A power material supplier auditing and evaluating method and device based on performance

Abstract

The application discloses a power material supplier auditing and evaluating method and device based on performance, which comprises the following steps: collecting full data of the performance of the power material supplier, dividing the data and processing the data by standardization feature engineering to obtain a space-time feature data set and a multi-modal image data set; combining a multi-branch multi-modal neural network and a double-model cross-validation strategy to generate image checking results and space-time checking results; reasoning and analyzing the image checking results and the space-time checking results based on checking reasoning rules to determine primary performance checking results; integrating the performance points according to the performance point rules of the supplier to determine the auditing and evaluating results of the power material supplier, realizing the full-process automation of auditing and evaluation, reducing the cost of manual auditing, improving the performance anomaly recognition rate, and then giving an early warning according to the evaluation results to promote the efficient cooperation of all parties in the power material supply chain and ensure the completion of the project on schedule.

Description

Technical Field

[0001] This invention belongs to the technical field of power data analysis, specifically relating to a method and device for auditing and evaluating power material suppliers based on performance behavior. Background Technology

[0002] As the scale of power grid construction gradually increases, the demand for power-related materials is also showing an upward trend. The supply of power-related materials directly affects supporting transmission lines, distributing conductor loads, improving transmission efficiency, and ensuring power grid stability. It is a key link in the safe power supply, stable operation, and integration of modern technologies in the power system.

[0003] In the process of supplying power materials, although power grid companies and suppliers establish corresponding performance agreements through contracts, the production capacity of materials and the actual needs of projects are subject to certain uncertainties over time, which affects the supplier's ability to perform its obligations. For example, suppliers face various internal risks during contract performance, such as corporate financial credit, production and operation, and transaction credit. At the same time, they are also affected by external factors such as raw material supply and weather, which greatly increase the likelihood of default.

[0004] Patent application CN121146494A discloses a supplier performance risk assessment system based on multi-source data fusion, comprising: a static data feature extraction module that uses a static feature extraction model based on an MLP architecture to generate supplier profile features from supplier status data; a dynamic data feature extraction module that uses a dynamic feature extraction model based on an LSTM architecture to generate supplier production status features from supplier production data; a basic risk identification module that uses the analytic hierarchy process (AHP) to generate a risk prior index; and a fusion evaluation module that uses a risk fusion mapping model based on gating mechanisms, attention mechanisms, and a fully connected network architecture to generate a multi-dimensional supplier performance risk assessment value. This invention accurately captures risk signs by dynamically allocating the importance of long-term static supplier features, dynamic supplier features, and the weights of the AHP with prior experience through a fusion model.

[0005] Currently, including the aforementioned technologies, the evaluation of supplier performance primarily relies on analyzing historical performance data. This lack of timely assessment of supplier capabilities during the current power material supply process leads to issues such as information lag, difficulty in identifying supplier violations, and a lack of dynamic evaluation mechanisms. Therefore, the urgent issue to be addressed is how to audit and evaluate the performance of power material suppliers, and then issue early warnings to project construction units based on the evaluation results. This will promote efficient collaboration among all parties in the power material supply chain and ensure timely project completion. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention provides a method and apparatus for auditing and evaluating power material suppliers based on performance behavior. The method includes: collecting full data on the performance process of power material suppliers; dividing the full data and performing standardized feature engineering processing to obtain a spatiotemporal feature dataset and a multimodal image dataset; using a multi-branch multimodal neural network to verify the image data in the multimodal image dataset and generate image verification results; analyzing and verifying the deviation of spatiotemporal features in the spatiotemporal feature dataset through a dual-model cross-validation strategy and providing spatiotemporal verification results; performing inference analysis on the image verification results and spatiotemporal verification results based on verification inference rules to determine the initial performance verification results; and performing integral fusion on the initial performance verification results of the power material supplier according to the supplier performance integration rules to determine the audit evaluation result of the power material supplier.

[0007] By employing a multi-branch, multi-modal neural network, image data is verified to generate image verification results. A dual-model cross-validation strategy is used to analyze and verify the deviations in spatiotemporal features, providing spatiotemporal verification results. Verification inference rules are used to analyze the verification results and determine the initial performance verification results. Integrating the supplier's performance scoring rules, the audit evaluation results for the power material supplier are determined, achieving full automation of the audit and evaluation process, reducing manual audit costs, and improving the identification rate of performance anomalies. Simultaneously, the results are used to iteratively optimize the thresholds of each detection model, continuously improving the adaptability of the audit evaluation method.

[0008] In a first aspect, the present invention provides a method for auditing and evaluating power material suppliers based on their performance behavior, specifically including the following steps: Collect full data on the performance of power supply providers during the contract fulfillment process; The entire dataset is divided into data segments and standardized feature engineering processes to obtain a spatiotemporal feature dataset and a multimodal image dataset. A multi-branch multimodal neural network is used to verify the image data in the multimodal image dataset and generate image verification results. By employing a dual-model cross-validation strategy, the deviation of spatiotemporal features in the spatiotemporal feature dataset is analyzed and verified, and spatiotemporal verification results are provided. Based on the verification reasoning rules, the image verification results and spatiotemporal verification results are reasoned and analyzed to determine the preliminary performance verification results; According to the supplier performance scoring rules, the initial performance verification results of the power material supplier are integrated to determine the audit evaluation result of the power material supplier.

[0009] Furthermore, a multi-branch multimodal neural network is employed to verify the image features in the multimodal image feature set, generating image verification results, specifically including: By using the image coding branch in a multi-branch multimodal neural network, image features are analyzed and core features are extracted to obtain a visual feature vector; Based on the metadata encoding branch in a multi-branch multimodal neural network, the metadata in the multimodal image dataset is encoded to determine the standard feature vector; Based on the cross-attention fusion mechanism in multi-branch multimodal neural networks, the visual feature vector and standard feature vector are correlated, matched, and mapped to generate image verification results.

[0010] Furthermore, based on the cross-attention fusion mechanism in multi-branch multimodal neural networks, the visual feature vector and standard feature vector are correlated, matched, and mapped to generate image verification results, specifically including: By performing linear transformations on the visual feature vector and the standard feature vector respectively, query vector and key vector are generated; A value vector is generated by performing a linear transformation on the vertically concatenated vector of the visual feature vector and the standard feature vector; Based on the query vector and key vector, calculate the cross-attention weight, analyze the correlation matching degree between the visual feature vector and the standard feature vector, and give the correlation matching parameters; The Hadamard product operation is performed on the associated matching parameters and value vectors, and the output is mapped by the activation function to obtain the image verification result.

[0011] Furthermore, using a dual-model cross-validation strategy, the bias of spatiotemporal features in the spatiotemporal feature dataset is analyzed and verified, and spatiotemporal verification results are provided, specifically including: Based on the Bayesian network model, the deviation of spatiotemporal features in the spatiotemporal feature dataset is analyzed, and the first spatiotemporal deviation probability is given. Based on the Hidden Markov Model, the spatiotemporal features in the spatiotemporal feature dataset are observed, and the second spatiotemporal deviation probability is given. Based on the cross-validation rules, the first spatiotemporal deviation probability and the second spatiotemporal deviation probability are fused and judged to determine the probability fusion strategy; By employing a probability fusion strategy, dynamic probability weights are generated, and the first and second spatiotemporal deviation probabilities are combined to provide spatiotemporal verification results.

[0012] Furthermore, through a probability fusion strategy, dynamic probability weights are generated, and by combining the first and second spatiotemporal deviation probabilities, spatiotemporal verification results are given, specifically including: The first spatiotemporal deviation probability and the second spatiotemporal deviation probability are normalized respectively to obtain the first standard deviation probability and the second standard deviation probability; Based on the scenario risk coefficient and combined with the dynamic basic weight relationship, the basic weights corresponding to the first standard deviation probability and the second standard deviation probability are adjusted to give the first basic weight and the second basic weight. By using the supply compliance correction coefficient, the first basic weight and the second basic weight are corrected respectively to generate the first dynamic weight and the second dynamic weight. Based on the first dynamic weight and the second dynamic weight, and combined with the first standard deviation probability and the second standard deviation probability, the target spatiotemporal deviation probability is determined; Based on the scenario risk coefficient, a dynamic deviation threshold is determined, and combined with the target spatiotemporal deviation probability, the spatiotemporal verification result is obtained.

[0013] Furthermore, based on the verification reasoning rules, the image verification results and spatiotemporal verification results are analyzed to determine the preliminary performance verification results, specifically including: Based on the power supply scenario, the scenario entropy under different verification rules is analyzed, and the preset initial weights corresponding to different verification rules are adjusted to determine the basic weights of the scenario under different verification rules. Based on the image verification results and spatiotemporal verification results, the scene base weights under different verification rules are nonlinearly attenuated and corrected, and the scene correction weights under different verification rules are given. Based on the conflict of different verification rules, the scenario correction weights are adjusted to generate verification rule weights corresponding to different verification rules; By triggering the indicator function according to the rule, combining the rule importance coefficient, and integrating the verification rule weights corresponding to different verification rules, the performance verification score is determined and the preliminary performance verification result is generated.

[0014] Furthermore, based on the supplier performance scoring rules, the initial performance verification results of the power material supplier are integrated to determine the audit evaluation result of the power material supplier, specifically including: Based on the information entropy of each verification rule, the initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule; Based on the initial performance verification results of the power supply provider, and combined with the global rule weights, global gain items and global penalty items are determined; By integrating global gain and global penalty terms, the audit evaluation results of power material suppliers are determined.

[0015] Furthermore, based on the information entropy of each verification rule, the initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule, specifically including: By analyzing the discriminative power of each verification rule, the information entropy of each verification rule is given; Calculate the objective weight of each verification rule by combining the sum of the information entropy of each verification rule; By using a weight balancing coefficient, the objective weights and initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule.

[0016] Furthermore, by integrating global gain and global penalty terms, the audit evaluation results for power material suppliers are determined, specifically including: Summing the global gain and global penalty terms yields the total integral change in the current performance process. Update the number of continuously compliant orders based on the total change in points during the current fulfillment process; The reward value is determined based on the updated number of consecutive compliant orders; Based on the audit evaluation results of the previous performance process, and by integrating the total change in points and reward value of the current performance process, the audit evaluation results of the power material supplier after the current performance process are determined.

[0017] Secondly, the present invention also provides an audit and evaluation device for power material suppliers based on performance behavior, employing any of the above-mentioned audit and evaluation methods for power material suppliers based on performance behavior, including: The data acquisition module is used to collect all data related to the performance of power supply contracts. The data processing module is used to partition the full dataset and perform standardized feature engineering to obtain spatiotemporal feature datasets and multimodal image datasets. The first verification module is used to verify the image data in the multimodal image dataset using a multi-branch multimodal neural network and generate image verification results. The second verification module is used to analyze and verify the deviation of spatiotemporal features in the spatiotemporal feature dataset through a dual-model cross-validation strategy, and to provide spatiotemporal verification results. The verification fusion module is used to perform reasoning analysis on image verification results and spatiotemporal verification results based on verification reasoning rules to determine the initial performance verification results; The points fusion module is used to integrate the initial performance verification results of power material suppliers according to the supplier performance points rules, and determine the audit evaluation results of power material suppliers.

[0018] The present invention provides a method and apparatus for auditing and evaluating power material suppliers based on performance behavior, which has at least the following beneficial effects: (1) By using a multi-branch multimodal neural network, image data is checked to generate image check results; combined with a dual-model cross-validation strategy, the deviation of spatiotemporal features is analyzed and verified to give spatiotemporal check results. Through check inference rules, the check results are analyzed and the primary performance check results are determined; combined with the supplier's performance integration rules, the audit evaluation results of the power material supplier are determined, realizing the automation of the entire audit and evaluation process, reducing the cost of manual audit, improving the performance anomaly identification rate, and continuously improving the adaptability of the audit evaluation method by iteratively optimizing the thresholds of each detection model through result feedback.

[0019] (2) By performing the Hadamard product operation, the attention weight (i.e. the association matching parameter) is multiplied element by element with the value vector corresponding to the vertical splicing vector to highlight the high correlation feature and suppress irrelevant background interference, and finally output the image verification result.

[0020] (3) By introducing a dynamic basic weight relationship, and combining the complexity of material categories with the risk level of delivery scenarios, the corresponding scenario risk coefficients are determined, and the first and second basic weights are determined to achieve scenario-based weight adaptation. By setting a dynamic deviation threshold, the threshold is appropriately lowered in high-risk scenarios to improve the detection rate of core material anomalies, while the threshold remains unchanged in low-risk scenarios to reduce false alarms.

[0021] (4) By cross-coupling verification of Bayesian network model and hidden Markov model, combined with adaptive weighted fusion mechanism, the model is first judged independently, then cross-verified to eliminate single model misjudgment, and finally dynamically weighted to output the spatiotemporal deviation probability of the target, so as to realize the complementary detection of static spatiotemporal probability and dynamic trajectory sequence and improve the accuracy of spatiotemporal anomaly detection.

[0022] (5) Adopt composite weighting logic to determine the weight of the verification rules, realize the real-time dynamic adjustment of the rule weight according to the scenario, the confidence level of the anomaly, and the cost of the conflict, and construct a multi-layer dynamic rule weight mechanism of scenario entropy driving + anomaly confidence level decay + rule conflict cost + dynamic priority scheduling to solve the problems of fixed weight, rigid conflict resolution, and rigid rule priority in traditional rule engines. Attached Figure Description

[0023] Figure 1 A flowchart of a power material supplier audit and evaluation method based on performance behavior provided in an embodiment of the present invention; Figure 2 A flowchart for determining the initial performance verification result provided in an embodiment of the present invention; Figure 3 A flowchart for determining audit evaluation results provided in an embodiment of the present invention; Figure 4This is a structural block diagram of a power material supplier audit and evaluation device based on performance behavior, provided in an embodiment of the present invention.

[0024] Among them, 201 is the data acquisition module; 202 is the data processing module; 203 is the first verification module; 204 is the second verification module; 205 is the verification fusion module; and 206 is the points fusion module. Detailed Implementation

[0025] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0027] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0028] Power grid companies face heavy construction tasks, involving a large number of material suppliers with diverse types and characteristics, leading to various risks in the supply of power materials. As the responsible parties for ensuring the supply and quality of materials, the timely fulfillment of contracts by suppliers is a major source of risk for the timely delivery of power materials, thus affecting the stability and reliability of the power grid's safe operation. Since suppliers are subject to various influences in raw material production and transportation, how to audit and evaluate the performance of power material suppliers to lay the foundation for timely project completion is a problem that needs to be solved.

[0029] Based on this, the present invention provides a method for auditing and evaluating power material suppliers based on performance behavior, comprising: collecting full data of the power material supplier's performance process; performing data partitioning and standardized feature engineering on the full data to obtain a spatiotemporal feature dataset and a multimodal image dataset; using a multi-branch multimodal neural network to verify the image data in the multimodal image dataset and generate image verification results; analyzing and verifying the deviation of spatiotemporal features in the spatiotemporal feature dataset through a dual-model cross-validation strategy and providing spatiotemporal verification results; performing inference analysis on the image verification results and spatiotemporal verification results based on verification inference rules to determine the initial performance verification results; and performing integral fusion on the initial performance verification results of the power material supplier according to the supplier performance integration rules to determine the audit evaluation results of the power material supplier.

[0030] By employing a multi-branch, multi-modal neural network, image data is verified to generate image verification results. A dual-model cross-validation strategy is used to analyze and verify the deviations in spatiotemporal features, providing spatiotemporal verification results. Verification inference rules are used to analyze the verification results and determine the initial performance verification results. Integrating the supplier's performance scoring rules, the audit evaluation results for the power material supplier are determined, achieving full automation of the audit and evaluation process, reducing manual audit costs, and improving the identification rate of performance anomalies. Simultaneously, the results are used to iteratively optimize the thresholds of each detection model, continuously improving the adaptability of the audit evaluation method.

[0031] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for auditing and evaluating power material suppliers based on performance behavior. The specific steps are as follows: S101: Collect full data on the performance of power supply providers.

[0032] For the entire fulfillment process of power equipment (such as 10kV transformers, high-voltage cables, and other core equipment), data barriers between multiple systems, including the power grid's internal ERP system, logistics management platform, acceptance management system, and supplier self-reporting portal, are broken down. Comprehensive core dynamic data is collected, covering key nodes throughout the supplier's fulfillment lifecycle. This comprehensive data includes basic order data, dynamic logistics tracking data, electronic image and document data, time-sensitive milestone data, and historical fulfillment data. The order data includes: basic order data such as purchase order number, power material category, technical parameters, contract quantity, delivery time limit, receiving area, target warehouse location, batch division, and contract compliance requirements; dynamic logistics data such as GPS / BeiDou real-time positioning of the transport vehicle, driving timestamp, unique ID of the on-board equipment, deviation between planned and actual routes, intermediate stops and durations, transportation mileage, and real-time vehicle speed; electronic data of documents such as photos of the shipping site, compliant photos of loading materials, photos of arrival and acceptance, product certificates, electronic delivery notes, quality inspection reports, photos of electronic seals, and document submission receipts; timeliness data such as delivery confirmation time, vehicle departure time, transit time, actual delivery time, acceptance initiation and completion time, and time of archiving of all documents; and historical performance data such as the supplier's past performance compliance rate, timeliness achievement rate, historical anomaly records, rectification closure status, and frequency of continuous compliant performance.

[0033] S102: Perform data partitioning and standardized feature engineering on the full dataset to obtain a spatiotemporal feature dataset and a multimodal image dataset.

[0034] By performing standardized feature engineering on the collected full data, including deduplication, noise reduction, missing value imputation, and normalization preprocessing, differences in units and dimensions are eliminated. After processing, spatiotemporal feature datasets and multimodal image datasets are extracted, transforming the data from scattered and disordered to standardized and ordered. The spatiotemporal feature dataset includes delivery-shipment time intervals, transportation mileage deviations, location latitude and longitude, regional delivery radius, vehicle speed ranges, dwell time, route deviation, and timeliness compliance rate, while the multimodal image dataset includes visual features of material images, image texture features, shooting timestamps, unique identifiers of shooting equipment, material category labels, delivery method labels, regional geographic labels, and document integrity labels.

[0035] By collecting multi-source dynamic heterogeneous data generated by suppliers throughout the entire process of fulfilling power material supply contracts, data cleaning, feature extraction, and standardization are completed. This provides a high-quality and quantifiable data foundation for subsequent multimodal detection, rule reasoning, and integral evaluation, eliminating audit bias and evaluation distortion caused by missing, distorted, or heterogeneous data from the source, and achieving full-dimensional traceability and analyzability of contract fulfillment behavior.

[0036] S103: A multi-branch, multi-modal neural network is used to verify the image data in the multi-modal image dataset and generate image verification results.

[0037] Specifically, a multi-branch, multi-modal neural network is used to verify image features in the multi-modal image feature set and generate image verification results, including: By using the image coding branch in a multi-branch multimodal neural network, image features are analyzed and core features are extracted to obtain a visual feature vector; Based on the metadata encoding branch in a multi-branch multimodal neural network, the metadata in the multimodal image dataset is encoded to determine the standard feature vector; Based on the cross-attention fusion mechanism in multi-branch multimodal neural networks, the visual feature vector and standard feature vector are correlated, matched, and mapped to generate image verification results.

[0038] For the delivery and acceptance image vouchers uploaded by suppliers, a multi-branch multimodal neural network is used to achieve dual closed-loop verification of image content and shooting metadata, preventing behaviors such as reuse of old images, shooting in different locations, and image tampering. The image encoding branch and metadata encoding branch are pre-trained using a dedicated dataset for power materials, and special optimizations are made for core features such as nameplates, lead seals, material appearance, and on-site environment.

[0039] In one specific implementation, a ResNet50 deep neural network is used as the image encoding branch. The bottom general feature layer is frozen, and the top classification layer is fine-tuned. Core visual features such as material appearance, nameplate parameters, seal numbers, loading / unloading environment, and warehouse markings are extracted from the multimodal image feature set. Images not related to this fulfillment, mismatched categories, or blurry / invalid images are filtered out, and a visual feature vector is output. In other examples, other models can be used to implement the image encoding and visual feature extraction process; this is not limited. A multilayer perceptron is used as the metadata encoding branch to convert metadata such as shooting time, device ID, latitude and longitude, and shooting angle from the multimodal image feature set into a standard feature vector. This verifies the consistency between the shooting time and space, device information, and the order fulfillment information, filtering out invalid documents with misaligned time and space or those taken by non-reported equipment. In other examples, other models can be used to implement the metadata conversion process; this is not limited.

[0040] Abandoning the traditional simple feature splicing, a cross-attention fusion mechanism for power material scenarios is constructed. It dynamically and weightedly fuses dual-modal features (visual feature vector and standard feature vector), strengthens the weight of core features, weakens background interference, realizes deep correlation verification between image content and metadata, and accurately identifies various image forgery behaviors.

[0041] Furthermore, based on the cross-attention fusion mechanism in multi-branch multimodal neural networks, the visual feature vector and standard feature vector are correlated, matched, and mapped to generate image verification results, specifically including: By performing linear transformations on the visual feature vector and the standard feature vector respectively, query vector and key vector are generated; A value vector is generated by performing a linear transformation on the vertically concatenated vector of the visual feature vector and the standard feature vector; Based on the query vector and key vector, calculate the cross-attention weight, analyze the correlation matching degree between the visual feature vector and the standard feature vector, and give the correlation matching parameters; The Hadamard product operation is performed on the associated matching parameters and value vectors, and the output is mapped by the activation function to obtain the image verification result.

[0042] Specifically, a query vector is generated based on the visual feature vector, combined with the query weight matrix and query bias term; a key vector is generated based on the standard feature vector, combined with the key weight matrix and key bias term; and a value vector is generated based on the vertical concatenation vector of the visual feature vector and the standard feature vector, combined with the value weight matrix and value bias term, as specifically represented as:

[0043] Where Q is the query vector, W Q To query the weight matrix, b Q To query the bias term, V visual Let K be the visual feature vector, and W be the key vector. K Let b be the key weight matrix. K For key bias terms, V sta V is the standard feature vector, and W is the value vector. V For the value weight matrix, b V For the value bias term, [V visual ;V sta [ ] is a vertically concatenated vector of visual feature vectors and standard feature vectors. The query weight matrix, key weight matrix, and value weight matrix are trainable weight matrices.

[0044] In one specific implementation, the visual feature vector encompasses core visual information such as the appearance of the materials, nameplates, seals, and the on-site environment; the standard feature vector includes spatiotemporal and equipment information such as shooting time, equipment ID, latitude and longitude, and shooting angle. A linear transformation maps the visual feature vector and the standard feature vector to the same dimensional space, generating query vectors, key vectors, and value vectors, providing a data foundation for subsequent feature association.

[0045] The associated matching parameters are specifically represented as follows:

[0046] Where α is the association matching parameter, softmax() is the activation function, Q is the query vector, and K is the match parameter. T d is the transpose of the key vector. k is the dimension of the key vector.

[0047] The association degree between the visual feature vector and the standard feature vector is calculated by performing an inner product operation on the query vector and the key vector. The result is then normalized to weight values ​​between 0 and 1 using the Softmax function, forming the association matching parameters. (Scaling factor) This is used to avoid the gradient vanishing of the Softmax function due to excessively large inner product values ​​of high-dimensional features, thus achieving adaptive weighting of core features.

[0048] The image verification results are specifically shown as follows:

[0049] Among them, P fake The image verification result represents the probability of image forgery, σ() is the activation function, α is the association matching parameter, and V is the value vector. For Hadama accumulation.

[0050] The Sigmoid activation function maps the output to the [0,1] interval. The Hadamard product operation multiplies the attention weights (i.e., the association matching parameters) element-wise with the value vector corresponding to the vertical concatenation vector, highlighting highly correlated features and suppressing irrelevant background interference. The final output is the image verification result; the closer the value is to 1, the higher the probability of image forgery. Combined with a preset forgery threshold, if the image verification result reaches the forgery threshold, it is directly determined that the image is forged, triggering an early warning and blocking subsequent fulfillment processes. In a specific example, the forgery threshold is 0.85.

[0051] S104: Using a dual-model cross-validation strategy, the deviation of spatiotemporal features in the spatiotemporal feature dataset is analyzed and verified, and spatiotemporal verification results are given.

[0052] Specifically, a dual-model cross-validation strategy is used to analyze and verify the bias of spatiotemporal features in the spatiotemporal feature dataset, and spatiotemporal verification results are provided, including: Based on the Bayesian network model, the deviation of spatiotemporal features in the spatiotemporal feature dataset is analyzed, and the first spatiotemporal deviation probability is given. Based on the Hidden Markov Model, the spatiotemporal features in the spatiotemporal feature dataset are observed, and the second spatiotemporal deviation probability is given. Based on the cross-validation rules, the first spatiotemporal deviation probability and the second spatiotemporal deviation probability are fused and judged to determine the probability fusion strategy; By employing a probability fusion strategy, dynamic probability weights are generated, and the first and second spatiotemporal deviation probabilities are combined to provide spatiotemporal verification results.

[0053] In one specific implementation, a Bayesian network model is used, with material category, delivery method, and receiving / sending area as prior conditions. Based on historical compliance data of power materials, a joint distribution of spatiotemporal features is fitted to calculate the spatiotemporal deviation probability, and the first spatiotemporal deviation probability is given, specifically expressed as follows:

[0054] Among them, P st,1 Let P be the first spatiotemporal deviation probability, t be the delivery time, d be the transportation distance, v be the average vehicle speed, C be the prior condition set, and P be the distance between the two points. st,1 ∈[0,1], the higher the value of the first spatiotemporal deviation probability, the greater the probability of spatiotemporal deviation anomaly.

[0055] Trajectory likelihood is calculated using a hidden markov model (HMM). An HMM is trained based on historical compliance data of power materials to obtain a power material compliance model. Continuous GPS positioning, vehicle speed, and timestamps are used as observation sequences to calculate the trajectory log-likelihood, specifically expressed as:

[0056] Where L(O|λ) is the trajectory log-likelihood, used to measure the degree of matching between the observed sequence O and the power material compliance model λ. The larger the value of the trajectory log-likelihood, the more compliant the trajectory; the smaller the value of the trajectory log-likelihood, the more abnormal the trajectory. A is the state transition matrix, B is the observation probability matrix, and π is the initial state probability vector. S is the hidden state sequence, and P(O,S|λ) is the probability that the observed sequence O and the hidden state sequence S occur simultaneously under the power material compliance model. The smaller the value of the trajectory log-likelihood, the higher the corresponding second spatiotemporal deviation probability.

[0057] According to the cross-validation rules, a three-level verification system is set up, including strong anomaly detection, valid anomaly detection, and suspected noise detection. Strong anomaly detection occurs when both the first and second spatiotemporal deviation probabilities reach a preset first deviation probability threshold; in this case, it is directly marked as an anomaly to be confirmed without probability fusion, indicating the existence of an anomaly that requires further manual determination of its cause. Valid anomaly detection occurs when both the first and second spatiotemporal deviation probabilities are within a preset second deviation probability range, and the difference between the first and second spatiotemporal deviation probabilities is less than a preset probability difference threshold; in this case, a valid anomaly exists, but further confirmation of its validity is needed, employing a probability fusion strategy to fuse the first and second spatiotemporal deviation probabilities. Suspected noise detection occurs when, apart from the cases of strong and valid anomaly detection, the anomaly is considered non-fraudulent noise (such as temporary detours, brief device offline times, etc.), and can be judged as compliant, filtering out invalid interference and reducing the false detection rate.

[0058] In a specific example, the first deviation probability threshold is ∈(0.9,1], the second deviation probability range is (0.6,0.9], and the probability difference threshold is 0.2. In other examples, adaptive adjustments can be made according to the actual situation, and no limitation is made on this.

[0059] Furthermore, through a probability fusion strategy, dynamic probability weights are generated, and by combining the first and second spatiotemporal deviation probabilities, spatiotemporal verification results are given, specifically including: The first spatiotemporal deviation probability and the second spatiotemporal deviation probability are normalized respectively to obtain the first standard deviation probability and the second standard deviation probability; Based on the scenario risk coefficient and combined with the dynamic basic weight relationship, the basic weights corresponding to the first standard deviation probability and the second standard deviation probability are adjusted to give the first basic weight and the second basic weight. By using the supply compliance correction coefficient, the first basic weight and the second basic weight are corrected respectively to generate the first dynamic weight and the second dynamic weight. Based on the first dynamic weight and the second dynamic weight, and combined with the first standard deviation probability and the second standard deviation probability, the target spatiotemporal deviation probability is determined; Based on the scenario risk coefficient, a dynamic deviation threshold is determined, and combined with the target spatiotemporal deviation probability, the spatiotemporal verification result is obtained.

[0060] When a valid anomaly is determined through cross-validation rules, a probability fusion strategy is adopted to fuse the first and second spatiotemporal deviation probabilities. Firstly, the first and second spatiotemporal deviation probabilities are normalized and calibrated to eliminate differences in the dimensions and output ranges of the two models, ensuring that the confidence levels of the two models are comparable in the same dimension and avoiding weight imbalance caused by different model output ranges.

[0061] A dynamic basic weighting formula is introduced, and the corresponding scenario risk coefficient is determined by combining the complexity of the material category with the risk level of the delivery scenario. This determines the first and second basic weights. For example, for core materials such as high-voltage cables and transformers, and remote delivery scenarios, the scenario risk coefficient is relatively high, requiring an increase in the weight corresponding to the probability of deviation from the second standard of the dynamic trajectory. Conversely, for regular materials and short-distance urban delivery, the scenario risk coefficient is relatively low, requiring an increase in the weight corresponding to the probability of deviation from the first standard of the static spatiotemporal trajectory. This achieves scenario-based weight adaptation.

[0062] The dynamic basic weight relationship is specifically expressed as follows:

[0063] Among them, w base,1 As the first basic weight, w b,1 Here, τ represents the base weight corresponding to the first standard deviation probability, γ represents the weight adjustment coefficient, and w represents the scenario risk coefficient. base,2 As the second basic weight, w b,2 Here, w represents the base weights corresponding to the second standard deviation probability. b,1 and w b,2 These are pre-set based on the actual mining activity; they can be the same or different. τ is pre-set to limit the adjustment range of the base weight, while γ is set according to different scenarios.

[0064] By introducing a supply compliance correction coefficient, the final first dynamic weight and second dynamic weight are constructed, specifically expressed as follows:

[0065] Among them, w base,1 w1 is the first basic weight, and w1 is the first dynamic weight. base,2 w1 is the second basic weight, and w2 is the second dynamic weight. The first standard deviation probability, β represents the second standard deviation probability, and β is the supply compliance correction coefficient, which is obtained through analysis of the supplier's historical performance data. The higher the supplier's historical compliance rate, the larger the value of the supply compliance correction coefficient, thus reducing the penalty for misjudging high-quality suppliers. The higher the supplier's historical anomaly rate, the smaller the value of the supply compliance correction coefficient, thus strengthening the sensitivity of anomaly detection and realizing differentiated weighting of suppliers.

[0066] The probability of spatiotemporal deviation from the target is specifically expressed as:

[0067] Among them, P joint Let w1 be the target spatiotemporal deviation probability, w2 be the first dynamic weight, and w1 be the second dynamic weight. The first standard deviation probability, This represents the second standard deviation probability.

[0068] The dynamic deviation threshold is specifically expressed as:

[0069] Where, θ dynamic For dynamic deviation from the threshold, θ max For deviations from the maximum threshold, θ d γ is the threshold adjustment coefficient, and γ is the scenario risk coefficient.

[0070] By setting dynamic deviation thresholds, the thresholds are appropriately lowered in high-risk scenarios to improve the detection rate of anomalies in core materials, while the thresholds remain unchanged in low-risk scenarios to reduce false alarms.

[0071] If the target spatiotemporal deviation probability is greater than or equal to the dynamic deviation threshold, the spatiotemporal verification result is confirmed as spatiotemporal fraud anomaly; if the target spatiotemporal deviation probability is less than the dynamic deviation threshold, the spatiotemporal verification result is normal and compliant. This reduces the false detection rate and also improves scenario adaptability.

[0072] By cross-coupling and validating Bayesian network models and Hidden Markov Models, combined with an adaptive weighted fusion mechanism, the system first makes independent judgments, then cross-validates to eliminate misjudgments from single models, and finally dynamically weights and outputs the spatiotemporal deviation probability of the target. This achieves complementary detection of static spatiotemporal probability and dynamic trajectory sequence, improving the accuracy of spatiotemporal anomaly detection. Through a composite weighted fusion mechanism of cross-calibration, historical compliance weighting correction, and dynamic threshold adaptation, the system addresses the problems of false detection by single models, bias from fixed weights, and incompatibility of fixed thresholds with different material scenarios.

[0073] S105: Based on the verification reasoning rules, perform reasoning analysis on the image verification results and spatiotemporal verification results to determine the primary performance verification results.

[0074] Specifically, based on verification reasoning rules, the image verification results and spatiotemporal verification results are analyzed to determine the preliminary performance verification results, with reference to... Figure 2 Specifically, it includes: Based on the power supply scenario, the scenario entropy under different verification rules is analyzed, and the preset initial weights corresponding to different verification rules are adjusted to determine the basic weights of the scenario under different verification rules. Based on the image verification results and spatiotemporal verification results, the scene base weights under different verification rules are nonlinearly attenuated and corrected, and the scene correction weights under different verification rules are given. Based on the conflict of different verification rules, the scenario correction weights are adjusted to generate verification rule weights corresponding to different verification rules; By triggering the indicator function according to the rule, combining the rule importance coefficient, and integrating the verification rule weights corresponding to different verification rules, the performance verification score is determined and the preliminary performance verification result is generated.

[0075] In one specific implementation, the basic weights of scenarios under different verification rules are determined, specifically as follows:

[0076] in, To verify the basic weights of the scenario corresponding to rule i, w i0 To verify the preset initial weights corresponding to rule i, H i To check the scene entropy corresponding to rule i.

[0077] Scenario entropy is calculated using data such as the type of goods, delivery area, and fulfillment batch of the current fulfillment order. The higher the scenario entropy value, the stronger the scenario uncertainty. The corresponding basic scenario weight will be appropriately reduced to ensure the core position of deterministic scenario rules and avoid scenario interference leading to weight imbalance.

[0078] Based on the image verification results and spatiotemporal verification results, the scene base weights under different verification rules are nonlinearly attenuated and corrected. The scene correction weights under different verification rules are given as follows:

[0079] Among them, w i1 Adjust the weights for the scenario corresponding to verification rule i. To verify the basic weights of the scenario corresponding to rule i, C i Let C be the anomaly confidence level corresponding to verification rule i. In a specific example, when verification rule i is a truth verification rule, C i =P fake P fake For image verification results; when verification rule i is a timeliness verification rule, C i =P joint P joint The probability of spatiotemporal deviation from the target.

[0080] Instead of the conventional linear multiplication method, a non-linear decay is adopted to achieve a slow increase in scene correction weights when the anomaly confidence is low and a rapid approach to saturation when the anomaly confidence is high. This avoids drastic changes in scene correction weights caused by small fluctuations in anomaly confidence and ensures inference stability.

[0081] Based on the conflict situations of different verification rules, the scenario correction weights are adjusted to generate verification rule weights corresponding to different verification rules, specifically expressed as follows:

[0082] Among them, W i w represents the weight of the verification rule corresponding to verification rule i. i1 To adjust the weights for the scenario corresponding to verification rule i, λ conflict η is the conflict penalty coefficient. i The degree of conflict between check rule i and other check rules.

[0083] When multiple verification rules conflict in parallel, the higher the degree of conflict, the greater the penalty. This reduces the weight of the corresponding verification rule, automatically resolving rule conflicts without manual intervention to adjust the priority of verification rules, thus achieving adaptive conflict resolution. Through a triple mechanism of entropy adaptation, nonlinear correction, and conflict penalty, the weight of verification rules is refined and dynamically adjusted adaptively, improving inference accuracy.

[0084] By triggering an indicator function based on rules, combining the rule importance coefficient, and integrating the weights of different verification rules, a performance verification score is determined, generating a preliminary performance verification result, specifically represented as follows:

[0085] Among them, L final As the initial performance verification result, W i Let δ(R) be the weight of the verification rule corresponding to verification rule i, where i is the number of the verification rule, δ() is the rule triggering indicator function, and δ(R) is the weight of the verification rule corresponding to the rule i. i ) represents the output value of the rule trigger indicator function corresponding to check rule i. If check rule i is triggered, then δ(R) i If )=1, and the verification rule i is not triggered, then δ(R) i )=0,k i To check the rule importance coefficient of rule i, S rule For the performance verification score, S1 is the preset first threshold for the verification result, and S2 is the preset second threshold for the verification result.

[0086] In a specific example, the rule importance coefficient corresponding to the authenticity verification rule is relatively high, while the rule importance coefficients corresponding to the timeliness verification rule and the defect verification rule are relatively low, further strengthening the weight of the core anomaly judgment. When the primary performance verification result is 0, it indicates that the anomaly level is normal; when the primary performance verification result is 1, it indicates that the anomaly level is slightly abnormal; when the primary performance verification result is 2, it indicates that the anomaly level is seriously abnormal. The preset first threshold and preset second threshold of the verification result are calibrated according to the actual situation, adapting to the score range of the performance verification score after composite weighting, so that the judgment result is more in line with the actual performance violation degree.

[0087] Employing a composite weighting logic, rule weights are dynamically adjusted in real-time based on scenario, anomaly confidence level, and conflict cost. This constructs a multi-layered dynamic rule weighting mechanism driven by scenario entropy, attenuated anomaly confidence, rule conflict cost, and dynamic priority scheduling, resolving the issues of fixed weights, rigid conflict resolution, and inflexible rule priorities in traditional rule engines. A closed-loop architecture combining parallel multi-dimensional detection and serial rule reasoning achieves deep integration of flexible feature recognition and rigid business rule constraints, completely replacing traditional manual auditing and resolving problems such as forged credentials, falsified tracking, timeliness violations, and missing documents.

[0088] S106: Based on the supplier performance points rules, the initial performance verification results of the power material supplier are integrated to determine the audit evaluation result of the power material supplier.

[0089] Specifically, based on the supplier performance scoring rules, the initial performance verification results of the power material supplier are scored and integrated to determine the audit evaluation result of the power material supplier, with reference to... Figure 3 Specifically, it includes: Based on the information entropy of each verification rule, the initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule; Based on the initial performance verification results of the power supply provider, and combined with the global rule weights, global gain items and global penalty items are determined; By integrating global gain and global penalty terms, the audit evaluation results of power material suppliers are determined.

[0090] Based on the information entropy of each verification rule, the initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule, specifically including: By analyzing the discriminative power of each verification rule, the information entropy of each verification rule is given; Calculate the objective weight of each verification rule by combining the sum of the information entropy of each verification rule; By using a weight balancing coefficient, the objective weights and initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule.

[0091] In one specific implementation, the global rule weights corresponding to each verification rule are generated, specifically represented as follows:

[0092] Among them, w global,i Let λ be the global rule weight for verification rule i, λ be the weight balancing coefficient, i be the verification rule number, j be the supplier number, n be the number of suppliers, and p be the global rule weight for verification rule i. ij Let E be the proportion of the value of the i-th verification rule for the j-th supplier to the total score, where ln() is the logarithmic function. iTo verify the information entropy of rule i, To verify the initial parameters of rule i, To verify the objective weight of rule i.

[0093] The initial parameters are initial weights set for each verification rule based on experience. Information entropy is used to measure the discriminative power of each verification rule. The smaller the information entropy value, the greater the difference between the corresponding verification rule and different suppliers, and the higher the discriminative power. Conversely, the larger the information entropy value, the smaller the difference between the corresponding verification rule and different suppliers, and the lower the discriminative power. The objective weight of each verification rule is calculated using information entropy. The verification rule with higher discriminative power corresponds to a larger objective weight. The calculation is based on historical performance data and is performed without human intervention, ensuring the objectivity of the objective weights. By setting a weight balancing coefficient, the business rationality of expert experience and the objectivity of data-driven approaches are taken into account, avoiding the one-sidedness of a single subjective weight and the detachment of a single objective weight from business applications, thus achieving dynamic optimization of the weights.

[0094] Based on the initial performance verification results of the power supply provider, and combined with the global rule weights, global gain and global penalty terms are determined, specifically as follows:

[0095] in, w is the global penalty term for the performance process k. global,i To verify the global rule weight of rule i, C i To check the anomaly confidence level corresponding to rule i, Score i To verify the basic deduction value of rule i, K l δ is the correlation coefficient for abnormality levels. first This is the first-order exemption function. The anomaly level linkage coefficient is set to different values ​​based on different actual situations. The first-order exemption function takes a value of 1 for minor anomalies during the first performance, and 0 for all other cases. By setting a global penalty item, the deduction is quantitatively calculated; the higher the anomaly confidence and the more severe the anomaly level, the more points are deducted. The first-order exemption function reflects the evaluation's fault tolerance, preventing minor errors in the first performance from affecting long-term evaluations.

[0096]

[0097] in, w is the global gain term for the performance process k. global,i To check the global rule weight of rule i, Score i This is the base deduction value for verification rule i.

[0098] If there are no abnormalities triggering compliance bonuses throughout the entire performance process, the global gain item will calculate compliance bonuses based on the integrated weight for compliant orders without any performance abnormalities, thereby incentivizing suppliers to perform in a standardized manner.

[0099] By integrating global gain and global penalty terms, the audit evaluation results for power material suppliers are determined, specifically including: Summing the global gain and global penalty terms yields the total integral change in the current performance process. Update the number of continuously compliant orders based on the total change in points during the current fulfillment process; The reward value is determined based on the updated number of consecutive compliant orders; Based on the audit evaluation results of the previous performance process, and by integrating the total change in points and reward value of the current performance process, the audit evaluation results of the power material supplier after the current performance process are determined.

[0100] Summing the global gain and global penalty terms yields the total integral change in the current performance process, specifically expressed as:

[0101] in. Let k be the total integral change value of the performance process. For the global penalty term in the performance process k, Let be the global gain term for the fulfillment process k. By combining the gain term and penalty term for a single order, we obtain the total change in points after the fulfillment of that order. A positive value represents an increase in points, and a negative value represents a decrease in points.

[0102]

[0103] Where, N c The number of consecutive compliant orders is accumulated when the points for a single order do not decrease. Once an abnormal point deduction occurs, the number of consecutive compliant orders is reset to zero and the count is restarted. This is to incentivize suppliers to maintain long-term stable compliance performance. The reward value is N consecutive times. max An extra point is awarded for each compliant order fulfillment; however, if the number of consecutive compliant orders does not reach N, the point will be awarded accordingly. max No reward is given for single or multiple violations, thus reinforcing the long-term compliance incentive effect. max Configure according to the actual situation.

[0104] By constructing a continuous compliance counting mechanism, we incentivize long-term compliance performance, do not accumulate rewards for perfect scores, and set hard boundaries to prevent points from exceeding limits, thus ensuring that the points range is stable and controllable.

[0105] By constructing a continuous compliance counting mechanism, we incentivize long-term compliance performance, do not accumulate rewards for perfect scores, and set hard boundaries to prevent points from exceeding limits, thus ensuring that the points range is stable and controllable.

[0106]

[0107] Among them, Sk To complete the audit evaluation results after the performance process k, Score full S represents the maximum score of the audit evaluation results. k-1 To complete the audit evaluation results after the performance process k-1, As a reward value, Let $k$ be the total integral change value of the performance process $k$, $max()$ be the maximum value function, and $min()$ be the minimum value function.

[0108] The integration interval [0, Score] is determined by using the max and min functions. full The min function prevents the integral from exceeding the maximum score. full Conversely, the integral of the max function is less than 0 points, ensuring a standardized and consistent integration range to facilitate subsequent supplier rating based on audit evaluation results. In a specific example, an audit evaluation result ≥90 points indicates an excellent supplier, enjoying benefits such as bonus points in bidding, priority in winning bids, simplified audits, and a fast-track process; an audit evaluation result between 80 and 89 points corresponds to a qualified supplier, subject to standard audit procedures without additional checks; an audit evaluation result between 70 and 79 points indicates a supplier requiring improvement, with increased audit frequency, focused monitoring, and a deadline for rectification; an audit evaluation result <70 points indicates an unqualified supplier, with restricted bidding privileges, suspension of cooperation, and restoration of qualifications upon rectification.

[0109] In a specific example, there are supplier A and supplier B. Supplier A is fulfilling its contract for the first time, with a minor anomaly level, and will be compliant after rectification. During supplier A's initial fulfillment, fulfillment data is collected, including uploaded shipping images, GPS tracks, order documents, and other full data. This data is then processed using feature engineering and analyzed. A multi-branch multimodal neural network is used for verification. The generated image verification result is 0.32 < 0.85, indicating that the image is normal and not falsified. A dual-model cross-validation strategy is used to provide spatiotemporal verification results. The delivery route has a slight detour; the first standard deviation probability corresponding to the Bayesian model is 0.42, and the second standard deviation probability corresponding to the HMM is 0.38. The resulting spatiotemporal deviation probability is 0.4 < 0.7, indicating that the spatiotemporal process is normal. The fulfillment verification score S is determined through verification inference rules. rule =55 < 60, indicating a minor anomaly in the initial performance verification result. According to the supplier performance points rules, the first performance triggers the first-order exemption function, resetting the deduction points to zero. Only the base deduction of 5 points is deducted, resulting in an initial score of 100 points and a final score S1 = 95 points, classifying the supplier as an excellent supplier. Subsequent two consecutive orders with full compliance will trigger a bonus of +1 point based on the number of consecutive compliant orders, granting priority in bidding.

[0110] Supplier B (serious fraud, abnormal score calculation) collected performance data, including uploading old shipping photos and GPS trajectory simulations of normal driving, and analyzed the data after processing. Verification was performed using a multi-branch multimodal neural network. The generated image verification result was 0.92 ≥ 0.85, indicating image fraud, with an abnormal confidence level of 0.92. A dual-model cross-validation strategy was used to provide spatiotemporal verification results. The normal delivery time for transformers in urban areas is 4-6 hours and the distance is 35-50 kilometers. This delivery took 1.5 hours and covered 20 kilometers. The first standard deviation probability corresponding to the Bayesian model was 0.98, and the second standard deviation probability corresponding to the HMM was 0.97. The resulting spatiotemporal deviation probability was 0.976 ≥ 0.95, confirming a spatiotemporal anomaly. Through verification inference rules, a double serious anomaly was triggered by both the authenticity verification rule and the timeliness verification rule. The verification rule weight W corresponding to verification rule i was... i =0.3×(1+0.92)+0.25×(1+0.976)=1.077, determine the performance verification score S. rule =98≥90, the initial performance verification result is judged as seriously abnormal. According to the supplier performance scoring rules, exemptions are not granted for non-first orders. The global rule weight fusion weight of verification rule i is 0.32 and 0.26, and the global penalty is -(0.32×15×0.92+0.26×8×0.976)=-6.624. The initial score is 100 points. After deducting the basic deduction value, the final score is 68 points, which is rated as unqualified supplier. The supplier is restricted from bidding for 6 months, the frequency of subsequent audits will be increased, and the supplier is not allowed to participate in new bidding until the rectification is completed.

[0111] By implementing the entire audit and evaluation process for power material suppliers, the audit and evaluation process is fully automated, reducing the cost of manual audits and improving the identification rate of contract performance anomalies. At the same time, the thresholds of each detection model are iteratively optimized through result feedback, continuously improving the adaptability of the audit and evaluation methods.

[0112] Reference Figure 4 This invention provides a power supply supplier audit and evaluation device based on performance behavior, comprising: Data acquisition module 201 is used to collect all data on the performance of power material suppliers. Data processing module 202 is used to perform data partitioning and standardized feature engineering on the full dataset to obtain spatiotemporal feature dataset and multimodal image dataset; The first verification module 203 is used to verify the image data in the multimodal image dataset using a multi-branch multimodal neural network and generate image verification results; The second verification module 204 is used to analyze and verify the deviation of spatiotemporal features in the spatiotemporal feature dataset through a dual-model cross-validation strategy, and to provide spatiotemporal verification results. The verification fusion module 205 is used to perform reasoning analysis on the image verification results and spatiotemporal verification results based on the verification reasoning rules to determine the primary performance verification results; The points fusion module 206 is used to integrate the initial performance verification results of the power material supplier according to the supplier's performance points rules, and determine the audit evaluation result of the power material supplier.

[0113] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0114] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for auditing and evaluating power material suppliers based on performance behavior, characterized in that, include: Collect full data on the performance of power supply providers during the contract fulfillment process; The entire dataset is divided into data segments and standardized feature engineering processes to obtain a spatiotemporal feature dataset and a multimodal image dataset. A multi-branch multimodal neural network is used to verify the image data in the multimodal image dataset and generate image verification results. By employing a dual-model cross-validation strategy, the deviation of spatiotemporal features in the spatiotemporal feature dataset is analyzed and verified, and spatiotemporal verification results are provided. Based on verification reasoning rules, the image verification results and spatiotemporal verification results are analyzed to determine the preliminary performance verification results. Specifically, this includes: analyzing the scenario entropy under different verification rules based on the power supply scenario, adjusting the preset initial weights corresponding to different verification rules, and determining the basic scenario weights under different verification rules; performing nonlinear attenuation correction on the basic scenario weights under different verification rules based on the image verification results and spatiotemporal verification results, and providing the scenario correction weights under different verification rules; adjusting the scenario correction weights based on the conflict situation of different verification rules to generate the verification rule weights corresponding to different verification rules; and determining the performance verification score by fusing the verification rule weights corresponding to different verification rules through the rule triggering indicator function and the rule importance coefficient, thus generating the preliminary performance verification result, specifically expressed as follows: ； Among them, L final As the initial performance verification result, W i Let δ(R) be the weight of the verification rule corresponding to verification rule i, where i is the number of the verification rule, δ() is the rule triggering indicator function, and δ(R) is the weight of the verification rule corresponding to verification rule i. i ) represents the output value of the indicator function that triggers the rule corresponding to check rule i, and k represents the output value of the function. i To check the rule importance coefficient of rule i, S rule For the performance verification score, S1 is the preset first threshold for the verification result, and S2 is the preset second threshold for the verification result; According to the supplier performance scoring rules, the initial performance verification results of the power material supplier are integrated to determine the audit evaluation result of the power material supplier.

2. The power supply supplier audit and evaluation method based on performance behavior as described in claim 1, characterized in that, A multi-branch, multi-modal neural network is used to verify image features in a multi-modal image feature set, generating image verification results, specifically including: By using the image coding branch in a multi-branch multimodal neural network, image features are analyzed and core features are extracted to obtain a visual feature vector; Based on the metadata encoding branch in a multi-branch multimodal neural network, the metadata in the multimodal image dataset is encoded to determine the standard feature vector; Based on the cross-attention fusion mechanism in multi-branch multimodal neural networks, the visual feature vector and standard feature vector are correlated, matched, and mapped to generate image verification results.

3. The power supply supplier audit and evaluation method based on performance behavior as described in claim 2, characterized in that, Based on the cross-attention fusion mechanism in multi-branch multimodal neural networks, the visual feature vector and standard feature vector are correlated, matched, and mapped to generate image verification results, specifically including: By performing linear transformations on the visual feature vector and the standard feature vector respectively, query vector and key vector are generated; A value vector is generated by performing a linear transformation on the vertically concatenated vector of the visual feature vector and the standard feature vector; Based on the query vector and key vector, calculate the cross-attention weight, analyze the correlation matching degree between the visual feature vector and the standard feature vector, and give the correlation matching parameters; The Hadamard product operation is performed on the associated matching parameters and value vectors, and the output is mapped by the activation function to obtain the image verification result.

4. The power supply supplier audit and evaluation method based on performance behavior as described in claim 1, characterized in that, Using a dual-model cross-validation strategy, the bias of spatiotemporal features in the spatiotemporal feature dataset is analyzed and verified, and spatiotemporal verification results are provided, including: Based on the Bayesian network model, the deviation of spatiotemporal features in the spatiotemporal feature dataset is analyzed, and the first spatiotemporal deviation probability is given. Based on the Hidden Markov Model, the spatiotemporal features in the spatiotemporal feature dataset are observed, and the second spatiotemporal deviation probability is given. Based on the cross-validation rules, the first spatiotemporal deviation probability and the second spatiotemporal deviation probability are fused and judged to determine the probability fusion strategy; By employing a probability fusion strategy, dynamic probability weights are generated, and the first and second spatiotemporal deviation probabilities are combined to provide spatiotemporal verification results.

5. The power supply supplier audit and evaluation method based on performance behavior as described in claim 4, characterized in that, By employing a probabilistic fusion strategy, dynamic probability weights are generated. These weights are then combined with the first and second spatiotemporal deviation probabilities to provide spatiotemporal verification results, specifically including: The first spatiotemporal deviation probability and the second spatiotemporal deviation probability are normalized respectively to obtain the first standard deviation probability and the second standard deviation probability; Based on the scenario risk coefficient and combined with the dynamic basic weight relationship, the basic weights corresponding to the first standard deviation probability and the second standard deviation probability are adjusted to give the first basic weight and the second basic weight. By using the supply compliance correction coefficient, the first basic weight and the second basic weight are corrected respectively to generate the first dynamic weight and the second dynamic weight. Based on the first dynamic weight and the second dynamic weight, and combined with the first standard deviation probability and the second standard deviation probability, the target spatiotemporal deviation probability is determined; Based on the scenario risk coefficient, a dynamic deviation threshold is determined, and combined with the target spatiotemporal deviation probability, the spatiotemporal verification result is obtained.

6. The power material supplier audit and evaluation method based on performance behavior as described in claim 1, characterized in that, According to the supplier performance scoring rules, the initial performance verification results of power material suppliers are scored and integrated to determine the audit evaluation results of power material suppliers, specifically including: Based on the information entropy of each verification rule, the initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule; Based on the initial performance verification results of the power supply provider, and combined with the global rule weights, global gain items and global penalty items are determined; By integrating global gain and global penalty terms, the audit evaluation results of power material suppliers are determined.

7. The power supply supplier audit and evaluation method based on performance behavior as described in claim 6, characterized in that, Based on the information entropy of each verification rule, the initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule, specifically including: By analyzing the discriminative power of each verification rule, the information entropy of each verification rule is given; Calculate the objective weight of each verification rule by combining the sum of the information entropy of each verification rule; By using a weight balancing coefficient, the objective weights and initial parameters of each verification rule are integrated to generate the global rule weight corresponding to each verification rule.

8. The power material supplier audit and evaluation method based on performance behavior as described in claim 6, characterized in that, By integrating global gain and global penalty terms, the audit evaluation results for power material suppliers are determined, specifically including: Summing the global gain and global penalty terms yields the total integral change in the current performance process. Update the number of continuously compliant orders based on the total change in points during the current fulfillment process; The reward value is determined based on the updated number of consecutive compliant orders; Based on the audit evaluation results of the previous performance process, and by integrating the total change in points and reward value of the current performance process, the audit evaluation results of the power material supplier after the current performance process are determined.

9. A power supply supplier audit and evaluation device based on performance behavior, characterized in that, The method for auditing and evaluating power material suppliers based on performance behavior as described in any one of claims 1-8 includes: The data acquisition module is used to collect all data related to the performance of power supply contracts. The data processing module is used to partition the full dataset and perform standardized feature engineering to obtain spatiotemporal feature datasets and multimodal image datasets. The first verification module is used to verify the image data in the multimodal image dataset using a multi-branch multimodal neural network and generate image verification results. The second verification module is used to analyze and verify the deviation of spatiotemporal features in the spatiotemporal feature dataset through a dual-model cross-validation strategy, and to provide spatiotemporal verification results. The verification fusion module is used to perform reasoning analysis on image verification results and spatiotemporal verification results based on verification reasoning rules to determine the initial performance verification results; The points fusion module is used to integrate the initial performance verification results of power material suppliers according to the supplier performance points rules, and determine the audit evaluation results of power material suppliers.

Images