Agricultural supply chain risk intelligent sensing and early warning system based on multi-source data fusion

Abstract

The invention relates to the technical field of agricultural informatization and supply chain risk control, and particularly discloses an agricultural supply chain risk intelligent sensing and early warning system based on multi-source data fusion. According to the system, six multi-source data including a producing area environment, an agricultural product category, a transaction behavior, a logistics link, credit and market information are integrated through a data acquisition layer; space-time alignment and cleaning are carried out through the data processing and alignment layer; through feature engineering and a multi-source fusion layer, depth features are constructed by comprehensively utilizing a knowledge graph and a graph attention network, Transform behavior sequence modeling, CNN-LSTM remote sensing time sequence analysis and Graph2Seq logistics trajectory prediction, and the Transform and the graph attention network are specially adapted for an agricultural scene; the dynamic risk scoring layer is used for comprehensively calculating five risks of producing areas, logistics, transactions, markets and credit, and a comprehensive risk score is output through weighted fusion; a decision basis is provided by integrating SHAP, attention visualization and map path tracking through an interpretability output layer; and finally, outputting a grading risk strategy by a strategy decision-making layer. According to the invention, dynamic, accurate and explainable intelligent assessment and early warning of full-link and multi-dimensional risks of the agricultural supply chain are realized.

Description

Technical Field

[0001] This invention relates to the fields of agricultural informatization, artificial intelligence, supply chain management and risk control, and in particular to an intelligent perception and early warning system for agricultural supply chain risks based on multi-source heterogeneous data fusion, knowledge graph, deep learning and explainable artificial intelligence. Background Technology

[0002] Agricultural supply chain risk management currently faces severe challenges: risks are diverse and intertwined, encompassing meteorological disasters and pests at the production end, logistical disruptions and quality spoilage at the distribution end, credit deficiencies and fraud at the transaction end, and drastic price fluctuations at the market end. This risk data is typically scattered across multiple independent information silos, including meteorological, IoT, trading platforms, and logistics systems, exhibiting heterogeneous formats and varying temporal and spatial scales, making unified correlation analysis difficult. Traditional risk control methods largely rely on expert experience or single-dimensional static rules, resulting in delayed responses and an inability to cope with complex and dynamic supply chain environments, lacking the ability to predict potential risks and interpret decision outcomes. Therefore, the agricultural industry urgently needs an intelligent sensing and early warning system that can integrate data across the entire supply chain, possess dynamic predictive capabilities, and provide clear insights into the causes of risks. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a system capable of end-to-end, multi-dimensional, dynamic, and interpretable risk perception and early warning in the agricultural supply chain. This system integrates six categories of multi-source data through a microservice architecture, utilizes advanced artificial intelligence technologies such as knowledge graphs, graph neural networks, Transformers, and time-series forecasting to construct composite risk characteristics, establishes a dynamically adjustable risk scoring model, and outputs risk reports with quantitative attribution and automated risk control strategies, thereby significantly improving the resilience, security, and decision-making transparency of the agricultural supply chain.

[0004] This system is based on a general deep learning model, but has been deeply adapted and innovated to address the specific characteristics of agricultural supply chain data.

[0005] In terms of transaction behavior sequence modeling, the Transformer encoder used is not a general structure, but introduces a dynamic time window and a perishability-related attention mechanism. The system can automatically identify the harvest season of agricultural products and adjust the time span of the input sequence. At the same time, it dynamically adjusts the attention to key risk behaviors (such as payment delays and address changes) according to the perishability level, thereby more accurately capturing the transaction risks caused by the characteristics of agricultural products.

[0006] In terms of knowledge graph feature extraction, the graph attention network innovatively integrates geospatial features and category dependencies into attention computation. By using the spatial distance between plot nodes as a feature and designing attention masks for agriculture-specific relationships such as "dependence," the model can effectively perceive the geographical transmission path of risks and category specificity, solving the problem that traditional methods struggle to address the spatial diffusion of agricultural risks and category dependencies. Attached Figure Description

[0007] Figure 1: System overall architecture and microservice deployment diagram.

[0008] Figure 2: Schematic diagram of multi-source data acquisition topology and data flow.

[0009] Figure 3: Data preprocessing and alignment (including spatiotemporal interpolation and anomaly detection).

[0010] Figure 4: Definition and example diagram of agricultural supply chain knowledge graph model.

[0011] Figure 5: Schematic diagram of graph feature extraction based on multi-head graph attention network.

[0012] Figure 6: Structure diagram of CNN-LSTM remote sensing image time series analysis model.

[0013] Figure 7: Graph2Seq-based model for logistics trajectory prediction and DTW deviation calculation.

[0014] Figure 8: Flowchart of dynamic risk scoring model (five sub-models and weighted fusion).

[0015] Figure 9: Schematic diagram of the interpretable output module (SHAP, attention, path tracing) in operation.

[0016] Figure 10: Flowchart of risk control strategy decision-making based on threshold, decision tree and reinforcement learning. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] This system adopts a layered, loosely coupled microservice architecture design, with seven core layers, as shown in Figure 1:

[0019] Data Acquisition Layer: As the system's data entry point, it enables real-time, near real-time, or batch acquisition of six types of data through customized adapters, API gateways, message middleware (such as Apache Kafka), and log collection tools deployed on the edge or cloud, ensuring the diversity and integrity of data sources.

[0020] Data Processing and Alignment Layer: This layer is crucial for ensuring data quality and consistency. Its main processes (as shown in Figure 3) include: ●Time alignment: Unify the timestamps of various data sources to a standard time base; ● Missing value handling: For sensor data, a spatiotemporal K-nearest neighbor interpolation algorithm is used to fill in the missing values ​​using data from neighboring spatiotemporal points; ● Anomaly detection: The isolated forest algorithm is applied to quickly identify anomalies by randomly partitioning the feature space; ●Standardization: Z-score standardization is applied to numerical features to eliminate the influence of units. The final output is a high-quality, spatiotemporally aligned unified dataset.

[0021] Feature Engineering and Multi-Source Fusion Layer: This layer is the core of the system's intelligence, responsible for extracting advanced risk features from raw data. ●Knowledge Graph Construction and GNN Feature Generation: As shown in Figures 4 and 5, a supply chain knowledge graph containing multiple types of entities and relationships is constructed. A multi-head graph attention network is used to learn node representations, capturing deep structured risk semantic information. Finally, a graph-structured risk feature vector F_graph is obtained through pooling operations. ● Behavioral sequence modeling: The Transformer encoder is used to model the user's transaction behavior sequence, and its self-attention mechanism is used to effectively capture long-term dependencies and abnormal patterns, and output the behavioral sequence encoding H_trade; ● Remote sensing image feature extraction: As shown in Figure 6, CNN is used to extract the spatial features of a single-period remote sensing image, and then the bidirectional LSTM is input to capture its temporal evolution pattern. Finally, the remote sensing risk feature F_rs is output to assess the yield and growth status. ● Logistics trajectory prediction: As shown in Figure 7, the planned route is encoded and the expected trajectory is predicted using the Graph2Seq model. The deviation between the actual trajectory and the expected trajectory is accurately calculated using the dynamic time warping algorithm. ●Feature fusion: All the above feature vectors are concatenated, and an attention mechanism is introduced to dynamically adjust the importance of features from different sources, forming a unified deep feature representation.

[0022] Dynamic Risk Scoring Layer: As shown in Figure 8, this layer integrates five models dedicated to calculating the five major risks in a functional format: ● R_origin = o( MLP([F_whether, F_rs, F_soil, F_ geol]) ● R_logistics = o( MLP([Deviation, Z(Temp_ Abnormality), Delay])) ● R_trade=o(Linear(H_trade)) ●R_market is calculated using the ARIMA-Transformer hybrid model. ●R_credit=MLP(F_credit) Finally, a weighted sum is calculated using dynamically adjustable weights: R_total = Z wi Ri.

[0023] Interpretable Output Layer: As shown in Figure 9, this layer immediately activates the interpretation engine after the risk score is generated. ●SHAP Analysis: Quantifies the contribution of each input feature to the final score; ●Attention Visualization: Shows which key points in the behavioral sequence the Transformer model focuses on when assessing risk; ● Knowledge Graph Path Tracing: Locates the association paths of high-risk entities within the knowledge graph. This ultimately generates a comprehensive report integrating quantitative attribution and semantic interpretation.

[0024] Strategy Decision Layer: As shown in Figure 10, based on the comprehensive risk score and interpretation report, a hybrid decision-making mechanism (threshold judgment, decision tree, reinforcement learning) automatically executes corresponding risk control actions, from "normal release" to "account freezing", etc.

[0025] System Service Layer: Provides underlying technical support for the above layers. All services support Docker containerized deployment and are orchestrated through Kubernetes to ensure high availability, scalability, and elasticity of the system. Methods and Steps

[0026] Step S1: Multi-source data acquisition: Through various components of the data acquisition layer, continuously and automatically acquire the six categories of data described in the background technology, and push them to the message queue for consumption.

[0027] Step S2: Data Preprocessing and Alignment: The data processing and alignment layer consumes the data stream and strictly performs cleaning, alignment, completion, denoising, and standardization operations to form a high-quality dataset that can be used for feature engineering.

[0028] Step S3: Feature Construction and Multi-Source Fusion: This step involves executing four major feature engineering processes. The knowledge graph is updated using GAT; transaction sequences are encoded using Transformer; remote sensing images are processed using CNN-LSTM; and logistics trajectories are predicted using Graph2Seq and their deviations are calculated using DTW. Finally, an attention mechanism is used to weightedly fuse all generated features.

[0029] Step S4: Dynamic Risk Scoring: Parallel reasoning across five risk sub-models. The system can dynamically adjust the weights of sub-risks based on business feedback (such as early warning accuracy) or through online learning, and finally calculate a weighted composite score.

[0030] Step S5: Explainability Output: The explanation engine is triggered, and by combining SHAP, attention visualization and graph path tracing technologies, a clear and quantitative risk cause analysis report is generated.

[0031] Step S6: Strategy Decision: The strategy engine receives R_total and the explanation report, automatically selects and executes the most appropriate risk control action based on the preset hybrid decision model, and records and feeds back the decision results for model optimization. Algorithm adaptation details

[0032] S3-1: Transformer adaptation for transaction behavior sequences ● The system incorporates a dynamic time window divider for the Transformer encoder. Taking strawberries as an example, when the system identifies the current time as May (harvest season), it automatically extends the observation window for the transaction behavior sequence from the usual 7 days to 14 days. Simultaneously, a seasonal weight mask is applied during the self-attention calculation, which strengthens the focus on the "order-cancellation" behavior pattern during the harvest season. ● In addition, the system maintains a dictionary of agricultural product perishability, assigning a "perishability coefficient" to each category (e.g., 0.5 for strawberries, 0.1 for corn). This coefficient acts as a multiplicative factor when calculating attention weights, affecting behavioral features related to logistics and payment timeliness. For example, for orders with high perishability coefficients, the "payment delay duration" feature will receive a significantly higher attention score in the Transformer.

[0033] S3-2: Adaptation of Knowledge Graphs and Graph Attention Networks ● When constructing a knowledge graph, the system calculates the spatial distance between each land entity and its surrounding land entities, and uses the reciprocal of this distance as a feature, which is then concatenated with the original features of the node. ● When message passing occurs in a graph attention network, a specialized category-dependent attention head is activated for the two strongly related relationships in agriculture: risk propagation and dependency. This attention head uses a predefined mask matrix when calculating attention coefficients, which significantly increases the weight of these specific relationship types. For example, when calculating the representation of the "rice production area risk" node, the "paddy field soil moisture" node, which is connected to it through a dependency relationship, will receive a much higher attention weight than its ordinary neighbors.

[0034] S3-3: Risk Transmission Path Analysis of Heterogeneous Graphs ●The knowledge graph constructed by this system, by introducing agricultural-specific entities such as land plots and perishable labels, as well as unique relationships such as planting on, dependence on, and transmission of risks, forms a unique agricultural risk transmission path; ●Example: When analyzing a suspicious strawberry order, the knowledge graph can reveal the following transmission path: High-risk account -- (equipment associated) → Farmer -- (planted in) → Plot A -- (depends on) → Strawberry variety -- (associated with perishability) → Perishability level 5 -- (transmits risk to) → Current order; ● Graph attention networks, through their adapted attention mechanisms, can automatically identify and reinforce the contribution of the key node at perishability level 5 in the path to the current order risk score, thereby achieving accurate insight into the complex risk pattern of "abnormal transactions of highly perishable goods." This entity-relationship system and risk transmission model are not available in existing general supply chain knowledge graphs. Specific Implementation

[0035] Example 1: Risk Assessment of Strawberry Trading ●Scenario: A strawberry supplier on an e-commerce platform launches a new batch of goods; ●Data and Application: The system collects rainfall and temperature data from its Jiangsu production area over the past week (production environment), combines this with data on the highly perishable strawberry category, analyzes the supplier's historical order cancellation rate (transaction behavior and credit), and queries the on-time rate of its commonly used logistics providers (logistics link). ●Analysis and Decision: Model calculations revealed that continuous rain in the production area (increased R_origin) increases the perishability of strawberries, and the supplier has a history of order cancellations (increased R_trade and R_credit). The transaction behavior model uses a Transformer adjusted for perishability-related attention. Due to the high perishability coefficient of strawberries, the model automatically increases its attention to the supplier's 'historical order cancellation rate' feature. Overall score R_total = 75; ●Output and Action: The system automatically labels the order as "medium risk". The strategy decision-making layer implements the strategy of "limiting the maximum daily transaction amount" and pushes an alert to the buyer's App via API, prompting "Due to the weather at the place of origin and the supplier's historical records, this order has medium risk. It is recommended to pay attention to logistics and timely acceptance".

[0036] Example 2: Risk Assessment of Bulk Grain Supply Chain Procurement ● Scenario: A grain company procures bulk corn from Northeast China; ●Data and Applications: The system integrates remote sensing data to monitor the growth of corn in major producing areas (producing environment), logistics data to track transportation routes up to 2,000 kilometers (logistics link), and market data to monitor spot price fluctuations (market conditions). ●Analysis and Decision-Making: The remote sensing model indicates a slight decline in the NDVI index (R_origin↑), the logistics model predicts a possible 3-day delay in transportation time due to weather (R_logistics↑), and the market model indicates an upward trend in prices (R_market↑). Overall score R_total = 65; ●Outputs and Actions: The system generates a report suggesting that the purchaser include a disclaimer clause for delayed delivery in the electronic contract and reminds the finance department to prepare cash flow contingency plans in case of price fluctuations.

[0037] Example 3: Prediction of the impact of frost on potatoes in high-altitude areas ● Scenario: A potato planting base in the Yunnan highlands; ●Data and Applications: The system is connected to the meteorological bureau's frost warning, and real-time temperature recorded by field IoT sensors and surface temperature retrieved from satellites (all of which are production area environmental data). ●Analysis and Decision-Making: The model identifies that the temperature in area A of plot A is expected to be below the potato frost critical value of -2℃ within the next 48 hours. The origin risk sub-model outputs a high risk R_origin=0.90. Through the geographic weight embedding of GAT, the system identifies that frost disasters have geographic diffusion, and the risk of adjacent plots can be transmitted to each other; ●Output and Action: The system sends an emergency alert via push notification and SMS to farmers and cooperating agricultural insurance companies: "[Frost Warning] Your plot A is expected to experience temperatures below -2℃ within 48 hours, posing an extremely high risk of frost damage to potatoes. Please take immediate frost prevention measures such as covering, fumigation, or water spraying." Simultaneously, the system triggers the insurance claims pre-assessment process.

[0038] Example 4: Identification of spoilage risk due to cold chain transportation deviation ● Scenario: A batch of refrigerated seafood is transported from a port to an inland city; ●Data and Application: The system monitors the GPS trajectory and temperature records of the carriage in real time and compares them with the predetermined route (logistics link data); ●Analysis and Decision-Making: The logistics trajectory prediction model detected that the vehicle deviated from the planned route by more than 50 kilometers (high Deviation value). Simultaneously, temperature records showed a 2-hour stop during the journey, during which the temperature inside the vehicle increased by 1.5℃ (Temp_Abnormality=1). The logistics risk score R_logistics rapidly increased to 0.89. The system used a knowledge graph to discover that the order was associated with a perishable tag (level 5) and, through the geographic weight embedding of GAT, identified a chain of risks that the deviation from the route might lead to delays in receiving goods in the surrounding area. ●Output and Action: The system triggers a "Logistics Anomaly" warning, notifying the recipient: "Warning! Order XX has deviated significantly from its route and exhibits temperature control anomalies, resulting in a high risk of spoilage (89%). Please conduct thorough quality inspection upon arrival." The explanation report clearly indicates that the deviation in the logistics trajectory is the primary cause.

[0039] Example 5: Identification of Scenarios of Abnormal Transaction Behavior Related to Weak Credit of Farmers ●Scenario: A farmer with a history of minor defaults suddenly places an order that is far higher than his historical levels; ●Data and Application: The system retrieves the farmer's credit file, analyzes the payment method and other behavioral sequences of this order, and uses knowledge graphs to analyze the correlation of its equipment; ●Analysis and Decision-Making: The knowledge graph shows that the farmer node is associated with multiple accounts that have previously been subject to risk control. The transaction behavior sequence model identifies an abnormal order placement pattern. Credit risk R_credit and transaction risk R_trade both increase. The knowledge graph reveals the complete risk transmission path from 'high-risk account' to 'current order', where the perishability association and dependency relationships are significantly strengthened under the category dependency attention head of GAT, providing an interpretable graph structure basis for the final risk decision; ●Output and Action: The strategy decision-making level implements the strategy of "mandating secondary verification of real-name authentication and submission of on-site photos of the source of goods", and routes the order to a human risk control specialist for key review.

[0040] Example 6: Multi-model integration and dynamic weight adjustment ● Scenario: Continuous optimization phase after system launch; ●Process: The system records the final result of each alert (true positive, false positive, etc.). After a period of time, analysis revealed that due to increased market volatility, the weight of R_market based on traditional rules was insufficient, resulting in some defaults caused by price crashes not being effectively alerted. ●Actions and Results: The system activated its online learning module and automatically increased the weight of market risk R_market (w_market) based on recent loss data. Following this adjustment, the system's ability to detect supply chain risks arising from market risks improved by 20%, and overall risk control accuracy continued to optimize itself.

Claims

1. An agricultural supply chain risk intelligent perception and early warning system based on multi-source data fusion, characterized in that, The system adopts a microservice architecture, including the following core levels: ● Data collection layer, for collecting agricultural supply chain related data from multiple heterogeneous data sources; ● Data processing and alignment layer, connected to the data collection layer, for time alignment, missing value processing, anomaly detection and standardization of the collected data, outputting a unified spatio-temporal dataset; ● Feature engineering and multi-source fusion layer, connected to the data processing and alignment layer, for constructing knowledge graphs, behavior sequences, remote sensing images, and logistics track features, and using attention mechanisms for deep fusion; ● Dynamic risk scoring layer, connected to the feature engineering and multi-source fusion layer, for calculating risk sub-scores in five dimensions of origin, logistics, transaction behavior, market, and credit based on the fused features, and performing weighted fusion to output a comprehensive risk score; ● Explainability output layer, connected to the dynamic risk scoring layer, for using model explainability techniques to perform attribution analysis on the risk score results and generate human-readable explanation reports; ● Strategy decision layer, connected to the explainability output layer and dynamic risk scoring layer, for executing pre-set risk strategies based on the comprehensive risk score and explanation report; ● System service layer, providing API interfaces, message queues, caches, and microservice deployment support for the above layers.

2. The system of claim 1, wherein, The data collected by the data collection layer includes: ● Origin environment data: including NDVI, EVI vegetation index, soil moisture, temperature, sunshine duration, rainfall, wind speed and direction, government-issued pest warning, and geographical polygon data representing the scope of the land; ● Agricultural product category data: including crop growth cycle characteristics, suitable environmental parameter range, harvesting time window, perishable level, historical price volatility, and classification tree structure for knowledge graph modeling; ● Transaction behavior data: including user order placement time series, average order placement amount, order cancellation or modification frequency, historical performance, and cross-account association behavior based on device fingerprint or network address; ● Logistics link data: including GPS positioning sequence of transportation tools, cold chain temperature record, predicted route from map API or historical model, actual driving trajectory point sequence, intermediate stop time, and temperature control anomaly frequency; ● Credit data: including historical default records of farmers or suppliers, authenticity score of submitted credentials, supply stability score, and historical delinquency rate; ● Market quote data: including prices of agricultural products in wholesale markets, spot markets, and retail markets, cross-regional price differences, and market index volatility data.

3. The system of claim 1, wherein, The data processing and alignment layer uses a spatio-temporal K-nearest neighbor based interpolation algorithm for missing value completion, and an isolation forest or local outlier factor algorithm for anomaly detection and processing, and finally performs Z-score standardization on numerical features.

4. The system of claim 1, wherein, The knowledge graph in the feature engineering and multi-source fusion layer is a directed heterogeneous graph, and its construction process includes the following specific entity and relationship definitions: ● In addition to general entities, specific entity types further include: ○ Parcel entity, attributes include geographic polygon coordinates, soil type, historical yield, current planting crop category; ○ Perishability label entity, attributes include perishability level (1-5 levels), preservation temperature range, maximum storage and transportation time; ○ Credit certificate entity, attributes include certificate type, authenticity score, validity period; ● Special agricultural relationship types, used to depict agricultural risk transmission paths, including: ○ Planted in, linking farmers and parcels, with a unique constraint: one farmer can plant multiple parcels, but one parcel is planted by only one farmer at a time; ○ Dependent on, linking crop categories and parcels, with a conditional constraint: this relationship is established only when the soil type and climate conditions of the parcel meet the suitable growth parameters of the crop category; ○ Associated with perishability, linking orders and perishability labels, with a unique constraint: one order is associated with only one perishability label determined by its crop category; ○ Verified by, linking credit and credit certificates, with a time constraint: the validity period of the credit certificate must cover the transaction time, otherwise the relationship is invalid; ○ Conduces risk to, linking weather events and production areas, with a spatial constraint: only when the impact area of the weather event and the geographic polygon of the production area have spatial overlap, this relationship is activated.

5. The system of claim 1, wherein, The feature engineering and multi-source fusion layer model the transaction behavior sequence, using a Transformer encoder adapted to the agricultural scene, with innovative adaptations including: ● Dynamic time window division mechanism: according to the characteristics of the harvest season of agricultural products, the time window of the input sequence is dynamically adjusted, and for crop categories in the harvest season, the window is automatically enlarged to capture high-frequency transaction patterns; ● Perishability-related attention adjustment: in the self-attention weight calculation, a "perishability coefficient" determined by the perishability level of agricultural products is introduced, automatically increasing the attention weight of high-risk behaviors such as payment delay and logistics change.

6. The system of claim 1, wherein, The feature engineering and multi-source fusion layer extract knowledge graph features, using a multi-head graph attention network adapted to the agricultural scene, with innovative adaptations including: ● Geographic weight embedding: geographic spatial distance information is integrated into node features, making spatially adjacent parcel nodes have higher weights in attention calculation to model the geographical spread of agricultural risks; ● Category-dependent mask mechanism: "category-dependent mask" is introduced in the attention head to strengthen the attention weight of certain relationship types such as "dependent on" to accurately depict the dependence of crop categories on the growing environment.

7. The system of claim 1, wherein, The feature engineering and multi-source fusion layer model the transaction behavior sequence using a Transformer encoder, which encodes the time series of transaction behaviors, with its core multi-head self-attention mechanism being: 。 8. The system of claim 1, wherein, The feature engineering and multi-source fusion layer extract remote sensing image features using a hybrid structure of convolutional neural networks and long short-term memory networks, which extracts spatial features and models temporal evolution of time-series remote sensing images, with the formula being: 。 9. The system of claim 1, wherein, The feature engineering and the logistics trajectory prediction model in the multi-source fusion layer use a sequence-to-sequence model based on a graph neural network to encode the planned route and predict the predicted trajectory recursively. The minimum alignment cost between the actual trajectory and the predicted trajectory is calculated by a dynamic time warping algorithm as the trajectory deviation value, and the formula is:

10. The system of claim 1, wherein, The risk sub-scores calculated by the dynamic risk scoring layer include: 1) Origin risk score, calculated by a multi-layer perceptron based on weather, remote sensing, soil, and geographical features; 2) Logistics risk score, calculated by a multi-layer perceptron based on trajectory deviation, temperature control abnormal times, and transportation delay; 3) Transaction behavior risk score, calculated by a linear transformation and a Sigmoid function based on the output of a Transformer encoder; 4) Market risk score, calculated by a hybrid model of ARIMA and Transformer based on price volatility and price prediction error, where the ARIMA model is 11. Credit risk score, calculated by a multi-layer perceptron based on historical credit data; The comprehensive risk score is the weighted sum of each risk sub-score R_total = Z w_i R_i, where the weights can be dynamically adjusted according to business feedback or online learning.

12. The system of claim 1, wherein, The explainable output layer integrates SHAP value analysis, attention weight visualization, knowledge graph path tracking, and rule engine; where the SHAP analysis quantifies the contribution value of feature j by calculating its average marginal contribution over all subsets of features:

13. The system of claim 1, wherein, The actions performed by the strategy decision layer include normal release, transaction limit, account freezing, submission of manual review, requirement of supplementary proof materials, and labeling of risk tags; the strategy is formulated by combining threshold judgment, decision tree, and reinforcement learning model.

14. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the steps of the functions of each level in the system of any one of claims 1 to 12.

15. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the steps of the functions of each level in the system of any one of claims 1 to 12. The program is executed by the processor to realize the steps of the functions of each level in the system of any one of claims 1 to 12.

Images