A method and system for human-robot collaborative nuclear power plant maintenance based on joint loss function

Abstract

The application discloses a kind of man-machine collaborative nuclear power plant maintenance method and system based on joint loss function, and relates to predictive maintenance technical field.The method includes the following steps: obtaining the multi-source data of the real-time running state of nuclear power plant equipment, obtaining input data set;Nuclear power plant maintenance model is processed to input data set by artificial review and model direct prediction cooperation;Based on joint loss function, consider artificial review error and model direct prediction error combined reinforcement learning mechanism to adaptively adjust the parameters of nuclear power plant maintenance model;According to the prediction result of nuclear power plant maintenance, combined with the historical risk level of equipment and operation context, automatically generate maintenance scheme.The application realizes the full integration of operator's knowledge and experience and AI model capability by constructing man-machine collaborative learning mechanism, improves the accuracy, adaptability and credibility of predictive maintenance, and forms a stable and reliable intelligent closed loop in the whole life cycle management of equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of predictive maintenance, in particular to a man-machine collaborative nuclear power plant maintenance method and system based on a joint loss function. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] In process industry sites such as nuclear power plants and thermal power plants, predictive maintenance technology has become a core means to ensure equipment reliability. Predictive maintenance generally continuously collects multi-source operating data such as vibration, temperature, acoustics, and electricity through sensors distributed throughout the unit, uses machine learning models to assess equipment health status and estimate remaining useful life in real time, and thus changes "after-repair" to "pre-maintenance".

[0004] With the development of deep learning, algorithms represented by autoencoders and attention networks have shown higher sensitivity in anomaly detection and life prediction. However, the following bottlenecks still restrict the deployment: current training data can only cover limited historical failure modes, when new failure mechanisms or working condition drifts occur on site, the model often gives unreliable confidence, leading to false positives or false negatives. Moreover, the model is "frozen" once it is put into operation and cannot be updated as the equipment ages, the process is adjusted, or the personnel's experience is accumulated, and the performance declines over time. In addition, the algorithm output usually lacks traceable physical or statistical explanations, and the operation shift workers have difficulty in judging its rationality, and tend to "prefer manual inspection" rather than accepting system recommendations. Even if some interfaces allow manual correction of the results, these corrections only remain at the log level and are not systematically fed back to the model training process, and cannot form a continuous evolution of man-machine collaborative mechanism.

[0005] For nuclear power plants, regulatory requirements further amplify the above-mentioned contradictions, all automated decisions must be "verifiable, traceable, and intervenable", and the final decision-making power is always in the hands of human operators, and any closed loop that bypasses human supervision is considered unacceptable. Although the existing public technology proposes a system framework in which humans and artificial intelligence jointly participate in scheduling, it focuses more on energy distribution control and does not solve the core problem of "how human feedback drives model adaptive update", and lacks a dynamic man-machine collaborative decision-making mechanism based on uncertainty. Therefore, the existing technology still has the fundamental defect of "not being able to continuously inject human experience into the model and not being able to let the model evolve on a compliant premise" in the harshly regulated scenario of nuclear power plants.

[0006] In summary, how to reasonably embed man-machine collaborative scheduling into the nuclear power plant maintenance process and achieve adaptive dynamic adjustment of predictive maintenance models in a high-safety scenario has become a problem to be solved by the existing technology. SUMMARY

[0007] In view of the deficiencies of the prior art, the purpose of the present application is to provide a man-machine collaborative nuclear power plant maintenance method and system based on a joint loss function, which realizes the full integration of operator knowledge and experience and AI model capabilities by constructing a man-machine collaborative learning mechanism, improves the accuracy, adaptability and reliability of predictive maintenance, and forms a stable and reliable intelligent closed loop in the whole life cycle management of equipment.

[0008] In order to achieve the above-mentioned purpose, the present application is realized by the following technical scheme:

[0009] The present application provides a man-machine collaborative nuclear power plant maintenance method based on a joint loss function in the first aspect, comprising the following steps:

[0010] Obtain multi-source data of the real-time running state of the nuclear power plant equipment, and preprocess the data to obtain an input data set;

[0011] Use the nuclear power plant maintenance model to process the input data set through artificial review and model direct prediction collaboration to obtain a nuclear power plant maintenance prediction result;

[0012] Based on the joint loss function, consider the artificial review error and model direct prediction error to combine a reinforcement learning mechanism to adaptively adjust the parameters of the nuclear power plant maintenance model;

[0013] According to the nuclear power plant maintenance prediction result, combine the historical risk level of the equipment and the operation context to automatically generate a maintenance scheme.

[0014] Further, the preprocessing includes filtering, normalization and time sequence alignment operations on the data.

[0015] Further, the nuclear power plant maintenance model includes an AI model and an artificial review interaction unit, wherein the AI model includes an anomaly detection model and a remaining life prediction model, the anomaly detection model is composed of a self-encoding neural network and is used for nuclear power plant equipment state evaluation, and the remaining life prediction model is composed of a time sequence-based attention mechanism network and is used for nuclear power plant equipment remaining service life estimation.

[0016] Further, the specific steps of using the nuclear power plant maintenance model to process the input data set through artificial review and model direct prediction collaboration are as follows:

[0017] Use the anomaly detection model and the remaining life prediction model to process the input data set to obtain a preliminary prediction result;

[0018] Evaluate the confidence of the preliminary prediction result, and according to a preset confidence threshold, evaluate whether to call the artificial review interaction unit for artificial review intervention.

[0019] Further, whether the manual review interaction unit is called or not, the processing process and result of the nuclear power plant maintenance model are recorded and fed back.

[0020] Further, the specific steps of manual review include:

[0021] Introducing a human expert agent model in the manual review interaction unit;

[0022] Training the human expert agent model using historical manual feedback data, learning and simulating the judgment tendency, preferred decision path and review mode of human experts in different operating states of nuclear power plant equipment;

[0023] Using the trained human expert agent model to review and correct the preliminary prediction result.

[0024] Further, the joint loss function includes a prediction error term, a manual feedback term and a regularization term, the prediction error term is used to measure the error between the predicted value and the true value of the nuclear power plant maintenance model, the manual feedback term is used to measure the difference between the judgment of the AI model and the manual review, and the regularization term is used to prevent model overfitting and constrain the parameter complexity.

[0025] The second aspect of the present application provides a man-machine collaborative nuclear power plant maintenance system based on a joint loss function, comprising:

[0026] The data acquisition module is configured to acquire multi-source data of the real-time operating state of the nuclear power plant equipment, and pre-process the data to obtain an input data set;

[0027] The model prediction module is configured to process the input data set by manual review and model direct prediction through the nuclear power plant maintenance model to obtain a nuclear power plant maintenance prediction result;

[0028] The dynamic updating module is configured to perform adaptive parameter adjustment on the nuclear power plant maintenance model based on the joint loss function, considering the manual review error and the model direct prediction error combined with the reinforcement learning mechanism;

[0029] The maintenance suggestion module is configured to automatically generate a maintenance scheme according to the nuclear power plant maintenance prediction result combined with the historical risk level of the equipment and the operation context.

[0030] The third aspect of the present application provides a computer readable storage medium, which stores a computer program, the computer program is suitable for being loaded and executed by a processor to perform the steps in the man-machine collaborative nuclear power plant maintenance method based on the joint loss function as described in the first aspect of the present application.

[0031] The fourth aspect of the present application provides a computer device, which comprises:

[0032] a processor adapted to execute the computer program;

[0033] a computer readable storage medium having stored therein a computer program, the computer program being executed by the processor to implement the method for human-machine collaborative nuclear power plant maintenance based on a joint loss function according to the first aspect of the present application.

[0034] The one or more technical solutions above have the following beneficial effects:

[0035] The present application discloses a kind of method and system for human-machine collaborative nuclear power plant maintenance based on joint loss function, to overcome the problems such as insufficient model self-adaptive ability, unable to systemically absorb artificial experience, lack of explainability and operator trust mechanism of existing predictive maintenance technology in high safety level scene such as nuclear power plant, the present application introduces structured artificial review process and artificial feedback collection mechanism, and establishes the feedback learning path of continuous interaction between AI model and field operator.In the present application, when AI model completes prediction task, the reliability of the prediction result is quantitatively evaluated by confidence calculation.System when confidence is higher than set threshold, model prediction result can directly enter execution link;When confidence is lower than threshold, the system automatically triggers artificial review request, allows operator or the thinking of analog operator agent model to intervene judgment, confirmation or correction model suggestion, and records its feedback data in structured way.

[0036] Unlike prior art, the present application not only retains the right of artificial review, but further returns the artificial judgment result as new training sample to model updating path, and constructs retraining mechanism including artificial experience learning ability.The mechanism integrates original prediction error, uncertainty penalty term and model complexity regular term by joint loss function, so that AI model continuously absorbs knowledge and experience of nuclear power plant operator in operation, realizes the continuous optimization and steady evolution of prediction performance.In addition, the present application also combines reinforcement learning mechanism and explainability mechanism, to enhance the response sensitivity and learning efficiency of model to human feedback, and output the causal basis of model decision through visualization means, so as to significantly improve the controllability, verifiability and operator trust degree of system in high risk working condition.

[0037] The application introduces a confidence-driven human-machine collaborative learning mechanism, constructs a feedback-driven learning closed-loop architecture, and realizes the deep integration between artificial judgment and AI model. Unlike the traditional "training-deployment-static use" paradigm, the system designed by the application still has online evolution ability in the running stage, that is, when the model makes a prediction, it evaluates its uncertainty level. When the system detects a low-confidence prediction result, it will actively trigger the artificial review path and invite the operator to confirm or correct. The artificial judgment result is not only used for current decision-making, but also as structured feedback collected, labeled and input to the retraining module, so as to continuously optimize the model performance.

[0038] To support the above mechanism, the application constructs a joint loss function with human feedback items. The function retains the basic prediction error term while adding a "human-machine distribution consistency term" based on KL divergence, and is supplemented by a regularization term to constrain the model complexity. This composite structure not only enhances the model's ability to absorb human judgment, but also improves its robustness and generalization ability, so that the model can quickly adjust and realign the human operation logic when facing new abnormalities or data drift.

[0039] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be understood through the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0041] Figure 1 The system architecture diagram of the human-machine collaborative nuclear power plant maintenance method based on the joint loss function in the embodiment of the application;

[0042] Figure 2 The flowchart of the human-machine collaborative nuclear power plant maintenance method based on the joint loss function in the embodiment of the application;

[0043] Figure 3 The schematic diagram of the total loss function in the embodiment of the application;

[0044] Figure 4 The decision schematic diagram of the human-machine collaborative nuclear power plant maintenance method based on the joint loss function in the embodiment of the application. DETAILED DESCRIPTION

[0045] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0046] It should be noted that the terms used herein are only intended to describe specific embodiments and are not intended to limit exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, they indicate the presence of the features, steps, operations, devices, components and / or combinations thereof;

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] Embodiment one:

[0049] Under the existing technical conditions, the predictive maintenance system deployed in high reliability industrial scenes such as nuclear power plants mostly relies on AI models trained offline, and adopts a pure automatic mode for equipment state evaluation and maintenance decision. Such models assume stable operating environment and consistent data distribution when designed, however, in actual working conditions, the equipment state is complex and variable, and new types of abnormality frequently occur, and the model is prone to "mismatch" problem. Once deployed, the existing system often lacks the ability to dynamically absorb the knowledge of operators, and cannot form a collaborative mechanism with human judgment, thereby limiting its adaptability, controllability and credibility in actual operation.

[0050] In view of the above technical defects of the prior art, the embodiment one of the present application provides a man-machine collaborative nuclear power plant maintenance method based on a joint loss function, as shown in Figure 1 、 Figure 2 、 Figure 3 and Figure 4 , comprising the following steps:

[0051] Step 1: Obtain multi-source data of real-time running state of nuclear power plant equipment, and pre-process the data to obtain an input data set.

[0052] In the embodiment, the multi-source data includes temperature, vibration, current and other signals, and the pre-processing includes filtering, normalization and time alignment operation on the data to obtain the input data set .

[0053] Step 2: The input data set is processed by manual review and direct prediction of the nuclear power plant maintenance model to obtain the nuclear power plant maintenance prediction result.

[0054] In this embodiment, the nuclear power plant maintenance model includes an AI model and a manual review interaction unit, wherein the AI model includes an anomaly detection model and a remaining life prediction model.

[0055] Step 2.1: The input data set is processed by the anomaly detection model and the remaining life prediction model to obtain a preliminary prediction result.

[0056] The anomaly detection model of this embodiment is composed of a self-encoding neural network, which is used for nuclear power plant equipment state evaluation, and the remaining life prediction model is composed of a time-based attention mechanism network, which is used for nuclear power plant equipment remaining service life estimation.

[0057] It should be particularly noted that using existing network structures to evaluate the state of nuclear power plant equipment and estimate the remaining service life is a common technical means for those skilled in the art, and the detailed process will not be repeated here.

[0058] Step 2.2: Evaluate the confidence of the preliminary prediction result, and evaluate whether to call the manual review interaction unit for manual review intervention according to the preset confidence threshold.

[0059] The input data set is input into the AI model for prediction to obtain the prediction result and its uncertainty distribution, denoted as the preliminary prediction result and the corresponding confidence .

[0060] Step 2.2.1: Set the confidence threshold.

[0061] Step 2.2.1.1: Data acquisition and preprocessing.

[0062] In this embodiment, multi-source sensor data generated by unit operation is collected, filtered, normalized, time-aligned, etc. to obtain an input sequence:

[0063] .

[0064] wherein, is the input sequence, is the th sensor data, , and T is the total number of sensor data.

[0065] Step 2.2.1.2: Model prediction and basic uncertainty evaluation.

[0066] In this embodiment, the preprocessed data is input into the AI prediction module to obtain a preliminary prediction result and the base uncertainty .

[0067] For classification tasks, the uncertainty estimate can be calculated by the entropy of the predicted probability distribution:

[0068] .

[0069] where, represents the predicted probability distribution of each class output by the model given the input . represents the predicted probability of the th class. represents the total number of classes. represents the entropy operator, which measures the degree of disorder of the predicted distribution. represents the normalization factor of the maximum entropy, which ensures the uncertainty value falls in the interval [0,1]. represents the uncertainty indicator for classification tasks, and the value closer to 1 indicates that the model is more uncertain.

[0070] For regression tasks, the uncertainty estimate can be calculated by the hybrid of the predicted variance and the residual:

[0071] .

[0072] where, represents the predicted variance of the model given the input , which measures the uncertainty of the model for the current input. represents the average predicted variance on the global or validation set, which serves as the normalization baseline. represents the true value. represents the predicted mean of the model for the input . represents the average predicted value on the global or validation set. represents the weighting coefficient, which is used to adjust the relative importance of the predicted variance and the residual in the uncertainty estimation. represents the Sigmoid activation function, which compresses the result to the interval [0,1]. represents the uncertainty indicator for regression tasks, which combines the predicted variance and the predicted bias of the model.

[0073] The unified uncertainty is obtained as:

[0074] .

[0075] where, represents the weight parameter, which is used to weight the fusion between the classification and regression uncertainty indicators. For the final uncertainty value, applicable to mixed tasks (or scenarios with both classification and regression outputs in the same system), the guaranteed value is in [0, 1].

[0076] Step 2.2.1.3: Nuclear power plant scenario confidence correction.

[0077] This embodiment introduces factors specific to nuclear power plants: equipment safety importance (mapped from nuclear power plant safety classification or importance analysis (such as safety class / important non-safety class / conventional class)), operating mode factor (reflecting the influence of different working conditions such as start-stop, power operation, hot shutdown, cold shutdown, refueling on safety importance), real-time configuration risk , LCO margin penalty .

[0078] Get the scenario confidence:

[0079] .

[0080] Where, is the comprehensive confidence after nuclear power plant scenario, which is corrected by combining basic uncertainty and specific factors of nuclear power plants, and is used to judge the credibility of the current prediction result. is the Sigmoid activation function, which compresses the input result to the [0, 1] interval, so that the output has a probability meaning and is convenient for threshold processing. is the bias term, which represents the baseline confidence or correction amount, used for the initial confidence level when there is no additional information. is the basic uncertainty or confidence index term, reflecting the original confidence level of the model to the input data, which is fused with the scenario factors after weighting. is the equipment safety importance weight term , which is obtained by mapping the nuclear power plant safety classification (such as safety class, important non-safety class, conventional class), used to reflect the influence of the safety importance of different equipment on the confidence. is the working condition coefficient term , which reflects the influence of different operating conditions (such as reactor start-stop, power operation, hot shutdown, cold shutdown, refueling) on safety importance and confidence. is the real-time configuration risk term , which is quantified in combination with the risk state of the current time or scenario, such as power fluctuation, key equipment failure rate, etc. is the technical specification or margin penalty term , which represents the current distance from the LCO (Technical Specification Limiting Condition for Operation) margin of the nuclear power plant. The closer to the limit value, the greater the penalty, and the confidence is weakened.

[0081] Step 2.2.1.4: Convert the confidence level (uncertainty) into the probability of error.

[0082] In this embodiment, under the scenario of predictive maintenance in nuclear power plants, the uncertainty of the model output is recorded by collecting a validation set. Based on the corresponding actual judgment results, a statistical table of error samples under different uncertainty intervals is constructed. For each uncertainty interval, the statistical model predicts the frequency of error samples (such as false alarms and false misses), forming an empirical error probability distribution. Then, through orthogonal-preserving regression, logistic regression, or other monotonic function fitting methods, a continuous and differentiable mapping function is obtained.

[0083] .

[0084] This makes the uncertainty value output by the model interpretable in the actual operating conditions of a nuclear power plant.

[0085] Through this calibration step, when the model outputs any confidence level... At that time, the system can automatically determine the corresponding error probability. .

[0086] Step 2.2.1.5: Threshold setting based on risk cost.

[0087] To minimize the expected risks in nuclear power scenarios, this embodiment adopts a cost-benefit approach of underreporting. Cost of misreporting Cost of manual review Construct a cost function and vary it with scene factors:

[0088] .

[0089] in, To underreport costs, The fixed baseline cost is the basic cost regardless of when a report is missed. The more critical the equipment, the greater the cost of underreporting; The Risk Index reflects the unit's operating status or external conditions at the time. If it is under high power, critical operating conditions, or maintenance window, the cost increases. The risk function is related to the remaining time of the technical specification (LCO, Limiting Condition for Operation). The closer to an LCO violation, the greater the cost of underreporting. To cover the cost of false reporting, To fix the baseline cost of false alarms, Costs of reactor shutdown, downtime, or maintenance caused by manual intervention or shutdown triggered by false alarms. cost of manual review, cost of fixed manual intervention; cost of radiation dose during manual review, higher if working in high dose area; cost of manual effort, such as man-hours, scheduling delay, additional training, etc.

[0090] Multiply the error probability with the scenario cost to get the expected risk. The formula is:

[0091] .

[0092] Select the minimum automatic execution risk:

[0093] .

[0094] where, is the false negative expected risk of automatic execution when the confidence is , i.e. the false negative probability multiplied by the false negative cost. is the false positive expected risk of automatic execution when the confidence is , i.e. the false positive probability multiplied by the false positive cost. is the manual review risk, which is the sum of the cost of manual review and the additional cost . is the overall expected risk of automatic execution when the confidence is , which is the smaller one between the false negative and false positive risks, representing the optimal choice of automatic execution under the current confidence.

[0095] This embodiment selects the one with smaller expected loss between automatic release and automatic alarm as the risk of automatic execution.

[0096] The meaning of the confidence threshold is that when the risk brought by model automatic review is exactly equal to the risk brought by manual review:

[0097] .

[0098] where, is the confidence threshold, which is the dividing line for making decisions between automatic execution and manual review.

[0099] Since is monotonic, the confidence threshold can be directly obtained by inverting:

[0100] .

[0101] where, is the inverse function of , which is used to deduce the corresponding confidence threshold from the risk. To dynamically adjust the automatic execution boundary of the model in time , the inverse risk formula is used to obtain a dynamic confidence threshold.

[0102] Step 2.2.2: Evaluate the confidence of the preliminary prediction result.

[0103] In this embodiment, the uncertainty of the preliminary prediction result is evaluated to obtain the confidence. Specifically, the confidence score of each prediction is output according to the uncertainty modeling method , which is used to evaluate the confidence degree of the model for the current prediction result.

[0104] In the embodiment of the present application, the model first calculates the prediction value and the corresponding confidence for the input sample. When the confidence is lower than the threshold , it indicates that the model has a large uncertainty or potential error risk for the current prediction. At this time, the system automatically triggers the human-machine collaborative process, and transmits the low-confidence sample to the external labeling source or the manual review unit.

[0105] In the manual review unit, the on-duty engineer or maintenance expert independently evaluates the sample and gives a correction judgment. The system converts the judgment result into a structured label , that is, the human experience is expressed in a form that can be recognized by model training. For classification tasks, a soft label (such as ) is used, and the rest of the categories are evenly distributed ) is used for regression or interval determination, and a Gaussian soft label is used to represent the mean and uncertainty range of the manual judgment:

[0106] .

[0107] Among them, is a Gaussian distribution, is the mean, is a fixed variance.

[0108] The essence of this approach is to convert the original "one-time judgment by human" into "training samples that can be absorbed by machines". The model uses as supervision in the next round of training, and introduces in the joint loss function, so as to systematically absorb human experience into model parameters, and realize the self-adaptation and online evolution of the model in operation.

[0109] Step 2.2.3: According to the preset confidence threshold, evaluate whether to call the manual review interaction unit for manual review intervention.

[0110] After obtaining the confidence calculation result, the confidence is compared with a preset confidence threshold to determine whether the confidence is lower than the preset threshold value , if the confidence is lower than the preset threshold value, it indicates that the AI model is not confident about the prediction, and then an artificial review process is activated, and the artificial review interaction unit is entered; if the confidence is greater than or equal to the preset threshold value, it indicates that the AI model has enough confidence, and the preliminary prediction result is directly output and the subsequent maintenance suggestion generation process is entered.

[0111] It should be particularly noted that, whether the artificial review interaction unit is called or not, the processing process and result of the nuclear power plant maintenance model are recorded and fed back for subsequent model retraining and parameter adjustment.

[0112] If the artificial review interaction unit is called for artificial review intervention, the following steps are included:

[0113] Step 2.3.3.1: Introduce a human expert agent model in the artificial review interaction unit.

[0114] In this embodiment, the feedback of the human operator can be directly input into the system through interface interaction, and in the form of labeled samples or evaluation signals, and enter the model updating process to realize human-machine collaborative learning closed loop. However, in some high complexity or high workload scenarios, the operator may not be able to continuously and directly participate in the feedback link, or the quality of artificial feedback may fluctuate and be subjective. In order to improve the stability and structure of the system feedback ability, the human expert agent model can also be introduced to realize indirect collaboration at the module level.

[0115] The human expert agent model can be constructed based on a graph neural network, a Bayesian network or a reinforcement learning framework.

[0116] In the running phase, the human expert agent model is embedded as an intermediate judgment module in the AI prediction main process to realize dynamic selection of the decision path, for example: if the prediction confidence is low but the agent model judges that the risk is high, the system automatically reports the operator or triggers an alarm.

[0117] In the retraining phase, the human expert agent model provides larger-scale and more consistent human behavior simulation samples to guide the optimization direction of the AI model.

[0118] In terms of model explainability, the human expert agent model provides experience rule-based auxiliary explanation output to help the operator understand the rationality or limitations of the AI suggestion.

[0119] Through the human expert agent model, the knowledge and experience of the human operator are embedded in the core reasoning mechanism of the system in the form of an "agent", realizing the combination of indirect feedback and efficient collaboration. This mode is especially suitable for high-frequency decision-making tasks or industrial control systems that require rapid response, effectively reducing the frequency of human participation while retaining the guiding role of human knowledge. Even if the direct interaction and retraining interface between the operator and the AI model is omitted, the human-machine collaborative learning mechanism can be realized through the setting of the human expert agent model.

[0120] Step 2.3.3.2: Train the human expert agent model using historical human feedback data to learn and simulate the judgment tendency, preferred decision path, and review mode of human experts in different operating states of nuclear power plant equipment.

[0121] Step 2.3.3.3: Use the trained human expert agent model to review and correct the preliminary prediction results.

[0122] This embodiment uses the trained human expert agent model or real operators to review, supplement, or correct the AI model prediction results. First, in the human review link, the input low confidence samples are confirmed to determine whether to adopt the scheduling suggestions proposed by the AI model, forming the human label . Among them, if , it is recorded as consistent feedback, and the sample and the original model output are recorded as "confirmation samples" together. If , it is recorded as correction feedback, and the model output is corrected with the human judgment result and recorded as "corrected samples", which are used as supervision signals for subsequent retraining. The system automatically records each human correction operation and its corresponding input features to form structured feedback data.

[0123] In this embodiment, the human operator or human expert agent model performs structured labeling on the input preliminary prediction results of the AI model, and the labeling results are used as supervision signals to participate in the model retraining process to build a human-machine collaborative learning mechanism. However, in some application scenarios, the operator or human expert agent model may not be able to or need to provide explicit labels (such as accurate abnormal categories or remaining life values), at which time weak labels or confidence scoring can be used for labeling.

[0124] Specifically, the system provides a feedback interface module for receiving the operator's score signal of the degree of adoption, trust, or subjective acceptance of the model prediction results. The signal can include but is not limited to:

[0125] Discrete scoring, such as the operator scoring the current prediction output in the range of 1 to 5;

[0126] Binary adoption feedback, such as whether the operator performed the maintenance action corresponding to the AI suggestion;

[0127] Time response indicators, such as the length of time the operator delayed responding to the model's suggestion, and other indirect behavioral signals.

[0128] The system converts such weakly supervised information into pseudo-labels or auxiliary indicators and integrates them into the joint loss function of the model to construct a feedback dataset for guiding training. Compared with the structured label scheme, this alternative reduces the operation threshold, improves the frequency and coverage of feedback acquisition, and is particularly suitable for industrial environments where labels are difficult to standardize and feedback needs to be quickly responded.

[0129] Step 3: Based on the joint loss function, consider the combination of artificial review errors and model direct prediction errors to strengthen the learning mechanism and adaptively adjust the parameters of the nuclear power plant maintenance model.

[0130] Step 3.1: Construct a joint loss function.

[0131] As shown in Figure 3 , the joint loss function of the present embodiment includes a prediction error term, an artificial feedback term, and a regularization term. The prediction error term is used to measure the error between the predicted value of the nuclear power plant maintenance model and the true value, continuously driving the model to align with the true label during the training phase and improving the prediction accuracy. The artificial feedback term is used to measure the difference between the AI model and the human review judgment. To enhance the model's ability to absorb human feedback, the inventors introduce a "human-machine feedback consistency term" defined based on KL divergence . This term takes the structured label generated by human review as the benchmark, measures the difference between the model prediction distribution and the human judgment distribution. By minimizing this divergence, the model can gradually align with human judgment on uncertain samples, essentially building a collaborative optimization mechanism guided by human experience, thereby giving the model the ability to continuously evolve. The regularization term is used to prevent model overfitting, constrain parameter complexity, and avoid overfitting of the model after introducing artificial signals, thereby ensuring the robustness and generalization ability of the learning process.

[0132] The application realizes the paradigm shift of the predictive maintenance model from "assisting human judgment" to "growing in feedback" by the closed-loop channel between the model output and human feedback at the algorithm level. Compared with the prior art, the application not only dynamically adapts to the complex working condition changes of the nuclear power plant, but also guides the AI model to continuously integrate and internalize the knowledge and judgment of the operator while ensuring the prediction accuracy, thereby fundamentally improving the safety, adaptability and trust of the prediction system. Such a mechanism design of human-computer collaboration is especially suitable for industrial key scenarios with high requirements for safety, transparency and long-term stability, and has wide engineering application prospects and technical popularization value.

[0133] Specifically, a joint loss function is constructed to fuse the artificial label correction signal and the original prediction target of the model. The joint loss function is in the form of:

[0134] .

[0135] Wherein: is a prediction error term, such as mean square error (MSE) or cross entropy (Cross Entropy); is a human feedback term, a human feedback loss for guided learning, such as Kullback-Leibler divergence (KL divergence); is a regularization term; is a weighting coefficient, satisfying The weighting coefficient can be adjusted according to the actual application scene and task demand to realize the precision-explanation-stability trade-off in different degrees.

[0136] Prediction error term is used to measure the error between the model prediction value and the true value y:

[0137] .

[0138] This term drives the model to continuously approach the true output in training and improves the prediction accuracy.

[0139] Artificial feedback term is used to fuse human feedback information and guide the model to learn and correct low-confidence samples. The embodiment preferably uses KL divergence as the core calculation form of the guiding term, that is:

[0140] .

[0141] Wherein, represents the label distribution confirmed by the artificial audit interaction unit, i.e. the artificial label, is the prediction distribution of the current output of the model, represents the KL divergence function.

[0142] The KL divergence (Kullback-Leibler Divergence) is a non-symmetric distance measure used to quantify the difference between two probability distributions. For two probability distributions P and Q defined on the same space and The expression of KL divergence is as follows:

[0143] .

[0144] In the present invention, generates a "target distribution" (i.e. artificial label ) usually represents human review, while represents the prediction distribution output by the model at present.

[0145] The KL divergence has the following properties:

[0146] Non-negativity: KL divergence is always greater than or equal to 0, and is 0 only when the two distributions are completely consistent;

[0147] Directionality: Therefore, it emphasizes "how much Q deviates from P";

[0148] Sensitivity to probability deviation: it can finely depict the structural deviation between model prediction and artificial correction.

[0149] In the human-machine collaborative learning mechanism of the present invention, the KL divergence is introduced as a loss term, which helps to measure and minimize the "false confidence" of the model on low-confidence samples. By continuously reducing the KL distance between the model prediction distribution and the artificial label distribution, the model gradually aligns with human judgment logic, thereby realizing continuous learning and autonomous optimization. This makes the KL divergence not only a tool for measuring differences, but also a bridge connecting AI models and operator knowledge, supporting the human-machine closed-loop feedback mechanism of the system. This loss measures the degree of difference between the model prediction distribution and the human correction opinion. By minimizing the KL divergence, the model can actively adjust its judgment of "low-confidence samples" during training, thereby gradually approaching the knowledge and experience of the operator and realizing continuous learning guided by human knowledge.

[0150] The design of the artificial feedback term makes the model no longer rely on fixed data labels for training when facing uncertain samples, but instead uses the feedback samples of the operator as a dynamic correction signal, overcoming the problem of insufficient adaptability of traditional models to environmental changes, significantly improving the generalization, interpretability, and adaptability of the model in real-world scenarios.

[0151] Regularization term The common form is L2 norm for constraining model parameter space to prevent overfitting:

[0152] = ||θ|| 2 .

[0153] This term can be understood as a mechanism to "limit model complexity" to maintain generalization ability.

[0154] This embodiment enables the model to have the ability to "request human assistance", "control model complexity" and "focus on prediction uncertainty" during the training phase, providing a foundation for human-machine collaboration and feedback retraining closed loop. At the same time, it also establishes a unified index interface for subsequent module switching, explainability mechanism and strategy selection.

[0155] Step 3.2: Incrementally update the AI model using the joint loss function.

[0156] Step 3.2.1: Construct the retraining sample set.

[0157] By collecting historical manual review samples, input data, model prediction results and manual correction results are constructed into a three-tuple data, which is part of the retraining sample set. "Confirm samples" and "correct samples" are unified into the retraining sample set to form the final retraining sample set.

[0158] Step 3.2.2: Incrementally update the AI model using the joint loss function to improve the model's ability to judge future low confidence situations.

[0159] Based on the joint loss function, the manual review error and the model direct prediction error are weighted and learned, and the reinforcement learning mechanism (such as the strategy optimization algorithm based on human feedback) is combined to realize self-adjustment and evolutionary update of the model, and the updated model parameters are obtained, realizing the absorption and optimization of human feedback in the running of the model. The confirmed human intervention samples are added to the "experience pool" and used for strategy evaluation in reinforcement learning, improving the model's discriminability when facing new samples.

[0160] The present application introduces nuclear power plant scene weight in constructing joint loss function, which makes it more suitable for high reliability industrial systems, aiming at the characteristics of high dimension, long periodicity, low frequency anomaly but high danger of nuclear power plant key equipment sensor data. The joint loss function includes prediction error term , human feedback loss term and regularization term , which is in the form of:

[0161] .

[0162] wherein, MSE or CrossEntropy can be used to measure the difference between the predicted value of the model and the true value; To introduce the loss term of artificial review feedback, KL divergence can be used to measure the difference between the artificial soft label and the model prediction probability distribution ; is used to control the complexity of the model to prevent overfitting. The weight coefficient is adaptively adjusted according to the safety level, operation and maintenance cycle, and risk tolerance of the nuclear power plant equipment, so that the prediction error and artificial feedback signal in the joint loss function under high-risk working conditions have a larger proportion, thereby prioritizing the learning and optimization of the model performance of key equipment and high-risk states.

[0163] In each time window, the updated model parameters are obtained according to the aforementioned joint loss function minimization , and the feedback sample buffer is dynamically maintained . The system adaptively updates the confidence threshold value in the nuclear power plant working condition data stream using a sliding window statistical method , and the update formula is:

[0164] .

[0165] wherein, and are the confidence threshold values at time and , respectively, and represent the false negative and false positive rates in the current window, respectively, and are the maximum tolerable values determined based on the safety constraints of the nuclear power plant; and are the learning rate coefficients. Through this dynamic threshold updating method, the system can automatically adjust the confidence threshold value according to the real-time feedback and safety constraints of the nuclear power plant, realizing the online evolution and safety adaptive ability of the predictive maintenance model under the operating conditions of the nuclear power plant.

[0166] The above retraining process directly incorporates the confirmed artificial review samples into the "experience pool" as the supervision signal for subsequent model training. The system records indicators such as ECE, FN / FP, artificial review rate, and delay cost, ensuring that in the high-reliability scenario of the nuclear power plant, the model not only continuously absorbs human feedback, improves prediction accuracy and robustness, but also evolves autonomously within the safety boundary, forming a complete closed-loop human-machine collaborative predictive maintenance mechanism.

[0167] threshold value As the dynamic confidence threshold of the present application, it is used to determine whether the current model output is reliable. Samples below the threshold are automatically transferred to the feedback sample buffer , and structured labels are generated after manual review . The samples in the feedback sample buffer are input into the joint loss function together with the original data , and the iterative update of the model parameters is realized by minimizing . Thus, a dynamic closed-loop learning mechanism is formed, with the threshold as the entrance, the feedback sample buffer as the relay, and the joint loss function as the core, so that the model can continuously absorb artificial feedback and optimize the prediction performance.

[0168] In a specific embodiment, the artificial feedback information obtained by the artificial review interaction unit is not directly used as explicit labels for supervised learning in model training, but is used to construct a pseudo-label dataset, thereby guiding the artificial intelligence model to carry out self-supervised learning (Self-Supervised Learning) to realize the continuous optimization and generalization enhancement of the model capability.

[0169] Specifically, when the operator or human expert agent makes a negative judgment on the system prediction result, the system marks the input sample as “to be reconstructed sample”, and constructs a specific self-supervised training task based on this sample. For example, contrastive learning (Contrastive Learning) or reconstruction loss (Reconstruction Loss) mechanism can be used to guide the model to distinguish “corrected samples” and “uncorrected samples” in the feature representation space. In the contrastive learning structure, frameworks such as SimCLR, MoCo, etc. can be selected, so that the model can make the feature vectors of the artificially confirmed positive samples close to each other in the representation space, and make the feature vectors of the negative samples inconsistent with each other far away.

[0170] In another embodiment, a training strategy based on masked prediction (Masked Prediction) can be used, only the artificially confirmed sample segment is retained, the rest is masked, and the training process is completed by predicting the content of the masked part. The above method does not require artificial to provide complete labels, but through the construction of proxy tasks (Proxy Task) to excavate the discriminant pattern implied in artificial feedback, so as to realize the implicit absorption of artificial experience.

[0171] The advantages of this step are: first, it does not rely on high-quality explicit label data and can work effectively in cases of label scarcity or high manual annotation cost; second, it has the ability to adapt to unstructured feedback data, especially suitable for industrial scenarios where operational data is abundant but human involvement is limited; third, in application scenarios with strong privacy constraints (such as nuclear power plant operation monitoring), the model can evolve based on human experience without exposing complete human annotation information.

[0172] Therefore, even if the direct labeling step of human feedback in traditional supervised learning is omitted, the integration of human experience and AI model capabilities can still be achieved through the introduction of a self-supervised learning mechanism.

[0173] Step 4: Explainable analysis of nuclear power plant maintenance prediction results.

[0174] This embodiment provides explanations for model predictions, such as attention visualization and feature importance ranking. Through modular design, an explainable analysis module and a visual feedback interface are constructed to assist operators in understanding model recommendations, improve system usability and trustworthiness, and enable operators to understand and challenge the model's prediction basis, thereby improving the overall explainability and operational trustworthiness of the system. In scenarios such as nuclear power plants where safety requirements are extremely high, this system retains human final decision-making authority, ensuring human intervention capabilities at critical moments and effectively reducing the system risks associated with black box models.

[0175] Specifically, a feature attribution-based method is used, which analyzes the role of each input variable (i.e., sensor measurement points, operating parameters, etc.) in the prediction results through mathematical or statistical means, thereby deducing why the model made such a judgment. Common methods include SHAP (Shapley Additive explanations) or LIME (Local Interpretable Model-agnostic Explanations), which provide model decision-making process visualization information to operators, improving system understandability and debuggability.

[0176] The retrained model is deployed again to realize a feedback-driven model self-improvement closed loop. The above process effectively integrates the efficient computing power of AI models and human experience judgments to form a dynamic adjustment of human-machine collaborative intelligent decision-making mechanism.

[0177] The steps in this embodiment are tightly coupled through data flow and feedback flow, forming a complete "prediction-audit-feedback-learning-evolution" human-machine collaborative intelligent maintenance closed loop, achieving the goal of continuously absorbing human knowledge, adapting to variable working conditions, and improving prediction performance in actual operation. This greatly improves the continuous evolution capability of the predictive maintenance system.

[0178] Step 5: Based on the maintenance prediction results of the nuclear power plant, combined with the historical risk level of the equipment and the operation context, automatically generate a maintenance plan. The maintenance plan includes maintenance suggestions and recommended execution time, etc., and supports manual approval, DCS interface access, and maintenance log recording.

[0179] The technical core of this embodiment is to build a predictive maintenance system with a human-machine collaborative feedback mechanism, especially suitable for industrial scenarios such as nuclear power plants with high requirements for safety and interpretability. Compared with traditional prediction models that rely only on static training data for reasoning, the proposed system can receive real-time human judgment feedback during operation and incorporate it into the retraining process, forming a dynamic and evolutionary AI capability enhancement path.

[0180] First, one of the key features of this embodiment is to use confidence evaluation results as the basis for driving human-machine diversion judgment. After the AI model completes the prediction, the system quantifies the confidence of the current prediction through the uncertainty evaluation module. When the confidence is higher than the set threshold, the prediction result automatically enters the execution path; when the confidence is insufficient, the human review process is triggered. This dynamic judgment mechanism based on the model's own confidence level significantly improves the safety and stability of the prediction system, avoiding uncontrollable decisions made by the model in low confidence situations.

[0181] Second, this embodiment places special emphasis on the reuse mechanism of human review results. Whether the human review adopts the AI model's suggestion or not, the judgment is structured as a label, labeled as "confirmation sample" or "correction sample". These samples are not auxiliary information for terminal operations, but are actively recycled by the system and used in subsequent model training to correct model bias and improve its learning ability for boundary cases. This design realizes a truly human-machine co-built knowledge system, enabling the AI model to continuously learn from human experience.

[0182] Another technical key of this embodiment is the design and application of the joint loss function. In the model retraining process, instead of only considering the traditional prediction error as the optimization target, the prediction error, uncertainty penalty term, and model complexity regularization term are considered in multiple dimensions to construct a joint loss function that can accommodate human feedback signals. Model training based on this not only optimizes accuracy, but also improves the confidence and interpretability of the output results.

[0183] In addition, the embodiment explicitly proposes a dynamic model optimization mechanism with artificial feedback as the retraining trigger signal. After the operator makes feedback on the prediction result, the system automatically incorporates it into the training process and updates the model parameters. This training mechanism can set the training pace according to the feedback frequency and confidence level changes, support online learning or periodic fine-tuning, and has high engineering deployability.

[0184] Finally, the embodiment also has protectable innovation points at the system structure design level. The entire system is composed of an AI prediction module, an uncertainty evaluation module, a manual review module, a sample feedback management module, and a model retraining module. The data paths between the modules form a closed-loop structure, ensuring smooth flow of information in the "prediction-judgment-feedback-learning" stages, supporting the continuous evolution capability of the system.

[0185] In summary, the embodiment scheme covers a complete composite technical solution from the prediction mechanism to the feedback path, and then to the learning method and system architecture. This scheme realizes the deep integration of artificial knowledge and AI models, and builds a predictive maintenance system with explainability, adaptability, and evolution capability.

[0186] The above technical solutions of the embodiment have obvious advantages in the following aspects:

[0187] Firstly, the accuracy of anomaly identification and the confidence management capability are improved. The system can automatically identify low-confidence predictions based on the model's uncertainty evaluation mechanism and actively switch to the manual path, thereby effectively reducing the risk of false positives and false negatives.

[0188] Secondly, the explainability and controllability of the AI model are enhanced. By designing an interactive interface for manual review and a traceable feedback mechanism, the system's prediction results are not only understandable to human operators but also can be corrected and incorporated into the learning process.

[0189] Thirdly, the model's self-evolution and continuous optimization are realized. Artificial feedback is incorporated into the joint loss function training process after being labeled, constituting a systematic retraining sample source, allowing the model to continuously evolve during operation and breaking through the limitations of traditional "static AI".

[0190] In addition, the system ensures the final judgment right of humans in high-risk scenarios. By introducing a manual review channel and judgment adoption mechanism, even if the model makes a mistake, it can still be taken over by humans, avoiding potential catastrophic consequences.

[0191] Finally, a set of system solutions of closed-loop deployment and online learning architecture are proposed. Including uncertainty-driven bifurcation path design, human-computer feedback recording module, joint loss function structure and standardized retraining process, so that the application not only has theoretical innovation, but also has practical landing feasibility and engineering promotion value.

[0192] In summary, the application realizes breakthrough innovation in multiple key technical points, including but not limited to decision confidence-driven mechanism, human-computer collaborative feedback structure, joint loss function design, retraining closed-loop construction and modular implementation of human-computer interface, thereby providing a systematic, growth-oriented and controllable solution path for the trusted deployment of predictive maintenance systems in high-security industrial scenarios, with wide application prospects and high technical promotion value.

[0193] Embodiment two:

[0194] The embodiment two of the application provides a human-computer collaborative nuclear power plant maintenance system based on a joint loss function, comprising:

[0195] A data acquisition module configured to acquire multi-source data of real-time running states of nuclear power plant equipment, and pre-process the data to obtain an input data set;

[0196] A model prediction module configured to process the input data set by artificial review and model direct prediction in cooperation through a nuclear power plant maintenance model to obtain a nuclear power plant maintenance prediction result;

[0197] A dynamic updating module configured to perform adaptive parameter adjustment on the nuclear power plant maintenance model based on a joint loss function, considering artificial review error and model direct prediction error in combination with a reinforcement learning mechanism;

[0198] A maintenance suggestion module configured to automatically generate a maintenance scheme according to the nuclear power plant maintenance prediction result in combination with equipment historical risk level and operation context.

[0199] Embodiment three:

[0200] The embodiment three of the application provides a computer readable storage medium, which stores a computer program, and the computer program is suitable for being loaded and executed by a processor to perform the steps in the human-computer collaborative nuclear power plant maintenance method based on a joint loss function as described in the embodiment one of the application.

[0201] Embodiment four:

[0202] The embodiment four of the application provides a computer device, which comprises:

[0203] A processor suitable for executing a computer program;

[0204] Computer readable storage medium, computer program is stored in computer readable storage medium, computer program is executed by the processor, and the step in the man-machine collaborative nuclear power plant maintenance method based on the joint loss function as the embodiment one of the application is realized.

[0205] The steps involved in the above embodiments two, three and four correspond to the method embodiment one, and the specific implementation can refer to the related description part of embodiment one.

[0206] The units and algorithm steps of each example described in combination with the embodiments disclosed in the present application can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The professional object can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0207] In the above embodiments, all or part can be realized by software, hardware, firmware or any combination thereof. When realized by software, all or part can be realized in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. Computer instructions can be stored in a computer readable storage medium or transmitted by a computer readable storage medium. Computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired or wireless means. The computer readable storage medium can be any available medium that the computer can access or a data processing device such as server, data center, etc. containing one or more available media sets. The available media can be magnetic media, optical media or semiconductor media, etc.

[0208] The above description is only a specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any skilled object in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A human-machine collaborative nuclear power plant maintenance method based on joint loss function, characterized in that, Includes the following steps: Acquire multi-source data on the real-time operating status of nuclear power plant equipment, and preprocess the data to obtain the input dataset; The nuclear power plant maintenance model is used to process the input dataset through a combination of manual review and direct model prediction to obtain the nuclear power plant maintenance prediction results. The nuclear power plant maintenance model includes an AI model and a manual review interaction unit. A human expert agent model is introduced into the manual review interaction unit, and the trained human expert agent model is used to review and correct the preliminary prediction results. Based on the joint loss function, a reinforcement learning mechanism is used to adaptively adjust the parameters of the nuclear power plant maintenance model, taking into account both manual review error and direct model prediction error. Joint loss function Including prediction error term Human feedback loss item and regularization terms Its form is: ， in, Used to measure the difference between model predictions and actual values. To incorporate loss items from manual review feedback, KL divergence is used to measure the impact of manual soft labeling. With model predicted probability distribution The differences between them Used to control model complexity and prevent overfitting. These are weighting coefficients. The system is adaptively adjusted based on the safety level of nuclear power plant equipment, operation and maintenance cycle, and risk tolerance. This represents the predicted distribution of the model's current output. For the true value, Manually labeled; Maintenance plans are automatically generated based on nuclear power plant maintenance forecasts, combined with equipment historical risk levels and operational context.

2. The human-machine collaborative nuclear power plant maintenance method based on joint loss function as described in claim 1, characterized in that, Preprocessing includes filtering, normalization, and time-series alignment of the data.

3. The human-machine collaborative nuclear power plant maintenance method based on joint loss function as described in claim 1, characterized in that, in, The AI ​​model includes an anomaly detection model and a remaining useful life prediction model. The anomaly detection model consists of an autoencoder neural network and is used to assess the condition of nuclear power plant equipment. The remaining useful life prediction model consists of a time-series-based attention mechanism network and is used to estimate the remaining useful life of nuclear power plant equipment.

4. The human-machine collaborative nuclear power plant maintenance method based on joint loss function as described in claim 3, characterized in that, The specific steps for processing the input dataset using a nuclear power plant maintenance model, combining manual review with direct model prediction, are as follows: The input dataset is processed using an anomaly detection model and a remaining useful life prediction model to obtain preliminary prediction results; Assess the confidence level of the preliminary prediction results, and determine whether to call the manual review interaction unit for manual review based on the preset confidence level threshold.

5. The human-machine collaborative nuclear power plant maintenance method based on joint loss function as described in claim 4, characterized in that, Regardless of whether the manual review interaction unit is invoked, the processing and results of the nuclear power plant maintenance model are recorded and fed back.

6. The human-machine collaborative nuclear power plant maintenance method based on joint loss function as described in claim 4, characterized in that, The specific steps for manual review include: By using historical human feedback data to train a human expert agent model, the model learns and simulates the judgment tendencies, preference decision-making paths, and review patterns of human experts under different operating conditions of nuclear power plant equipment.

7. The human-machine collaborative nuclear power plant maintenance method based on joint loss function as described in claim 1, characterized in that, The joint loss function includes a prediction error term, a human feedback term, and a regularization term. The prediction error term measures the error between the predicted value and the actual value of the nuclear power plant maintenance model. The human feedback term measures the difference between the judgment of the AI ​​model and the human review. The regularization term is used to prevent the model from overfitting and to constrain the parameter complexity.

8. A human-machine collaborative nuclear power plant maintenance system based on a joint loss function, characterized in that, include: The data acquisition module is configured to acquire multi-source data on the real-time operating status of nuclear power plant equipment and preprocess the data to obtain the input dataset; The model prediction module is configured to use the nuclear power plant maintenance model to process the input dataset through collaboration between manual review and direct model prediction to obtain the nuclear power plant maintenance prediction results. The nuclear power plant maintenance model includes an AI model and a manual review interaction unit. A human expert agent model is introduced into the manual review interaction unit, and the trained human expert agent model is used to review and correct the preliminary prediction results. The dynamic update module is configured to adaptively adjust the parameters of the nuclear power plant maintenance model based on a joint loss function, taking into account both human review errors and direct model prediction errors, and using a reinforcement learning mechanism. Joint loss function Including prediction error term Human feedback loss item and regularization terms Its form is: ， in, Used to measure the difference between model predictions and actual values. To incorporate loss items from manual review feedback, KL divergence is used to measure the impact of manual soft labeling. With model predicted probability distribution The differences between them Used to control model complexity and prevent overfitting. These are weighting coefficients. The system is adaptively adjusted based on the safety level of nuclear power plant equipment, operation and maintenance cycle, and risk tolerance. This represents the predicted distribution of the model's current output. For the true value, Manually labeled; The maintenance recommendation module is configured to automatically generate maintenance plans based on nuclear power plant maintenance prediction results, combined with equipment historical risk levels and operational context.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-7: a human-machine collaborative nuclear power plant maintenance method based on a joint loss function.

10. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the human-machine collaborative nuclear power plant maintenance method based on a joint loss function as described in any one of claims 1-7.

Images