Supply chain risk dynamic early warning method and system based on time series data association analysis

Abstract

The invention discloses a supply chain risk dynamic early warning method and system based on time series data association analysis. The system comprises a sensing layer, an analysis layer and an application layer. The sensing layer comprises sensors applied to all links of the supply chain; the analysis layer comprises a time sequence data processing module, a dynamic dependency modeling module and a real-time risk calculation module; the application layer comprises an intelligent early warning module, and when the comprehensive risk score obtained by the real-time risk calculation module reaches a corresponding early warning threshold value, the intelligent early warning module is triggered, and early warning information is sent out in time by integrating different communication interfaces. According to the invention, an intelligent and closed-loop industrial monitoring and diagnosis system is formed through all frontier technologies, transition from passive response to active prediction, accurate positioning, rapid disposal and root cause treatment is realized, and powerful technical support is provided for guaranteeing safe, stable and efficient operation of a complex system.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and deep learning, and in particular to a method and system for dynamic early warning of supply chain risks based on time-series data correlation analysis. Background Technology

[0002] Traditional supply chain operations and maintenance management systems refer to information systems used by enterprises to manage and coordinate the operations of various links in the supply chain. They mainly include functions such as warehouse management, manufacturing execution, and product management. Currently, the main shortcomings in supply chain operations and maintenance management include:

[0003] I. The Lag of Static Risk Models

[0004] Traditional supply chain operation and maintenance management systems rely on statistical analysis of historical anomalies to build risk models, focusing only on static deviations of data points (such as parameter exceeding limits at a single point in time) and ignoring the temporal variation characteristics of monitored data (such as continuous fluctuation trends of parameters and abrupt changes in process time). For example, for highly sensitive products in cold chain transportation, existing models cannot identify a gradual anomaly in a certain indicator that rises continuously within 2 hours, resulting in a false negative rate of up to 40% for sudden anomalies.

[0005] II. The crudeness of the analysis of related links

[0006] Existing solutions bind related processes through the intersection of data types (e.g., matching only the "temperature and humidity" data type for warehousing and processing), without quantifying the temporal dependencies between processes. In real-world scenarios, abnormal temperature and humidity in warehousing and quality inspection failures in processing have a significant time lag (e.g., high temperatures in warehousing typically lead to microbial contamination in processing 6-12 hours later). However, traditional methods cannot identify such dynamic causal relationships, resulting in an accuracy of only 65% ​​in assigning responsibility, often leading to misjudgments of liability.

[0007] III. Application Bottlenecks Due to Insufficient Real-Time Performance

[0008] Risk bias coefficient calculations are based on offline aggregated data (such as daily or hourly updates), which cannot meet the real-time monitoring needs of highly time-sensitive products (such as fresh produce and electronic components). For example, for highly sensitive products in transit, a sudden drop in a key indicator needs to trigger an alert within 2 minutes, but the existing system's response delay exceeds 10 minutes, resulting in a product loss rate increase of more than 20%.

[0009] IV. Blind Spots in Tracing the Sources of Anomalies Across Multiple Links

[0010] When multiple processes experience anomalies simultaneously (such as abnormal warehouse parameters + transportation route congestion + increased return rate at the sales end), the existing system lacks effective causal analysis tools, making it impossible to determine the source of the abnormal event, resulting in a chaotic handling process and an average anomaly handling time exceeding 45 minutes. Summary of the Invention

[0011] In view of this, one of the objectives of this invention is to provide a dynamic early warning method for supply chain risks based on time-series data correlation analysis. By extracting the time-series features of monitoring data and constructing a link dependency weight matrix, this method solves the problems of missed judgments in static models, coarse correlation analysis, and insufficient real-time performance, thereby achieving real-time monitoring and accurate identification of responsibilities for abnormal events throughout the entire supply chain.

[0012] One of the objectives of this invention is achieved through the following technical solution:

[0013] A dynamic early warning method for supply chain risks based on time-series data correlation analysis, characterized in that the method includes the following steps:

[0014] Step S1: Data preprocessing for time-series feature enhancement: Variable sliding window processing is applied to sensor data from each stage of the supply chain to extract relevant features. Combined with product category sensitivity thresholds, the time-series features are normalized to generate a time-series feature matrix T(n,t) = [f1,f2,…,f...]. k ], where n is the stage number, t is the timestamp, and f k This is the k-th eigenvalue;

[0015] Step S2: Modeling of the association links with dynamic dependence weights: Based on historical abnormal event data, statistically analyze the probability distribution of the time difference between the occurrence of abnormal events between link n and link j, and calculate the time lag association degree; introduce an attention mechanism and combine the cosine similarity of time series features to calculate the dynamic association weight a(n,j);

[0016] Step S3: Real-time risk score calculation and early warning: combining product category sensitivity coefficient λ i The system calculates anomaly scores for individual processes and integrates dynamic dependency weights to calculate a comprehensive risk score. When the comprehensive risk score reaches the warning threshold, an alarm is triggered.

[0017] Furthermore, it also includes step S4, namely, the determination of responsibility based on Granger causality test: for the sequence of links with continuous early warning, a vector autoregression model is constructed, and the causal relationship p-value between links is calculated through Granger causality test. If the p-value is <0.05 and the time of occurrence of the anomaly is sequential, then the responsible source link is determined.

[0018] Furthermore, in step S1, the relevant features include trend features, mean squared error, and outlier features;

[0019] Trend characteristics: including the slope and mean square error of the data within the window;

[0020] Where x iFor the i-th data point within the window, The mean;

[0021] Abnormal characteristics: the number of extreme points exceeding the preset threshold and the percentage of consecutive abnormal duration, i.e., the ratio of abnormal duration to the total window duration.

[0022] Furthermore, in step S2, the lag correlation degree W(n,j) is calculated as follows:

[0023]

[0024] The correlation degree ranges from [0,1] and is used to reflect the degree of time lag correlation between abnormal events in different processes;

[0025] The dynamic association weight a(n,j) is calculated as follows:

[0026]

[0027] Among them G n Let T(n,t), T(k,t)) be the set of links associated with link n, and let cosine(T(n,t),T(k,t)) be the cosine similarity of the temporal feature vectors.

[0028] Furthermore, in step S3, the calculation method for the process anomaly score S(n) is as follows:

[0029]

[0030] The comprehensive risk score R(n) is calculated as follows:

[0031]

[0032] Furthermore, in step S3, a multi-level early warning method is adopted, which is set according to the severity from smallest to largest, specifically as follows:

[0033] When 0.3 ≤ R(n) < 0.6, the first-level warning is triggered;

[0034] When 0.6 ≤ R(n) < 0.8, a second-level warning is triggered;

[0035] When 0.8 ≤ R(n), the third-level warning is triggered.

[0036] The second objective of this invention is to provide a dynamic early warning system for supply chain risks based on time-series data correlation analysis, the system comprising a perception layer, an analysis layer, and an application layer;

[0037] The sensing layer includes sensors applied to various stages of the supply chain; the collected monitoring data includes warehouse temperature and humidity, transportation speed, and processing equipment operating parameters.

[0038] The analysis layer includes a time-series data processing module, a dynamic dependency modeling module, and a real-time risk calculation module. The time-series data processing module processes monitoring data according to a preset sliding window size and step size. Within each sliding window, the module extracts multiple features and normalizes these features based on their corresponding sensitivity thresholds, converting feature values ​​of different magnitudes and ranges to a unified standard scale, ultimately generating a time-series feature matrix with timestamps. The dynamic dependency modeling module constructs the time-series dependencies between links and outputs dynamic correlation weights through model training. The real-time risk calculation module calculates the risk score for each link in the supply chain and triggers corresponding early warnings based on the score results.

[0039] The application layer includes an intelligent early warning module. When the comprehensive risk score obtained by the real-time risk calculation module reaches the corresponding early warning threshold, the intelligent early warning module is triggered and the early warning information is sent out in a timely manner through the integration of different communication interfaces.

[0040] Furthermore, the application layer also includes a causal analysis tool. After receiving an early warning signal, the causal analysis tool performs a Granger causality test and analyzes the causal relationship between abnormal events in each stage by constructing a vector autoregression model.

[0041] Furthermore, the system also includes a visualization management module, which generates anomaly event tracing reports and presents them through a visual interface. The reports include:

[0042] Causal relationship diagrams visually represent the causal relationships between different stages;

[0043] Time-series characteristic comparison curves compare the trends of key indicators between abnormal and related stages, helping to understand the development and impact of abnormal situations;

[0044] Historical case matching results: Find cases similar to the current abnormal event from historical data.

[0045] The beneficial effects of this invention are:

[0046] This invention deeply mines the temporal dynamic characteristics of monitoring data, extracts these characteristics, and identifies "gradual" anomalies that slowly deviate from the normal pattern and are easily missed by traditional methods (such as slow equipment deterioration, slight drift in process parameters, and hidden leaks). This can significantly reduce the false alarm rate, achieve early warning of faults or risks, and buy valuable time for preventive maintenance and proactive intervention, preventing small problems from evolving into major accidents.

[0047] In complex production chains or equipment networks, anomalies in upstream links often propagate to downstream links after a certain time delay. This invention uses time-delay correlation analysis and transmission path modeling to accurately quantify the magnitude of the time lag and the strength of the causal relationship in the propagation of anomalies between links. When an anomaly occurs in the system, it can not only locate the "symptom point" that directly triggers the alarm, but also accurately trace back to the "root cause point" that initially caused the chain reaction. It clearly defines the responsibility of different links or equipment in the abnormal event, solves the problem of "passing the buck" in multi-link coupled scenarios, and provides objective data basis for assessment, optimization, and accountability.

[0048] This invention integrates the aforementioned time-series characteristics, real-time operating conditions, and environmental parameters based on real-time access to high-frequency (minute-level or even second-level) monitoring data to construct a lightweight, low-latency online risk assessment model. This model can continuously and dynamically calculate the comprehensive risk index or failure probability of key equipment, thereby providing near real-time risk situation awareness and transforming "post-event analysis" into "in-event response." Once the risk index exceeds a preset threshold, the system can trigger differentiated alarms and predefined emergency response plans within seconds, greatly improving the response speed and accuracy of handling emergencies and effectively controlling the development of the situation.

[0049] This invention addresses complex anomaly patterns involving multiple simultaneous or successive events by constructing a causal reasoning engine. It analyzes the conditional dependencies and causal directions among multiple variables, infers the complete causal chain of anomaly propagation, and transcends single-point localization to achieve root cause diagnosis of complex anomaly scenarios. This clearly depicts the complete picture, which not only serves to accurately assign responsibility but, more importantly, provides an effective means for system optimization (such as strengthening weak links and improving control logic) and preventing the recurrence of similar problems, thereby fundamentally improving system resilience and reliability.

[0050] This invention constructs an intelligent, closed-loop industrial monitoring and diagnostic system through various cutting-edge technologies: from precise perception (detecting gradual anomalies), clarifying relationships (quantifying time-delay dependencies), and real-time decision-making (minute-level risk assessment) to deep cognition (causal tracing and accountability), it achieves a leap from "passive response" to "proactive prediction, precise location, rapid handling, and root cause management," providing strong technical support for ensuring the safe, stable, and efficient operation of complex systems.

[0051] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a system architecture diagram of the present invention;

[0055] Figure 3 Here is a flowchart for time-series feature extraction;

[0056] Figure 4 Graph of the dynamic dependency weight calculation model;

[0057] Figure 5 This is a time series diagram for multi-level early warning and time-series causal analysis. Detailed Implementation

[0058] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the preferred embodiments are for illustrative purposes only and not for limiting the scope of protection of the invention. Similar reference numerals are used throughout the drawings to denote similar elements, and the structures and apparatus illustrated herein are not necessarily drawn to scale. As used herein, the terms “component,” “system,” “interface,” etc., are intended to refer to computer-related entities, hardware, software (e.g., software in execution), and / or firmware. For example, a component may be a processor, a process operating on a processor, a controller, an object, an executable file, a program, a storage device, and / or a computer having processing means. An application operating on a server and a server, illustrated herein, may also be a component. One or more components may reside within a process, and components may be localized within a single computer and / or distributed across two or more computers. Groups of elements or other components may be described herein, wherein the term “group” may be interpreted as “one or more.”

[0059] like Figure 1 As shown, this invention provides a dynamic early warning method for supply chain risks based on time-series data correlation analysis. The method includes the following steps:

[0060] Step S1: Data preprocessing for time-series feature enhancement: Variable sliding window processing is applied to sensor data from each stage of the supply chain. This sensor data includes warehouse temperature and humidity, transportation speed, and processing equipment operating parameters. The window length for variable sliding window processing supports dynamic configuration from 5 minutes to 24 hours. By extracting relevant features and combining them with product category sensitivity thresholds, the time-series features are normalized to generate a time-series feature matrix T(n,t) = [f1,f2,…,f...]. k ], where n is the stage number, t is the timestamp, and f k This is the k-th eigenvalue;

[0061] The relevant features include trend features, mean squared error, and outlier features; each feature is described in detail below:

[0062] Trend characteristics: including the slope and mean square error of the data within the window;

[0063] Where x i For the i-th data point within the window, The mean;

[0064] Abnormal characteristics: the number of extreme points exceeding the preset threshold and the percentage of consecutive abnormal duration, i.e., the ratio of abnormal duration to the total window duration.

[0065] Step S2: Modeling the Relationship of Dynamically Dependent Weights: Based on historical abnormal event data, the probability distribution of the time difference between abnormal events occurring in link n and link j is statistically analyzed, and the time lag correlation degree is calculated; an attention mechanism is introduced, and the dynamic correlation weight a(n,j) is calculated by combining the cosine similarity of temporal features; where,

[0066] The lag correlation degree W(n,j) is calculated as follows:

[0067]

[0068] The correlation degree ranges from [0,1] and is used to reflect the degree of time lag correlation between abnormal events in different processes;

[0069] The dynamic association weight a(n,j) is calculated as follows:

[0070]

[0071] Among them G n Let T(n,t), T(k,t)) be the set of links associated with link n, and let cosine(T(n,t),T(k,t)) be the cosine similarity of the temporal feature vectors.

[0072] Step S3: Real-time risk score calculation and early warning: combining product category sensitivity coefficient λ i The system calculates anomaly scores for each stage and integrates dynamic dependency weights to calculate a comprehensive risk score. When the comprehensive risk score reaches the warning threshold, an alarm is triggered.

[0073] Furthermore, the calculation method for the process anomaly score S(n) is as follows:

[0074]

[0075] The comprehensive risk score R(n) is calculated as follows:

[0076]

[0077] As a further improvement, the alarm adopts a multi-level early warning method, specifically set as follows:

[0078] When 0.3 ≤ R(n) < 0.6, the first-level warning is triggered;

[0079] When 0.6 ≤ R(n) < 0.8, a second-level warning is triggered;

[0080] When 0.8 ≤ R(n), a level 3 warning is triggered.

[0081] Step S4, namely, the determination of responsibility based on Granger causality test: For the sequence of links with continuous early warning, construct a vector autoregression model, calculate the causal relationship p-value between links through Granger causality test, and if the p-value < 0.05 and the time of occurrence of the anomaly is sequential, then determine the link that is responsible.

[0082] Based on the design concept of the supply chain risk dynamic early warning method based on time-series data correlation analysis described above, this invention also provides a supply chain risk dynamic early warning system based on time-series data correlation analysis, such as... Figure 2 As shown, the system includes a perception layer, an analysis layer, and an application layer;

[0083] The perception layer includes sensors applied to various links in the supply chain, such as temperature and humidity sensors, GPS positioning modules, industrial cameras, weighing sensors, and RFID readers as shown in the figure. The collected monitoring data includes warehouse temperature and humidity, transportation speed, and processing equipment operating parameters.

[0084] The analysis layer includes a time-series data processing module, a dynamic dependency modeling module, and a real-time risk calculation module. The time-series data processing module processes monitoring data according to a preset sliding window size and step size. Within each sliding window, the module extracts multiple features and normalizes these features based on their corresponding sensitivity thresholds, converting feature values ​​of different magnitudes and ranges to a unified standard scale, ultimately generating a time-series feature matrix with timestamps. The dynamic dependency modeling module constructs the time-series dependencies between links and outputs dynamic correlation weights through model training. The real-time risk calculation module calculates the risk score for each link in the supply chain and triggers corresponding early warnings based on the score results.

[0085] The application layer includes an intelligent early warning module. When the comprehensive risk score obtained by the real-time risk calculation module reaches the corresponding early warning threshold, the intelligent early warning module is triggered and sends out the early warning information in a timely manner through the integration of different communication interfaces.

[0086] The application layer also includes a causal analysis tool. After receiving an early warning signal, the causal analysis tool performs a Granger causality test and analyzes the causal relationship between abnormal events in each stage by constructing a vector autoregression model.

[0087] The application layer also includes a visualization management module, which generates an abnormal event tracing report and presents it in a visualization interface. The report includes: (1) a causal relationship map: visually showing the causal relationship between each link; (2) a time series feature comparison curve: comparing the key indicator trends of the abnormal link and the related link to help understand the development and impact of the abnormal situation; (3) historical case matching results: finding cases similar to the current abnormal event from historical data.

[0088] The following will further illustrate the specific implementation methods of the above modules through several specific embodiments.

[0089] Example 1

[0090] like Figure 3 As shown in this embodiment, the time-series data processing module will be described in detail. This module is responsible for acquiring data collected by sensors at various stages of the supply chain in real time. The sensor monitoring content covers warehouse temperature and humidity, transportation speed, and operating parameters of processing equipment. To ensure stable data access, multiple protocols such as Modbus and MQTT are supported, and it is compatible with sensor devices from different manufacturers and of different types. Modbus (RTU / TCP) is the de facto standard protocol in the field of industrial automation, and its support is crucial. It allows the system to seamlessly connect to a large number of existing devices, such as PLCs, DCSs, HMIs, and various industrial sensors (temperature, pressure, flow meters, etc.), covering RS-485 serial bus and Ethernet TCP / IP network environments. MQTT, as a lightweight, publish / subscribe model-based preferred protocol for IoT, solves the problem of efficient and reliable data transmission in wide area network, low bandwidth, or unstable network environments (such as 4G / 5G, NB-IoT). It is particularly suitable for distributed sensor networks (such as environmental monitoring and energy metering) and cloud-edge collaborative architectures.

[0091] After acquiring the data, the module processes the data according to the preset sliding window size and step size. For example, the window size is set to 15 minutes and the step size is 5 minutes. A segment of data is extracted every 5 minutes for feature extraction.

[0092] Within each sliding window, the module extracts various features, including trend features, outlier features, and mean squared error. For trend features, the slope reflects the trend of data change (such as the rate of temperature increase or decrease); the mean squared error measures the degree of data fluctuation; for outlier features, it counts the number of extreme points exceeding preset thresholds and the percentage of consecutive outlier durations.

[0093] The module standardizes the combined product category characteristics because different product categories have varying degrees of sensitivity to environmental factors. For example, highly sensitive product categories are more sensitive to changes in temperature and humidity. Therefore, the extracted features are normalized based on their corresponding sensitivity thresholds, transforming feature values ​​of different magnitudes and ranges to a unified standard scale. This ultimately generates a time-series feature matrix with timestamps. This matrix integrates feature data from different time points in each stage, providing a structured data foundation for subsequent analysis.

[0094] At the technical level, this embodiment uses Python's Pandas library to implement sliding window computation. Pandas provides rich data processing and analysis functions, enabling efficient slicing, computation, and transformation of time series data. By writing the corresponding code logic, and utilizing Pandas' DataFrame data structure and function methods, the sliding window partitioning, feature calculation, and data standardization are implemented.

[0095] To ensure the timeliness and stability of data access, the following optimization measures were taken:

[0096] (1) Hybrid transmission protocol settings for network transmission: In order to meet the real-time requirements, this embodiment adopts a hybrid transmission protocol of UDP and TCP for network transmission. For sensor data with extremely high real-time requirements (such as temperature data in cold chain transportation), UDP protocol is used first to reduce transmission delay; for data with high accuracy requirements and relatively low real-time requirements (such as timed reporting data of processing equipment operating parameters), TCP protocol is used to ensure reliable data transmission.

[0097] (2) A double buffering mechanism is adopted: a data buffer is set up at the data receiving end, one buffer is used to receive new data, and the other buffer is used for data processing. When one buffer is full of data, the process immediately switches to the other buffer for data processing, while the first buffer starts to receive new data, thereby improving data processing efficiency;

[0098] (3) Network optimization: Regularly detect and optimize network connections. If the network latency exceeds a certain threshold, automatically re-establish the connection or switch the network channel to ensure that the data access latency is controlled within ≤50ms, providing real-time and accurate data support for subsequent risk assessment.

[0099] Example 2

[0100] In this embodiment, the dynamic dependency modeling module will be described in detail. This module is mainly used to construct the temporal dependencies between stages and output dynamic association weights through model training. Figure 4As shown, firstly, historical anomaly event data is retrieved from a Hadoop-based distributed storage system. Hadoop's distributed nature enables the efficient storage and management of large amounts of historical data. This historical data contains detailed information about past anomalies at each stage of the supply chain, such as the time of occurrence, the involved stages, and related monitoring data. Using this data, the module calculates the time lag correlation between stages and constructs a time-series dependency matrix. For example, for the warehousing and processing stages, the probability distribution of the time difference between the occurrence of anomalies in both stages is statistically analyzed to determine the time lag correlation. If it is found that 80% of microbial contamination events in the processing stage occur within 6-12 hours after high temperature (>10℃) in the warehousing stage, the dependency between the two stages is calculated and recorded in the time-series dependency matrix. For example, regarding the transportation and sales stages, the probability distribution of the time difference between transportation delays and abnormal inventory buildup in the sales stage is statistically analyzed. If it is found that 70% of abnormal inventory buildup in the sales stage occurs more than 12 hours after a transportation delay, the corresponding dependency is recorded in the matrix to demonstrate the universality of this method across different stage combinations. Next, an attention mechanism is introduced, combining the cosine similarity of the temporal features of the current stage and related stages to assign dynamic weights to the paired related stages. This process considers the feature similarity of different stages at different time points, making the weight allocation more reasonable.

[0101] In this embodiment, an attention mechanism model based on the Transformer architecture is adopted. The model consists of three Transformer layers, each containing eight attention heads. The input is the temporal feature vectors of the current stage and related stages. After processing through multi-head attention layers, feedforward neural network layers, and residual connections, dynamic weights are output. The model is trained using the Adam optimizer with a learning rate of 0.001 and a mean squared error loss function. By continuously training the attention mechanism model, accurate dynamic correlation weights between stages are output. These weights can reflect the degree of correlation and influence of different stages in abnormal events.

[0102] At the technical level, this embodiment uses the TensorFlow deep learning framework for model training. TensorFlow provides rich tools for building and training neural networks, facilitating the construction of attention mechanism models. The model's structure, inputs, outputs, and training algorithms are defined by writing corresponding code. The model input consists of historical anomaly event data obtained from the Hadoop distributed storage system and time-series feature data of the current stage and related stages. The output is the dynamic correlation weights between stages. Daily online incremental learning is supported during model training. Newly generated historical data is retrieved from the Hadoop system daily, merged with previously trained data, and the model is retrained. During training, the learning rate is dynamically adjusted. When the model loss decreases by less than 0.001 within five consecutive training epochs, the learning rate is reduced to 0.8 times its original value to accelerate model convergence, enabling the model to adapt to dynamic data changes and more accurately reflect the dynamic dependencies between stages.

[0103] Example 3

[0104] This embodiment will provide a detailed description of the real-time risk calculation module. The core function of this module is to calculate the risk score for each link in the supply chain and trigger corresponding warnings based on the score results. When calculating the risk score, the single-link anomaly score is first calculated based on the single-link anomaly scoring model. Combining the product category sensitivity coefficient (e.g., a sensitivity coefficient of 1.5 for high-sensitivity products and 1.0 for low-sensitivity products), the ratio of the number of extreme points within the window to the window length, and the ratio of the duration of continuous anomalies to the window length, are comprehensively considered to determine the single-link anomaly score. For example, within a 15-minute window, if the number of extreme points where the temperature in the warehousing link exceeds the preset threshold is large, and the duration of continuous anomalies is significant, and this link deals with high-sensitivity products, then its single-link anomaly score will be correspondingly high.

[0105] The method for determining the sensitivity coefficient of each product category is as follows: A large amount of historical data on different product categories under various environmental conditions is collected, including the frequency and severity of various anomalies. For each product category, the impact on quality and supply chain risks is calculated when different characteristic indicators (such as temperature, humidity, and transportation time) exceed normal ranges. Through statistical analysis and machine learning algorithms (such as logistic regression), the sensitivity coefficient for each product category for different characteristic indicators is determined. For example, for highly sensitive products, changes in temperature and humidity have a significant impact on their quality; after data analysis, their temperature sensitivity coefficient is determined to be 1.8, and their humidity sensitivity coefficient to be 1.6. For low-sensitivity products, which are relatively less sensitive to humidity changes, the humidity sensitivity coefficient is set to 1.2.

[0106] After obtaining the anomaly score for each individual stage, the overall risk score for the related stages is calculated by combining the dynamic correlation weights output by the dynamic dependency modeling module. The anomaly scores for the individual stages and the related stages are then weighted and summed using dynamic weights to obtain the overall risk score. For example, if the warehousing stage is related to the transportation stage, and if the warehousing stage has a high anomaly score, and the transportation stage also exhibits some anomalies, and the dynamic correlation weight between the two is large, then the overall risk score for the warehousing stage will increase significantly.

[0107] In addition, as shown in the figure Figure 5 As shown, the module supports multi-level warning threshold configuration. Different warning levels can be set according to different ranges of the comprehensive risk score. The table below shows the specific multi-level warning threshold settings. For easy identification, they are divided into three levels according to severity: yellow warning, orange warning, and red warning.

[0108] Table 1 Multi-level Early Warning Threshold Setting Table

[0109]

[0110] Examples of alert levels include yellow (0.3≤R(n)<0.6), orange (0.6≤R(n)<0.8), and red (R(n)≥0.8). Different alert levels correspond to different response measures: During a yellow alert, the system automatically pushes detailed abnormal data information to the administrator's mobile app, including the abnormal process, abnormal indicators, and degree of abnormality; simultaneously, it initiates high-frequency data collection of relevant processes, increasing the collection frequency to once per minute to more closely monitor the development of the abnormal situation. During an orange alert, in addition to pushing information and high-frequency data collection, some non-critical operational permissions related to the abnormal process are frozen, such as prohibiting manual adjustment of warehouse environment parameters and suspending non-emergency dispatch of transport vehicles, pending manual confirmation and further processing. During a red alert, the alert information is broadcast throughout the network to notify all relevant personnel; all operations in related processes are automatically suspended to prevent further escalation of the abnormal situation; and drones or robots are immediately dispatched to the site for verification to obtain more accurate on-site information.

[0111] At the technical level, this embodiment utilizes an Apache Flink in-memory computing cluster to achieve minute-level risk scoring calculations. Apache Flink is a distributed stream processing framework characterized by high concurrency and low latency. By deploying corresponding computing tasks within the cluster, Flink's stream processing capabilities are leveraged to rapidly calculate the real-time input time-series feature matrix and dynamic correlation weights. For algorithm optimization, parallel computing and distributed caching technologies are employed. For single-stage anomaly scoring calculations, data from different stages is distributed across different computing nodes for parallel processing, improving computational efficiency. For the calculation results of dynamic correlation weights, distributed caching is used for storage, reducing redundant calculations. Regarding cluster parameter configuration, parameters such as parallelism and memory allocation are rationally set based on server hardware resources and actual business volume. For example, for a computing node with an 8-core CPU and 32GB of memory, the parallelism is set to 4, and each task is allocated 8GB of memory. Through these optimized algorithms and configured cluster parameters, single-stage scoring latency is ensured to be ≤10ms, enabling real-time risk assessment and early warning, providing strong support for supply chain risk management.

[0112] Example 4

[0113] This embodiment will provide a detailed description of the intelligent early warning module, which plays a crucial role in decision-making and response within the entire system. This module is triggered when the comprehensive risk score derived by the real-time risk calculation module reaches the corresponding early warning threshold. It possesses the capability to trigger multi-channel early warnings, and by integrating different communication interfaces, it can promptly send early warning information to relevant personnel via more than 10 methods, including SMS, WeChat, and API, ensuring widespread information dissemination and timely awareness.

[0114] Upon receiving an early warning signal, the module performs a Granger causality test, constructing a vector autoregressive (VAR) model to analyze the causal relationships between abnormal events in each stage. For example, when abnormalities occur simultaneously in the warehousing and transportation stages, the Granger causality test can determine whether the abnormality in the warehousing stage caused the abnormality in the transportation stage, or whether the abnormality in the transportation stage affected the warehousing stage, or whether there is no causal relationship between the two.

[0115] Simultaneously, the module generates an anomaly incident tracing report. The report includes a causal relationship graph, visually displaying the causal connections between each stage; a time-series feature comparison curve, comparing the trends of key indicators between the anomaly stage and related stages, helping users intuitively understand the development and impact of the anomaly; and historical case matching results, finding cases similar to the current anomaly from historical data, providing reference experience for handling the current anomaly, and assisting managers in quickly locating the responsible stage and formulating effective response measures.

[0116] In terms of technical implementation, this embodiment adopts the Granger test library integrated with the R language. R language has powerful functions in statistical analysis and causal relationship testing. Granger causality testing is implemented by calling relevant R language functions and algorithms within the system. Specifically, an R language runtime environment is set up within the system, and R language script files are called through the system interface. The data requiring causal testing is organized and transmitted according to the R language format requirements, and the `grangertest` function in R is called to perform the causal relationship test. During the call, appropriate parameters, such as the lag order, are set, determined based on the actual data and business needs; generally, the lag order is set to 1-3. Simultaneously, a multi-channel early warning interface is developed, utilizing existing communication technologies and platforms to integrate various early warning methods such as SMS and WeChat. For SMS early warnings, the API interface provided by the SMS gateway provider is used to encapsulate and send early warning information according to the specified format; for WeChat early warnings, the early warning information is pushed to relevant users who follow the WeChat official account through the WeChat official account development interface. During development, technical difficulties such as interface compatibility with different communication platforms and data format conversion were resolved, ensuring the stability and accuracy of information push.

[0117] Example 5

[0118] This embodiment provides a detailed explanation of the system linkage after an early warning is triggered, supporting on-site verification triggered by drones / robots. Specifically, it interfaces with the control system of the drone and / or robot. When an early warning is triggered, it automatically sends instructions to the drone or robot to proceed to the relevant site for verification, obtaining more accurate on-site information and improving the efficiency and accuracy of handling abnormal events. The interface uses the MQTT protocol to establish a stable communication connection between the system and the drone / robot control system. When an early warning is triggered, the system sends instructions containing information such as the site location and verification task requirements to the drone / robot according to the MQTT protocol format. After receiving the instructions, the drone / robot proceeds to the designated site according to a preset program to perform verification tasks such as image acquisition and data monitoring. After completing the verification, the collected data and image information are transmitted back to the system via the same MQTT protocol. The system analyzes and processes the transmitted data to assist management personnel in making decisions.

[0119] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0120] Furthermore, the procedures described herein may be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by the context. The procedures described herein (or variations and / or combinations thereof) may be executed under the control of one or more computer systems configured with executable instructions, and may be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program comprises a plurality of instructions executable by one or more processors.

[0121] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices, etc. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention described herein includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques described in the invention, the invention also includes the computer itself.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A dynamic early warning method for supply chain risks based on time-series data correlation analysis, characterized by: The method includes the following steps: Step S1: Data preprocessing for time-series feature enhancement: Variable sliding window processing is applied to sensor data from each stage of the supply chain to extract relevant features. Combined with product category sensitivity thresholds, the time-series features are normalized to generate a time-series feature matrix T(n,t) = [f1,f2,…,f...]. k ], where n is the stage number, t is the timestamp, and f k This is the k-th eigenvalue; Step S2: Modeling of the association links with dynamic dependence weights: Based on historical abnormal event data, statistically analyze the probability distribution of the time difference between the occurrence of abnormal events in link n and link j, and calculate the time lag association degree; introduce an attention mechanism and combine the cosine similarity of time series features to calculate the dynamic association weight a(n,j); Step S3: Real-time risk score calculation and early warning: combining product category sensitivity coefficient λ i The system calculates anomaly scores for individual processes and integrates dynamic dependency weights to calculate a comprehensive risk score. When the comprehensive risk score reaches the warning threshold, an alarm is triggered.

2. The supply chain risk dynamic early warning method and system based on time-series data correlation analysis according to claim 1, characterized in that: It also includes step S4, namely, the determination of responsibility based on Granger causality test: For the sequence of links with continuous early warning, a vector autoregression model is constructed, and the causal relationship p-value between links is calculated through Granger causality test. If the p-value is <0.05 and the time of occurrence of the anomaly is sequential, the link that is responsible is determined.

3. The supply chain risk dynamic early warning method based on time-series data correlation analysis according to claim 1, characterized in that: In step S1, the relevant features include trend features, mean squared error, and outlier features. Trend characteristics: including the slope and mean square error of the data within the window; Where x i For the i-th data point within the window, The mean; Abnormal characteristics: the number of extreme points exceeding the preset threshold and the percentage of consecutive abnormal duration, i.e., the ratio of abnormal duration to the total window duration.

4. The supply chain risk dynamic early warning method based on time-series data correlation analysis according to claim 1, characterized in that: In step S2, the lag correlation degree W(n,j) is calculated as follows: The correlation degree ranges from [0,1] and is used to reflect the degree of time lag correlation between abnormal events in different processes; The dynamic association weight a(n,j) is calculated as follows: Among them G n Let T(n,t), T(k,t)) be the set of links associated with link n, and let cosine(T(n,t),T(k,t)) be the cosine similarity of the temporal feature vectors.

5. The supply chain risk dynamic early warning method based on time-series data correlation analysis according to claim 4, characterized in that: In step S3, the abnormality score S(n) is calculated as follows: The comprehensive risk score R(n) is calculated as follows:

6. The supply chain risk dynamic early warning method and system based on time-series data correlation analysis according to claim 5, characterized in that: In step S3, a multi-level early warning method is adopted, set according to the severity from smallest to largest, specifically as follows: When 0.3 ≤ R(n) < 0.6, the first-level warning is triggered; When 0.6 ≤ R(n) < 0.8, a second-level warning is triggered; When 0.8 ≤ R(n), a level 3 warning is triggered.

7. A dynamic early warning system for supply chain risks based on time-series data correlation analysis, characterized in that: The system includes a perception layer, an analysis layer, and an application layer; The sensing layer includes sensors applied to various stages of the supply chain; the collected monitoring data includes warehouse temperature and humidity, transportation speed, and processing equipment operating parameters. The analysis layer includes a time series data processing module, a dynamic dependency modeling module, and a real-time risk calculation module. The time series data processing module processes the monitoring data according to a preset sliding window size and step size. Within each sliding window, the time series data processing module extracts multiple features and normalizes the extracted features according to their corresponding sensitivity thresholds. It converts feature values ​​of different magnitudes and ranges to a unified standard scale and finally generates a time series feature matrix with timestamps. The dynamic dependency modeling module is used to construct the temporal dependency relationship between links and output dynamic association weights through training the model. The real-time risk calculation module is used to calculate the risk score of each link in the supply chain and trigger corresponding early warnings based on the score results; The application layer includes an intelligent early warning module. When the comprehensive risk score obtained by the real-time risk calculation module reaches the corresponding early warning threshold, the intelligent early warning module is triggered and the early warning information is sent out in a timely manner through the integration of different communication interfaces.

8. The supply chain risk dynamic early warning system based on time-series data correlation analysis according to claim 7, characterized in that: The application layer also includes a causal analysis tool. After receiving an early warning signal, the causal analysis tool performs a Granger causality test and analyzes the causal relationship between abnormal events in each stage by constructing a vector autoregression model.

9. The supply chain risk dynamic early warning system based on time-series data correlation analysis according to claim 7, characterized in that: The system also includes a visualization management module, which generates anomaly event tracing reports and presents them through a visual interface. The reports include: Causal relationship diagrams visually represent the causal relationships between different stages; Time-series characteristic comparison curves compare the trends of key indicators between abnormal and related stages, helping to understand the development and impact of abnormal situations; Historical case matching results: Find cases similar to the current abnormal event from historical data.

Images