A method and related equipment for scoring the behavioral risks of power grid employees based on multi-model collaboration and uncertainty perception.

Abstract

This application provides a method and related equipment for scoring the behavioral risk of power grid employees based on multi-model collaboration and uncertainty perception, belonging to the field of multi-model collaborative optimization and intelligent risk assessment technology. The method includes: constructing a unified behavioral sample; inputting the unified behavioral sample in parallel into several different risk assessment models to obtain the risk scores, model confidence, and data completeness output by the risk assessment models; dynamically calculating the collaborative weights of each risk assessment model for the current unified behavioral sample; performing a weighted fusion operation on each risk score using the collaborative weights to generate an initial behavioral risk score representing the overall risk level; constructing a collaborative uncertainty index; determining whether the collaborative uncertainty index exceeds a preset threshold; if it does, triggering a reassessment model to perform a secondary scoring on the current unified behavioral sample to output the final behavioral risk score. This application can achieve more accurate and stable employee behavioral risk level determination and retraining task generation.

Description

Technical Field

[0001] This application relates to the fields of multi-model collaborative optimization and intelligent risk assessment, and in particular to a method and related equipment for scoring the behavioral risks of power grid employees based on multi-model collaboration and uncertainty perception. Background Technology

[0002] In related technologies, information such as training, examinations, job execution, violation handling, equipment operation, and team collaboration is often scattered across different systems. When data is missing, samples are sparse, job differences are significant, or there are many boundary samples, the risk scores output by the model often lack reliable measurement, making the scoring results difficult to interpret. This, in turn, affects the efficiency of platform-side early warning threshold setting, retraining task arrangement, and manual review.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a method and related equipment for scoring the behavioral risks of power grid employees based on multi-model collaboration and uncertainty perception. This method can solve the problems of data fragmentation, difficulty in handling model conflicts, insufficient credibility of scoring results, and weak re-judgment ability in high-risk scenarios in training evaluation and behavioral scoring schemes. It also highlights the technical contributions of scoring, re-judgment, knowledge injection, and dynamic weighting in a purely algorithmic manner.

[0005] To achieve the above objectives, one aspect of this application proposes a method for scoring the behavioral risks of power grid employees based on multi-model collaboration and uncertainty perception. The method includes the following steps: Acquire multi-source heterogeneous employee behavior data and preprocess it to construct a unified behavior sample; The unified behavior sample is input into several different risk assessment models in parallel to obtain the risk score, model confidence and data completeness of the corresponding output of the risk assessment model; Based on the model confidence, the data completeness, and the historical calibration error in similar scenarios that meet preset conditions, the collaborative weight of each risk assessment model for the current unified behavior sample is dynamically calculated. The risk scores are weighted and fused using the collaborative weights to generate an initial behavioral risk score that represents the overall risk level. Based on the statistical dispersion of each risk score, the cumulative deviation of each risk score from the initial behavioral risk score, and the statistical dispersion of each model confidence level, a collaborative uncertainty index is constructed. Determine whether the collaborative uncertainty index exceeds a preset threshold. If it does, trigger the re-judgment model to perform a secondary scoring on the current unified behavior sample to output the final behavior risk score.

[0006] In some embodiments, acquiring multi-source heterogeneous employee behavior data and preprocessing it to construct a unified behavior sample includes: Several data sources are collected from the training management system, examination system, work ticket system, safety supervision system, equipment monitoring system, and human resources system. The data sources include training course completion records, theoretical examination scores and time taken, practical assessment scores and evaluation levels, work ticket issuance and execution details, violation records and deduction values, equipment operation sequence logs, access control records, job qualification certificate information, and team affiliation data. The collected data sources are cleaned at the field level to remove records with outlier formats and logical conflicts. Missing fields are processed using at least one of the following methods: missing mask marking, filling with the mean of similar samples, or filling with preset business default values. The different data sources are aligned by time windows according to employee identifier, timestamp, and business event type; the time windows include fixed statistical period windows, tracking and observation windows starting from preset training events, or front and back slice windows centered on preset risk events. Based on the aligned data, data features are extracted and structured to obtain unified behavioral samples. These data features include time-series features, job-related features, relationship features, and anomaly features. The time-series features include the performance trends of employees within a continuous time window, the fluctuation patterns of operation frequency, and the distribution of violation interval durations. The job-related features include the matching status of job qualifications with practical tasks and the compliance status of retraining cycles. The relationship features include graph structure features extracted based on the association graph between employees and job positions, equipment, job types, and work groups. The anomaly features include the degree of deviation of behavioral patterns from the group benchmark and the identification markers of risk events.

[0007] In some embodiments, the risk assessment model includes a rule-constrained model, a time-series behavior model, a relationship graph model, and an anomaly detection model; The rule constraint model constructs a rule engine based on a preset rule base, job qualification constraints, and system trigger items. It obtains the risk score of the rule constraint model by calculating the weighted score of each triggered rule and performing normalization processing. The model confidence of the rule constraint model is determined based on the consistency ratio or conflict ratio among the triggered rules. The data completeness of the rule constraint model is calculated based on the proportion of valid fields that meet the preset conditions of the input fields required for rule triggering. The time-series behavior model uses a bidirectional gated recurrent unit network or a Transformer encoder to encode the behavioral sequence data of employees within a continuous time window. The risk score of the time-series behavior model is output through a fully connected layer and a Sigmoid activation function. The model confidence of the time-series behavior model is determined based on the normalized entropy value of the output probability distribution. The data completeness of the time-series behavior model is calculated based on the proportion of valid observation points in the time-series input field that meet preset conditions. The relationship graph model constructs a heterogeneous graph structure based on the relationships between employees, positions, equipment, job types, and work groups. It uses any one of graph attention networks, GraphSAGE, or relationship graph convolutional networks to perform message propagation and node representation learning on the heterogeneous graph structure. The risk score of the relationship graph model is output through node embedding vectors and a classification layer. The model confidence of the relationship graph model is determined based on the normalized entropy value of the output probability. The data completeness of the relationship graph model is calculated based on the proportion of valid edges and valid node attributes that meet preset conditions in the heterogeneous graph structure. The anomaly detection model uses any one of the following algorithms—isolated forest, autoencoder network, or support vector machine—to measure the deviation of the uniform behavior sample from the group baseline pattern. The anomaly score is normalized and then mapped to the risk score of the anomaly detection model. The model confidence of the anomaly detection model is determined based on the consistency of several perturbation inferences or the distance between the anomaly score and a preset anomaly threshold. The data completeness of the anomaly detection model is calculated based on the proportion of valid values ​​of the continuous feature fields required for anomaly detection that meet preset conditions.

[0008] In some embodiments, dynamically calculating the collaborative weights of each risk assessment model for the current unified behavior sample based on the model confidence, the data completeness, and historical calibration errors in similar scenarios meeting preset conditions includes: Obtain the scene attribute vector of the current unified behavior sample; the scene attribute vector includes job category code, job type code, voltage level code, work group type code, time window length value, and data missing rate interval code; Calculate the weighted similarity between the scene attribute vector of the current unified behavior sample and the scene attribute vector of each historical sample in the historical sample set, and classify the historical samples with a weighted similarity greater than a preset similarity threshold into the similar scene sample set; Based on the set of similar scenario samples, the local historical calibration error of each of the risk assessment models is calculated; the local historical calibration error is the mean square error or mean absolute error between the predicted risk score and the actual risk label of the risk assessment model on each sample in the set of similar scenario samples. The local historical calibration error is weighted and fused with the global historical calibration error of each risk assessment model on the global validation set to obtain the scenario-based historical calibration error for the current unified behavior sample. The fusion weight is dynamically determined according to the ratio of the number of samples in the similar scenario sample set to a preset sample number threshold. When the number of similar scenario samples is insufficient, the fusion ratio of the global historical calibration error is automatically increased to a preset range. For each of the risk assessment models, the original weight energy value of the risk assessment model is calculated. The original weight energy value is equal to the confidence adjustment coefficient multiplied by the model confidence of the risk assessment model, plus the completeness adjustment coefficient multiplied by the data completeness of the risk assessment model, minus the error penalty coefficient multiplied by the scenario-based historical calibration error of the risk assessment model, wherein the confidence adjustment coefficient, the completeness adjustment coefficient, and the error penalty coefficient are all positive real numbers. A normalized exponential function operation is performed on the original weight energy values ​​of each risk assessment model to obtain the collaborative weights corresponding to each model. The normalized exponential function operation includes: calculating the exponential value of the original weight energy value of each risk assessment model, dividing the exponential value of the risk assessment model by the sum of the exponential values ​​of all risk assessment models, and the resulting ratio is the collaborative weight corresponding to the risk assessment model. The sum of the collaborative weights of all risk assessment models is equal to one.

[0009] In some embodiments, the step of performing a weighted fusion operation on each of the risk scores using the collaborative weights to generate an initial behavioral risk score characterizing the overall risk level includes: The risk score output by each risk assessment model is multiplied by the corresponding collaborative weight of the risk assessment model to obtain the weighted risk contribution value of the risk assessment model. The weighted risk contribution values ​​of all the risk assessment models are summed up, and the summation result is used as the initial behavioral risk score; the value range of the initial behavioral risk score is consistent with the value range of each of the risk scores, which are all normalized intervals. During the weighted fusion process, if the data completeness of a certain risk assessment model is lower than the preset minimum completeness threshold, the collaborative weight of the risk assessment model is set to zero, and the weight of the risk assessment model is proportionally redistributed to the other risk assessment models that meet the completeness requirements. If the initial behavioral risk score obtained after weighted fusion exceeds the preset risk score upper limit, the initial behavioral risk score will be truncated to the risk score upper limit.

[0010] In some embodiments, the construction of a collaborative uncertainty index based on the statistical dispersion of each of the risk scores, the cumulative deviation of each of the risk scores from the initial behavioral risk score, and the statistical dispersion of each of the model confidence scores includes: Calculate the first statistical dispersion of the risk scores output by each of the risk assessment models; the first statistical dispersion uses variance or standard deviation as a measure to characterize the degree of disagreement between the risk assessment models for the same sample. Calculate the cumulative deviation between each of the risk scores and the initial behavioral risk score; the cumulative deviation is measured by the sum of the absolute values ​​of the differences between each of the risk scores and the initial behavioral risk score, or by the square root of the sum of the squares of the differences between each of the risk scores and the initial behavioral risk score, to characterize the overall deviation between the scoring result of a single risk assessment model and the initial behavioral risk score. Calculate the second statistical dispersion of the confidence scores of each model; the second statistical dispersion uses variance or standard deviation as a measure to characterize the consistency of each risk assessment model in judging the reliability of its own output; The first statistical dispersion, the cumulative deviation, and the second statistical dispersion are each multiplied by their respective preset weighting coefficients and then linearly weighted and summed. The summation result is used as the collaborative uncertainty index.

[0011] In some embodiments, determining whether the collaborative uncertainty index exceeds a preset threshold, and if so, triggering a reassessment model to perform a secondary scoring on the current unified behavior sample to output a final behavior risk score, includes: A threshold is preset based on the performance indicators of the risk assessment model on the validation set; the performance indicators include false positive rate and false negative rate; the threshold is negatively correlated with the risk tolerance of the business scenario. The collaborative uncertainty index is compared with the threshold. When the collaborative uncertainty index is less than or equal to the threshold, it indicates that the initial behavioral risk score has sufficient credibility, and the initial behavioral risk score is output as the final behavioral risk score. When the collaborative uncertainty index is greater than the threshold, it indicates that the risk assessment models have inconsistent judgments on the reliability of their own outputs, triggering the re-judgment model to perform a second scoring on the current unified behavior sample. After triggering the reassessment model, a joint feature vector is constructed for input to the reassessment model. The joint feature vector is composed of the following components concatenated in a preset order: the original feature semantic vector extracted by the feature encoder from the unified behavior sample; the risk score sequence vector formed by concatenating the risk scores output by each of the risk assessment models in a preset model order; the confidence score sequence vector formed by concatenating the model confidence scores output by each of the risk assessment models in the same preset model order; the data completeness sequence vector formed by concatenating the data completeness output by each of the risk assessment models in the same preset model order; and an extended dimension formed by appending the collaborative uncertainty index as an independent dimension to the end of the vector. The joint feature vector is input into the pre-built re-judgment model for forward inference. The re-judgment model is implemented using any one of the following: a lightweight multilayer perceptron network, a logistic regression classifier, a gradient boosting tree model, or a rule-based classifier. When the re-judgment model is implemented using a lightweight multilayer perceptron network, the lightweight multilayer perceptron network includes an input layer, at least one fully connected hidden layer and an output layer. The hidden layer uses a linear rectified function as the activation function, and the output layer uses a sigmoid activation function to map the output value to the same normalized risk score range as the initial behavioral risk score. When the re-judgment model is implemented using a gradient boosting tree model, the gradient boosting tree model uses a decision tree as a base learner and gradually optimizes the model performance by iteratively fitting the prediction residuals of the previous round, and finally outputs the risk score result mapped after weighted voting by multiple decision trees. The revised final behavioral risk score is output using the aforementioned reassessment model. The final behavioral risk score is a continuous risk score, and its value range is consistent with that of the initial behavioral risk score.

[0012] In some embodiments, the risk assessment model and the re-judgment model employ a multi-loss adaptive weighting mechanism during the training process, the multi-loss adaptive weighting mechanism comprising: Construct a joint optimization total loss function; the total loss function is composed of a weighted sum of the main risk scoring loss term, the knowledge consistency constraint loss term, the confidence calibration loss term, and the uncertainty constraint loss term. In each iteration of the training process, the average collaborative uncertainty index of all samples in the current training batch is calculated, and the ratio of the average collaborative uncertainty index of the current round to the average collaborative uncertainty index of the initial training round is used as a dynamic adjustment factor. The initial weight coefficients of the knowledge consistency constraint loss term, the initial weight coefficients of the confidence calibration loss term, and the initial weight coefficients of the uncertainty constraint loss term are multiplied by the dynamic adjustment factor to obtain the actual weight coefficients of each auxiliary loss term in the current training round. The total loss function is synchronously backpropagated to update the network parameters of the risk assessment model and the network parameters of the re-judgment model until the total loss function converges to a preset condition or reaches the upper limit of a preset training round.

[0013] In some embodiments, the main risk score loss term is used to measure the degree of deviation between the final behavioral risk score and the true risk label. When the true risk label is a binary label, the main risk score loss term is calculated using weighted cross-entropy loss, assigning a first weight coefficient to positive samples and a second weight coefficient to negative samples. The first weight coefficient and the second weight coefficient are adjusted to balance the ratio of positive to negative samples. When the true risk label is a continuous risk score, the main risk score loss term is calculated using mean squared error loss or mean absolute error loss. The knowledge consistency constraint loss term is used to measure the semantic alignment between teacher representations and student representations across multiple business semantic feature blocks. The teacher representations are obtained by forward reasoning of the unified behavior samples based on a preset rule knowledge base or expert experience model. The student representations are extracted from the output of the intermediate hidden layer of at least one of the risk assessment models. The knowledge consistency constraint loss term is obtained by calculating the negative value of the cosine similarity between the teacher representation and the student representation on the corresponding business semantic feature blocks and performing batch averaging. The smaller the value of the knowledge consistency constraint loss term, the higher the semantic alignment between the student representation and the teacher representation. The confidence calibration loss term is used to measure the consistency between the fusion confidence and the prediction accuracy index output by each of the risk assessment models. The fusion confidence is obtained by weighting and summing the model confidence of each of the risk assessment models according to the collaborative weight coefficient. The prediction accuracy index is an indicator function value, which takes a value of one when the predicted risk category corresponding to the final behavioral risk score is consistent with the actual risk label, and a value of zero otherwise. The confidence calibration loss term is calculated by squared the difference between the fusion confidence and the prediction accuracy index. The uncertainty constraint loss term is used to measure the consistency between the collaborative uncertainty index and the prediction error. The prediction error is the absolute value of the difference between the final behavioral risk score and the true risk label. The uncertainty constraint loss term is calculated by squared the difference between the collaborative uncertainty index and the prediction error. By minimizing the uncertainty constraint loss term, the collaborative uncertainty index reflects the prediction error level of the model for the current sample.

[0014] To achieve the above objectives, another aspect of this application proposes a power grid employee behavior risk scoring device based on multi-model collaboration and uncertainty perception, the device comprising: The first module is used to acquire multi-source heterogeneous employee behavior data and preprocess it to build a unified behavior sample. The second module is used to input the unified behavior sample into several different risk assessment models in parallel to obtain the risk score, model confidence and data completeness output by the risk assessment model. The third module is used to dynamically calculate the collaborative weight of each risk assessment model for the current unified behavior sample based on the model confidence, the data completeness, and the historical calibration error in similar scenarios that meet preset conditions. The fourth module is used to perform a weighted fusion operation on each of the risk scores using the collaborative weights to generate an initial behavioral risk score that represents the overall risk level. The fifth module is used to construct a collaborative uncertainty index based on the statistical dispersion of each of the risk scores, the cumulative deviation of each of the risk scores from the initial behavioral risk score, and the statistical dispersion of each of the model confidence scores. The sixth module is used to determine whether the collaborative uncertainty index exceeds a preset threshold. If it does, the re-judgment model is triggered to perform a secondary scoring on the current unified behavior sample to output the final behavior risk score.

[0015] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0017] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0018] The embodiments of this application include at least the following beneficial effects: This application provides a method, device, electronic device, storage medium, and program product for scoring the behavioral risk of power grid employees based on multi-model collaboration and uncertainty perception. The method includes: constructing a unified behavioral sample; inputting the unified behavioral sample into several different risk assessment models in parallel to obtain the risk score, model confidence, and data completeness output by the risk assessment models; dynamically calculating the collaborative weights of each risk assessment model for the current unified behavioral sample; performing a weighted fusion operation on each risk score using the collaborative weights to generate an initial behavioral risk score representing the comprehensive risk level; constructing a collaborative uncertainty index; determining whether the collaborative uncertainty index exceeds a preset threshold, and if it does, triggering a reassessment model to perform a secondary scoring on the current unified behavioral sample to output the final behavioral risk score. This application can achieve more accurate and stable employee behavioral risk level determination and retraining task generation. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for scoring the behavioral risks of power grid employees based on multi-model collaboration and uncertainty perception, provided in an embodiment of this application. Figure 2 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0021] 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. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0023] 1) Transformer, a neural network architecture based on self-attention mechanism.

[0024] 2) Sigmoid, a type of S-type activation function.

[0025] 3) GraphSAGE, an inductive node representation learning method in graph neural networks.

[0026] 4) BiGRU, a bidirectional gated recurrent unit, is a variant of a recurrent neural network that scans sequence data in both forward and reverse directions, capturing contextual information from the past and future.

[0027] 5) GAT, Graph Attention Network, introduces an attention mechanism into graph neural networks, enabling the model to automatically learn the differentiated importance weights of different neighboring nodes to the target node.

[0028] 6) R-GCN, Relational Graph Convolutional Network, extends the ability to model various edge relationship types on the basis of graph convolutional networks. Different weight matrices are used for message passing for different types of edges.

[0029] In related technologies, information such as training, examinations, job execution, violation handling, equipment operation, and team collaboration is often scattered across different systems, lacking a unified modeling mechanism. Even when machine learning models are introduced, they often employ single-model scoring or averaging of multiple models with fixed weights, failing to explicitly address discrepancies, uncertainties, and scenario applicability between models. This application addresses the shortcomings of related technologies: The lack of a unified representation of multi-source behavioral evidence makes it impossible to establish a stable association between training performance, field behavior, and job constraints; Multi-model scoring lacks a dynamic collaboration mechanism and cannot distinguish the reliability of different models under different job positions, scenarios, and data quality conditions. There is a lack of dedicated processing mechanisms for model divergence samples and boundary samples, and high-risk but low-confidence samples are prone to misjudgment. Without the input of teacher knowledge or rule knowledge, a purely data-driven model is prone to learning superficial correlations that are inconsistent with safety regulations. The training phase lacks dynamic balance for scoring loss, consistency loss, calibration loss, and uncertainty constraints, resulting in insufficient model convergence stability.

[0030] In view of this, this application provides a method and related equipment for scoring the behavioral risk of power grid employees based on multi-model collaboration and uncertainty perception. This scheme achieves employee behavioral risk scoring based on pure data input, without relying on image input, video input, or dedicated hardware acquisition structures. It constructs employee behavioral samples, calls multiple risk scoring models in parallel, dynamically collaborates and fuses the outputs of multiple models, calculates collaborative uncertainty indicators, and triggers a re-judgment model for high-uncertainty samples, ultimately outputting the employee behavioral risk score. Simultaneously, this invention uses teacher-student consistency constraints and a multi-loss adaptive weighting mechanism to jointly train the models.

[0031] The power grid employee behavior risk scoring method based on multi-model collaboration and uncertainty perception provided in this application relates to the fields of multi-model collaborative optimization and intelligent risk assessment. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, or desktop computer, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the power grid employee behavior risk scoring method based on multi-model collaboration and uncertainty perception, but is not limited to the above forms.

[0032] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0033] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0034] Figure 1 This is an optional flowchart of the power grid employee behavior risk scoring method based on multi-model collaboration and uncertainty perception provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S106.

[0035] Step S101: Obtain multi-source heterogeneous employee behavior data and preprocess it to construct a unified behavior sample; Step S102: Input the uniform behavior sample into several different risk assessment models in parallel to obtain the risk score, model confidence and data completeness of the corresponding output of the risk assessment model; Step S103: Based on model confidence, data completeness, and historical calibration errors in similar scenarios that meet preset conditions, dynamically calculate the collaborative weights of each risk assessment model for the current unified behavior sample. Step S104: Perform a weighted fusion operation on each risk score using collaborative weights to generate an initial behavioral risk score that represents the overall risk level; Step S105: Construct a collaborative uncertainty index based on the statistical dispersion of each risk score, the cumulative deviation of each risk score from the initial behavioral risk score, and the statistical dispersion of each model confidence level. Step S106: Determine whether the collaborative uncertainty index exceeds the preset threshold. If it does, trigger the re-judgment model to perform a secondary scoring on the current unified behavior sample to output the final behavior risk score.

[0036] Steps S101 to S106 of this application embodiment address the problems of biased evaluation and data fragmentation caused by traditional single exam scores or rule-triggered evaluations by constructing unified behavioral samples from multi-source heterogeneous data. By parallel invoking multiple heterogeneous models such as rule constraints, temporal behavior, relationship graphs, and anomaly detection, and dynamically calculating collaborative weights based on confidence, completeness, and scenario-based historical errors, automatic optimization and fusion of scoring sources under different job positions and data quality scenarios are achieved, improving the stability and reliability of scoring results. Furthermore, by constructing collaborative uncertainty indicators for fusion score variance, cumulative deviation, and confidence variance, and triggering a re-judgment model to perform secondary recalculation on high-uncertainty samples, the false positive and false negative rates of boundary and conflict samples are reduced. Finally, by embedding knowledge consistency constraints and multi-loss adaptive weighting mechanisms into the model joint training, safety regulations and expert experience are injected into the data-driven model, enhancing the interpretability and convergence stability of high-risk behavior identification, and achieving accurate employee behavior risk assessment suitable for retraining screening, risk warning, and intervention task generation.

[0037] In some embodiments, step S101 may include, but is not limited to, steps S111 to S114: Step S111: Collect several data sources from the training management system, examination system, work ticket system, safety supervision system, equipment monitoring system, and human resources system. The data sources include training course completion records, theoretical exam scores and time taken, practical assessment scores and evaluation levels, work ticket issuance and execution details, violation records and deduction values, equipment operation sequence logs, access control records, job qualification certificate information, and team affiliation data. Step S112: Perform field-level cleaning on the collected data source, remove records with abnormal format values ​​and logical conflicts, and process missing fields using at least one of the following methods: missing mask marking, filling with the mean of similar samples, or filling with preset business default values. Step S113: Align different data sources with time windows according to employee identifier, timestamp, and business event type; the time windows include fixed statistical period windows, tracking and observation windows starting from preset training events, or front and back slice windows centered on preset risk events. Step S114: Based on the aligned data, extract and structure the data features to obtain a unified behavior sample. The data features include time-series features, job features, relationship features, and anomaly features. The time-series features include the trend of employee performance changes, operation frequency fluctuation patterns, and distribution of violation interval durations within a continuous time window. The job features include the matching status of job qualifications and practical tasks, and the compliance status of retraining cycles. The relationship features include graph structure features extracted based on the association graph between employees and job positions, equipment, work types, and work groups. The anomaly features include the degree of deviation of behavior patterns from the group benchmark and the identification markers of risk events.

[0038] Steps S111 to S114, as illustrated in this embodiment, collect data from multiple heterogeneous systems such as training management, examinations, work tickets, safety supervision, equipment monitoring, and human resources. They then perform field-level cleaning, time window alignment, and structured feature extraction on the data from different sources. This addresses the problem of one-sided evaluations caused by traditional methods relying solely on single exam scores or isolated violation records. It achieves a multi-dimensional, full-cycle unified representation of employee behavioral risks, providing a high-quality, standardized data foundation for subsequent multi-model collaborative scoring. For example, employee A of a power supply company's maintenance team would traditionally be evaluated based solely on their safety regulations exam score of 85, which would be considered satisfactory. However, through the unified behavioral sample constructed in this step, the system simultaneously collects: employee A has two work ticket execution timeout records within the past three months, three non-standard operation sequences shown in the equipment operation log, and one unauthorized entry into an access control area. After time alignment, this data from different systems forms a complete behavioral profile including time-series trends, rule triggers, relationship graphs, and anomaly patterns. This allows subsequent models to comprehensively assess the employee's true risk level, rather than making misjudgments based solely on exam scores.

[0039] In some embodiments, the risk assessment model in step S102 includes a rule constraint model, a time-series behavior model, a relationship graph model, and an anomaly detection model; The rule constraint model constructs a rule engine based on a preset rule base, job qualification constraints, and system triggers. It obtains the risk score of the rule constraint model by calculating the weighted score of each triggered rule and performing normalization. The model confidence of the rule constraint model is determined based on the consistency ratio or conflict ratio among the triggered rules. The data completeness of the rule constraint model is calculated based on the proportion of valid fields that meet the preset conditions of the input fields required for rule triggering. The temporal behavior model uses a bidirectional gated recurrent unit network or a Transformer encoder to encode the behavioral sequence data of employees within a continuous time window. The risk score of the temporal behavior model is output through a fully connected layer and a Sigmoid activation function. The model confidence of the temporal behavior model is determined based on the normalized entropy value of the output probability distribution. The data completeness of the temporal behavior model is calculated based on the proportion of valid observation points that meet the preset conditions in the temporal input field. The relational graph model constructs a heterogeneous graph structure based on the relationships between employees, positions, equipment, job types, and work groups. It uses any one of graph attention networks, GraphSAGE, or relational graph convolutional networks to perform message propagation and node representation learning on the heterogeneous graph structure. The risk score of the relational graph model is output through node embedding vectors and classification layers. The model confidence of the relational graph model is determined based on the normalized entropy value of the output probability. The data completeness of the relational graph model is calculated based on the proportion of valid edges and valid node attributes that meet the preset conditions in the heterogeneous graph structure. The anomaly detection model uses any one of the following algorithms—isolated forest, autoencoder network, or support vector machine—to measure the deviation of uniform behavior samples from the group baseline pattern. The anomaly score is normalized and then mapped to the risk score of the anomaly detection model. The model confidence of the anomaly detection model is determined based on the consistency of several perturbation inferences or the distance between the anomaly score and the preset anomaly threshold. The data completeness of the anomaly detection model is calculated based on the proportion of valid values ​​of the continuous feature fields required for anomaly detection that meet preset conditions.

[0040] In step S102 of some embodiments, multiple heterogeneous risk assessment models, such as rule constraint models, time-series behavior models, relationship graph models, and anomaly detection models, are invoked in parallel. This fully leverages the different data perception capabilities of different model architectures—the rule model accurately captures regulatory red lines, the time-series model perceives behavioral evolution trends, the relationship graph model mines associated risk transmission, and the anomaly detection model identifies low-frequency outlier patterns—achieving multi-perspective and complementary risk assessments of the same employee sample. Simultaneously, each model outputs a triplet of risk score, confidence level, and data completeness, providing explicit reliability criteria for subsequent dynamic fusion. Taking employee A as an example, after parallel inference, the models output: Rule-constrained model: Due to the triggering of the "unauthorized area intrusion" rule, a high-risk score of 0.82 is output, with a confidence level of 0.95 (the rule is explicit) and a completeness of 0.9. Time-series behavioral model: Analysis of the operation sequence over the past three months revealed a downward trend in operational standardization, with a risk score of 0.65, confidence level of 0.7 (due to some missing time-series data), and completeness of 0.6. Relationship graph model: It was found that employee A and employee B from the same shift who recently committed a violation have a close equipment sharing relationship. The output risk score is 0.72, confidence level is 0.8, and completeness is 0.85. Anomaly detection model: By comparing employee A's operating mode with the maintenance team's benchmark, it was found that the distribution of his operating intervals deviated from the normal mode, and the output risk score was 0.78, confidence level was 0.6 (anomaly detection itself has high uncertainty), and completeness was 0.75.

[0041] The four models each identified risk signals for employee A from four perspectives: “facts of violation,” “deteriorating trend,” “relationship transmission,” and “abnormal pattern.” None of the single models could fully cover all of these risks.

[0042] In some embodiments, step S103 may include, but is not limited to, steps S301 to S306: Step S301: Obtain the scene attribute vector of the current unified behavior sample; the scene attribute vector includes job category code, job type code, voltage level code, work group type code, time window length value, and data missing rate interval code; Step S302: Calculate the weighted similarity between the scene attribute vector of the current unified behavior sample and the scene attribute vector of each historical sample in the historical sample set, and classify the historical samples with a weighted similarity greater than the preset similarity threshold into the similar scene sample set. Step S303: Based on the sample set of similar scenarios, calculate the local historical calibration error of each risk assessment model; the local historical calibration error is the mean square error or mean absolute error between the predicted risk score and the actual risk label of the risk assessment model on each sample in the sample set of similar scenarios. Step S304: The local historical calibration error is weighted and fused with the global historical calibration error of each risk assessment model on the global validation set to obtain the scenario-based historical calibration error for the current unified behavior sample. The fusion weight is dynamically determined according to the ratio of the number of samples in the similar scenario sample set to the preset sample number threshold. When the number of similar scenario samples is insufficient, the fusion ratio of the global historical calibration error is automatically increased to the preset range. Step S305: For each risk assessment model, calculate the original weight energy value of the risk assessment model. The original weight energy value is equal to the confidence adjustment coefficient multiplied by the model confidence of the risk assessment model, plus the completeness adjustment coefficient multiplied by the data completeness of the risk assessment model, minus the error penalty coefficient multiplied by the scenario-based historical calibration error of the risk assessment model, where the confidence adjustment coefficient, completeness adjustment coefficient and error penalty coefficient are all positive real numbers. Step S306: Perform a normalized exponential function operation on the original weight energy values ​​of each risk assessment model to obtain the collaborative weights corresponding to each model. The normalized exponential function operation includes: calculating the exponential value of the original weight energy value of each risk assessment model, dividing the exponential value of the risk assessment model by the sum of the exponential values ​​of all risk assessment models, and the resulting ratio is the collaborative weight corresponding to the risk assessment model. The sum of the collaborative weights of all risk assessment models is equal to one.

[0043] In steps S301 to S306 of some embodiments, collaborative weights are dynamically calculated based on model confidence, data completeness, and historical calibration errors in similar scenarios to achieve scenario-adaptive adjustment of the fusion weights. Unlike fixed-weight averaging, this step can automatically switch to a more reliable scoring source based on the data quality of the current sample, model confidence, and the model's historical performance in similar scenarios—automatically reducing the weight when a model has low data completeness or large historical errors, effectively avoiding interference from low-quality inputs or unreliable models on the fusion results, and improving the scenario adaptability and robustness of the scoring. For example, the system calculation process for employee A sample: Time-series behavior model: Although it captured the signal of deteriorating trend, the weight was reduced from the default 0.25 to 0.12 because the data completeness was only 0.6 (some operation logs were missing) and the historical calibration error of the model in similar maintenance positions reached 0.15. Anomaly detection model: Confidence level only 0.6 (lack of confidence in its own judgment), weight reduced to 0.15; Rule-constrained model: high confidence 0.95, high completeness 0.9, historical error of only 0.03 for similar scenarios, weight increased to 0.48; Relationship graph model: medium confidence 0.8, completeness 0.85, historical error 0.07, weight 0.25.

[0044] The system automatically allocates discourse power towards rule-based and relational graph models, rather than blindly averaging or favoring time-series models with poor data quality. If the sample is replaced with a new employee sample containing complete data, the weight allocation will dynamically adjust. This adaptability is something that fixed-weight schemes cannot achieve.

[0045] In some embodiments, step S104 may include, but is not limited to, steps S401 to S404: Step S401: Multiply the risk score output by each risk assessment model by the corresponding collaborative weight of the risk assessment model to obtain the weighted risk contribution value of the risk assessment model; Step S402: Sum the weighted risk contribution values ​​of all risk assessment models and use the sum as the initial behavioral risk score; the range of the initial behavioral risk score is consistent with the range of each risk score, which is a normalized interval. Step S403: During the weighted fusion process, if the data completeness of a certain risk assessment model is lower than the preset minimum completeness threshold, the collaborative weight of the risk assessment model is set to zero, and the weight of the risk assessment model is proportionally redistributed to the other risk assessment models that meet the completeness requirements. Step S404: If the initial behavioral risk score obtained after weighted fusion exceeds the preset risk score upper limit, the initial behavioral risk score is truncated to the risk score upper limit.

[0046] In steps S401 to S404 of some embodiments, a weighted fusion of the risk scores of each model is performed using dynamic collaborative weights to generate an initial score that comprehensively represents the employee's risk level. This score retains the risk signals captured by each model from different perspectives, while suppressing noise interference from low-quality models through weight adjustment, thus achieving organic integration of multi-source scores. Simultaneously, the introduction of low-completeness elimination and score truncation protection mechanisms during the fusion process further enhances the numerical stability and business usability of the score. For example, the initial score calculation for employee A's case: Rule model: 0.82 × 0.48 = 0.394; Relationship graph model: 0.72 × 0.25 = 0.180; Anomaly detection model: 0.78 × 0.15 = 0.117; Time series model: 0.65 × 0.12 = 0.078; The weighted summation yielded an initial behavioral risk score of 0.769.

[0047] This score integrates the two most reliable risk signals: employee A's "clear violation of unauthorized entry" (rule capture) and "close association with the violator" (graph capture), while also partially incorporating supplementary information on "abnormal operating patterns" and "deteriorating trends." Using a simple arithmetic average would only yield a score of 0.74, and would be hampered by the poor timeliness and low completeness of the time-series model; relying solely on the rule model would miss two important dimensions: relationship transmission and abnormal patterns. Weighted fusion achieves a balance between information fidelity and noise suppression.

[0048] In some embodiments, step S105 may include, but is not limited to, steps S501 to S504: Step S501: Calculate the first statistical dispersion of the risk scores output by each risk assessment model; the first statistical dispersion uses variance or standard deviation as a measure to characterize the degree of disagreement between the risk judgment results of each risk assessment model for the same sample. Step S502: Calculate the cumulative deviation between each risk score and the initial behavioral risk score. The cumulative deviation is measured by the sum of the absolute values ​​of the differences between each risk score and the initial behavioral risk score, or by the square root of the sum of the squares of the differences between each risk score and the initial behavioral risk score. This measure is used to characterize the overall deviation between the scoring result of a single risk assessment model and the initial behavioral risk score. Step S503: Calculate the second statistical dispersion of the confidence scores of each model; the second statistical dispersion uses variance or standard deviation as a measure to characterize the consistency of each risk assessment model's judgment on the reliability of its own output; Step S504: Multiply the first statistical dispersion, the cumulative deviation, and the second statistical dispersion by their respective preset weighting coefficients and then perform a linear weighted summation. Use the summation result as a collaborative uncertainty index.

[0049] In steps S501 to S504 of some embodiments, a collaborative uncertainty index is constructed based on the statistical dispersion of each model's risk score, the cumulative deviation of each score from the initial score, and the statistical dispersion of each model's confidence level. This index quantifies the difficulty of sample judgment. It effectively identifies three types of high-risk scenarios: boundary samples where model scores show significant discrepancies, contradictory samples where model confidence levels are highly inconsistent, and outlier samples where scores deviate excessively from the fusion center. By explicitly modeling uncertainty, an objective decision-making basis is provided for whether to trigger a reassessment. For example, the uncertainty index is calculated using employee A as an example: The scores for the four models are [0.82, 0.65, 0.72, 0.78], with a variance of 0.0038. The cumulative absolute deviation of each score from the initial score of 0.769 is: (|0.82-0.769| + |0.65-0.769| + |0.72-0.769| + |0.78-0.769|) = 0.051 + 0.119 + 0.049 + 0.011 = 0.230; The confidence scores for the four models are [0.95, 0.70, 0.80, 0.60], with a variance of 0.017.

[0050] The weighted composite uncertainty index is 0.48, which is lower than the preset threshold of 0.5, indicating that although there are differences in the judgments of the four models, they are still within an acceptable range, and the initial score is reliable.

[0051] Comparison: For employee B, the rule-based model outputs a low-risk score of 0.15 (confidence 0.9) due to the absence of explicit violations. However, the anomaly detection model captures a highly subtle anomaly pattern and outputs a high-risk score of 0.92 (confidence 0.55). The difference between the two scores is 0.77, and the confidence variance is as high as 0.06. The uncertainty index surges to 0.82, exceeding the threshold. The system identifies this as a typical boundary conflict sample of "rule compliance but abnormal behavior," and marks it as requiring re-evaluation.

[0052] In some embodiments, step S106 may include, but is not limited to, steps S601 to S606: Step S601: Preset thresholds based on the performance metrics of the risk assessment model on the validation set; performance metrics include false positive rate and false negative rate; the thresholds are negatively correlated with the risk tolerance of the business scenario. Step S602: Compare the collaborative uncertainty index with the threshold. When the collaborative uncertainty index is less than or equal to the threshold, it indicates that the initial behavioral risk score has sufficient credibility. The initial behavioral risk score is then used as the final behavioral risk score output. Step S603: When the collaborative uncertainty index is greater than the threshold, it indicates that the risk assessment models have inconsistent judgments on the reliability of their own outputs, triggering the re-judgment model to perform a second scoring on the current unified behavior sample. Step S604: After triggering the reassessment model, a joint feature vector for the input of the reassessment model is constructed. The joint feature vector is composed of the following components in a preset order: the original feature semantic vector extracted from the unified behavior sample by the feature encoder, the risk score sequence vector formed by concatenating the risk scores output by each risk assessment model in a preset order, the confidence sequence vector formed by concatenating the model confidence scores output by each risk assessment model in the same preset order, the completeness sequence vector formed by concatenating the data completeness output by each risk assessment model in the same preset order, and the extended dimension formed by appending the collaborative uncertainty index as an independent dimension to the end of the vector. Step S605: Input the joint feature vector into the pre-built re-judgment model for forward inference. The re-judgment model is implemented using any one of the following: lightweight multilayer perceptron network, logistic regression classifier, gradient boosting tree model, or rule-based classifier. When the re-judgment model is implemented using a lightweight multilayer perceptron network, the lightweight multilayer perceptron network contains an input layer, at least one fully connected hidden layer and an output layer. The hidden layer uses a linear rectified function as the activation function, and the output layer uses a sigmoid activation function to map the output value to the same normalized risk score range as the initial behavioral risk score. When the re-judgment model is implemented using the gradient boosting tree model, the gradient boosting tree model uses decision trees as base learners and gradually optimizes the model performance by iteratively fitting the prediction residuals of the previous round. Finally, it outputs the risk score result mapped after weighted voting by multiple decision trees. Step S606: Use the re-judgment model to output the corrected final behavioral risk score. The final behavioral risk score is a continuous risk score, and its value range is consistent with the value range of the initial behavioral risk score.

[0053] In steps S601 to S606 of some embodiments, by determining whether the collaborative uncertainty index exceeds a preset threshold, a secondary actuation is triggered on high-uncertainty samples to perform a re-judgment model, thereby prioritizing the allocation of computing resources to high-risk, difficult-to-determine samples. The re-judgment model performs in-depth analysis by comprehensively considering the original features, scores of each model, confidence levels, completeness, and global information of the uncertainty index. This effectively corrects the tendency of multi-model fusion to misjudge boundary samples, reducing false positive and false negative rates. Simultaneously, the uncertainty index is linked to the business risk tolerance threshold to adjust the threshold, enabling the system to be flexibly configured to adapt to the security control requirements of different positions and scenarios. For example, in the case of employee B, the uncertainty index of 0.82 exceeds the threshold of 0.5, triggering the re-judgment model. The re-judgment model concatenates employee B's original features (operation log sequence, device interaction mode), scores and confidence sequences of the four models, data completeness information, and uncertainty index into a joint feature vector, which is then input into a lightweight multilayer perceptron for secondary analysis. The reassessment model found that although employee B did not trigger any explicit rules, their equipment operation sequence pattern was highly similar to the precursor pattern of a type of "hidden violation" in the past, and the relationship graph showed that they indirectly shared equipment with several other personnel who had previously violated regulations. Integrating these deeper signals, the reassessment model revised the initial score from 0.39 to 0.81, outputting a high-risk judgment.

[0054] Without a reassessment, employee B would be mistakenly classified as low-risk in the initial assessment and escape control; if a reassessment is called for all samples (instead of only for highly uncertain samples), a large amount of computing resources will be wasted.

[0055] In some embodiments of this application, the risk assessment model and the re-judgment model employ a multi-loss adaptive weighting mechanism during the training process. The multi-loss adaptive weighting mechanism includes: Step S107: Construct the total loss function for joint optimization; the total loss function is composed of a weighted sum of the main risk scoring loss term, the knowledge consistency constraint loss term, the confidence calibration loss term, and the uncertainty constraint loss term. Step S108: In each iteration of the training process, calculate the average collaborative uncertainty index of all samples in the current training batch, and use the ratio of the average collaborative uncertainty index of the current round to the average collaborative uncertainty index of the initial training round as a dynamic adjustment factor. Step S109: Multiply the initial weight coefficients of the knowledge consistency constraint loss term, the initial weight coefficients of the confidence calibration loss term, and the initial weight coefficients of the uncertainty constraint loss term by the dynamic adjustment factor to obtain the actual weight coefficients of each auxiliary loss term in the current training round. The total loss function is updated synchronously through backpropagation of the network parameters of the risk assessment model and the re-judgment model until the total loss function converges to the preset condition or reaches the preset training round limit.

[0056] In steps S107 to S109 of some embodiments, end-to-end joint optimization among multiple models is achieved by performing synchronous backpropagation updates on the network parameters of the risk assessment model and the re-judgment model. Unlike training each model independently and then simply combining them, joint training allows each model to perceive each other's existence and output characteristics during training, spontaneously forming a complementary division of labor—some models focus on identifying rule-based risks, while others focus on capturing patterns and anomalies. Simultaneously, the re-judgment model, as one of the optimization targets of the total loss function, is also guided by knowledge consistency and uncertainty constraints in its parameter updates, enabling the re-judgment model's judgment on boundary samples to form a closed loop with the collaborative mechanism of the front-end multiple models. After training converges, the entire scoring system becomes an organic whole, rather than a simple splicing of multiple independent components. For example, employee C of a substation maintenance team skipped a non-mandatory safety confirmation step during operation, for which there was no corresponding clause in the rule base; the rule model judged it as low risk. However, the anomaly detection model captured the subtle deviation of its operation sequence from the standard procedure, outputting medium-high risk. In independent training mode, the scores of the two models were 0.2 and 0.7 respectively, and 0.45 after fusion. An uncertainty of 0.6 triggered a re-judgment. Because the re-judgment model had not seen enough samples of this "rule blind spot" during training, it only slightly adjusted the score to 0.52, which is still considered a medium risk. In joint training mode, the re-judgment model learned through backpropagation of the total loss function during training that "when there is a large discrepancy between the rule model and the abnormal model, the judgment of the abnormal model should be given priority." It output a corrected score of 0.78 for employee C sample, accurately identifying the hidden risk.

[0057] In some embodiments, the main risk score loss term in step S107 is used to measure the degree of deviation between the final behavioral risk score and the true risk label. When the true risk label is a binary label, the main risk score loss term is calculated using weighted cross-entropy loss, assigning a first weight coefficient to positive samples and a second weight coefficient to negative samples. The first weight coefficient and the second weight coefficient are adjusted to balance the ratio of positive to negative samples. When the true risk label is a continuous risk score, the main risk score loss term is calculated using mean squared error loss or mean absolute error loss. The knowledge consistency constraint loss term is used to measure the semantic alignment between teacher representations and student representations across multiple business semantic feature blocks. Teacher representations are obtained by forward reasoning on uniform behavior samples based on a pre-defined rule knowledge base or expert experience model. Student representations are extracted from the output of the intermediate hidden layer of at least one risk assessment model. The knowledge consistency constraint loss term is obtained by calculating the negative value of the cosine similarity between teacher representations and student representations on the corresponding business semantic feature blocks and performing batch averaging. The smaller the value of the knowledge consistency constraint loss term, the higher the semantic alignment between student representations and teacher representations. The confidence calibration loss term is used to measure the consistency between the fusion confidence and the prediction accuracy index output by each risk assessment model. The fusion confidence is obtained by weighting and summing the model confidence of each risk assessment model according to the synergistic weight coefficient. The prediction accuracy index is an indicator function value. It takes a value of one when the predicted risk category corresponding to the final behavioral risk score is consistent with the actual risk label, and a value of zero otherwise. The confidence calibration loss term is calculated by squared the difference between the fusion confidence and the prediction accuracy index. The uncertainty constraint loss term is used to measure the consistency between the collaborative uncertainty index and the prediction error. The prediction error is the absolute value of the difference between the final behavioral risk score and the true risk label. The uncertainty constraint loss term is calculated by using the square of the difference between the collaborative uncertainty index and the prediction error. By minimizing the uncertainty constraint loss term, the collaborative uncertainty index reflects the level of the model's prediction error for the current sample.

[0058] Optionally, the technical fields of this application include employee behavior data processing, machine learning, multi-model collaborative scoring and uncertainty estimation, the problem objects include risk scoring and credibility assessment of multi-source heterogeneous employee behavior samples, the innovations include dynamic collaborative weighting, uncertainty-driven re-judgment, teacher-student consistency constraints and multi-loss adaptive weighting, the main inputs include training records, exam results, practical operation records, work ticket execution, violation records, equipment operation logs and team / position relationships, etc., and the main outputs include risk score, risk level, confidence level, retraining suggestions and manual review mark.

[0059] The algorithm in this application embodiment includes: 1. Construction of multi-source heterogeneous behavioral samples: The system acquires multi-source data related to employee behavioral risks, including but not limited to training records, theoretical exam scores, practical assessment results, work order execution details, safety violation records, equipment operation logs, access control or trajectory information, job qualification information, and team relationship data. The data from different sources is then cleaned, time-aligned, field-standardized, and correlated to form a unified sample of employee behavior.

[0060] At the sample level, each employee behavior sample can correspond to comprehensive behavior within a certain time window, or to a specific task, follow-up observation after a certain type of training, or behavioral slices before and after a risk event. A unified sample representation enables subsequent models to simultaneously handle temporal features, rule-constrained features, graph structure relationship features, and anomaly pattern features.

[0061] 2. Multi-model parallel scoring mechanism: For the same employee behavior sample, the system calls multiple scoring models in parallel, preferably including: rule constraint model, time-series behavior model, relationship graph model, and anomaly detection model. The rule constraint model is used to identify strong constraint violations and policy triggers; the time-series behavior model is used to identify behavioral evolution trends; the relationship graph model is used to make inferences based on the relationships between employees, positions, equipment, job types, and work groups; and the anomaly detection model is used to identify low-frequency but high-risk abnormal patterns.

[0062] Each model outputs not only a risk score, but also model confidence and data completeness, to explicitly reflect the model's applicability to the current sample and the reliability of the input data.

[0063] Preferably, the temporal behavior model uses a bidirectional gated recurrent unit (BiGRU) or a Transformer Encoder to encode the employee behavior sequence and outputs the risk probability through a fully connected layer; the relationship graph model uses a graph attention network (GAT), GraphSAGE, or a relationship graph convolutional network (R-GCN) to propagate the relationship graph of "employee-position-equipment-job type-team"; and the anomaly detection model uses an isolated forest, an autoencoder, or a support vector machine to generate anomaly scores.

[0064] 3. Reliability-based dynamic collaborative fusion: This invention does not employ a fixed-weight averaging method. Instead, it dynamically calculates the collaborative weight of the current sample based on the confidence level of each model's output, the completeness of the sample data, and the model's historical calibration error in similar scenarios. This allows for the automatic switching of more reliable scoring sources based on job type, task type, data sparsity, and scenario complexity.

[0065] 4. Collaborative uncertainty perception and reassessment: After obtaining the fusion score, the system further constructs a collaborative uncertainty index based on the dispersion of multi-model scores, the deviation between each model and the fusion score, and the dispersion of model confidence. When the collaborative uncertainty index exceeds a preset threshold, it indicates that the current sample is a complex sample with significant model divergence, ambiguous boundaries, or insufficient data quality. At this time, a re-evaluation model is triggered to perform a second scoring.

[0066] Preferably, the input of the re-judgment model includes at least the original feature encoding vector, the multi-model risk score vector, the model confidence vector, the data completeness vector, and the collaborative uncertainty index. The re-judgment model can be implemented using a lightweight multilayer perceptron, logistic regression, gradient boosting tree, or rule-enhanced classifier.

[0067] This design enables the system to prioritize the allocation of computing resources to high-risk and difficult-to-determine samples, thereby reducing false positives and false negatives and improving the platform's robustness in complex scenarios.

[0068] 5. Teacher-student consistency constraint: To prevent data-driven models from deviating from the rule-based knowledge and expert experience of power grid companies, this invention uses rule-based knowledge representations or expert model outputs as teacher representations and behavioral semantic representations from at least one data-driven model as student representations, aligning them through block-level cosine similarity or other semantic consistency losses. This mechanism can inject safety regulations, job constraints, and experiential judgments into the student model, improving the reliability of high-risk behavior identification.

[0069] Preferably, the feature blocks are divided according to business semantics, including at least two or more of the following: training and assessment feature blocks, job execution feature blocks, violation and safety incident feature blocks, job qualification feature blocks, team relationship feature blocks, and time-series trend feature blocks; or the hidden vectors of teacher representation and student representation are divided into N sub-blocks of equal length according to a preset dimension.

[0070] 6. Multi-loss adaptive weighting: During the training phase, the system jointly optimizes the risk scoring loss, consistency constraint loss, confidence calibration loss, and uncertainty constraint loss. The system dynamically adjusts the weights of each loss term based on statistical changes in collaborative uncertainty or sample scenario perturbation during training. In the early stages of training, the weights of consistency and calibration-related losses are increased to quickly stabilize the collaborative relationships among multiple models; in the later stages of training, the weights of auxiliary constraints are gradually reduced, shifting the optimization focus to the main scoring task.

[0071] Preferably, the risk scoring loss is a weighted cross-entropy loss or a mean squared error loss, the confidence calibration loss is a mean squared calibration loss based on prediction correctness, and the uncertainty constraint loss is a regression constraint loss that makes the uncertainty consistent with the true error.

[0072] 7. Methodology and Flow Overview: S1: Acquire multi-source heterogeneous employee behavior data and construct a unified behavior sample; S2: Extract time-series behavioral features, job constraint features, relationship graph features, and anomaly features; S3: Call multiple risk scoring models for parallel inference to obtain risk scores, model confidence, and data completeness; S4: Calculate dynamic collaborative weights based on confidence level, completeness, and historical calibration errors to generate an initial risk score; S5: Construct a collaborative uncertainty index and use the re-judgment model to perform secondary scoring on high uncertainty samples; S6: Implement teacher-student consistency constraints and multi-loss adaptive weighting during the training phase; S7: Output the final risk score, risk level, retraining recommendations, and manual review mark.

[0073] The following formulas in the embodiments of this application are used to describe the preferred implementation of the present invention. In specific implementation, equivalent substitutions or parameter adjustments can be made according to the deployment environment, model type and training data characteristics.

[0074] 1. Single model output: ; in, For the first A sample of employee behavior; No. The risk score output by each model; Model confidence; For data completeness. Preferably, risk score. This indicates that the m-th model is paired with the sample. The normalized output representing the risk state, with values ​​ranging from [0,1]; model confidence. This represents the estimate of the stability and correctness of the current output by the m-th model, with a value range of [0,1]; data completeness. This represents the percentage of valid fields that meet the input requirements of the m-th model, with a value range of [0,1].

[0075] Preferably, data completeness can be calculated as follows: ; in, This is the set of fields required for the m-th model. For field weights, This is an indicator function.

[0076] The meaning is: Check the d-th data field (e.g., "exam score", "practical record") in the current sample. Check if the field has a value, is not missing, and is free of garbled characters. If the field is valid, the indicator function value = 1, and the weight of that field is added to the numerator. If the field is missing or invalid, the indicator function value = 0, and the weight of that field is not included in the numerator calculation.

[0077] For rule-constrained models, the risk score can be obtained by weighting and normalizing the triggering rules, i.e.: ; in, This indicates whether rule r is triggered; its model confidence can be determined based on the conflict or consistency ratio of the applicable rules. Indicates the first The weighting coefficient of each rule.

[0078] For time-series behavioral models and relational graph models, the risk score can be given by the output probability of the Sigmoid or Softmax algorithm; the model confidence can be determined by the normalized entropy of the predicted probability, for example: ; in, Let C be the entropy function and C be the number of categories. For anomaly detection models, the risk score can be obtained by normalizing the anomaly scores, and the model confidence can be determined by the consistency of multiple perturbation inferences or the distance from the anomaly threshold. This indicates that the m-th model is paired with the sample. The output is the predicted probability distribution vector.

[0079] 2. Dynamic collaborative weights: ; in, This represents the historical calibration error of the m-th model in the same or similar scenarios as the current sample; , , This is the adjustment coefficient. The formula reflects that models with "high confidence, high completeness, and low error" have higher fusion weights. Preferably, similar scenarios are determined by at least two of the following: job category, work type, voltage level, work group type, time window length, and missing rate interval; when the above key fields are consistent, they are considered similar scenarios. 'm' specifically refers to "the model currently calculating the weights." It is the index variable in the summation formula, which means traversing all risk assessment models involved in the fusion.

[0080] Preferably, similar scenes are identified through scene vectors. Compared with historical sample scene vectors Weighted similarity Make a judgment when When the value is greater than the preset threshold η, a determination is made. The set of similar scenes belonging to the current sample. .

[0081] Historical calibration error can be calculated as follows: ; ; ; ; in, Provide historical samples with true labels or expert-labeled risk values. For the validation set, This represents the threshold for the number of local samples. Therefore, in the early stages of training or when there are insufficient local samples, the global calibration error, cross-validation error, or a preset prior error can be used as the initial value.

[0082] 3. Initial fusion risk score: ; in, For the sample Initial fusion risk score.

[0083] 4. Collaborative uncertainty indicators: ; The first term describes the dispersion of model scores, the second term describes the degree of deviation between each model and the fusion result, and the third term describes the dispersion of confidence scores. The larger the value, the more likely the current sample is to be a boundary sample or a conflict sample.

[0084] 5. Uncertainty-driven reassessment: ; in, The preset threshold; This is a re-judgment model. This mechanism is used to perform targeted optimization on highly uncertain samples. Preferably, the re-judgment model is implemented using any one of a lightweight multilayer perceptron, logistic regression, gradient boosting tree, or rule-enhanced classifier.

[0085] In a preferred implementation, the complex judgment input vector is first constructed: ; in, The final score is then output by the re-judgment model, which extracts the fused representation from the original behavioral features. ; threshold The threshold can be determined based on the false positive rate, false negative rate, or the combined F1 score on the validation set. When interpretability needs to be enhanced, the re-judgment model can also be overlaid with manual rules to correct the thresholds for key positions, high-risk consequences, and critical violations. It is the first layer weight matrix (the weight matrix of the first fully connected layer), which takes the input vector A linear transformation from the original dimension to the hidden layer dimension. Each element is a parameter learned during the model training phase using the backpropagation algorithm; It is the bias vector of the first layer (the bias vector of the first fully connected layer), which, together with the weight matrix, completes the linear transformation. This introduces translation capability into the model. It is a linear rectified activation function that performs a nonlinear mapping on the result of the linear transformation, enabling the neural network to fit complex nonlinear relationships. It is the weight matrix of the second layer (the weight matrix of the second fully connected layer). It is the second layer bias scalar (the bias term of the second fully connected layer). This is the Sigmoid function, a function that generates an S-shaped growth curve.

[0086] 6. Teacher-student consistency constraint: ; in, This represents the j-th feature block obtained from the teacher model or rule-based knowledge mapping. The student model is represented by feature blocks, where B is the batch size and N is the number of feature blocks. By minimizing this loss, the consistency between teacher and student representations can be maximized. Preferably, the feature blocks are divided according to business semantics, including at least two or more of the following: training and assessment feature blocks, job execution feature blocks, violation and safety incident feature blocks, job qualification feature blocks, team relationship feature blocks, and time-series trend feature blocks.

[0087] When the teacher model and student model output hidden vectors of length d, each hidden vector can be divided into N sub-blocks of equal length according to a preset dimension, with each sub-block having a dimension of d / N; or, semantic grouping can be completed first according to the data source, and then each group can be mapped to a unified dimension before block-level alignment.

[0088] 7. Total Loss Function: ; in, For risk scoring loss, To calibrate the loss for confidence level, Loss due to uncertainty constraints , , The weights for each loss item are defined. Preferably, when the monitoring label is a binary risk label, Weighted cross-entropy loss is used; when the supervision label is a continuous risk score, Mean squared error loss is used.

[0089] In binary label scenarios, the following options are possible: ; in, For sample weights, This is a true risk label.

[0090] The confidence level calibration loss is acceptable: ; in, To integrate confidence levels, This indicates the predicted label based on the final risk score. This is an indicator function. It is a dynamic collaborative weight, where the subscript i corresponds to the i-th employee behavior sample and the subscript m corresponds to the m-th risk assessment model (e.g., rule model, time series model, etc.).

[0091] Uncertainty-constrained loss is acceptable: ; This aims to ensure that the sample uncertainty is consistent with the actual prediction error of the sample, thereby improving the interpretability and calibrability of the uncertainty index.

[0092] 8. Dynamic weight update: ; ; ; ; in, Let be the average collaborative uncertainty during the t-th round of training. To prevent extremely small constants with a denominator of zero, this design allows for adaptive adjustment of the strength of each auxiliary loss term during training. These are the initial weighting coefficients for the knowledge consistency constraint loss term; These are the initial weights for the confidence level calibration loss; These are the initial weights for the uncertainty constraint loss; It represents the weight of the knowledge consistency constraint loss during the t-th round of training; These are the weights of the confidence calibration loss during the t-th training round; It is the weight of the uncertainty constraint loss during the t-th round of training.

[0093] Optionally, the power grid employee behavior risk scoring method based on multi-model collaboration and uncertainty perception in this application includes: acquiring multi-source heterogeneous data related to employee behavior and constructing employee behavior samples; inputting the employee behavior samples into multiple risk scoring models for parallel inference to obtain the risk score, model confidence, and data completeness output by each risk scoring model; determining the dynamic collaboration weight of each risk scoring model under the current employee behavior samples based on model confidence, data completeness, and historical calibration errors, and performing weighted fusion of each risk score to obtain an initial behavior risk score; determining a collaboration uncertainty index based on the dispersion between each risk score, the deviation between each risk score and the initial behavior risk score, and the dispersion between each model confidence; and when the collaboration uncertainty index is greater than a preset threshold, calling a re-judgment model for secondary scoring to obtain the final behavior risk score.

[0094] Multi-source heterogeneous data includes at least two or more of the following: training records, exam results, practical operation records, work order execution records, violation records, equipment operation logs, job information, and work group relationship data.

[0095] Multiple risk scoring models include two or more of the following: rule-constrained models, time-series behavior models, relationship graph models, and anomaly detection models.

[0096] The dynamic collaborative weights are determined based on the model confidence, data completeness, and historical calibration error of the same or similar scenarios as the current sample for each risk scoring model; the same or similar scenarios are determined based on at least two of the following: job category, job type, voltage level, work group type, time window length, and data missing rate interval.

[0097] Collaborative uncertainty indicators include at least one or more of the following: multi-model risk score variance, the degree of deviation of each model risk score from the initial behavioral risk score, and model confidence variance.

[0098] The training phase further includes: training the behavioral semantic representation of at least one data-driven model with consistency constraints based on the rule knowledge representation or expert knowledge representation output by the teacher model.

[0099] Consistency-constrained training is achieved by minimizing the semantic bias loss between teacher and student representations.

[0100] During the training phase, a total loss function is constructed based on risk scoring loss, consistency constraint loss, confidence calibration loss, and uncertainty constraint loss. The weights of each loss term are dynamically adjusted according to the statistical results of collaborative uncertainty during the training process. Among them, risk scoring loss is weighted cross-entropy loss or mean squared error loss, confidence calibration loss is used to constrain the consistency relationship between fusion confidence and prediction correctness, and uncertainty constraint loss is used to constrain the consistency relationship between collaborative uncertainty index and prediction error.

[0101] The following section, using a specific application example, illustrates the implementation process of this invention through the scenario of "post-training on-the-job behavior risk assessment," and provides a detailed introduction and explanation of the solutions in this application's embodiments: 1. Data Acquisition and Sample Generation: The system retrieves the target employee's training course completion status, theoretical exam scores, practical assessment scores, work order execution records, on-site violation records, equipment operation logs, and job qualification information for the most recent statistical period, and generates employee behavior samples according to a preset time window. For missing fields, the system can handle them using missing masking, experience-based imputation, or separate coding.

[0102] 2. Feature extraction and parallel model inference: The temporal behavior model encodes continuous behavior sequences and outputs behavioral evolution characteristics; the rule constraint model identifies triggers based on safety regulations and training requirements; the relationship graph model performs propagation calculations on the "employee-position-equipment-job type-work group" graph structure; and the anomaly detection model outputs anomaly scores and anomaly types. Each model outputs risk scores, model confidence, and data completeness, respectively.

[0103] 3. Dynamic fusion and reassessment: The system uses a dynamic collaborative weighting formula to calculate the weights of each model and obtain an initial risk score. Subsequently, the collaborative uncertainty index was calculated. .when When the threshold is exceeded, the reassessment model is invoked to perform a second scoring based on the original features, multi-model outputs, and uncertainty indicators, resulting in the final risk score. .

[0104] 4. Risk level output and task generation: Based on the final risk score, the system can map samples into different levels such as low risk, medium risk, high risk, and key concern. It then combines the criticality of the position and the availability of training resources to generate retraining suggestions, a list of key observation points, or manual review tasks. If a sample falls under a high-risk and highly uncertain category, manual review can be prioritized.

[0105] 5. Model Training and Updates: During the offline training phase, the system uses historical behavior labels, violation events, or expert judgment results as supervision information to jointly train multiple models and the re-judgment model. During training, the auxiliary loss weights are dynamically updated based on average collaborative uncertainty, enabling the model to prioritize learning collaborative relationships and knowledge consistency in the early stages of training, and to focus on optimizing scoring accuracy and calibration capabilities in the later stages.

[0106] 6. Extensible Implementation: This invention can also be extended to scenarios such as screening of retraining candidates, recommending recertification for key positions, profiling work team behavior risks, assessing contractor personnel behavior risks, and providing early warnings of on-site safety regulations. The model type, input data source, and specific loss form can all be adjusted according to business-side conditions.

[0107] The beneficial effects of this application are: 1. By using a unified representation of multi-source behavioral evidence, the problem of biased evaluation caused by traditional single test scores or rule-triggered results is overcome; 2. Through dynamic collaborative weighting, a more reliable scoring model can be selected for different positions, sample quality, and scenarios, thereby improving the stability of the fusion results; 3. By coordinating uncertainty indicators and reassessment mechanisms, we can focus on processing boundary samples and conflict samples, thereby reducing false alarms and false negatives; 4. By incorporating rule knowledge and expert experience into the data-driven model through teacher-student consistency constraints, the interpretability and credibility of high-risk behavior identification are improved; 5. Improve the convergence stability and generalization ability of the model in small samples, sparse samples and complex scenarios by using multi-loss adaptive weighting; 6. Emphasizing the algorithm processing flow through a methodologically independent approach helps to focus innovation on the multi-model collaborative scoring mechanism itself.

[0108] Alternative embodiments of this application: 1. Multiple model types can be replaced with any combination of two or more scoring models, not limited to rule models, time series models, graph models and anomaly detection models; 2. The calculation of dynamic collaborative weights can employ equivalent methods such as Softmax normalization, gating networks, attention mechanisms, or Bayesian fusion; 3. In addition to variance and deviation, collaborative uncertainty indicators can also include entropy, calibration bias, confidence interval width, and other characterization indicators. 4. The re-judgment model can employ a lightweight classifier, a ranking model, a large model discriminator, or a rule-enhancing model; 5. Teachers' knowledge sources can include rule bases, expert models, knowledge graph reasoning results, or historical manual review results; 6. Without changing the core algorithm process, the storage medium claim can be supplemented later according to the proxy strategy.

[0109] This application embodiment also provides a power grid employee behavior risk scoring device based on multi-model collaboration and uncertainty perception, which can implement the above method. The device includes: The first module is used to acquire multi-source heterogeneous employee behavior data and preprocess it to build a unified behavior sample. The second module is used to input uniform behavioral samples into several different risk assessment models in parallel to obtain the risk score, model confidence and data completeness of the corresponding output of the risk assessment models. The third module is used to dynamically calculate the collaborative weights of each risk assessment model for the current unified behavior sample based on model confidence, data completeness, and historical calibration errors in similar scenarios that meet preset conditions. The fourth module is used to perform a weighted fusion operation on each risk score using collaborative weights to generate an initial behavioral risk score that represents the overall risk level. The fifth module is used to construct a collaborative uncertainty index based on the statistical dispersion of each risk score, the cumulative deviation of each risk score from the initial behavioral risk score, and the statistical dispersion of each model confidence level. The sixth module is used to determine whether the collaborative uncertainty index exceeds the preset threshold. If it does, the re-judgment model is triggered to perform a second scoring on the current unified behavior sample to output the final behavior risk score.

[0110] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0111] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0112] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0113] Please see Figure 2 , Figure 2 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 201 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 202 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 202 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 202 and is called and executed by the processor 201 using the methods described in the embodiments of this application. Input / output interface 203 is used to implement information input and output; The communication interface 204 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 205 transmits information between various components of the device (e.g., processor 201, memory 202, input / output interface 203, and communication interface 204); The processor 201, memory 202, input / output interface 203 and communication interface 204 are connected to each other within the device via bus 205.

[0114] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0115] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0116] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0117] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0118] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0119] This application provides a method, device, electronic device, storage medium, and program product for power grid employee behavior risk scoring based on multi-model collaboration and uncertainty perception. The method acquires multi-source heterogeneous data related to employee behavior and constructs employee behavior samples. It performs parallel inference on the employee behavior samples based on multiple risk scoring models to obtain risk scores, model confidence, and data completeness. Dynamic collaborative weights are determined based on model confidence, data completeness, and historical calibration errors, and multiple risk scores are fused to obtain an initial behavior risk score. Further, a collaborative uncertainty index is determined based on the dispersion and deviation between the outputs of multiple models. When the collaborative uncertainty index exceeds a preset threshold, a re-judgment model is triggered for secondary scoring to obtain the final behavior risk score. During the training phase, the model is jointly optimized using teacher-student consistency constraints and a multi-loss adaptive weighting mechanism. This scheme can improve the accuracy, stability, and interpretability of power grid employee behavior risk scoring and is suitable for scenarios such as retraining screening, risk warning, and intervention task generation. This application does not simply weight and average the scores of multiple models. Instead, it uses a dynamic coordination of model reliability based on power grid employee behavior samples, constructs uncertainty by utilizing model divergence and triggers re-judgment, and combines rule knowledge injection and dynamic loss weighting to achieve interpretable, calibrable and closed-loop optimized behavioral risk scoring.

[0120] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0121] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0122] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0124] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0125] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for scoring the behavioral risk of power grid employees based on multi-model collaboration and uncertainty perception, characterized in that, The method includes the following steps: Acquire multi-source heterogeneous employee behavior data and preprocess it to construct a unified behavior sample; The unified behavior sample is input into several different risk assessment models in parallel to obtain the risk score, model confidence and data completeness of the corresponding output of the risk assessment model; Based on the model confidence, the data completeness, and the historical calibration error in similar scenarios that meet preset conditions, the collaborative weight of each risk assessment model for the current unified behavior sample is dynamically calculated. The risk scores are weighted and fused using the collaborative weights to generate an initial behavioral risk score that represents the overall risk level. Based on the statistical dispersion of each risk score, the cumulative deviation of each risk score from the initial behavioral risk score, and the statistical dispersion of each model confidence level, a collaborative uncertainty index is constructed. Determine whether the collaborative uncertainty index exceeds a preset threshold. If it does, trigger the re-judgment model to perform a secondary scoring on the current unified behavior sample to output the final behavior risk score.

2. The method according to claim 1, characterized in that, The process of acquiring and preprocessing multi-source heterogeneous employee behavior data to construct a unified behavior sample includes: Several data sources are collected from the training management system, examination system, work ticket system, safety supervision system, equipment monitoring system, and human resources system. The data sources include training course completion records, theoretical examination scores and time taken, practical assessment scores and evaluation levels, work ticket issuance and execution details, violation records and deduction values, equipment operation sequence logs, access control records, job qualification certificate information, and team affiliation data. The collected data sources are cleaned at the field level to remove records with outlier formats and logical conflicts. Missing fields are processed using at least one of the following methods: missing mask marking, filling with the mean of similar samples, or filling with preset business default values. The different data sources are aligned by time windows according to employee identifier, timestamp, and business event type; the time windows include fixed statistical period windows, tracking and observation windows starting from preset training events, or front and back slice windows centered on preset risk events. Based on the aligned data, data features are extracted and structured to obtain unified behavioral samples. These data features include time-series features, job-related features, relationship features, and anomaly features. The time-series features include the performance trends of employees within a continuous time window, the fluctuation patterns of operation frequency, and the distribution of violation interval durations. The job-related features include the matching status of job qualifications with practical tasks and the compliance status of retraining cycles. The relationship features include graph structure features extracted based on the association graph between employees and job positions, equipment, job types, and work groups. The anomaly features include the degree of deviation of behavioral patterns from the group benchmark and the identification markers of risk events.

3. The method according to claim 1, characterized in that, The risk assessment model includes a rule-constraint model, a time-series behavior model, a relationship graph model, and an anomaly detection model; The rule constraint model constructs a rule engine based on a preset rule base, job qualification constraints, and system trigger items. It obtains the risk score of the rule constraint model by calculating the weighted score of each triggered rule and performing normalization processing. The model confidence of the rule constraint model is determined based on the consistency ratio or conflict ratio among the triggered rules. The data completeness of the rule constraint model is calculated based on the proportion of valid fields that meet the preset conditions of the input fields required for rule triggering. The time-series behavior model uses a bidirectional gated recurrent unit network or a Transformer encoder to encode the behavioral sequence data of employees within a continuous time window. The risk score of the time-series behavior model is output through a fully connected layer and a Sigmoid activation function. The model confidence of the time-series behavior model is determined based on the normalized entropy value of the output probability distribution. The data completeness of the time-series behavior model is calculated based on the proportion of valid observation points in the time-series input field that meet preset conditions. The relationship graph model constructs a heterogeneous graph structure based on the relationships between employees, positions, equipment, job types, and work groups. It uses any one of graph attention networks, GraphSAGE, or relationship graph convolutional networks to perform message propagation and node representation learning on the heterogeneous graph structure. The risk score of the relationship graph model is output through node embedding vectors and a classification layer. The model confidence of the relationship graph model is determined based on the normalized entropy value of the output probability. The data completeness of the relationship graph model is calculated based on the proportion of valid edges and valid node attributes that meet preset conditions in the heterogeneous graph structure. The anomaly detection model uses any one of the following algorithms—isolated forest, autoencoder network, or support vector machine—to measure the deviation of the uniform behavior sample from the group baseline pattern. The anomaly score is normalized and then mapped to the risk score of the anomaly detection model. The model confidence of the anomaly detection model is determined based on the consistency of several perturbation inferences or the distance between the anomaly score and a preset anomaly threshold. The data completeness of the anomaly detection model is calculated based on the proportion of valid values ​​of the continuous feature fields required for anomaly detection that meet preset conditions.

4. The method according to claim 1, characterized in that, The step of dynamically calculating the collaborative weights of each risk assessment model for the current unified behavior sample based on the model confidence, the data completeness, and the historical calibration error under similar scenarios meeting preset conditions includes: Obtain the scene attribute vector of the current unified behavior sample; the scene attribute vector includes job category code, job type code, voltage level code, work group type code, time window length value, and data missing rate interval code; Calculate the weighted similarity between the scene attribute vector of the current unified behavior sample and the scene attribute vector of each historical sample in the historical sample set, and classify the historical samples with a weighted similarity greater than a preset similarity threshold into the similar scene sample set; Based on the set of similar scenario samples, the local historical calibration error of each of the risk assessment models is calculated; the local historical calibration error is the mean square error or mean absolute error between the predicted risk score and the actual risk label of the risk assessment model on each sample in the set of similar scenario samples. The local historical calibration error is weighted and fused with the global historical calibration error of each risk assessment model on the global validation set to obtain the scenario-based historical calibration error for the current unified behavior sample. The fusion weight is dynamically determined according to the ratio of the number of samples in the similar scenario sample set to a preset sample number threshold. When the number of similar scenario samples is insufficient, the fusion ratio of the global historical calibration error is automatically increased to a preset range. For each of the risk assessment models, the original weight energy value of the risk assessment model is calculated. The original weight energy value is equal to the confidence adjustment coefficient multiplied by the model confidence of the risk assessment model, plus the completeness adjustment coefficient multiplied by the data completeness of the risk assessment model, minus the error penalty coefficient multiplied by the scenario-based historical calibration error of the risk assessment model, wherein the confidence adjustment coefficient, the completeness adjustment coefficient, and the error penalty coefficient are all positive real numbers. A normalized exponential function operation is performed on the original weight energy values ​​of each risk assessment model to obtain the collaborative weights corresponding to each model. The normalized exponential function operation includes: calculating the exponential value of the original weight energy value of each risk assessment model, dividing the exponential value of the risk assessment model by the sum of the exponential values ​​of all risk assessment models, and the resulting ratio is the collaborative weight corresponding to the risk assessment model. The sum of the collaborative weights of all risk assessment models is equal to one.

5. The method according to claim 1, characterized in that, The step of performing a weighted fusion operation on each of the risk scores using the collaborative weights to generate an initial behavioral risk score representing the overall risk level includes: The risk score output by each risk assessment model is multiplied by the corresponding collaborative weight of the risk assessment model to obtain the weighted risk contribution value of the risk assessment model. The weighted risk contribution values ​​of all the risk assessment models are summed up, and the summation result is used as the initial behavioral risk score; the value range of the initial behavioral risk score is consistent with the value range of each of the risk scores, which are all normalized intervals. During the weighted fusion process, if the data completeness of a certain risk assessment model is lower than the preset minimum completeness threshold, the collaborative weight of the risk assessment model is set to zero, and the weight of the risk assessment model is proportionally redistributed to the other risk assessment models that meet the completeness requirements. If the initial behavioral risk score obtained after weighted fusion exceeds the preset risk score upper limit, the initial behavioral risk score will be truncated to the risk score upper limit.

6. The method according to claim 1, characterized in that, The collaborative uncertainty index is constructed based on the statistical dispersion of each risk score, the cumulative deviation of each risk score from the initial behavioral risk score, and the statistical dispersion of each model confidence level, including: Calculate the first statistical dispersion of the risk scores output by each of the risk assessment models; the first statistical dispersion uses variance or standard deviation as a measure to characterize the degree of disagreement between the risk assessment models for the same sample. Calculate the cumulative deviation between each of the risk scores and the initial behavioral risk score; the cumulative deviation is measured by the sum of the absolute values ​​of the differences between each of the risk scores and the initial behavioral risk score, or by the square root of the sum of the squares of the differences between each of the risk scores and the initial behavioral risk score, to characterize the overall deviation between the scoring result of a single risk assessment model and the initial behavioral risk score. Calculate the second statistical dispersion of the confidence scores of each model; the second statistical dispersion uses variance or standard deviation as a measure to characterize the consistency of each risk assessment model in judging the reliability of its own output; The first statistical dispersion, the cumulative deviation, and the second statistical dispersion are each multiplied by their respective preset weighting coefficients and then linearly weighted and summed. The summation result is used as the collaborative uncertainty index.

7. The method according to claim 1, characterized in that, The step of determining whether the collaborative uncertainty index exceeds a preset threshold, and if it does, triggering a re-evaluation model to perform a secondary scoring on the current unified behavior sample to output a final behavior risk score, includes: A threshold is preset based on the performance indicators of the risk assessment model on the validation set; the performance indicators include false positive rate and false negative rate; the threshold is negatively correlated with the risk tolerance of the business scenario. The collaborative uncertainty index is compared with the threshold. When the collaborative uncertainty index is less than or equal to the threshold, it indicates that the initial behavioral risk score has sufficient credibility, and the initial behavioral risk score is output as the final behavioral risk score. When the collaborative uncertainty index is greater than the threshold, it indicates that the risk assessment models have inconsistent judgments on the reliability of their own outputs, triggering the re-judgment model to perform a second scoring on the current unified behavior sample. After triggering the reassessment model, a joint feature vector is constructed for input to the reassessment model. The joint feature vector is composed of the following components concatenated in a preset order: the original feature semantic vector extracted by the feature encoder from the unified behavior sample; the risk score sequence vector formed by concatenating the risk scores output by each of the risk assessment models in a preset model order; the confidence score sequence vector formed by concatenating the model confidence scores output by each of the risk assessment models in the same preset model order; the data completeness sequence vector formed by concatenating the data completeness output by each of the risk assessment models in the same preset model order; and an extended dimension formed by appending the collaborative uncertainty index as an independent dimension to the end of the vector. The joint feature vector is input into the pre-built re-judgment model for forward inference. The re-judgment model is implemented using any one of the following: a lightweight multilayer perceptron network, a logistic regression classifier, a gradient boosting tree model, or a rule-based classifier. When the re-judgment model is implemented using a lightweight multilayer perceptron network, the lightweight multilayer perceptron network includes an input layer, at least one fully connected hidden layer and an output layer. The hidden layer uses a linear rectified function as the activation function, and the output layer uses a sigmoid activation function to map the output value to the same normalized risk score range as the initial behavioral risk score. When the re-judgment model is implemented using a gradient boosting tree model, the gradient boosting tree model uses a decision tree as a base learner and gradually optimizes the model performance by iteratively fitting the prediction residuals of the previous round, and finally outputs the risk score result mapped after weighted voting by multiple decision trees. The revised final behavioral risk score is output using the aforementioned reassessment model. The final behavioral risk score is a continuous risk score, and its value range is consistent with that of the initial behavioral risk score.

8. The method according to claim 1, characterized in that, The risk assessment model and the re-judgment model employ a multi-loss adaptive weighting mechanism during the training process. This multi-loss adaptive weighting mechanism includes: Construct a joint optimization total loss function; the total loss function is composed of a weighted sum of the main risk scoring loss term, the knowledge consistency constraint loss term, the confidence calibration loss term, and the uncertainty constraint loss term. In each iteration of the training process, the average collaborative uncertainty index of all samples in the current training batch is calculated, and the ratio of the average collaborative uncertainty index of the current round to the average collaborative uncertainty index of the initial training round is used as a dynamic adjustment factor. The initial weight coefficients of the knowledge consistency constraint loss term, the initial weight coefficients of the confidence calibration loss term, and the initial weight coefficients of the uncertainty constraint loss term are multiplied by the dynamic adjustment factor to obtain the actual weight coefficients of each auxiliary loss term in the current training round. The total loss function is synchronously backpropagated to update the network parameters of the risk assessment model and the network parameters of the re-judgment model until the total loss function converges to a preset condition or reaches the upper limit of a preset training round.

9. The method according to claim 8, characterized in that, The main risk score loss term is used to measure the degree of deviation between the final behavioral risk score and the true risk label. When the true risk label is a binary label, the main risk score loss term is calculated using weighted cross-entropy loss. A first weight coefficient is assigned to positive samples and a second weight coefficient is assigned to negative samples. The first weight coefficient and the second weight coefficient are adjusted to balance the ratio of positive to negative samples. When the true risk label is a continuous risk score, the main risk score loss item is calculated using mean squared error loss or mean absolute error loss. The knowledge consistency constraint loss term is used to measure the semantic alignment between teacher representations and student representations across multiple business semantic feature blocks. The teacher representations are obtained by forward reasoning of the unified behavior samples based on a preset rule knowledge base or expert experience model. The student representations are extracted from the output of the intermediate hidden layer of at least one of the risk assessment models. The knowledge consistency constraint loss term is obtained by calculating the negative value of the cosine similarity between the teacher representation and the student representation on the corresponding business semantic feature blocks and performing batch averaging. The smaller the value of the knowledge consistency constraint loss term, the higher the semantic alignment between the student representation and the teacher representation. The confidence calibration loss term is used to measure the consistency between the fusion confidence and the prediction accuracy index output by each of the risk assessment models. The fusion confidence is obtained by weighting and summing the model confidence of each of the risk assessment models according to the collaborative weight coefficient. The prediction accuracy index is an indicator function value, which takes a value of one when the predicted risk category corresponding to the final behavioral risk score is consistent with the actual risk label, and a value of zero otherwise. The confidence calibration loss term is calculated by squared the difference between the fusion confidence and the prediction accuracy index. The uncertainty constraint loss term is used to measure the consistency between the collaborative uncertainty index and the prediction error. The prediction error is the absolute value of the difference between the final behavioral risk score and the true risk label. The uncertainty constraint loss term is calculated by squared the difference between the collaborative uncertainty index and the prediction error. By minimizing the uncertainty constraint loss term, the collaborative uncertainty index reflects the prediction error level of the model for the current sample.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 9.

Images