A fault diagnosis method and system based on human-in-the-loop strategy and power set confidence rule base

Abstract

This invention discloses a fault diagnosis method and system based on a human-in-the-loop strategy and a power-set confidence rule base. This invention significantly improves the accuracy and reliability of fault diagnosis for wireless sensor network nodes. By effectively representing the fuzziness and uncertainty of fault types through a power-set framework, it accurately distinguishes between local and global ignorance, enhancing the credibility of the diagnostic results. The introduction of a human-in-the-loop strategy and an adaptive optimization mechanism greatly reduces the dependence on expert prior knowledge and large amounts of high-quality labeled data, enabling the model to continuously learn and optimize parameters from new data, exhibiting good environmental adaptability and robustness. Furthermore, this scheme effectively controls the computational cost of model training and updates while ensuring diagnostic performance through an intelligent sample selection strategy, achieving a balance between diagnostic accuracy and operational efficiency, and is suitable for large-scale, dynamically changing wireless sensor network environments.

Description

Technical Field

[0001] This invention relates to the field of wireless sensor fault diagnosis, specifically to a fault diagnosis method and system based on human-in-the-loop strategy and power set confidence rule base. Background Technology

[0002] Wireless sensor networks (WSNs) are widely used in environmental monitoring, industrial control, and intelligent transportation due to their advantages such as flexible deployment and low cost. However, sensor nodes in these networks often operate in complex, harsh, or unattended environments, making them susceptible to various failures due to energy depletion, hardware aging, communication interference, or environmental factors, affecting data reliability and overall network performance. Therefore, accurate and efficient node fault diagnosis is one of the key technologies for ensuring the reliable operation of wireless sensor networks. Existing fault diagnosis methods, such as those based on statistical analysis, machine learning, or deep learning, still face challenges in addressing the uncertainty and ambiguity of node fault data. In particular, fault characteristics may be highly similar in time and space, and the boundaries between different types of faults are unclear, leading to misjudgments or omissions in diagnostic models, resulting in the "unknown information" problem.

[0003] While traditional confidence rule base methods can handle some uncertainty, their identification framework has limitations in distinguishing between local and global ignorance, affecting the accuracy of diagnostic conclusions. Furthermore, existing methods often heavily rely on expert experience to set initial model parameters or require large amounts of high-quality labeled data for training, resulting in insufficient adaptability and generalization ability when data is scarce or the environment is dynamically changing. Simultaneously, efficiently selecting the most valuable samples for model optimization from the continuously generated massive amounts of monitoring data to reduce computational overhead and enable continuous online model learning is also a pressing problem in practical applications. Therefore, there is an urgent need for an intelligent fault diagnosis method that can more precisely characterize fault uncertainty, reduce reliance on prior knowledge, and adaptively optimize and update. Summary of the Invention

[0004] To address the technical problems mentioned above, the present invention provides the following technical solution: A fault diagnosis method based on human-in-the-loop strategy and power-set confidence rule base includes the following steps: S1. Based on the characteristic data in the fault diagnosis of wireless sensor network nodes, construct the PBRB-H fault diagnosis model and provide a formulaic description of the problems in fault diagnosis. S2. Based on the PBRB-H fault diagnosis model constructed in S1, define the basic structure of the model, execute the model inference process, and optimize the model parameters using the P-CMA-ES algorithm; S3. Construct a human-in-the-loop model, divide the data samples into error sample set, minority class sample set, weak sample set and correct strong sample set, and update the training set based on the Euclidean distance algorithm with attribute weights. S4. Based on the updated training set from S3, relearn the model parameters and use the PBRB-H fault diagnosis model to complete the fault diagnosis of the wireless sensor network.

[0005] Preferably, S1 includes: Based on the feature data in wireless sensor network node fault diagnosis, a feature set for model input is generated. Based on the feature set and fault type set, a fuzzy representation of fault type based on the power set framework is generated; Based on the fuzzy representation of fault types, the inference function and model parameter set of the PBRB-H model are generated; The model parameters are optimized based on the model parameter set and loss function.

[0006] Preferably, S2 includes: Define a PBRB model under the power set identification framework, where the set of fault types is Ω, and use its power set framework for modeling; Define the rule form of the PBRB model, including: rule weights, attribute weights, and confidence distributions.

[0007] Preferably, the model inference process in S2 includes: Based on the input attribute values, generate matching results between the input sample information and the confidence rules; Based on the matching results, the activation weight of each rule is generated; Based on activation weights, an evidence reasoning algorithm is used to generate the confidence level of the model output; The output values ​​of the PBRB-H fault diagnosis model are generated based on the confidence level and the utility values ​​at different confidence levels.

[0008] Preferably, the optimization of model parameters using the P-CMA-ES algorithm in step S2 includes: Based on the initial parameter set to be optimized, initialize the mean vector, covariance matrix, and step size; Candidate solutions are generated by sampling from a normal distribution based on the mean vector, covariance matrix, and step size; Based on candidate solutions and preset constraints, feasible solutions that satisfy the constraints are generated through projection operations. Update the mean vector and covariance matrix based on the fitness values ​​of feasible solutions; Based on the updated mean vector and covariance matrix, candidate solutions are repeatedly generated and updated until the preset termination condition is met.

[0009] Preferably, S3 includes: Based on the number of new data samples monitored by sentinel variables and a preset threshold, a model update trigger signal is generated. Based on the model update trigger signal and the preset data sample classification function, an error sample set, a minority class sample set, a weak sample set, and a correct strong sample set are generated; Based on the samples in the error sample set, minority class sample set, weak sample set, and correct strong sample set, and the attribute weights of each input attribute, the closest sample set is generated by the Euclidean distance algorithm with attribute weights. An updated training set is generated based on the closest sample set, the error sample set, the minority class sample set, and the weak sample set.

[0010] Preferably, the relearned model parameters in S4 include: Based on the updated training set, the P-CMA-ES algorithm is used to re-optimize the rule weights, attribute weights, and confidence assignment parameters of the PBRB-H model; The optimized model is used to diagnose faults in wireless sensor network nodes.

[0011] This invention also provides a fault diagnosis system based on human-in-the-loop strategy and power set confidence rule base, the system being used to implement the above method, including: The module is used to build a PBRB-H fault diagnosis model based on the feature data in the fault diagnosis of wireless sensor network nodes, and to provide a formulaic description of the problems in fault diagnosis. The optimization module is used to define the basic structure of the PBRB-H fault diagnosis model constructed based on S1, execute the model inference process, and optimize the model parameters through the P-CMA-ES algorithm. The update module is used to build the human-in-the-loop model, which divides the data samples into error sample set, minority class sample set, weak sample set and correct strong sample set, and updates the training set based on the Euclidean distance algorithm with attribute weights. The diagnostic module is used to relearn the model parameters based on the updated training set of S3 and to complete the fault diagnosis of the wireless sensor network using the PBRB-H fault diagnosis model.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention significantly improves the accuracy and reliability of fault diagnosis in wireless sensor network nodes. By effectively characterizing the fuzziness and uncertainty of fault types through a power-set framework, it accurately distinguishes between local and global ignorance, enhancing the credibility of diagnostic results. The introduction of a human-in-the-loop strategy and adaptive optimization mechanism greatly reduces reliance on expert prior knowledge and large amounts of high-quality labeled data, enabling the model to continuously learn and optimize parameters from new data, exhibiting good environmental adaptability and robustness. Furthermore, through an intelligent sample selection strategy, this scheme effectively controls the computational cost of model training and updates while ensuring diagnostic performance, achieving a balance between diagnostic accuracy and operational efficiency, and is suitable for large-scale, dynamically changing wireless sensor network environments. Attached Figure Description

[0013] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the PBRB-H model according to an embodiment of the present invention; Figure 2 This is a flowchart of the P-CMA-ES optimization algorithm according to an embodiment of the present invention; Figure 3 This is a flowchart of the human-in-the-loop algorithm according to an embodiment of the present invention; Figure 4 This is a flowchart of a comparative experiment to verify the practicality and advancement of the method in this invention. Figure 5 This is a graph showing the accuracy change of the PBRB-H model during the iteration process in an embodiment of the present invention. Figure 6 This is a graph showing the variation of some rule weights with iteration in the PBRB-H model of this invention. Figure 7 This is a graph showing the change of attribute weights with iteration in the PBRB-H model of this invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] Example 1 S1. Based on the characteristic data in the fault diagnosis of wireless sensor network nodes, construct the PBRB-H fault diagnosis model and provide a formulaic description of the problems in fault diagnosis.

[0018] In fault diagnosis of wireless sensor networks, the characteristic data of node faults often exhibit a certain degree of uncertainty and ambiguity. To construct a fault diagnosis method based on the human-in-the-loop policy and power-set confidence rule base, this step systematically and formally describes the problems in fault diagnosis, laying a theoretical foundation for the subsequent model design and implementation. When constructing the PBRB-H fault diagnosis model, the following five key issues need to be addressed: Problem 1: Data Feature Extraction. In fault diagnosis of wireless sensor network nodes, input features are usually extracted from the raw data collected by the sensors. However, due to the significant spatiotemporal correlation between different nodes, these features may exhibit high similarity, making it difficult to accurately distinguish fault types. Therefore, it is necessary to design an efficient feature extraction method to obtain spatiotemporal features with discriminative power. This process can be described by formula (1): (1) in, Represents the original dataset. It is the parameter set for feature extraction. It extracts the feature set used as input to the model.

[0019] Question 2: The problem of fuzzy representation of fault types. In the fault diagnosis of wireless sensor network nodes, the data characteristics of different types of faults may be highly similar in some cases, making it difficult to accurately distinguish the specific fault type during the diagnosis process. This phenomenon is called unknowable information, which mainly includes two forms: local unknowable information and global unknowable information. To effectively characterize this kind of unknowable information, the fault type set is defined as Ω, and its power set framework is used for modeling. The power set can comprehensively describe the fuzziness and uncertainty between fault types, thus providing a more flexible and accurate information expression for the diagnostic model, which can be specifically described by equations (2) and (3): (2) (3) in, This indicates no fault. This indicates that the current fault may be or .

[0020] Question 3: Model inference problem. The PBRB-H model achieves fault diagnosis inference by mapping feature inputs and fuzzy information. The inference process of this model is shown in equation (4): (4) in, For inference functions, This represents the model parameter set, including rule weights, attribute weights, and confidence distributions. This represents the result of the model's inference.

[0021] Question 4: Parameter optimization. Initial parameters are usually determined by expert experience, but due to data uncertainty and distribution differences, these initial parameters may not be fully adapted to actual diagnostic needs. Therefore, it is necessary to continuously adjust the parameters through optimization strategies to improve the accuracy and adaptability of the model and ensure that the model can continuously improve its diagnostic performance in different environments. The optimization objective can be described as Equation (5): (5) in, It is the initial parameter set of the fault diagnosis model. This represents the optimized parameter set. Indicates the actual fault type. This represents the loss function (e.g., mean squared error) used to measure the deviation between the model's diagnosis and the true value. By iteratively minimizing the loss function, the parameters can be progressively optimized to improve the model's diagnostic accuracy.

[0022] Question 5: Sample selection problem. Data sample selection algorithms can effectively control the ever-expanding training set size in fault diagnosis models, reducing training and running costs without sacrificing model accuracy. By selecting more targeted training samples, the model can focus on the shortcomings of previous models in each iteration, thereby improving diagnostic performance. This process can be described by formula (6): (6) in, Indicates the first Data samples, Indicators representing the weaknesses of the model in the generation process (such as error rate or variance). It is a data sample filtering function that selects a sample set with higher potential for improvement by analyzing the performance of previous models on different samples. .

[0023] S2. Based on the PBRB-H fault diagnosis model constructed in S1, define the basic structure of the model, execute the model inference process, and optimize the model parameters using the P-CMA-ES algorithm.

[0024] By analyzing S1, a PBRB-H fault diagnosis model is constructed, and its overall structure is as follows: Figure 1 As shown.

[0025] (1) Definition of the basic structure of the model In the BRB model, rules are designed by domain experts combining professional knowledge with historical data and are expressed in numerous IF-THEN forms. These rules effectively handle fuzzy and uncertain information in the system by introducing confidence levels. The rules in BRB can be represented by formula (7).

[0026] (7) in Indicates the first in BRB Rules. Indicates the first Rule number 1 Reference values ​​for each prior attribute. Indicates the first The number of prior attributes in the rule. When the... The input vector of the rule satisfies hour, Corresponding confidence level . Indicates the first The weight of a rule reflects its importance relative to other rules. Indicates the first The weight of an attribute reflects its importance relative to other attributes.

[0027] However, the BRB identification framework, consisting of some states or propositions and a complete set containing all states or propositions, is incomplete. Global ignorance and local ignorance cannot be distinguished, which may lead to inaccurate diagnostic results. To distinguish between local and global ignorance information, a BRB model with a power-set identification framework, namely the PBRB model, is defined, where the set of fault types is defined as... and using its idempotent framework Modeling is performed. The PBRB model can be used as a fault diagnosis model. Its rules can be described by equation (8): (8) in The output result is or Confidence level is .

[0028] (2) The reasoning process of the model The reasoning process of the PBRB-H fault diagnosis model is based on the Evidence-Based Reasoning (ER) algorithm. This algorithm transforms confidence levels into potential probability quality. The specific reasoning steps of the Evidence-Based Reasoning algorithm are as follows: Step 1: Map the input attribute values ​​to confidence assignments. In the inference process of the PBRB-H model, the input information needs to be transformed first to calculate the degree of matching between the input sample information and the confidence rules. This process can be represented by formula (9): (9) in Represents the normalized i-th One input value. Indicates the first Rule No. Matching results under each input condition. and These represent two adjacent reference values ​​of the input attribute.

[0029] Step 2: After determining the degree of rule matching, calculate the activation weight of each rule to measure the impact of input information on the degree of rule activation. The calculation process of activation weight can be expressed by formula (10).

[0030] (10) in Indicates the first The activation weight of each rule. Indicates the rule weight. Indicates attribute weight. This represents the relative attribute weight.

[0031] Step 3: Use the ER inference algorithm to calculate the confidence level of the model output, which can be represented by formula (11).

[0032] (11) in Indicates the first Confidence level of each output reference value It is a normalization factor used to adjust the overall confidence level of the model output.

[0033] Step 4: Finally, the output of the model can be represented by equation (12).

[0034] (12) in This represents the output value of the model. This represents the utility value at different confidence levels.

[0035] (3) Model optimization process The initial parameters of the model are set by domain experts based on their professional knowledge and practical experience. Due to the limited knowledge of experts and the complexity of the fault diagnosis mechanism, it is difficult to determine the accurate and reasonable parameters of the untrained model. Therefore, it is necessary to adjust the parameters through optimization algorithms to improve the model accuracy. Model parameter optimization directly affects the diagnostic accuracy and stability, and the design of the optimization process will be directly related to the model output effect. Therefore, choosing a suitable optimization algorithm is crucial for parameter optimization. The optimization objective has been given by formula (5).

[0036] Currently, the advantages of the Projected Covariance Matrix Adaptive Evolutionary Strategy (P-CMA-ES) in BRB have been verified by multiple studies. For example, Sun et al. demonstrated experimentally that in the problem of fault node diagnosis in wireless sensor networks, the optimization capability of the P-CMA-ES algorithm is superior to the differential evolution algorithm and the FMINCON function of the MATLAB optimization toolbox. Ming et al. used the P-CMA-ES algorithm to reduce model errors caused by expert subjectivity. Wang et al. expanded the search range based on the degree of constraint violation of infeasible solutions, allowing feasible and infeasible solutions containing important information to enter the next generation population together. Subsequently, the projection operator in the P-CMA-ES algorithm was used to correct solutions that did not meet the constraints. Liu et al. established an optimization model for the developed flywheel system fault diagnosis model based on the P-CMA-ES algorithm, thereby reducing the impact of model uncertainty. As mentioned above, this embodiment uses the P-CMA-ES algorithm as the optimization algorithm for the online HBRB inventory prediction model because of its significant advantages. The Projected Covariance Matrix Adaptive Evolutionary Strategy (P-CMA-ES) is introduced to optimize the rule weights, attribute weights, and confidence assignment parameters in the HBRB model. To address the dependence of traditional models on initial parameters, this optimization algorithm reduces sensitivity to initial parameters, improving the model's adaptability and inference efficiency in multidimensional and complex variable environments. The simplified steps of the P-CMA-ES algorithm are as follows: Figure 2 As shown.

[0037] Step 1: Parameter initialization. Initialize the mean vector, covariance matrix, and step size. Input the initial parameter set to be optimized into the optimization algorithm. The initial parameter set can be represented by equation (13).

[0038] (13) Step 2: Sampling to generate candidate solutions. In each generation, samples are taken from the normal distribution and candidate solutions are generated. candidate solutions The process can be represented by equation (14).

[0039] (14) in Representation based on covariance matrix Multivariate normal distribution sampling.

[0040] Step 3: Projection Operation. Perform a projection operation on each candidate solution to ensure that the solution satisfies the constraints. Define the projection operation as follows: This process can be described by equation (15).

[0041] (15) in It is Projection operation on the feasible region. express Projection in the feasible region. express Constraint variables in. This represents a parameter vector. express The number of constraints.

[0042] Step 4: Update the mean vector. Select the optimal samples (e.g., the solutions with the best fitness values) and update the mean vector based on these samples. This process can be represented by Equation (16).

[0043] (16) in The sample weight coefficients typically satisfy the following conditions: . It is the updated mean vector.

[0044] Step 5: Update the covariance matrix. Update the covariance matrix based on the selected samples. This process adapts to changes in the search space. It can be represented by equation (17).

[0045] (17) in and It's the learning rate. It is the step control vector, representing the direction information in the covariance update. This represents the updated covariance matrix.

[0046] Step 6: Assign the updated values ​​of the mean vector and covariance matrix to the mean vector respectively. m Covariance Matrix C The algorithm is then terminated based on a termination condition (such as reaching the maximum number of iterations or the target accuracy). If the termination condition is not met, the process returns to step 2 to continue iterating.

[0047] S3. Construct a human-in-the-loop model, divide the data samples into error sample set, minority class sample set, weak sample set and correct strong sample set, and update the training set based on the Euclidean distance algorithm with attribute weights.

[0048] This step proposes a human-in-the-loop model for fault diagnosis. The model divides available data samples into four categories: misclassified samples, minority class samples, weak class samples, and correct strong class samples. For samples in the misclassified, minority class, and weak class samples, the algorithm uses Euclidean distance with attribute weights to calculate the similarity between the correct strong class samples and the aforementioned samples, and selects the sample from the correct strong class samples that is closest to the aforementioned sample in multidimensional space. The selected sample, together with the misclassified, minority class, and weak class samples, constitutes a new data sample set, which is added to the training set to update model parameters. Unselected correct strong class samples are removed. The human-in-the-loop algorithm flow is as follows: Figure 3 As shown, the specific steps are as follows: Step 1: Sentinel variables are used to monitor the availability of data samples.

[0049] During model training, the availability of data samples is monitored by a monitoring variable G. When the number of new data samples reaches a preset threshold, the model will trigger an alarm to prompt an update. The mechanism of this monitoring variable can be described by equation (18).

[0050] (18) in This indicates the number of new data samples currently available. It is the threshold that triggers the update.

[0051] Step 2: Data Sample Grouping. Before updating the model, the Human-in-the-Loop algorithm first divides the data samples into four categories: misclassified samples, minority class samples, weakly classified samples, and correctly classified strong samples. The grouping of data samples can be described by equations (19)-(22): (19) in This represents the minority class sample set. Indicates the first The number of class samples.

[0052] (20) in This represents a weak sample set. Indicates the first The sample accuracy of each category is used to determine the category with the lowest accuracy, which is then selected as a weak classifier.

[0053] (twenty one) in This represents the set of samples that were misjudged. This represents a function used to detect classification errors in new data.

[0054] (twenty two) in This represents the correct strong sample set, which is all the data remaining after removing the incorrect sample set, the minority sample set, and the weak sample set.

[0055] Step 3: Weighted Euclidean Distance Algorithm. After obtaining the four datasets from Step 2, the Human-in-the-Loop algorithm invokes the weighted Euclidean distance algorithm. Data from the correct strong sample set that is closest to the error sample set, minority sample set, and weak sample set are selected. The weighted Euclidean distance algorithm can be described by Equation (23).

[0056] (twenty three) in Represents the first, second, and third instances in the error set, minority set, and weak set. The data sample and the first data sample in the correct strong sample set The Euclidean distance between data samples with attribute weights. in This represents the total number of data samples in the error set, minority set, and weak set. Represents the first, second, and third samples in the error sample set, minority sample set, and weak sample set, respectively. The first data sample One input attribute. This indicates the first correct strong sample set. The first data point One input attribute. Indicates the first The attribute weights of each input attribute.

[0057] Step 4: Update the model's training set by integrating the sample set obtained in Step 3 with the error sample set, minority class sample set, and weak sample set. Ensure that the training sample set contains all the necessary data to improve model accuracy.

[0058] The training set update is represented by formula (24).

[0059] (twenty four) in This represents the data sample set obtained in step 3.

[0060] S4. Based on the updated training set from S3, relearn the model parameters and use the PBRB-H fault diagnosis model to complete the fault diagnosis of the wireless sensor network.

[0061] After obtaining a new training sample set, the parameters of the PBRB-H fault diagnosis model are updated according to the formula defined in S2. Finally, the PBRB-H fault diagnosis model is used to complete the fault diagnosis of the wireless sensor network.

[0062] Example 2 To verify the practicality and advancement of the present invention, this embodiment is set up as a comparative experiment.

[0063] The experiment mainly includes dataset download, data preprocessing, model building and training, and comparative analysis of experimental results. The experimental procedure is as follows: Figure 4 As shown.

[0064] This embodiment selects a wireless sensor network dataset collected and released by Intel Berkeley Labs. This dataset contains data collected by 54 sensors deployed at Intel Berkeley Labs between February 28th and April 5th, 2004. Each sensor node collected data including temperature, humidity, light intensity, and voltage. Analysis of the dataset revealed data gaps during specific time periods, and the amount of data from sensors 1 to 4 varied inconsistently over time. Therefore, an averaging method was used to fill in the missing data based on the original data. Finally, data collected every 5 minutes from sensors 1 to 4 between March 1st and March 7th was processed, totaling 2016 data points per sensor. After data preprocessing, 1897 valid data points were extracted.

[0065] (1) Data preprocessing Since raw sensor data cannot directly reflect the fault status of sensor network nodes, data features need to be extracted as model input. Wireless sensor networks (WSNs), as distributed data acquisition networks, have numerous sensors with the same function—collecting information about the object being measured. This sensor information exhibits the following spatiotemporal similarity characteristics: First, temporal correlation. If the overall trend of the object being measured remains consistent over time, the collected data will show a similar trend. For example, when a wireless sensor network is used to monitor temperature changes in a certain area, if the overall temperature of that area shows an upward trend, the data collected in that area will also show a similar upward trend. Second, spatial correlation. The smaller the distance between two monitoring points, the closer they are to the object being monitored. Therefore, closely distributed sensors have strong similarity in their monitoring data. For example, when monitoring temperature changes in a certain area, when the distance between two sensors is less than one meter, the monitoring data is almost identical. When a sensor network node fails, the temporal and spatial correlation characteristics of the data change, and the data characteristics differ for different types of faults. Based on this, the temporal and spatial correlation data features of wireless sensor network nodes are selected as model input.

[0066] Trend correlation is a form of time-series correlation, reflecting the degree of similarity in the trends of collected data within a specific time period. Its calculation is described by formula (25): (25) in Indicates sensor In time The collected data. Indicates sensor From time Time The average value of the collected data.

[0067] The residual characteristics represent spatial correlation and can be described by equation (26): (26) in Indicates sensor In time The collected data, Indicates other sensors in time Average data collected. Indicates the number of sensors.

[0068] (2) Model building Step 1: Extract data features, including trend correlation features and residual features. To construct a fault diagnosis model for wireless sensor network nodes, features need to be extracted from the original sensor data first. This paper uses the method shown in formula (25) to calculate the trend correlation features, denoted as... The residual characteristics are calculated using the method shown in formula (26), denoted as... .

[0069] Step 2: Define the reference point and reference value for the input attributes. This is done by analyzing the input attributes. and The data features and graphics allow for the definition of reference points for two input attributes. First, There are seven reference points that can be described by formula (27): very low (VL), relatively low (RL), low (L), medium (M), high (H), relatively high (RH) and very high (VH), with corresponding reference values ​​shown in Table 1. There are seven reference points, which can be described by Equation (28), namely, extremely low (VL), relatively low (RL), low (L), medium (M), high (H), relatively high (RH), and extremely high (VH), and the reference values ​​of the corresponding reference points are shown in Table 2. Finally, five reference points are set for the model output results, which can be described by Equation (29), namely, normal (N), offset fault (OSF), high noise fault (HNF), outlier fault (OF), and fixed value fault (FVF), and their reference values ​​are shown in Table 3.

[0070] (27) (28) (29) Table 1 Table 2 Table 3 The model construction process according to Implementation Example 1 needs to be based on the set of fault types. Construct an idempotent framework. This idempotent framework is a set of fault types defined in the modeling process. However, observations of the data feature images show that local ignorance exists only in the neighboring faults within the defined fault. Therefore, the local ignorance and global ignorance that are absent in the power set framework can be mitigated. Finally, the model's discriminative framework... As shown in equation (30).

[0071] (30) Step 3: The data features extracted in Step 1, together with the reference points and reference values ​​determined in Step 2, constitute the initial confidence rule base, containing 49 rules. The initial parameters of the model are optimized using an optimization algorithm. The optimized rule weights and confidence levels are shown in Table 4. The NO. column represents the rule number; the current model contains a total of 49 rules. The rule weight column represents the weight value of the current rule. The attribute columns... Indicates a reference point for trend correlation. This represents the reference point for residual characteristics.

[0072] Table 4 (3) Model training In this embodiment, the first 497 samples are selected as the initial training set, and the remaining samples are used as the test set. Since this method has a dynamic growth characteristic, the selection of the initial training samples can be flexibly adjusted according to actual needs. Assuming that the model provides new sample data for learning every 200 predictions, the test set is divided into 7 groups, each containing approximately 200 samples. First, the model is initially trained using the first 497 training samples. Then, in each iteration, the model dynamically updates the confidence distribution based on the newly added sample data, undergoing a total of 7 iterations. With the increase in training samples, the model parameters are enhanced in adapting to different data samples and improving prediction accuracy and stability. To ensure the effectiveness of the experimental analysis, establishing clear and objective evaluation criteria is crucial. In this embodiment, the overall sample accuracy is selected as the indicator for evaluating the model's precision and effectiveness, and its calculation method is shown in formula (31).

[0073] (31) in This represents the overall accuracy of the model. This indicates the number of samples with correct output results out of all samples. This represents the total number of all samples.

[0074] Figure 5 The graph illustrates the performance of the PBRB-H model during the iteration process. As can be seen from the figure, in the first round of sample testing, the model's diagnostic results differed significantly from the true values, with an accuracy of only 65%. With each iteration, the accuracy significantly improved to 84% in the third training round. In subsequent iterations, the model's accuracy continued to improve and showed an overall upward trend, reaching a final sample accuracy of 98% after seven rounds of learning.

[0075] The rule weights for different rounds during model iteration are shown in Table 5. Figure 6 The iterative change curves of the weights for some randomly selected rules are shown. The attribute weights for different rounds are shown in Table 6. Figure 7 The iterative change process of attribute weights is demonstrated. It is evident that the model continuously adjusts its parameters during iteration to adapt to dynamically changing data, effectively enhancing the model's adaptability to dynamic environments.

[0076] Table 5 Table 6 To more comprehensively evaluate model performance, the PBRB-H model proposed in this embodiment was compared with five mainstream fault diagnosis models (Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Extreme Learning Machine (ELM), and Backpropagation Neural Network (BPNN)). Each model underwent twenty rounds of experiments. In each round, 100 samples were randomly selected as the test set, and the remaining samples were used for subsequent validation. Table 7 shows the average accuracy of each model after 20 rounds of experiments.

[0077] Table 7 Experimental results show that the PBRB-H fault diagnosis model achieves an average accuracy of 92.15%, significantly outperforming comparative models such as RF, KNN, SVM, ELM, and BPNN. Furthermore, its accuracy consistently remains above 90% with minimal fluctuations, demonstrating exceptional robustness. In contrast, traditional models like SVM and BPNN exhibit accuracy below 80% and significant fluctuations in some rounds, while KNN and ELM even reach as low as 76% in certain rounds, reflecting the poor adaptability of traditional methods in complex data environments. By introducing a human-in-the-loop strategy and a power-set framework, the PBRB-H model not only significantly enhances its ability to express and reason about multi-state and unknown information but also dynamically adjusts diagnostic rules and parameters, maintaining excellent diagnostic performance even in data-imbalanced and high-noise environments, fully demonstrating its application value in complex industrial fault scenarios.

[0078] Example 3 This embodiment also provides a fault diagnosis system based on the human-in-the-loop strategy and power set confidence rule base, including: a construction module, used to construct a PBRB-H fault diagnosis model based on feature data in the fault diagnosis of wireless sensor network nodes, and to provide a formulaic description of the problems in fault diagnosis; an optimization module, used to define the basic structure of the constructed PBRB-H fault diagnosis model, execute the model inference process, and optimize the model parameters using the P-CMA-ES algorithm; an update module, used to construct a human-in-the-loop model, divide the data samples into an error sample set, a minority class sample set, a weak sample set, and a correct strong sample set, and update the training set based on the Euclidean distance algorithm with attribute weights; and a diagnosis module, used to relearn the model parameters based on the updated training set, and use the PBRB-H fault diagnosis model to complete the fault diagnosis of the wireless sensor network.

[0079] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for fault diagnosis based on human-in-the-loop strategy and power set confidence rule base, characterized in that, Includes the following steps: S1. Based on the characteristic data in the fault diagnosis of wireless sensor network nodes, construct the PBRB-H fault diagnosis model and provide a formulaic description of the problems in fault diagnosis. S2. Based on the PBRB-H fault diagnosis model constructed in S1, define the basic structure of the model, execute the model inference process, and optimize the model parameters using the P-CMA-ES algorithm; S3. Construct a human-in-the-loop model, divide the data samples into error sample set, minority class sample set, weak sample set and correct strong sample set, and update the training set based on the Euclidean distance algorithm with attribute weights. S4. Based on the updated training set from S3, relearn the model parameters and use the PBRB-H fault diagnosis model to complete the fault diagnosis of the wireless sensor network.

2. The human-in-the-loop strategy and power set confidence rule base based fault diagnostic method according to claim 1, characterized in that, S1 includes: Based on the feature data in wireless sensor network node fault diagnosis, a feature set for model input is generated. Based on the feature set and fault type set, a fuzzy representation of fault type based on the power set framework is generated; Based on the fuzzy representation of fault types, the inference function and model parameter set of the PBRB-H model are generated; The model parameters are optimized based on the model parameter set and loss function.

3. The human-in-the-loop strategy and power set confidence rule base based fault diagnostic method according to claim 2, characterized in that, S2 includes: Define a PBRB model under the power set identification framework, where the set of fault types is Ω, and use its power set framework for modeling; Define the rule form of the PBRB model, including: rule weights, attribute weights, and confidence distributions.

4. The human-in-the-loop strategy and power set confidence rule base based fault diagnostic method according to claim 3, characterized in that, The execution model inference process in S2 includes: Based on the input attribute values, generate matching results between the input sample information and the confidence rules; Based on the matching results, the activation weight of each rule is generated; Based on activation weights, an evidence reasoning algorithm is used to generate the confidence level of the model output; The output values ​​of the PBRB-H fault diagnosis model are generated based on the confidence level and the utility values ​​at different confidence levels.

5. The human-in-the-loop strategy and power set confidence rule base based fault diagnostic method according to claim 4, characterized in that, The optimization of model parameters using the P-CMA-ES algorithm in S2 includes: Based on the initial parameter set to be optimized, initialize the mean vector, covariance matrix, and step size; Candidate solutions are generated by sampling from a normal distribution based on the mean vector, covariance matrix, and step size; Based on candidate solutions and preset constraints, feasible solutions that satisfy the constraints are generated through projection operations. Update the mean vector and covariance matrix based on the fitness values ​​of feasible solutions; Based on the updated mean vector and covariance matrix, candidate solutions are repeatedly generated and updated until the preset termination condition is met.

6. The human-in-the-loop policy and power set confidence rule base based fault diagnostic method according to claim 1, wherein, S3 includes: Based on the number of new data samples monitored by sentinel variables and a preset threshold, a model update trigger signal is generated. Based on the model update trigger signal and the preset data sample classification function, an error sample set, a minority class sample set, a weak sample set, and a correct strong sample set are generated; Based on the samples in the error sample set, minority class sample set, weak sample set, and correct strong sample set, and the attribute weights of each input attribute, the closest sample set is generated by the Euclidean distance algorithm with attribute weights. An updated training set is generated based on the closest sample set, the error sample set, the minority class sample set, and the weak sample set.

7. The human-in-the-loop strategy and power set confidence rule base based fault diagnostic method according to claim 6, characterized in that, The relearned model parameters in S4 include: Based on the updated training set, the P-CMA-ES algorithm is used to re-optimize the rule weights, attribute weights, and confidence assignment parameters of the PBRB-H model; The optimized model is used to diagnose faults in wireless sensor network nodes.

8. A fault diagnosis system based on human-in-the-loop strategy and power-set confidence rule base, the system being used to implement the method described in any one of claims 1-7, characterized in that, include: The module is used to build a PBRB-H fault diagnosis model based on the feature data in the fault diagnosis of wireless sensor network nodes, and to provide a formulaic description of the problems in fault diagnosis. The optimization module is used to define the basic structure of the PBRB-H fault diagnosis model constructed based on S1, execute the model inference process, and optimize the model parameters through the P-CMA-ES algorithm. The update module is used to build the human-in-the-loop model, which divides the data samples into error sample set, minority class sample set, weak sample set and correct strong sample set, and updates the training set based on the Euclidean distance algorithm with attribute weights. The diagnostic module is used to relearn the model parameters based on the updated training set of S3 and to complete the fault diagnosis of the wireless sensor network using the PBRB-H fault diagnosis model.

Images