence threshold; Represents the interval cutoff function; Indicates time step The average diagnostic accuracy of the labeled reference sample set; This indicates the adjustment coefficient for the acceptance deficit; and These represent the lower limit and upper limit of the confidence threshold, respectively; Indicates time step The uncertainty threshold; The quantile parameter is quantile function; Indicates time step quantile parameters; Indicates time step The pool of unlabeled faulty samples; Indicates the pool of unlabeled faulty samples The first in One unlabeled fault sample; Indicates time step The candidate acceptance sample set; Indicates time step The number of trial batch samples; Indicates time step The trial batch.

5. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 4, characterized in that, The distillation weight and trial loss in the trial update respectively satisfy: ， ， ， ， ， in, Indicates the first A trial sample at time step Distillation weight; Indicates a trial batch The number of samples in; Indicates a trial batch The sample index in; Indicates time step The loss of supervision due to false labels; This indicates that the student fault diagnosis network is for the first... The strongly augmented sample of the trial sample outputs its pseudo-label. The predicted probability; Indicates time step The loss of consistency due to weak and strong reinforcement; and These represent the student fault diagnosis network for the first... The class prediction probability vectors output by weakly enhanced and strongly enhanced samples of the trial sample; Indicates time step Distillation loss; Indicates the Kullback-Leibler divergence; This represents the softmax function; This indicates that the teacher fault diagnosis network is for the first... The unnormalized classification output vector of the weakly augmented sample output of the trial sample; This indicates that the student fault diagnosis network is for the first... The unnormalized classification output vector of the strongly augmented sample output of the trial sample; Indicates the distillation temperature coefficient; Indicates time step The total loss from the trial update; and This represents the loss weighting coefficient.

6. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 5, characterized in that, The acceptance conditions and parameter differential write-back satisfy the following: ， when and hour, ， ， when or hour, in, Indicates time step The parameter difference write-back coefficients; Indicates time step The acceptance threshold in the acceptance criteria; and Representing time steps and time step The official acceptance sample set; and Representing time steps and time step Student fault diagnosis network parameters; This indicates that the shadow student network branch is at time step Complete the trial and update of network parameters; This represents a small constant to prevent the denominator from being zero.

7. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 6, characterized in that, The updates to the screening threshold and acceptance criteria satisfy the following: ， ， ， in, Indicates time step The acceptance threshold in the acceptance criteria; Indicates an indicator function; Indicates time step pseudo-tag categories Category confidence threshold; Indicates time step pseudo-tag categories Category acceptance of vacancies; Indicates time step The uncertainty threshold; , , , , and This indicates the update step size.

8. The adaptive semi-supervised unbalanced fault diagnosis method according to claim 7, characterized in that, The formal training satisfies: ， ， ， ， in, Indicates time step The officially accepted sample pseudo-label supervision loss; Indicates formal acceptance of the sample set The number of samples in; Indicates the category acceptance deficit amplification factor; Indicates time step The relationship maintains the distillation loss; Indicates inner product operation; and These represent the teacher fault diagnosis network and the student fault diagnosis network, respectively, for the first... Intermediate representation of the output of a formally accepted sample; and These represent the teacher fault diagnosis network and the student fault diagnosis network, respectively, for the first... Intermediate representation of the output of a formally accepted sample; and Indicates formal acceptance of the sample set The sample index in; Indicates time step The supervised loss of labeled faulty samples; Indicates time step Total losses from formal training; and Indicates the loss weighting coefficient; and Representing time steps and time step Teacher fault diagnosis network parameters; This represents the momentum coefficient.

9. An adaptive semi-supervised unbalanced fault diagnosis system, used to implement the adaptive semi-supervised unbalanced fault diagnosis method according to any one of claims 1-8, characterized in that, It includes a teacher fault diagnosis network, a student fault diagnosis network, a sample construction module, a candidate acceptance ranking module, a shadow trial learning module, an acceptance contribution evaluation module, an acceptance write-back module, a rule update module, and a formal training module; The sample construction module is used to construct a labeled reference sample set, as well as weakly and strongly enhanced samples of unlabeled fault samples, based on the labeled fault sample set. The teacher fault diagnosis network is used to output pseudo-labels, class prediction probabilities, intermediate representations, and prediction uncertainties for weakly enhanced samples. The candidate acceptance ranking module is used to calculate the candidate acceptance ranking score based on the category prediction probability, prediction uncertainty, category acceptance gap, and the representation distance of the intermediate representation relative to the pseudo-label category prototype, and to form a trial batch by combining the screening threshold. The shadow trial learning module is used to replicate the student fault diagnosis network to obtain the shadow student network branch, and to perform trial updates with pseudo-label supervision, weak-strong reinforcement consistency constraints and teacher-student distillation using trial batches; The Admission Contribution Assessment module is used to calculate the admission contribution score based on the accuracy increment and loss reduction on the labeled reference sample set before and after the trial update, the improvement in the consistency of weak and strong enhancement predictions in the trial batch, and the reduction in the category admission deficit. The Acceptance Writeback Module is used to write the unmarked fault samples corresponding to the trial batch into the formal acceptance sample set when the acceptance contribution score meets the acceptance conditions, and to write back the parameter difference to the student fault diagnosis network in weighted according to the acceptance contribution score; otherwise, the update of the shadow student network branch is discarded. The rule update module is used to update the calculation rules, screening thresholds, admission conditions, and category admission gaps of candidate admission ranking scores based on the admission contribution scores. The formal training module is used to formally train the student fault diagnosis network based on the labeled fault sample set and the formally accepted sample set, and output the fault diagnosis results.

Images