A supply chain production and marketing coordination abnormal event real-time detection and visualization method

Abstract

This application discloses a real-time detection and visualization method for abnormal events in supply chain production and sales collaboration, belonging to the field of computer technology. By synchronously collecting and deeply fusing multimodal data, it eliminates the data fragmentation phenomenon in traditional monitoring systems, enabling coupled analysis of abnormal indicators across physiological, collaborative, and equipment dimensions within a unified framework. The introduction of a dynamic graph neural network not only quantifies the abnormal propagation path between supply chain entities but also reveals the relative impact weight of each modality indicator on overall risk through contribution distribution, thereby improving the real-time performance and accuracy of abnormal detection. The visualization interface design transforms abstract abnormal probabilities into an interactive multimodal evidence chain, allowing managers to locate the root cause of abnormalities without relying on subjective experience. This solves the problems of delayed early warning and difficulty in tracing caused by data silos in existing technologies, ultimately achieving efficient handling of abnormal events in the supply chain production and sales collaboration process.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for real-time detection and visualization of abnormal events in supply chain production and sales coordination. Background Technology

[0002] In modern supply chain management, the stability and efficiency of the production and sales coordination process are crucial. However, this process involves multiple entities such as production, warehousing, logistics, and sales, and its operation is subject to complex influences from multiple dimensions, including the physiological state of operators, the efficiency of inter-departmental collaboration, and the operating status of production line equipment.

[0003] Currently, the industry generally uses independent monitoring systems, such as monitoring the physical parameters of production line equipment or issuing simple keyword alerts for business communications. These methods have significant limitations: first, they can only perform isolated analysis on data of a single modality and cannot capture the inherent coupling relationship between abnormal events of different dimensions; second, due to the lack of deep fusion and collaborative analysis of multi-source heterogeneous data, existing technologies struggle to locate the root cause of anomalies in complex data in real time and accurately, and to quantify the impact of each factor on the overall anomaly, resulting in delayed early warnings and difficulties in tracing the source.

[0004] Therefore, there is an urgent need in this field for a technical solution that can break down data silos and achieve real-time coupled analysis and visual traceability of multimodal anomalies, so as to comprehensively improve the anomaly detection and handling capabilities in the production and sales collaboration process of the supply chain. Summary of the Invention

[0005] This application provides a method for real-time detection and visualization of abnormal events in supply chain production and sales collaboration, which can improve the ability to detect and handle abnormalities in the supply chain production and sales collaboration process. The technical solution is as follows: On the one hand, a method for real-time detection and visualization of abnormal events in supply chain production and sales coordination is provided, the method including: By deploying biosensors, enterprise communication system interfaces, and spectrum acquisition devices at key workstations on the production line, operator eye movement trajectories and heart rate variability sequences, cross-departmental business communication texts, and production line electromagnetic spectrum signals are acquired simultaneously to form a multimodal data stream. Physiological state analysis is performed on the eye movement trajectory and heart rate variability sequence to obtain the real-time attention distraction index and physiological stress level; sentiment semantic analysis is performed on the cross-departmental business communication text to obtain the collaborative health score for specific production work orders; spectral features are extracted from the electromagnetic spectrum signal of the production line to obtain the spectral anomaly offset and production line beat disturbance coefficient. Based on the real-time attention distraction index, the physiological stress level, the collaborative health score, the spectrum anomaly offset, and the production line beat disturbance coefficient, a multimodal anomaly coupling analysis is performed using a dynamic graph neural network to output the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modal indicator. Based on the comprehensive anomaly probability and the contribution distribution of each modal indicator, the corresponding eye-tracking trajectory fragments, communication text fragments, and electromagnetic spectrum fragments are extracted from the multimodal data stream through time window matching and entity identification association. A visualization interface that integrates business topology and multimodal source tracing evidence is generated, in which high-risk entities are synchronously associated with and displayed with the extracted data fragments.

[0006] On the one hand, a computer device is provided, the computer device including one or more processors and one or more memories, the one or more memories storing at least one computer program, the computer program being loaded and executed by the one or more processors to realize the method for real-time detection and visualization of abnormal events in supply chain production and sales coordination.

[0007] On the one hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to realize a method for real-time detection and visualization of abnormal events in supply chain production and sales coordination.

[0008] On the one hand, a computer program product or computer program is provided, which includes program code stored in a computer-readable storage medium. The processor of a computer device reads the program code from the computer-readable storage medium and executes the program code, causing the computer device to execute the aforementioned method for real-time detection and visualization of abnormal events in supply chain production and sales coordination. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the implementation environment of a method for real-time detection and visualization of abnormal events in supply chain production and sales collaboration provided in an embodiment of this application; Figure 2 This is a flowchart of a method for real-time detection and visualization of abnormal events in supply chain production and sales coordination, provided in an embodiment of this application. Figure 3This is a partial flowchart of a method for real-time detection and visualization of abnormal events in supply chain production and sales coordination provided in an embodiment of this application; Figure 4 This is a partial flowchart of another method for real-time detection and visualization of abnormal events in supply chain production and sales coordination provided in an embodiment of this application; Figure 5 This is a partial flowchart of another method for real-time detection and visualization of abnormal events in supply chain production and sales coordination provided in the embodiments of this application; Figure 6 This is a partial flowchart of another method for real-time detection and visualization of abnormal events in supply chain production and sales coordination provided in the embodiments of this application; Figure 7 This is a partial flowchart of another method for real-time detection and visualization of abnormal events in supply chain production and sales collaboration provided in the embodiments of this application. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0012] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0013] Artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain better results.

[0014] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge sub-models to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0015] Spectrum acquisition equipment: refers to a device used to capture, record, and analyze electromagnetic wave signals within a specific frequency range. In this application, it specifically refers to a broadband radio frequency sensor deployed around production equipment to monitor the electromagnetic radiation generated during production line operation in order to obtain the electromagnetic spectrum signal of the production line. Its function is to sense the electromagnetic state of the production environment as an indirect indicator of the equipment's operating status.

[0016] Real-time Attention Distraction Index: A numerical indicator that quantifies the degree to which an operator's visual attention deviates from their primary task within a specific time window. This index is calculated by analyzing eye movement data (such as fixation points and saccades). For example, a higher fixation point distribution entropy value indicates a more dispersed fixation point and less focused attention. It is a dynamically updated variable that reflects the operator's instantaneous cognitive load and concentration.

[0017] Physiological stress level: A physiological indicator that quantitatively characterizes the activation state of an operator's autonomic nervous system, particularly the level of sympathetic nerve excitation. This level is primarily derived through analysis of heart rate variability sequences. An increased ratio of low-frequency power to high-frequency power, and a decreased standard deviation of heart rate intervals, are generally markers of elevated physiological stress levels. It reflects the operator's intrinsic physiological tension caused by time pressure, work complexity, and other factors.

[0018] Sentiment semantic analysis: a technique combining sentiment computing and natural language processing, aimed at extracting sentiment orientation and semantic content from text. In this application, the method not only calculates basic sentiment polarity values ​​(positive / negative) and sentiment intensity values ​​(strong / weak), but also incorporates contextual information such as dialogue thread structure and semantic role labeling results for refinement, in order to more accurately understand the emotions and intentions conveyed in business communications.

[0019] Collaboration Health Score: A comprehensive quantitative score assessing the efficiency and quality of cross-departmental business collaboration for a specific production work order. This score is calculated by weighting and integrating multiple dimensions, including sentiment polarity, sentiment intensity, communication response latency, interaction frequency, and semantic conflict level. A lower score indicates greater communication barriers between departments during the completion of the work order, and a less healthy collaboration status.

[0020] Spectral Anomaly Offset: A quantized value characterizing the degree of deviation of the current electromagnetic spectrum signal from a reference normal spectrum template. This offset is calculated by comparing the real-time spectrum with the reference template in terms of differences in characteristic frequency band energy distribution, spectral shape, etc. It is used to indicate anomalies in the electromagnetic environment of the production line, which may originate from equipment failure, external interference, or non-standard operation.

[0021] Production line cycle time disturbance coefficient: a quantitative indicator characterizing the stability of the production process rhythm. This coefficient is extracted by analyzing the periodicity or volatility of the autocorrelation coefficient sequence of the electromagnetic spectrum signal envelope. Disruption of the regularity of the correlation coefficient (e.g., increased variance) indicates abnormalities in the start-up, shutdown, and operation cycles of production equipment, reflecting disturbances in the overall production line cycle time.

[0022] Multimodal anomaly coupling analysis: an analytical method aimed at exploring the interactions, enhancements, or cancellations among multiple anomaly indicators from different data modalities, rather than performing isolated analyses. In this application, it specifically refers to using dynamic graph neural networks to fuse anomaly indicators from physiological, collaborative, and spectral modalities, analyze how they collectively influence and lead to the overall anomaly state of supply chain entities, and output the contribution distribution of each modality.

[0023] The quality inspection station is a specific work area on the production line responsible for inspecting and testing semi-finished or finished products to determine whether they meet quality standards. Distractions in the operator's attention at this station can directly lead to missed inspections or misjudgments.

[0024] Precision assembly stations are specific work areas on production lines where components are assembled with high precision and strict requirements. Physiological stress or attention problems of operators at these stations can lead to assembly errors and serious quality defects.

[0025] A non-invasive eye tracker is a device that tracks eye movements using remote optical sensors (such as near-infrared cameras) without direct contact with the human body or the implantation of sensors. This ensures that deployment in industrial settings will not interfere with the operator's normal work.

[0026] A heart rate sensor is a device used to detect and record heartbeat activity. In the context of this application, it generally refers to a sensor capable of non-invasively measuring heart rate variability sequences via photoplethysmography (PPG), such as a wristband or chest patch device.

[0027] Fixation point distribution entropy: Based on the concept of information entropy, this index quantifies the degree of disorder in the spatial distribution of fixations in eye movement trajectories. A higher entropy value indicates a more random and scattered distribution of fixations on the screen or within the visual field, suggesting a lack of focus. A lower entropy value indicates that fixations are concentrated in a few key areas for an extended period, indicating better focus.

[0028] Pupil diameter coefficient of variation: The ratio of the standard deviation to the mean of pupil diameter over a certain period of time. This coefficient is a sensitive indicator reflecting cognitive load and physiological arousal. Generally, higher cognitive load or emotional fluctuations will cause larger changes in pupil diameter, thus leading to an increase in this coefficient.

[0029] Power spectral density in two specific frequency bands after frequency domain analysis of heart rate variability sequences using low-frequency and high-frequency power.

[0030] Low-frequency power: usually corresponds to the joint regulatory activity of the sympathetic and parasympathetic nervous systems, and is related to stress and mood regulation.

[0031] High-frequency power: Primarily synchronized with respiratory rhythm, reflecting parasympathetic (vagus nerve) activity. The ratio of these two values ​​is a key indicator for assessing autonomic balance and stress state.

[0032] Sentiment polarity value: Indicates the basic orientation of the text's sentiment, such as positive, negative, or neutral. It is usually a discrete label or a numerical value within a continuous range (e.g., -1 to +1).

[0033] Emotional intensity value: This indicates the intensity of the emotion expressed in the text. It is usually a positive value, and the higher the value, the stronger the emotion.

[0034] Semantic conflict level: A metric that quantifies the degree of contradiction or inconsistency in logic, fact, or intent between different statements or semantic units in a text or dialogue. It is achieved by analyzing the topology of the semantic association network (such as whether there are mutually exclusive concept nodes) or by using a pre-trained language model to detect contradictions.

[0035] Dialogue thread structure: refers to the tree-like or chain-like structure formed by dividing and organizing a continuous communication record (such as an email chain or chat log) according to topics or reply relationships. It clarifies which statements are responses to which previous statements.

[0036] Semantic role labeling results: A natural language processing task used to label the semantic role of each component in a sentence, such as "who" (agent) did "what" (action) to "who" (patient). This helps in understanding the responsible parties and key actions in communication content.

[0037] Signal envelope: refers to the outline of an oscillating signal, reflecting the overall trend of signal amplitude change over time. For electromagnetic spectrum signals, its envelope can reflect the overall energy output cycle of production line equipment.

[0038] Autocorrelation coefficient sequence: A sequence obtained by calculating the similarity of a signal to itself at different time delays. The peak value and periodicity of this sequence can effectively reveal hidden repetitive patterns or beats in the signal.

[0039] Variational Mode Decomposition (VMD) is a completely non-recursive signal decomposition method used to adaptively decompose the original signal into multiple quasi-orthogonal eigenmode components. Compared with classical empirical mode decomposition, VMD has a more robust mathematical foundation and better noise robustness, making it suitable for extracting characteristic frequency band components with practical physical meaning from complex electromagnetic spectrum signals from production lines.

[0040] Hadamard product: a matrix operation, also known as element-wise product. When two matrices of the same dimension are subjected to the Hadamard product, each element of the resulting matrix is ​​equal to the product of the corresponding elements of the original two matrices. In this embodiment, it is used to fuse a standardized edge weight matrix with the original graph adjacency matrix, thereby achieving dynamic and refined adjustment of the graph structure.

[0041] Entity identification and resolution technology refers to the technical process of identifying and extracting information (such as ID, code) that uniquely identifies a specific supply chain entity (such as a specific machine, work order, operator) from a multimodal data stream.

[0042] Entity identification mapping technology refers to the technique of establishing and maintaining the association between entity identifiers across different systems or data views. For example, associating a graphical node (representing an entity) in a visualization interface with a record in the backend database and its related eye-tracking trajectory fragments and communication text fragments can enable interactive responses between views.

[0043] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0044] Figure 1 This is a schematic diagram illustrating the implementation environment of a method for real-time detection and visualization of abnormal events in supply chain production and sales collaboration, as provided in an embodiment of this application. (See attached diagram.) Figure 1 This implementation environment may include node 110 and server 140.

[0045] Node 110 is connected to server 140 via a wireless or wired network. Node 110 has an application installed and running that supports real-time detection and visualization of abnormal events in supply chain production and sales coordination.

[0046] Server 140 is a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Server 140 can provide background services for applications running on node 110.

[0047] Since the embodiments of this application involve a relatively complex calculation process, node 110 may not be able to meet the computing power requirements. Therefore, node 110 will send the relevant data to server 140, and server 140 with high computing power will execute the technical solution provided in the embodiments of this application.

[0048] In traditional supply chain production and sales collaboration processes, the problem of multimodal data silos prevents the effective capture of the inherent coupling relationships between anomalies. Existing technologies only perform isolated analysis on single-dimensional data, failing to achieve deep fusion and collaborative processing of multi-source heterogeneous data. This results in a lack of real-time coupled analysis and visual tracing capabilities for anomalies, leading to delayed early warning information and difficulties in locating the root causes of anomalies, thus hindering the stability and response efficiency of the supply chain collaboration process.

[0049] For example, in the production execution phase of the electronics manufacturing supply chain, eye-tracking of operators at key workstations on the production line was detected by biosensors as a sign of distraction, and analysis of heart rate variability sequences revealed abnormally elevated levels of physiological stress. Simultaneously, sentiment analysis of cross-departmental business communication texts obtained through the enterprise communication system interface revealed a continuous decline in collaborative health scores for specific production work orders. Electromagnetic spectrum signals from the production line were captured by spectrum acquisition equipment, and spectrum feature extraction indicated that abnormal spectrum offsets exceeded thresholds and that the production line cycle time disturbance coefficients underwent abrupt changes. Furthermore, existing monitoring systems process these data streams separately, failing to establish a correlation mechanism between abnormal operator physiological states and abnormal equipment operation. This results in isolated alarms for abnormal events, an inability to identify the coupling propagation path between the distraction index and abnormal spectrum offsets, and thus, an accumulation of production interruption risks.

[0050] If the above problems are not addressed, abnormal events in the supply chain production and sales coordination process will not be able to be analyzed in real time, the delayed early warning information will make it difficult to trace the root cause of the anomalies, decision-making delays and reduced efficiency in handling will continue, the overall stability of the supply chain will be under systemic threat, and the reliability and continuity of the coordination process will face uncontrollable risks.

[0051] To address this, this application proposes a method for real-time detection and visualization of abnormal events in supply chain production and sales coordination, see [link to relevant documentation]. Figure 2 Taking the server as the executing entity as an example, the following steps are included.

[0052] 201. By deploying biosensors, enterprise communication system interfaces, and spectrum acquisition devices at key workstations on the production line, operator eye movement trajectories and heart rate variability sequences, cross-departmental business communication texts, and production line electromagnetic spectrum signals are simultaneously acquired to form a multimodal data stream.

[0053] 202. Perform physiological state analysis on eye movement trajectories and heart rate variability sequences to obtain real-time attention distraction index and physiological stress level. Perform sentiment semantic analysis on cross-departmental business communication texts to obtain collaborative health scores for specific production work orders. Extract spectral features from production line electromagnetic spectrum signals to obtain spectral anomaly offset and production line cycle time disturbance coefficient.

[0054] 203. Based on real-time attention distraction index, physiological stress level, collaborative health score, spectrum anomaly offset and production line beat disturbance coefficient, multimodal anomaly coupling analysis is performed through dynamic graph neural network to output the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modality indicator.

[0055] 204. Based on the comprehensive anomaly probability and the contribution distribution of each modality indicator, the corresponding eye-tracking trajectory fragments, communication text fragments and electromagnetic spectrum fragments are extracted from the multimodal data stream through time window matching and entity identification association, and a visualization interface that integrates business topology and multimodal source tracing evidence is generated.

[0056] High-risk entities are displayed in sync with the extracted data segments.

[0057] This embodiment relates to a method for real-time detection and visualization of abnormal events in supply chain production and sales collaboration. In practical applications, biosensors can be implemented using wearable physiological monitoring devices, such as photoplethysmography (PPG) sensors or electrodermal response (EDR) sensors, which are used to collect the operator's eye movement trajectory and heart rate variability sequences. This implementation continuously monitors changes in physiological parameters through non-contact signal acquisition technology. Furthermore, physiological state analysis can be implemented using time series analysis methods, such as wavelet transform or autoregressive moving average models, which are used to extract features from the eye movement trajectory and heart rate variability sequences and calculate the real-time attention distraction index and physiological stress level. This implementation quantifies the operator's working state through frequency domain decomposition and statistical feature extraction. The dynamic graph neural network can be implemented using graph attention networks or spatiotemporal graph convolutional networks, which are used to perform anomaly coupling analysis on multimodal indicators and output the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modality indicator. For example, it processes the temporal changes of graph structure data through node embedding and dynamic edge weight update mechanisms. Specifically, time window matching and entity identification association can be implemented using a sliding window algorithm or dynamic time warping technology. This is used to extract corresponding eye-tracking fragments, communication text fragments, and electromagnetic spectrum fragments from the multimodal data stream. This implementation establishes the association between data fragments and supply chain entities based on timestamp alignment and entity identification mapping. As a preferred implementation, generating a visualization interface that integrates business topology and multimodal traceability evidence can be implemented using a WebGL rendering engine or an interactive charting library, such as D3.js or ECharts. This interface is used to synchronously associate and display high-risk entities with the extracted data fragments. This implementation achieves dynamic interaction between business topology and traceability evidence through view linkage and event listening mechanisms. Therefore, this application realizes real-time detection and visual tracing of abnormal events in the production and sales collaboration process of the supply chain through the synchronous acquisition of multimodal data and the coupled analysis framework of dynamic graph neural network. It overcomes the shortcomings of the existing technology, which can only analyze single-dimensional data in isolation, resulting in delayed early warning and difficulty in root cause location. The formation of multimodal data flow and the abnormal coupling analysis process are designed to break down data silos, so that abnormal indicators of physiological, collaborative and equipment dimensions can be quantitatively correlated and propagated under a unified framework.

[0058] In the process of supply chain production and sales collaboration, the problem of multimodal data silos has long constrained the ability to detect and trace abnormal events in real time. Existing technologies can only analyze single-dimensional data in isolation, resulting in delayed anomaly warnings and difficulties in root cause identification. To solve this problem, this method systematically integrates multi-dimensional data such as operator physiological status, cross-departmental collaboration efficiency, and production line equipment operation status to construct a complete real-time detection and visualization framework. In specific implementation, biosensors, enterprise communication system interfaces, and spectrum acquisition devices are first deployed at key workstations on the production line to simultaneously collect operator eye movement trajectories and heart rate variability sequences, cross-departmental business communication text, and production line electromagnetic spectrum signals, forming a spatiotemporally aligned multimodal data stream. Among them, biosensors are used to continuously monitor operator visual attention and autonomic nervous system activity, enterprise communication system interfaces realize automated capture and de-identification processing of business text, and spectrum acquisition devices capture dynamic changes in the production line electromagnetic environment, thereby ensuring that multi-source heterogeneous data maintain strict correlation at the timestamp and work order identification levels. Furthermore, physiological state analysis is performed on the acquired eye movement trajectory and heart rate variability sequences. By comprehensively calculating the fixation point distribution entropy, pupil diameter variation coefficient, low-frequency to high-frequency power ratio, and heart rate interval standard deviation, a real-time attention distraction index and physiological stress level are generated, thus objectively quantifying the operator's work status. Simultaneously, cross-departmental business communication texts undergo sentiment semantic analysis. Combining sentiment polarity, sentiment intensity, communication response time delay, and semantic association network topology features, a collaborative health score is output for specific production work orders, transforming unstructured text into quantifiable collaborative quality indicators. In addition, electromagnetic spectrum signals from the production line are extracted using spectral features. Spectral anisotropy index, fluctuation variance, and periodicity metrics are used to determine spectral anomaly offset and production line beat disturbance coefficient, establishing a quantitative correlation between electromagnetic environment anomalies and production rhythm fluctuations. Based on the aforementioned multimodal feature indicators, a dynamic graph neural network is applied to multimodal anomaly coupling analysis. Real-time attention distraction index and physiological stress level serve as physiological dimension inputs, which, together with health score, spectral anomaly offset, and production line beat disturbance coefficient, construct an edge weight matrix. Anomaly propagation is achieved through a cross-modal graph attention mechanism, ultimately outputting the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modal indicator, making the attribution process of anomaly root causes interpretable. Consequently, the comprehensive anomaly probability and contribution distribution trigger a time window matching and entity identification association mechanism, extracting corresponding eye-tracking trajectory fragments, communication text fragments, and electromagnetic spectrum fragments from the multimodal data stream. This generates a visual interface that integrates business topology and multimodal tracing evidence. High-risk entities and original data fragments are synchronously associated and displayed in this interface, allowing managers to intuitively trace the multi-dimensional causal links of anomaly events.

[0059] In one specific implementation, within the smartphone assembly stage of the electronics manufacturing supply chain, biosensors are implemented as wearable photoplethysmography (PPG) heart rate monitoring straps and infrared eye-tracking devices, deployed at key workstations on the assembly line to continuously collect operator physiological data. The enterprise communication system interface connects to the enterprise instant messaging platform via API, automatically retrieving emails and chat logs associated with specific production work orders. The spectrum acquisition equipment employs a wideband radio frequency sensor array, capturing electromagnetic spectrum signals in real time around the production equipment. After analysis, the resulting multimodal data stream is analyzed, with the physiological state analysis module outputting an operator attention distraction index of 0.78 (higher values ​​indicate greater attention distraction) and a physiological stress level of 0.65 (higher values ​​indicate stronger stress). The sentiment semantic analysis module generates a collaborative health score of 0.82 based on the communication text (lower scores indicate higher collaborative risk). The spectrum feature extraction module determines a spectrum anomaly offset of 1.25 and a production line cycle disturbance coefficient of 0.35. Based on this, the dynamic graph neural network performs multimodal anomaly coupling analysis, outputting a comprehensive anomaly probability of 0.91 for a certain warehouse entity, and showing that the physiological modality contributes 45%, the collaborative modality 30%, and the spectral modality 25%. Based on this result, the system automatically extracts eye-tracking trajectory segments (showing operators frequently deviating from the work area), communication text segments (including dialogues with timeouts in inter-departmental responses), and electromagnetic spectrum segments (showing the spectral shift when equipment starts up) during the anomaly occurrence period, and establishes a bidirectional association with the warehouse entity in the visualization interface to achieve real-time presentation of anomaly evidence.

[0060] This method eliminates data fragmentation in traditional monitoring systems through synchronous acquisition and deep fusion of multimodal data, enabling coupled analysis of anomaly indicators across physiological, collaborative, and equipment dimensions within a unified framework. The introduction of a dynamic graph neural network not only quantifies the anomaly propagation path between supply chain entities but also reveals the relative impact weight of each modality indicator on overall risk through contribution distribution, thereby improving the real-time performance and accuracy of anomaly detection. The visual interface transforms abstract anomaly probabilities into an interactive multimodal evidence chain, allowing managers to pinpoint the root cause of anomalies without relying on subjective experience. This solves the problems of delayed early warnings and difficulty in tracing caused by data silos in existing technologies, ultimately achieving efficient handling of anomalies in the supply chain production and sales collaboration process.

[0061] In modern supply chain management, the stability and efficiency of the production-sales collaboration process are crucial. However, this process involves multiple entities such as production, warehousing, logistics, and sales, and its operational status is complexly affected by multiple factors, including operator physiological state, inter-departmental collaboration efficiency, and production line equipment operating conditions. In some embodiments described above, this application proposes using biosensors, enterprise communication system interfaces, and spectrum acquisition devices to synchronously acquire multimodal data streams for real-time detection of abnormal events in supply chain production-sales collaboration. However, in its implementation, due to the lack of targeted sensor deployment, the absence of specific work order scenarios for data acquisition, and the lack of multi-source data processing steps, the collected physiological data may be disconnected from actual production line operations, communication text may contain irrelevant noise, electromagnetic spectrum signals may lack a reference standard, and the spatiotemporal asynchrony of various modal data and weak correlation with work orders may exist. Consequently, subsequent multimodal anomaly coupling analysis cannot accurately capture the root cause of anomalies due to unreliable input data.

[0062] To this end, this application further proposes to simultaneously acquire operator eye-tracking trajectories and heart rate variability sequences, cross-departmental business communication text, and production line electromagnetic spectrum signals by deploying biosensors, enterprise communication system interfaces, and spectrum acquisition devices at key workstations on the production line, forming a multimodal data stream. See [link to relevant documentation]. Figure 3 Taking the server as the executing entity as an example, the following steps are included.

[0063] 301. By deploying non-invasive eye trackers and heart rate sensors at quality inspection and precision assembly stations, eye movement trajectory data and heart rate variability data of operators during the execution of specific production work orders are collected.

[0064] 302. Obtain cross-departmental email exchanges and instant messaging records associated with the target production work order through the enterprise communication system interface, and perform anonymization processing and topic clustering on the communication records.

[0065] 303. By deploying broadband radio frequency sensors around the production equipment, electromagnetic spectrum signals of the production line are collected during the standard production cycle, and a reference spectrum template for each production line equipment under normal operating conditions is established.

[0066] 304. The collected eye-tracking data, heart rate variability data, communication record text, and electromagnetic spectrum signals are timestamped and associated with work order identifiers to generate a multimodal data stream with spatiotemporal alignment markers.

[0067] In practical applications, non-invasive eye trackers and heart rate sensors refer to sensing devices that acquire physiological data without interfering with the operator's normal work. These can be implemented using eye trackers with infrared optical tracking technology or heart rate sensors using photoplethysmography. The aim is to avoid distortion of operational behavior due to invasive equipment and ensure that the collected data reflects the true working state. Specific production work order execution processes can be understood as limiting the time and scope of data collection. Specifically, the work order initiation signal triggered by the production execution system can be used as the collection start identifier. The purpose is to bind physiological data with specific production tasks and eliminate interference from irrelevant operations. Anonymization and topic clustering refer to privacy protection and content structuring of communication text. This can be achieved by using regular expressions to replace sensitive fields or clustering algorithms based on word embedding models. The aim is to eliminate privacy risks and extract semantic topics strongly related to the current production task. Wideband radio frequency sensors can be understood as electromagnetic signal capture devices covering a wide frequency range. Specifically, they can use superheterodyne receiver architectures or software-defined radio platforms. The purpose is to fully capture the full-band characteristics of the device's operation and avoid missing key frequency bands. A baseline spectrum template refers to an electromagnetic characteristic reference system for equipment under normal operating conditions. It can be implemented based on statistical averaging of historical data or typical spectrum models generated by machine learning, with the aim of providing a comparable basis for subsequent anomaly detection. Timestamp alignment and work order identifier association refer to the unified processing of spatiotemporal and business dimensions of multi-source data. Specifically, work order association can be established by synchronizing clocks through network time protocols or database mapping tables. Its purpose is to eliminate timing deviations between devices and ensure that heterogeneous data strictly corresponds at the business level.

[0068] Specifically, this application's solution involves the targeted deployment of non-invasive eye trackers and heart rate sensors at quality inspection and precision assembly stations. These stations require extremely high operator concentration and are susceptible to interference, enabling the capture of physiological state changes directly related to production quality. Simultaneously, the data collection scope is limited to the execution process of a specific production work order, tightly binding eye movement trajectories and heart rate variability data to the specific task scenario, improving the work order-level accuracy of subsequent attention distraction index analysis. Communication records associated with the target production work order are obtained through the enterprise communication system interface. Data is filtered based on the target work order to ensure that text content focuses on the current production task, avoiding cross-work order information mixing. Anonymization eliminates privacy risks to meet compliance requirements, and topic clustering merges scattered emails and instant messages by topic, providing structured input for sentiment semantic analysis. Broadband radio frequency sensors collect electromagnetic spectrum signals around the production equipment, utilizing their wideband coverage to capture the full-frequency characteristics of equipment operation. A benchmark spectrum template is collected and established within a standard production cycle, constructing a reference system based on normal operating condition data, making the calculation of spectrum anomaly offsets comparable. Finally, the multi-source data is timestamped and associated with work order identifiers. The timing deviation between devices is corrected according to a unified time base. The heterogeneous data is mapped to the same business dimension by combining the work order identifiers, generating a multimodal data stream with spatiotemporal alignment markers. This ensures that eye-tracking trajectories, communication texts and spectral signals are strictly synchronized at the spatiotemporal and business levels, providing a traceable associated data foundation for the multimodal coupling analysis of dynamic graph neural networks.

[0069] As a specific implementation method, the solution of this application is implemented as follows: A Tobii Pro Fusion eye tracker and a Polar H10 heart rate sensor are deployed at the quality inspection station of the automobile manufacturing production line. Data acquisition is initiated when the production execution system triggers a production work order for a specific vehicle model. Emails and chat records associated with the work order number are obtained through the WeChat API interface. Employee names are replaced using regular expressions, and topic clustering is performed based on the BERT model. A Rohde & Schwarz FSW spectrum analyzer is deployed around the stamping equipment to collect electromagnetic signals within a standard production cycle and generate a reference spectrum template based on historical data. The timestamps of each device are aligned using an NTP server, and eye-tracking data, communication text, and spectrum signals are associated through a work order database mapping table to generate a multimodal data stream with spatiotemporal markers.

[0070] Through the above technical solutions, this application ensures a close correlation between physiological data and actual production line operations, eliminates irrelevant noise in communication text, provides a benchmark reference for electromagnetic spectrum signals, and achieves strict synchronization of multimodal data in the spatiotemporal and business dimensions. This provides a reliable data input basis for subsequent multimodal anomaly coupling analysis and effectively solves the problem of inaccurate anomaly root cause capture caused by unreliable data.

[0071] In some embodiments described above, this application proposes physiological state analysis of eye-tracking trajectories and heart rate variability sequences to obtain real-time attention distraction indices and physiological stress levels. However, during implementation, eye-tracking data is susceptible to variations in device sampling frequencies and environmental noise interference, leading to distortion of attention indicators. Without frequency domain standardization analysis, heart rate variability sequences cannot accurately quantify the state of autonomic nervous activity. Without multi-feature fusion, relying solely on a single physiological parameter will fail to comprehensively reflect the operator's real-time state, resulting in misjudgments in subsequent anomaly detection due to unreliable input data.

[0072] In response, this application further proposes the following technical solution, see [link to technical solution]. Figure 4 Taking the server as the executing entity as an example, the following steps are included.

[0073] 401. Perform sampling frequency normalization and noise filtering on the eye movement trajectory data, and extract the gaze point distribution entropy and pupil diameter variation coefficient per unit time.

[0074] 402. Perform frequency domain feature analysis on the heart rate variability sequence to determine the ratio of low-frequency power to high-frequency power, and extract the standard deviation of heart rate interval between adjacent normal heart rate intervals.

[0075] 403. Based on the gaze point distribution entropy, pupil diameter variation coefficient, low-frequency to high-frequency power ratio, and heart rate interval standard deviation, the data are processed by a pre-trained physiological state assessment model to output the real-time attention distraction index and physiological stress level.

[0076] Among them, the real-time attention distraction index comprehensively reflects the operator's visual attention concentration, while the physiological stress level quantitatively represents the operator's autonomic nervous system activation state.

[0077] In practical applications, sampling frequency normalization refers to unifying raw eye-tracking data from different acquisition devices to a standard time reference. This can be achieved using linear interpolation or resampling techniques, aiming to eliminate timing inconsistencies caused by differences in device hardware and ensure time synchronization in subsequent analyses. Noise filtering removes random fluctuations introduced by changes in ambient lighting or device jitter, which can be achieved using wavelet thresholding or Kalman filtering methods. This aims to improve the signal-to-noise ratio of eye-tracking data and avoid the generation of false attention indicators. Fixation point distribution entropy quantifies the spatial dispersion of visual attention. It can be calculated based on the Shannon entropy formula to determine the spatial distribution probability of fixation points per unit time, aiming to objectively characterize the concentration or dispersion of operator attention. The pupil diameter variation coefficient refers to the relative dispersion of pupil size changes, which can be calculated as the ratio of the standard deviation of pupil diameter to the mean. This aims to dynamically capture the impact of cognitive load fluctuations on physiological responses. Frequency domain feature analysis refers to converting the heart rate variability sequence to the frequency domain for energy distribution analysis. This can be achieved using Fast Fourier Transform or wavelet packet decomposition techniques, aiming to separate the spectral contributions of sympathetic and parasympathetic nerve activity. The low-frequency to high-frequency power ratio refers to the ratio of power in the low-frequency band (0.04-0.15Hz) to that in the high-frequency band (0.15-0.4Hz), reflecting the balance of the autonomic nervous system and quantifying physiological stress levels. The standard deviation of heart rate intervals refers to the standard deviation of adjacent normal heart rate intervals, calculated based on the RR interval sequence to assess the stability of heart rate variability. A pre-trained physiological state assessment model is a machine learning model pre-trained on historical physiological datasets, implemented using multilayer perceptrons or gradient boosting tree structures, aiming to integrate multi-dimensional heterogeneous features and output a comprehensive physiological state index.

[0078] Specifically, the proposed solution first normalizes the sampling frequency of eye-tracking data, eliminating the influence of device differences by using a unified time reference, and then removes environmental interference through noise filtering to obtain high-fidelity eye-tracking data. Based on this, the fixation point distribution entropy is extracted from the normalized eye-tracking data to quantify the spatial dispersion of visual attention, while the pupil diameter variation coefficient is calculated to characterize the dynamic changes in cognitive load. Simultaneously, frequency domain feature analysis is performed on the heart rate variability sequence, capturing the balance of autonomic nervous activity by determining the low-frequency to high-frequency power ratio, and extracting the standard deviation of heart rate intervals to assess heart rate stability. Finally, the above four types of features are input into a pre-trained physiological state assessment model. This model integrates multi-dimensional physiological parameters through a nonlinear mapping mechanism, simultaneously outputting a real-time attention dispersion index and physiological stress level, achieving a comprehensive analysis of the operator's physiological state. This process ensures data quality at each stage through phased processing, avoids the limitations of single parameters by utilizing a multi-feature collaborative mechanism, and ultimately provides reliable physiological dimension input for anomaly detection.

[0079] As a specific implementation method, the scheme of this application is implemented as follows: Eye movement trajectory data is collected by a non-invasive eye tracker. Linear interpolation technology is used to normalize the original 250Hz sampling frequency data to a uniform time step of 100Hz, and wavelet thresholding is applied to filter out environmental noise. The gaze point distribution entropy is calculated based on the Shannon entropy formula to determine the spatial distribution probability of the gaze point per unit time. The pupil diameter variation coefficient is obtained by the ratio of the standard deviation to the mean. The heart rate variability sequence is acquired by a heart rate sensor, and frequency domain decomposition is performed using Fast Fourier Transform to determine the power ratio of the low-frequency and high-frequency bands, while simultaneously calculating the standard deviation of adjacent normal heartbeat intervals. The pre-trained physiological state assessment model adopts a three-layer fully connected neural network structure, trained on a historical physiological dataset containing operators from various trades. After inputting a four-dimensional feature vector, it outputs a real-time attention distraction index and physiological stress level.

[0080] By employing the aforementioned approach, this application eliminates the interference of device sampling frequency differences and environmental noise on eye-tracking data, ensuring the accuracy of attention indicators. Frequency domain feature analysis quantifies the activity state of the autonomic nervous system, avoiding misjudgments of physiological stress caused by simple time-domain statistics. Simultaneously, based on a multi-dimensional feature fusion mechanism, it comprehensively reflects the operator's real-time physiological state, overcoming the limitations of single physiological parameters and providing highly reliable physiological dimension input data for the real-time detection of abnormal events in supply chain production and sales coordination.

[0081] In practical applications, some of the embodiments described above in this application propose to perform sentiment and semantic analysis on cross-departmental business communication texts to obtain a collaborative health score for specific production work orders. However, in its implementation, it relies solely on basic sentiment analysis without fully integrating the communication context, response time dynamics, and semantic network topology. This results in the collaborative health score failing to accurately quantify the true health status of inter-departmental collaboration. Specifically, sentiment analysis ignores the continuity of dialogue threads, response delays and interaction frequencies are not included in the evaluation, and semantic conflicts are not effectively identified. Consequently, the score becomes one-sided and fails to support the coupled analysis and real-time tracing of supply chain anomalies.

[0082] To address this, this application further proposes a step of performing sentiment and semantic analysis on cross-departmental business communication texts to obtain a collaborative health score for a specific production work order, see [link to relevant documentation]. Figure 5 Taking the server as the executing entity as an example, the following steps are included.

[0083] 501. Perform word segmentation and named entity recognition on the cross-departmental business communication text to extract key communication content associated with the production work order number.

[0084] 502. Determine the emotional polarity value and emotional intensity value of the key communication content, and count the communication response time delay and communication interaction frequency of the key communication content.

[0085] 503. The emotional polarity value, the emotional intensity value, the communication response time delay, and the communication interaction frequency are weighted and calculated based on preset weights to obtain an emotional consistency index.

[0086] 504. Construct a semantic association network based on the key communication content, determine the topological features of the semantic association network, and determine the semantic conflict degree based on the topological features.

[0087] 505. Based on the sentiment consistency index, the topological features of the semantic association network, and the semantic conflict degree, the collaborative health score is generated through normalization and scaling transformation.

[0088] Word segmentation refers to dividing continuous text into meaningful word units. This can be achieved using rule-based forward maximum matching or statistically driven Hidden Markov Models, aiming to accurately identify lexical boundaries in the text and provide a structured foundation for subsequent analysis. Named entity recognition identifies proper nouns of specific categories in the text. This can be achieved using Conditional Random Fields or Transformer-based pre-trained sequence labeling models, aiming to locate key entities related to production order numbers and avoid interference from irrelevant text. The determination of sentiment polarity and intensity values ​​can be achieved by directly outputting sentiment tendency and intensity using an end-to-end sentiment analysis neural network, without relying on sentiment dictionary matching. This aims to dynamically adapt to the semantic expression of different business scenarios. Communication response time latency refers to the time interval between adjacent communication records. It can be automatically calculated based on message timestamps and mapped to discrete levels, aiming to quantify inter-departmental response efficiency. Communication interaction frequency refers to the number of communications within a preset time window. This can be achieved through message counting combined with a time decay factor, aiming to reflect the activity and continuity of communication. Emotional consistency index refers to a quantitative indicator that integrates emotional expression and communication timeliness. It can be achieved through a nonlinear weighted fusion function, aiming to balance emotional coherence and collaborative smoothness. Semantic association network refers to a graph structure built based on textual semantic relationships. It can generate nodes and edges using dependency parsing, aiming to represent the structured relationships of communication content. Topological features refer to the structural attributes of the network, which can include clustering coefficients and average path lengths, aiming to reveal the connectivity and information transmission efficiency of the communication network. Semantic conflict degree is a measure of the degree of semantic contradiction, which can be calculated using word vector cosine similarity, aiming to identify inconsistencies in expression. The generation of collaborative health score refers to the fusion of multi-dimensional indicators into a single health value, which can be achieved through Min-Max normalization, aiming to provide a comparable collaborative health status indicator.

[0089] Specifically, this application's solution first extracts key communication content associated with production work order numbers by segmenting and naming entity recognition of cross-departmental business communication texts, ensuring that the analysis focuses on specific production tasks. Second, it determines the sentiment polarity and intensity values ​​of key communication content and statistically analyzes communication response time delays and communication interaction frequencies, simultaneously capturing the collaborative status from both sentiment and timeliness dimensions. Next, it weights these indicators based on preset weights to obtain a sentiment consistency index, achieving a preliminary fusion of sentiment dynamics and communication efficiency. Further, it constructs a semantic association network based on key communication content, determines topological features, and calculates semantic conflict degree, identifying potential contradictions at the semantic structure level. Finally, based on the sentiment consistency index, topological features, and semantic conflict degree, it generates a collaborative health score through normalization, completing the comprehensive quantification of multi-dimensional information. This process, through phased and multi-faceted analysis, organically integrates text sentiment, communication timeliness, and semantic structure to form a comprehensive assessment of the health status of inter-departmental collaboration, effectively solving the problem of one-sided scores caused by single-dimensional analysis.

[0090] As a specific implementation method in the automotive manufacturing supply chain, for the email communication records of production work order W12345, the system first uses a BERT-based named entity recognition model to extract email content related to the work order. Then, a pre-trained sentiment analysis model directly outputs the sentiment polarity and intensity values ​​of each email, while simultaneously calculating the average time delay of email replies and the daily communication frequency. Next, the sentiment index and timeliness index are fused into a sentiment consistency index according to preset weights. Then, a semantic association network of email content is constructed based on semantic role labeling, the clustering coefficient of the network is calculated as a topological feature, and semantic conflict points are detected through word vector similarity. Finally, the sentiment consistency index, clustering coefficient, and semantic conflict degree are normalized and weighted to generate a collaborative health score, which is used to evaluate the departmental collaborative health status of the production work order.

[0091] Through the above technical solutions, this application can accurately quantify the true health status of inter-departmental collaboration, effectively capture the impact of dialogue thread continuity, response time dynamics, and semantic network topology on collaboration, and enable the collaboration health score to comprehensively reflect the degree of collaboration health, thereby providing a reliable basis for coupled analysis and real-time tracing of abnormal events in the supply chain.

[0092] Specifically, in some of the embodiments described above in this application, the sentiment polarity value and sentiment intensity value of key communication content are proposed to calculate the collaborative health score. However, in its implementation, relying solely on simple matching of sentiment dictionaries cannot fully consider the dynamic influence of the communication context, such as the changes in the continuity of sentiment in the dialogue thread and the modification effect of semantic structure. This makes the basic sentiment value susceptible to interference from isolated words or local expressions, resulting in distorted sentiment analysis results and thus affecting the accuracy and reliability of supply chain collaborative health assessment.

[0093] In response, this application further proposes to determine the sentiment polarity and sentiment intensity values ​​of key communication content, and to statistically analyze the communication response time delay and communication interaction frequency of key communication content, including: Based on the sentiment dictionary, sentiment words are matched and intensity is calculated for key communication content to obtain the basic sentiment polarity value and basic sentiment intensity value of each sentence.

[0094] Based on the communication timing relationship of key communication content, the time interval between adjacent communication records is determined as the communication response time delay, and the number of communication interactions within a preset time window is counted as the communication interaction frequency.

[0095] Extract the dialogue thread structure and semantic role annotation results from key communication content to generate communication context features.

[0096] Based on the basic emotional polarity value and the basic emotional intensity value, the emotional correction is performed by integrating the contextual features of the communication context to obtain the emotional polarity value and the emotional intensity value.

[0097] The sentiment lexicon refers to a predefined vocabulary resource library used to quantify the sentiment tendency of text. It can be implemented using open-source sentiment lexicons such as HowNet or SentiWordNet, providing initial quantitative basis for sentiment analysis through term matching and intensity weighting, aiming to establish an objective benchmark for basic sentiment assessment. Communication temporal relationships can be understood as the time-series dependency characteristics between communication records. Specific implementation methods include interval calculation and frequency statistics based on timestamp sequences, aiming to supplement the contextual information of sentiment analysis from the perspective of behavioral efficiency. Dialogue thread structure refers to the coherence framework of dialogue logic in the communication content, which can be generated through dependency parsing or dialogue behavior annotation technology, aiming to capture the continuous characteristics of sentiment evolution. Semantic role annotation results can be understood as the structured analysis of the relationship between predicates and arguments in a sentence. Typical implementations use the PropBank framework or the FrameNet semantic framework, aiming to quantify the modifying strength of subject-verb-object structure on sentiment expression. The sentiment correction process refers to the computational mechanism that dynamically adjusts the basic sentiment value based on contextual features. It can be implemented through weighted fusion algorithms or neural network mapping, aiming to eliminate interference from isolated words and reflect the true sentiment state.

[0098] Specifically, the proposed solution first obtains basic sentiment values ​​through sentiment dictionary matching, providing initial input for subsequent corrections. Simultaneously, response delay and interaction frequency are extracted based on communication temporal relationships to form auxiliary features for the behavioral efficiency dimension. Next, the dialogue thread structure is analyzed from key communication content to quantify changes in sentiment continuity, and semantic role labeling results are extracted to determine negation word density and subject-verb-object sentiment weights. Finally, the basic sentiment values ​​are dynamically fused with contextual features, smoothing sentiment intensity values ​​through a sentiment consistency factor and correcting sentiment polarity values ​​using a sentiment correction coefficient. This sequence of steps is designed based on the context-dependent nature of sentiment expression. Since dialogue sentiment has temporal continuity and semantic structural modification, a phased feature extraction and fusion mechanism is adopted, enabling sentiment analysis to adapt to the dynamic evolution of the communication context, thereby avoiding evaluation bias caused by local expressions.

[0099] As a specific implementation method, the solution of this application is implemented as follows: In the supply chain management system, when processing cross-departmental email communication, the HowNet sentiment dictionary is used to segment the email body into sentences, and the weighted score of sentiment words in each sentence is calculated as the basic sentiment polarity value and basic sentiment intensity value. By parsing the timestamp information in the email header, the time interval between adjacent emails is determined as the communication response time delay, and the number of email exchanges within a 24-hour window is counted as the communication interaction frequency. The Stanford CoreNLP toolkit is used to extract the dialogue thread structure, identify the changes in sentiment gradient in consecutive emails, and simultaneously parse the subject-verb-object structure through semantic role labeling to calculate the distribution density of negative words in sentences. Based on the extracted contextual features, the basic sentiment intensity value is smoothed by a sliding window to suppress variability, and a correction coefficient based on the density of negative words is used to weight and adjust the basic sentiment polarity value, finally outputting the sentiment polarity value and sentiment intensity value that reflect the actual collaborative state.

[0100] Through the above-mentioned scheme, this application can effectively suppress the fluctuation of emotional intensity caused by emotional jumps between sentences, correct the distortion of emotional polarity by negative words or semantic roles, and ensure that the sentiment analysis results truly reflect the actual emotional state in cross-departmental collaboration. This provides high-precision input for the calculation of collaboration health score and improves the accuracy and reliability of supply chain anomaly detection.

[0101] In some of the embodiments described above in this application, a method for sentiment correction based on the fusion of basic sentiment polarity value and basic sentiment intensity value with communication context features is proposed. However, in its implementation, the basic sentiment analysis relies solely on sentiment dictionary matching to generate initial sentiment values, failing to fully capture the continuous characteristics of sentiment evolution in the dialogue thread. At the same time, it ignores the deep influence of semantic structure on sentiment expression, resulting in the sentiment polarity value and intensity value failing to accurately reflect the true emotional state in cross-departmental collaboration, thereby affecting the reliability of collaboration health assessment.

[0102] In response, this application further proposes steps for emotion correction, including: Based on the basic sentiment polarity value and the basic sentiment intensity value, the sentiment gradient change between adjacent sentences in the key communication content is determined, and the sentiment consistency factor is generated. The sentiment consistency factor is used to quantify the degree of continuity of sentiment changes in the dialogue thread.

[0103] By utilizing the semantic role labeling results in the contextual features of communication, the sentiment carrying weight of the subject-verb-object structure is extracted. At the same time, a sentiment correction coefficient is constructed based on the density distribution of negative words. The sentiment correction coefficient is used to characterize the intensity of the modification of the semantic structure on the sentiment expression.

[0104] The emotional polarity value is obtained by smoothing the basic emotional intensity value through an emotional consistency factor.

[0105] The emotional intensity value is obtained by weighting the basic emotional polarity value using an emotional correction coefficient.

[0106] Specifically, the sentiment consistency factor is an indicator that quantifies the continuity of sentiment changes within a dialogue thread. It can be implemented by calculating the standard deviation of the sentiment value difference sequence between adjacent sentences or by using dynamic time warping algorithms. The aim is to capture the gradual trend of sentiment evolution and avoid the lack of continuity caused by evaluating isolated sentences. The sentiment correction coefficient can be understood as a parameter characterizing the intensity of semantic structure modification. It can be implemented based on the distribution of negation word density using a piecewise linear function or an exponential decay model. Its purpose is to compensate for the distortion effect of semantic modification on sentiment expression, such as the difference in the subject's sentiment dominance in a subject-verb-object structure or the reversal effect of negation words. In practical applications, smoothing is specifically an operation that applies continuity constraints to the basic sentiment intensity value. For example, sliding window averaging filtering or adaptive Kalman filtering algorithms can be used. Its purpose is to suppress random noise in sentiment fluctuations and ensure that the sentiment polarity value stably reflects the overall dialogue tendency. Furthermore, weighted adjustment is the process of dynamically correcting the basic sentiment polarity value through the sentiment correction coefficient. This can be understood as allocating sentiment carrying weights according to the semantic role labeling results. Its purpose is to accurately restore the intensity level of sentiment expression and avoid the deviation between literal sentiment and actual intention.

[0107] Specifically, the proposed solution first determines the sentiment gradient changes between adjacent sentences in key communication content based on basic sentiment polarity and basic sentiment intensity values, generating a sentiment consistency factor. This factor provides a continuity measurement basis for subsequent smoothing by quantifying the continuity of sentiment evolution. Next, it extracts the sentiment carrying weights of the subject-verb-object structure using semantic role labeling results from the communication context features, and constructs a sentiment correction coefficient based on the density distribution of negative words. This coefficient provides a correction basis for weighted adjustment by characterizing the modification intensity of semantic structure. Then, it smooths the basic sentiment intensity values ​​using the sentiment consistency factor to obtain sentiment polarity values. This process uses continuity measurement to suppress random noise, ensuring that the sentiment polarity values ​​reflect the overall sentiment tendency of the dialogue thread. Finally, it uses the sentiment correction coefficient to weight and adjust the basic sentiment polarity values ​​to obtain sentiment intensity values. This process dynamically responds to changes in modification intensity based on semantic structure features, realistically restoring the intensity levels of sentiment expression. These steps, through the synergistic effect of sentiment gradient continuity modeling and semantic structure modification compensation, upgrade sentiment analysis from static dictionary matching to dynamic context awareness.

[0108] As a specific implementation method, when analyzing email exchanges between the purchasing and production departments, the system first obtains the basic sentiment polarity and basic sentiment intensity values ​​of each sentence in the email text. Then, it calculates the sentiment gradient changes between adjacent sentences, generating a sentiment consistency factor. For example, when the sentiment values ​​of consecutive sentences show a decreasing trend, this factor reflects the continuity of gradually negative sentiment. Simultaneously, based on semantic role labeling results, the system identifies the subject-verb-object structure, determining that the subject "supplier" carries a higher sentiment weight, and detecting negative words such as "no," thus constructing a sentiment correction coefficient. Next, the sentiment consistency factor is used to smooth the basic sentiment intensity values, resulting in more stable sentiment polarity values. Finally, the sentiment correction coefficient is used to weight and adjust the basic sentiment polarity values; for example, when the density of negative words is high, the correction magnitude for the sentiment polarity values ​​is increased, resulting in accurate sentiment intensity values.

[0109] Through the above scheme, this application can accurately capture the continuity of emotional evolution in the dialogue thread and effectively compensate for the deep influence of semantic structure on emotional expression, so that the emotional polarity value and intensity value can truly reflect the actual emotional state in cross-departmental collaboration, thereby improving the reliability of collaboration health assessment.

[0110] In existing technologies, when extracting spectral features from electromagnetic spectrum signals of production lines to obtain spectral anomaly offset and production line cycle disturbance coefficients, traditional spectral analysis methods only focus on single frequency domain feature indicators and fail to fully integrate the multidimensional dynamic characteristics of the spectrum. This results in the inability to accurately capture subtle abnormal changes in electromagnetic signals and implicit disturbances in production cycle, making the feature extraction results one-sided and lacking a quantitative correlation between the degree of electromagnetic environment anomaly and the stability of production rhythm. Consequently, this affects the real-time performance and accuracy of subsequent multimodal anomaly coupling analysis.

[0111] To address this, this application further proposes to extract spectral features from the electromagnetic spectrum signals of the production line to obtain the spectral anomaly offset and the production line cycle time disturbance coefficient, see [link to relevant documentation]. Figure 6 Taking the server as the executing entity as an example, the following steps are included.

[0112] 601. Perform variational mode decomposition and frequency domain structure analysis on the electromagnetic spectrum signal of the production line to obtain the spectrum anisotropy index and characteristic frequency band energy distribution.

[0113] 602. Calculate the autocorrelation coefficient sequence of the signal envelope based on the energy distribution of the characteristic frequency band, and extract the fluctuation variance and periodicity measure of the autocorrelation coefficient sequence.

[0114] 603. Based on the spectral anisotropy index, fluctuation variance, and periodicity metric, the spectral anomaly offset and production line cycle disturbance coefficient are calculated synchronously using a nonlinear mapping function.

[0115] Among them, the spectrum anomaly offset characterizes the degree of anomaly in the electromagnetic environment, and the production line cycle disturbance coefficient characterizes the stability of the production rhythm.

[0116] In practical applications, variational mode decomposition (VMD) is an adaptive signal decomposition method. It can be implemented using a variational mode decomposition algorithm, which separates non-stationary components in a signal through optimization, aiming to effectively suppress mode aliasing and extract the signal's intrinsic modal characteristics. The spectral anisotropy index quantifies the non-uniformity of electromagnetic energy distribution in the frequency domain, and can be implemented using eigenvalue decomposition of the frequency domain covariance matrix, aiming to characterize the spatial distribution characteristics of electromagnetic energy. Specifically, the characteristic frequency band energy distribution refers to the distribution vector reflecting the dynamic changes of energy in key frequency bands, which can be implemented using energy distribution weighting and normalization, aiming to comprehensively capture the dynamic characteristics of the spectrum's energy. In practical applications, the autocorrelation coefficient sequence refers to the autocorrelation characteristic sequence of the signal envelope in the time domain, which can be implemented using the autocorrelation function, aiming to analyze the periodicity of the signal. The fluctuation variance is a statistical measure quantifying the degree of random fluctuation in the autocorrelation coefficient sequence, which can be implemented using variance calculation, aiming to reflect the instantaneous disturbance intensity of the electromagnetic environment. Specifically, periodicity metrics refer to indicators that assess the periodicity of autocorrelation coefficient sequences. These can be implemented using periodicity detection algorithms, aiming to characterize the stability of production cycles. In practical applications, nonlinear mapping functions are functions that map multidimensional features to target metrics. These can be implemented using neural networks or polynomial functions, aiming to integrate heterogeneous features and avoid oversimplification of complex coupling relationships by linear simplification.

[0117] Specifically, the proposed solution first performs variational mode decomposition on the electromagnetic spectrum signal of the production line, effectively separating non-stationary components and suppressing mode aliasing, providing clean mode components for subsequent analysis. Then, through frequency domain structure analysis, a spectral anisotropy index is generated based on the eigenvalue decomposition of the frequency domain covariance matrix to quantify the non-uniformity of electromagnetic energy distribution. Simultaneously, the energy distribution of characteristic frequency bands is combined to capture dynamic energy changes, avoiding missed anomalies caused by single-frequency features. Based on this, the autocorrelation coefficient sequence of the signal envelope is calculated using the energy distribution of characteristic frequency bands. By extracting fluctuation variance and periodicity metrics, the frequency domain features are correlated with time domain disturbances to identify implicit shifts in production rhythm. Finally, the spectral anisotropy index, fluctuation variance, and periodicity metrics are integrated through a nonlinear mapping function to simultaneously output the spectral anomaly shift and production line rhythm disturbance coefficient, ensuring that the correlation between electromagnetic environment anomalies and production rhythm disturbances is not severed, thus providing a coherent quantitative basis for multimodal anomaly coupling analysis.

[0118] As a specific implementation method, the solution of this application is implemented as follows: In the production line, a wideband radio frequency sensor is used to collect electromagnetic spectrum signals, and the signal is decomposed into multiple intrinsic mode components using a variational mode decomposition algorithm. The spectral anisotropy index is calculated using the eigenvalue decomposition of the frequency domain covariance matrix. The energy distribution of the characteristic frequency bands is weighted, where the weight coefficients are dynamically adjusted based on the spectral energy distribution. The autocorrelation coefficient sequence of the signal envelope is calculated, and the fluctuation variance and periodicity metrics are extracted using statistical methods. Finally, a multilayer perceptron is used as a nonlinear mapping function, inputting the above features, and outputting the spectral anomaly offset and the production line cycle time disturbance coefficient.

[0119] Through the above scheme, this application can comprehensively capture subtle abnormal changes in electromagnetic signals and hidden disturbances in production rhythm, realize the quantitative correlation between the degree of electromagnetic environment abnormality and the stability of production rhythm, thereby improving the real-time performance and accuracy of multimodal anomaly coupling analysis.

[0120] In modern supply chain management, the stability and efficiency of the production-sales collaboration process are crucial. However, this process involves multiple entities such as production, warehousing, logistics, and sales, and its operational status is complexly influenced by multiple factors, including operator physiological conditions, inter-departmental collaboration efficiency, and the operating status of production line equipment. Currently, the industry commonly uses independent monitoring systems, such as monitoring the physical parameters of production line equipment or issuing simple keyword alerts for business communications. These methods have significant limitations: First, they can only perform isolated analysis on data of a single modality, failing to capture the inherent coupling relationships between anomalies of different dimensions. Second, due to the lack of deep fusion and collaborative analysis of multi-source heterogeneous data, existing technologies struggle to accurately and in real-time locate the root cause of anomalies from complex data and quantify the impact of each factor on the overall anomaly, resulting in delayed early warnings and difficulties in tracing the source. Therefore, there is an urgent need in this field for a technical solution that can break down data silos and achieve real-time coupled analysis and visualized tracing of multi-modal anomalies to comprehensively improve the anomaly detection and handling capabilities of the supply chain production-sales collaboration process.

[0121] In some of the embodiments described above in this application, variational mode decomposition and frequency domain structure analysis of the electromagnetic spectrum signal of the production line are proposed to obtain the spectral anisotropy index and characteristic frequency band energy distribution. However, in its implementation, the determination of the characteristic frequency band energy distribution only depends on the calculation of static energy values, without considering the influence of the spectral centroid offset on the dynamic changes of energy distribution. As a result, when the electromagnetic environment fluctuates or the equipment operating state changes abruptly, the characteristic frequency band energy distribution cannot accurately characterize the actual spectral characteristics, which in turn affects the calculation accuracy of the spectral anomaly offset and the production line cycle disturbance coefficient, thus reducing the real-time performance and reliability of abnormal event detection.

[0122] In response, this application further proposes variational mode decomposition and frequency domain structure analysis of the electromagnetic spectrum signal of the production line to obtain the spectral anisotropy index and characteristic frequency band energy distribution, including: Variational mode decomposition is performed on the electromagnetic spectrum signal of the production line to obtain multiple intrinsic mode components, and the center frequency and energy value of each intrinsic mode component are extracted.

[0123] A frequency domain covariance matrix is ​​constructed based on the center frequency and energy value. Eigenvalue decomposition is performed on the frequency domain covariance matrix to obtain the spectral anisotropy index. The spectral anisotropy index characterizes the non-uniformity of electromagnetic energy distribution in the frequency domain space.

[0124] The energy distribution of the main characteristic frequency bands is determined based on the energy value.

[0125] Based on the spectral centroid offset and the energy distribution of the main characteristic frequency bands, the energy distribution of the characteristic frequency bands is determined. The spectral centroid offset is determined based on the center frequency.

[0126] In practical applications, variational mode decomposition (VMD) is an adaptive signal decomposition method. It can be implemented using variational mode decomposition algorithms or improved variants of empirical mode decomposition (EMD). Its purpose is to decompose non-stationary electromagnetic spectrum signals into multiple physically meaningful intrinsic mode components (EMCs), avoiding the mode aliasing problem caused by traditional Fourier transforms, thereby obtaining a high-resolution time-frequency representation. EMCs can be understood as basic components in a signal with specific oscillatory characteristics. They can be separated from the original signal through an iterative optimization process, aiming to characterize frequency components at different scales in the electromagnetic spectrum. Specifically, the center frequency refers to the main frequency location where the energy of the EMC is concentrated. It can be determined by calculating the centroid or peak frequency of the component's power spectrum, aiming to identify the core position of the component in the frequency domain. The energy value can be understood as the energy intensity of the EMC in the frequency domain, obtained by integrating the component's power spectrum, aiming to quantify the contribution of this frequency component to the overall signal. The frequency domain covariance matrix is ​​a matrix structure describing the correlation of frequency domain signals. It can be constructed using the joint statistical properties of the center frequency and energy value, aiming to characterize the distribution correlation of electromagnetic energy in the frequency domain space. Eigenvalue decomposition (EVD) can be understood as a matrix factorization technique, implemented using the QR algorithm or the Jacobi method. Its purpose is to extract dominant eigenvectors from the frequency domain covariance matrix to quantify the degree of non-uniformity in frequency domain energy distribution. The spectral anisotropy index is a quantitative indicator characterizing the directionality of electromagnetic energy distribution. It can be calculated by the ratio of the maximum to the minimum eigenvalue after EVD, reflecting potential electromagnetic interference or fault symptoms during equipment operation. The main characteristic frequency band can be understood as a frequency range strongly correlated with the equipment's operating state. It can be determined through energy threshold screening or cluster analysis, focusing on key spectral regions and eliminating noise interference from irrelevant frequency bands. The spectral centroid offset refers to the dynamic shift of the spectral energy centroid relative to a reference position. It can be calculated by weighted averaging of the center frequencies, capturing the impact of electromagnetic environment fluctuations on the spectral distribution in real time. The characteristic frequency band energy distribution refers to the normalized distribution of energy within the key frequency band, obtained through statistical normalization of energy values, providing spectral characteristic representation for subsequent anomaly detection.

[0127] Specifically, the proposed solution decomposes the electromagnetic spectrum signal of the production line into multiple intrinsic mode components through variational mode decomposition, effectively separating multi-scale frequency components in the signal and providing a high-resolution time-frequency basis for frequency domain analysis. A frequency domain covariance matrix is ​​constructed based on the extracted center frequency and energy values, and a spectral anisotropy index is obtained through eigenvalue decomposition. This index quantifies the non-uniformity of electromagnetic energy distribution in the frequency domain, thus reflecting potential electromagnetic interference during equipment operation. Simultaneously, the energy distribution of the main characteristic frequency bands is determined based on the energy values, focusing on spectral regions strongly correlated with the equipment's operating state and excluding interference from irrelevant frequency bands. Crucially, by introducing a spectral centroid offset as a dynamic adjustment factor, the spectral centroid offset is combined with the energy distribution of the main characteristic frequency bands, enabling the characteristic frequency band energy distribution to respond in real-time to the dynamic changes in the spectral centroid. This dynamic adjustment mechanism avoids the distortion defects of static energy distribution calculations under environmental disturbances, ensuring that the characteristic frequency band energy distribution can still accurately characterize the actual spectral characteristics during electromagnetic environment fluctuations or sudden changes in equipment state, thus providing a reliable basis for calculating spectral anomaly offsets and production line cycle disturbance coefficients.

[0128] As a specific implementation method, the solution of this application is implemented as follows: At the SMT placement station of an electronic manufacturing production line, a wideband RF sensor is used to collect the electromagnetic spectrum signal of the production line. A variational mode decomposition algorithm is used to decompose the signal into five intrinsic mode components, and the center frequency and energy value of each component are extracted. A 3×3 frequency domain covariance matrix is ​​constructed based on the center frequency and energy value, and the spectral anisotropy index is calculated through eigenvalue decomposition. Three main characteristic frequency bands with an energy share exceeding 70% are selected based on the energy value, and their initial energy distribution is determined. After calculating the spectral centroid offset, it is used as a weighting coefficient to dynamically adjust the energy distribution of the main characteristic frequency bands, generating a standardized characteristic frequency band energy distribution vector. Finally, this energy distribution vector is input into a nonlinear mapping function to simultaneously calculate the spectral anomaly offset and the production line cycle time disturbance coefficient.

[0129] Through the above technical solution, this application solves the problem of inaccuracy in the calculation of characteristic frequency band energy distribution when the electromagnetic environment changes dynamically. It enables the characteristic frequency band energy distribution to respond in real time to the shift of the spectrum centroid, so that it can still accurately characterize the actual spectrum characteristics when the electromagnetic environment fluctuates or the equipment operating state changes abruptly. It effectively improves the calculation accuracy of spectrum anomaly offset and production line cycle disturbance coefficient, and enhances the real-time performance and reliability of abnormal event detection in the production and sales collaboration process of the supply chain.

[0130] In some of the embodiments described above in this application, a method for determining the energy distribution of a characteristic frequency band based on the spectral centroid offset and the energy distribution of the main characteristic frequency bands is proposed to characterize the degree of electromagnetic environment anomaly. However, in its implementation, the correlation mechanism between the spectral centroid offset and the energy distribution of the characteristic frequency band fails to fully quantify the influence intensity of the dynamic offset of the spectral centroid on the energy distribution. This results in the calculation of the energy distribution of the characteristic frequency band lacking adaptive response capability to the degree of spectral offset, which makes the sensitivity and accuracy of spectral feature extraction in subsequent abnormal event detection insufficient, and makes it difficult to capture subtle abnormal changes in the electromagnetic environment of the production line.

[0131] In response, this application further proposes a method for determining the energy distribution of characteristic frequency bands based on spectral centroid offset and the energy distribution of major characteristic frequency bands, including: The weighting coefficients of each characteristic frequency band are calculated based on the spectral centroid offset. The weighting coefficients are positively correlated with the spectral centroid offset. The weighting coefficients are calculated using a linear transformation function based on the spectral centroid offset.

[0132] The energy distribution of the main characteristic frequency bands is weighted and adjusted based on the weighting coefficients to generate a weighted characteristic frequency band energy distribution vector.

[0133] The weighted characteristic frequency band energy distribution vector is normalized and scaled to obtain a standardized characteristic frequency band energy distribution.

[0134] In practical applications, the spectral centroid offset refers to the dynamic shift of the electromagnetic spectrum energy distribution center relative to a reference position. It can be quantified by the difference between the first moment of the spectral energy and the reference spectral centroid. Specifically, it can be calculated by performing a discrete Fourier transform on the real-time acquired electromagnetic spectrum signal, aiming to dynamically characterize the changing trend of the electromagnetic energy distribution center. The weighting coefficient can be understood as a quantitative parameter reflecting the intensity of the spectral offset's influence. It can be implemented using a linear proportionality coefficient or a piecewise linear function, specifically generated by inputting the spectral centroid offset into a preset linear mapping model. Its purpose is to establish a quantitative correlation between spectral offset and energy distribution adjustment. Specifically, the linear transformation function refers to the mathematical relationship that maps the spectral centroid offset to the weighting coefficient. It can be implemented using a linear equation with adjustable slope or a linear interpolation method with threshold constraints. Specifically, different proportionality coefficients can be configured to adapt to the electromagnetic characteristics of different production line equipment, aiming to ensure the linear response characteristics of the weighting coefficient as the offset changes. Furthermore, normalization refers to the standardization operation that eliminates dimensional differences. It can be achieved using maximum value normalization or L2 norm normalization methods, specifically by dividing by the vector magnitude or the maximum element value. Its purpose is to ensure the numerical stability of the characteristic frequency band energy distribution vector. Scale transformation can be understood as a transformation operation that adjusts the numerical range. It can be achieved using logarithmic transformation or piecewise scaling methods. Specifically, it can map the data to the target interval using a preset nonlinear function, aiming to adapt to the input requirements of the subsequent anomaly analysis module.

[0135] This application's solution dynamically transforms the spectral centroid offset into weighting coefficients, enabling real-time calculation of characteristic frequency band energy distribution to respond to changes in the electromagnetic environment. First, weighting coefficients are calculated based on the spectral centroid offset. A linear transformation function is used to establish a positive correlation between the offset and the weights, ensuring that the larger the spectral distribution center offset, the higher the weighting coefficient, thus quantifying the impact of the offset on the energy distribution. Second, the weighting coefficients are used to weight and adjust the energy distribution of the main characteristic frequency bands, differentially amplifying the energy in the frequency bands affected by the offset, preventing key anomaly information from being homogenized and generating a weighted vector that highlights anomaly characteristics. Finally, the weighted vector is normalized and scaled to eliminate differences in energy magnitude between different production line equipment, outputting a standardized characteristic frequency band energy distribution with a unified numerical range, providing stable input for subsequent calculations of spectral anomaly offsets. This process, through the close integration of three stages—dynamically quantifying the impact of spectral offset, strengthening the characterization of key frequency bands, and unifying data scales—forms a complete feature optimization chain, effectively solving the problem of insufficient sensitivity of spectral feature extraction to subtle anomaly changes.

[0136] As a specific implementation method, the solution of this application is implemented as follows: During the electromagnetic spectrum signal processing of the production line, when an increase in the centroid offset of the spectrum is detected, the linear transformation function automatically generates corresponding weighting coefficients. For example, the offset is mapped to the interval [0.1, 1.0] using the linear equation y = 0.8x + 0.1. Subsequently, the weighting coefficients are used to perform element-wise multiplication on the energy distribution of the main characteristic frequency bands, thereby amplifying the energy value in the high-frequency band region. Finally, normalization is achieved by dividing by the maximum value of the vector, and a logarithmic function is used for scaling transformation to compress the energy distribution vector to the range [0, 1]. In this process, the electromagnetic spectrum signal continuously acquired by the broadband RF sensor is subjected to variational mode decomposition, and the dynamic adjustment of its characteristic frequency band energy distribution can reflect the subtle fluctuations in the operating status of the production line equipment in real time. For example, when the equipment motor experiences slight vibration, abnormal energy changes in a specific frequency band can be captured and highlighted.

[0137] Through the above technical solution, this application achieves an adaptive response of characteristic frequency band energy distribution to dynamic shift of the spectral centroid, improving the sensitivity and accuracy of spectral feature extraction. Specifically, the dynamic quantization mechanism of the spectral centroid shift enables energy distribution calculation to track changes in the electromagnetic environment in real time. The weighted adjustment process effectively highlights key frequency band regions affected by anomalies, while normalization and scaling transformation ensure the comparability of feature data under different operating conditions. This allows for the identification of subtle abnormal changes in the electromagnetic environment of the production line, providing reliable spectral feature input for real-time detection of supply chain anomalies.

[0138] Specifically, in some of the embodiments described above in this application, supply chain anomaly detection is proposed through multimodal data streams. However, in this process, although multi-dimensional data such as operator physiological status, business communication text, and production line electromagnetic spectrum are acquired and independently analyzed, there is a lack of dynamic modeling mechanism for the abnormal coupling relationship between supply chain entities. It is impossible to quantify the contribution of each modal indicator to the overall anomaly in real time, which makes it difficult to locate the root cause of the anomaly and assess the comprehensive impact of multi-dimensional factors.

[0139] To address this, this application further proposes a method based on real-time attention distraction index, physiological stress level, collaborative health score, spectral anomaly offset, and production line cycle time disturbance coefficient. This method utilizes a dynamic graph neural network to perform multimodal anomaly coupling analysis, outputting the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modality indicator. (See [link to relevant documentation]). Figure 7 Taking the server as the executing entity as an example, the following steps are included.

[0140] 701. Based on real-time attention distraction index and physiological stress level, the abnormal state vector of supply chain entities in the physiological dimension is generated through the first graph neural layer of dynamic graph neural network.

[0141] 702. Based on the collaborative health score, spectrum anomaly offset and production line cycle disturbance coefficient, a multimodal coupled edge weight matrix is ​​constructed, and the edge weight matrix is ​​used to update the graph structure representation of the dynamic graph neural network.

[0142] 703. Input the abnormal state vector and the updated graph structure representation into the second graph neural layer of the dynamic graph neural network, perform anomaly propagation calculation through the cross-modal graph attention mechanism, and synchronously output the comprehensive anomaly probability of each supply chain entity and the contribution distribution of each modality indicator.

[0143] In this context, a dynamic graph neural network (GNN) refers to a neural network model capable of adapting to dynamic changes in the graph structure. It can be implemented using variants of graph convolutional networks or graph attention networks, aiming to handle the characteristics of entity relationships in a supply chain network evolving over time. The first graph neural layer can be understood as the initial processing unit of the dynamic graph neural network, implemented using a multilayer perceptron or graph convolution operations, aiming to transform physiological indicators into structured vector representations. The abnormal state vector specifically refers to a vector characterizing the degree of abnormality in supply chain entities, which can be represented as a real-number vector, aiming to quantify abnormal states caused by human factors. The edge weight matrix defines the connection strength between nodes in the graph, which can be calculated based on multimodal data, aiming to dynamically quantify the intensity of anomaly propagation between entities. The graph structure representation can be understood as the topological model of the supply chain network, implemented using adjacency matrices or graph embeddings, aiming to reflect the business relationships between entities. The second graph neural layer specifically refers to the subsequent processing unit of the dynamic graph neural network, implemented using a gated graph neural network or attention mechanism, aiming to perform cross-modal anomaly propagation calculations. Cross-modal graph attention mechanisms refer to attention mechanisms that can fuse multi-source data. They can be implemented using multi-head attention networks, aiming to adaptively weight anomaly information from different dimensions. The comprehensive anomaly probability specifically refers to the probability value of anomalies occurring in supply chain entities, which can be output using a sigmoid function, with the aim of locating high-risk nodes. The contribution distribution can be understood as the proportion of each modal indicator's contribution to the anomaly state; it can be obtained through gradient analysis, aiming to reveal the multi-source causes of anomalies.

[0144] Specifically, the proposed solution first processes the real-time attention distraction index and physiological stress level through a first graph neural layer to generate a structured vector reflecting human-related anomalies, thereby transforming the operator's physiological state into quantifiable node features. Second, an edge weight matrix is ​​constructed based on collaborative health scores, spectral anomaly offsets, and production line rhythm disturbance coefficients. This matrix integrates business collaboration health, electromagnetic environment anomaly levels, and production rhythm stability to dynamically adjust the connection strength between supply chain entities, enabling the graph structure to update in real time with changes in multimodal data. Finally, the abnormal state vector and the updated graph structure representation are input into a second graph neural layer. Anomaly propagation is performed on the enhanced graph adjacency matrix using a cross-modal graph attention mechanism. The node abnormal state representation is updated through a gating mechanism, and the comprehensive anomaly probability and the contribution distribution of each modality are output simultaneously, forming a complete analysis chain from the physiological dimension to multimodal coupling. This process achieves deep coupling of multi-dimensional anomalies through hierarchical dynamic modeling, avoiding the shortcomings of static modeling of entity relationships in traditional methods, and enabling the anomaly propagation path and contribution analysis to be adjusted in real time with dynamic changes in the supply chain.

[0145] As a preferred embodiment, the solution of this application is specifically implemented as follows: The dynamic graph neural network can be specifically implemented as a model built based on the PyTorch Geometric framework. The first graph neural layer uses graph convolutional layers to process physiological indicators, mapping supply chain entities as graph nodes, with business department relationships as initial edges. The edge weight matrix is ​​dynamically generated by calculating the correlation factor between the collaborative health score and the spectral anomaly offset, and an exponential decay function is used to adjust the retention degree of historical graph structure information. The second graph neural layer deploys a multi-head attention network to adaptively weight anomaly information in physiological and non-physiological dimensions. The output layer generates a comprehensive anomaly probability using a sigmoid activation function, and simultaneously uses the attention gradient matrix to analyze the contribution of each modality. In actual deployment, supply chain entities include production work orders, warehousing nodes, and logistics units. Their business topology relationships are obtained in real time through the enterprise resource planning system. The dynamic graph neural network updates the graph structure representation every 5 minutes to ensure that anomaly analysis is synchronized with the supply chain operation status.

[0146] Through the above scheme, this application can quantify the contribution of each modal indicator to the overall anomaly in real time, locate the root cause of the anomaly and assess the comprehensive impact of multi-dimensional factors, effectively solve the problem of anomaly tracing caused by isolated processing of multi-dimensional data, and improve the timeliness and interpretability of anomaly detection in the production and sales collaboration process of the supply chain.

[0147] Specifically, in some of the embodiments described above in this application, an abnormal state vector of the supply chain entity in the physiological dimension is generated based on the real-time attention distraction index and physiological stress level. However, in its implementation, the dynamic coupling relationship and statistical correlation between the two are not fully considered, which results in the generated abnormal state vector failing to accurately capture the comprehensive abnormal pattern of the operator's physiological state, thereby affecting the reliability and real-time performance of multimodal abnormal coupling analysis.

[0148] In response, this application further proposes a method for generating abnormal state vectors of supply chain entities in the physiological dimension through the first graph layer of a dynamic graph neural network, based on real-time attention distraction index and physiological stress level, including: The real-time attention distraction index and physiological stress level were standardized for time series processing, and the statistical correlation coefficient and covariance matrix of the two were calculated.

[0149] Based on the statistical correlation coefficient and covariance matrix, an attention stress coupling feature matrix is ​​constructed, and the principal component eigenvectors of the attention stress coupling feature matrix are calculated.

[0150] The principal component feature vector is concatenated with the basic attribute features of the supply chain entities and input into the graph convolutional network of the first graph neural layer. Through neighbor node information aggregation and activation function transformation, an abnormal state vector is generated. The dimension of the abnormal state vector is consistent with the number of supply chain entities, and the value of each dimension reflects the degree of abnormality of the corresponding entity in the physiological dimension.

[0151] Time series standardization refers to the process of converting physiological indicator data of different dimensions into a uniform scale. This can be achieved using Z-score standardization or Min-Max normalization methods, aiming to eliminate differences in data dimensions and ensure the accuracy of subsequent correlation analysis. Statistical correlation coefficients and covariance matrices are statistical measures that quantify the linear correlation between real-time attention distraction index and physiological stress level. They can be calculated based on Pearson correlation coefficients or Spearman rank correlation coefficients, aiming to capture the dynamic coupling relationship between the two. The attention-stress coupling feature matrix is ​​a matrix representation integrating the correlation information between attention distraction and physiological stress. It can be achieved by mapping the elements of the statistical correlation coefficients and covariance matrix to matrix elements, aiming to construct structured data reflecting the coupling pattern of physiological indicators. Principal component eigenvectors refer to the main variation directions of the attention-stress coupling feature matrix extracted through principal component analysis. They can be calculated using eigenvalue decomposition or singular value decomposition algorithms, aiming to reduce dimensionality, highlight key anomalous features, and filter out noise interference. Neighbor node information aggregation refers to the operation of collecting and integrating the state information of adjacent supply chain entities in a graph neural network. This can be achieved through weighted averaging or max pooling, with the aim of capturing the propagation effect of anomalies between entities using the supply chain network topology. Activation function transformation refers to the nonlinear mapping of the aggregated node features, which can be implemented using ReLU or Sigmoid functions. The purpose is to enhance the expressive power of the model, enabling the abnormal state vector to more accurately reflect the degree of physiological abnormality.

[0152] Specifically, the proposed solution first performs time-series standardization on the real-time attention distraction index and physiological stress level to eliminate dimensional differences and establish a unified analytical benchmark. Based on this, the statistical correlation coefficient and covariance matrix of the two are calculated to quantify the dynamic correlation strength between operator attention distraction and physiological stress. Subsequently, an attention-stress coupling feature matrix is ​​constructed based on the statistical correlation coefficient and covariance matrix, and its principal component feature vector is extracted through principal component analysis to effectively filter noise interference and highlight key abnormal features. Next, the principal component feature vector is concatenated with the basic attribute features of the supply chain entities, so that the abnormal state vector simultaneously contains real-time physiological dynamic information and the entity's inherent business attributes. Finally, the concatenated features are input into the graph convolutional network of the first graph neural layer. The physiological state information of associated supply chain entities is collected through neighbor node information aggregation operations, and the final abnormal state vector is generated through activation function transformation. This ensures that the values ​​of each dimension can intuitively quantify the degree of abnormality of the corresponding entity in the physiological dimension, thus providing input for multimodal abnormal coupling analysis.

[0153] As a specific implementation method, the solution of this application is implemented as follows: In the supply chain production and sales collaboration system, after obtaining the real-time attention distraction index and physiological stress level data of operators, the time series data is first processed using the Z-score normalization method. Then, the Pearson correlation coefficient and covariance matrix of the two are calculated. Based on the calculation results, a 3×3 dimension attention stress coupling feature matrix is ​​constructed, and principal component feature vectors are obtained through eigenvalue decomposition. This feature vector is concatenated with the basic attribute feature vectors representing the entity's job type and work order complexity. Finally, the concatenated vector is input into a graph neural network module composed of graph convolutional layers and ReLU activation functions. By aggregating the physiological state information of adjacent workstation entities, an abnormal state vector consistent with the number of supply chain entities is generated.

[0154] Through the above scheme, this application can systematically integrate the dynamic coupling characteristics of physiological indicators, accurately capture the comprehensive abnormal patterns of operator physiological state, effectively improve the reliability and real-time performance of multimodal abnormal coupling analysis, and thus provide more physiological dimension basis for the detection of abnormal events in the supply chain production and sales collaboration process.

[0155] In practical applications, when constructing edge weight matrices based on multimodal data to support anomaly propagation calculations in dynamic graph neural networks, if the edge weights rely solely on static business topology distances and fail to effectively integrate the dynamic coupling relationships between collaborative health, spectral anomaly offset, and production line cycle disturbance coefficients, the anomaly propagation intensity between supply chain entities cannot accurately reflect the real-time interactive impact of multimodal anomalies. This leads to distortion of the contribution of key modes during anomaly tracing, reducing the reliability of the overall anomaly probability and the accuracy of visualized tracing.

[0156] In response, this application further proposes a multimodal coupling edge weight matrix based on collaborative health score, spectral anomaly offset, and production line cycle time disturbance coefficient, including: The first correlation factor between the collaborative health score and the spectrum anomaly offset was determined, and the second correlation factor between the production line cycle disturbance coefficient and the spectrum anomaly offset was determined.

[0157] Based on the first and second correlation factors, the relationship strength weights between supply chain entities are determined, and the relationship strength weights are used to fuse the business topology distances between entities to generate an initial edge weight matrix.

[0158] The multimodal coupling coefficient is obtained by fusing the collaborative health score, spectral anomaly offset, and production line cycle disturbance coefficient.

[0159] The initial edge weight matrix is ​​dynamically adjusted using multimodal coupling coefficients to obtain a multimodal coupling edge weight matrix. Each element of the edge weight matrix reflects the intensity of anomaly propagation based on multimodal data among the corresponding supply chain entities.

[0160] The first correlation factor refers to the quantitative correlation index between the collaborative health score and the spectrum anomaly offset. It can be calculated using the Pearson correlation coefficient or mutual information, aiming to capture the coupling pattern between cross-departmental collaboration anomalies and equipment electromagnetic environment anomalies. The second correlation factor refers to the quantitative correlation index between the production line cycle time disturbance coefficient and the spectrum anomaly offset. It can be determined using the Spearman rank correlation coefficient or dynamic time warping distance, aiming to characterize the dynamic correlation characteristics between production rhythm anomalies and electromagnetic environment anomalies. The relationship strength weight refers to the anomaly transmission strength index between supply chain entities calculated based on the multimodal correlation factors. It can be generated through weighted averaging or a fuzzy logic system, aiming to quantify the potential strength of anomaly propagation between entities. The initial edge weight matrix refers to the graph structure base matrix constructed by integrating business topology distance and relationship strength weight. It can be implemented through matrix weighting or graph embedding techniques, aiming to preserve the inherent hierarchical relationships of the supply chain while introducing data-driven anomaly transmission paths. The multimodal coupling coefficient is a comprehensive fusion index of the collaborative health score, spectral anomaly offset, and production line cycle disturbance coefficient. It can be generated using principal component analysis or neural network fusion models, with the aim of uniformly characterizing the degree of joint impact of multimodal anomalies.

[0161] Specifically, the proposed solution first identifies a first correlation factor between the collaborative health score and the spectrum anomaly offset, and a second correlation factor between the production line rhythm disturbance coefficient and the spectrum anomaly offset. This effectively captures specific coupling patterns between cross-departmental collaboration anomalies and equipment electromagnetic environment anomalies, as well as between production rhythm anomalies and electromagnetic environment anomalies, avoiding the correlation blind spots caused by single-modal analysis. Subsequently, based on these correlation factors, the relationship strength weights between supply chain entities are determined, and these weights are fused with the business topology distance between entities to generate an initial edge weight matrix. This achieves synergy between business logic constraints and data-driven correlations, preserving the inherent hierarchical relationships of the supply chain while dynamically correcting implicit anomaly propagation paths not reflected in the business topology. Furthermore, the collaborative health score, spectrum anomaly offset, and production line rhythm disturbance coefficient are fused to obtain a multimodal coupling coefficient, which serves as a unified adjustment parameter to ensure that the synergistic effects of multimodal anomalies are considered when adjusting edge weights. Finally, the initial edge weight matrix is ​​dynamically adjusted using the multimodal coupling coefficient, so that the edge weights can be adaptively optimized with the real-time changes of multimodal data. This quantifies the instantaneous intensity of anomaly propagation among supply chain entities and provides a graph structure foundation that conforms to the multimodal coupling characteristics for subsequent anomaly tracing.

[0162] As a preferred embodiment, the solution of this application is implemented as follows: When the system detects an increase in the abnormal spectral offset of a precision assembly station, a first correlation factor between the anomaly and the collaborative health score of the adjacent quality inspection department can be calculated. If the analysis shows a strong negative correlation between the two, the connection weight between the assembly station and the quality inspection department is enhanced in the initial edge weight matrix. Simultaneously, a second correlation factor between the production line cycle disturbance coefficient and the abnormal spectral offset is calculated. If the cycle disturbance and the abnormal spectral offset are found to intensify synchronously, the weights between related entities are further adjusted. Subsequently, a multimodal coupling coefficient is obtained by fusing three modal indicators, and the initial edge weight matrix is ​​dynamically scaled so that the final edge weight matrix accurately reflects the coupling state of the current multimodal anomaly, thereby achieving more accurate anomaly propagation calculation in the dynamic graph neural network.

[0163] Through the above scheme, this application enables the intensity of anomaly propagation among supply chain entities to accurately reflect the real-time interactive impact of multimodal anomalies, avoids distortion of key modality contribution during anomaly tracing, and improves the reliability of comprehensive anomaly probability and the accuracy of visualized tracing.

[0164] In modern supply chain management, the stability and efficiency of the production-sales collaboration process are crucial. However, this process involves multiple entities, including production, warehousing, logistics, and sales, and its operational status is complexly influenced by multiple factors, such as operator physiological conditions, inter-departmental collaboration efficiency, and the operating status of production line equipment. Currently, the industry commonly uses independent monitoring systems, such as monitoring the physical parameters of production line equipment or issuing simple keyword alerts for business communications. These methods have significant limitations: First, they can only perform isolated analysis on data of a single modality, failing to capture the inherent coupling relationships between anomalies of different dimensions. Second, due to the lack of deep fusion and collaborative analysis of multi-source heterogeneous data, existing technologies struggle to accurately and in real-time locate the root cause of anomalies from complex data and quantify the impact of each factor on the overall anomaly, resulting in delayed early warnings and difficulties in tracing the source. In some of the embodiments described above in this application, a graph structure representation of a dynamic graph neural network updated with an edge weight matrix is ​​proposed to fuse multimodal data to achieve dynamic modeling of anomaly propagation among supply chain entities. However, in its implementation, the update of the edge weight matrix lacks an adaptive processing mechanism for the dynamic characteristics of time, which causes the graph structure representation to be unable to effectively respond to the real-time fluctuations of edge weights in the supply chain environment. The retention degree of historical graph structure information is fixed and disconnected from the current abnormal state, resulting in unreasonable distribution of message weights among nodes during anomaly propagation calculation, thereby affecting the timeliness and accuracy of the comprehensive anomaly probability output.

[0165] In response, this application further proposes a graph structure representation for updating a dynamic graph neural network using an edge weight matrix, including: The edge weight matrix is ​​normalized to obtain a standardized edge weight matrix.

[0166] The standardized edge weight matrix is ​​combined with the original graph adjacency matrix using the Hadamard product to generate an enhanced graph adjacency matrix.

[0167] Based on the enhanced graph adjacency matrix, the graph structure representation of the dynamic graph neural network is updated, and a dynamic decay factor of the graph structure representation is calculated to adjust the message propagation weights between nodes. The dynamic decay factor is determined based on the time-series rate of change of edge weights and is used to control the degree to which historical graph structure information is retained in the current update.

[0168] In practical applications, the edge weight matrix is ​​a numerical matrix characterizing the intensity of anomaly propagation among supply chain entities. It can be implemented using a dynamically generated weight matrix based on multimodal coupling coefficients, aiming to quantify the real-time strength of anomaly associations between entities. Normalization refers to mapping the element values ​​of the edge weight matrix to a uniform numerical range. This can be achieved using maximum value normalization or Z-score standardization, aiming to eliminate dimensional differences in edge weights from different sources and ensure the comparability of subsequent fusion calculations. Specifically, the Hadamard product operation is an algebraic operation of multiplying corresponding elements of two matrices. It can be implemented using a parallel computing framework to perform element-wise multiplication, aiming to fuse the structural constraints of the original business topology with edge weights driven by real-time multimodal data. In practical applications, the dynamic decay factor is an adaptive parameter that regulates the degree of retention of historical graph structural information. It can be generated based on a nonlinear function of the time series rate of change, aiming to dynamically balance the system's response sensitivity and stability according to the fluctuation characteristics of edge weights.

[0169] Specifically, the proposed solution eliminates dimensional differences in the edge weight matrix through normalization, making edge weights from different modalities comparable and providing a reliable foundation for subsequent Hadamard product operations. Based on this, the standardized edge weight matrix is ​​combined with the original graph adjacency matrix using Hadamard product operations, achieving element-level fusion of business topology logic and real-time anomaly propagation strength, generating an enhanced graph adjacency matrix that retains the inherent connections within the supply chain while injecting multimodal dynamic information. Based on this enhanced graph adjacency matrix, the system simultaneously updates the graph structure representation and calculates the dynamic decay factor: on the one hand, it reconstructs the connection strength between nodes using the enhanced graph adjacency matrix; on the other hand, it determines the dynamic decay factor by quantifying the time-series change rate of edge weights (such as variance or derivative within a sliding window). This factor adaptively adjusts the degree of historical information retention based on the magnitude of the change rate—increasing the decay strength to quickly respond to sudden anomalies when edge weights fluctuate drastically, and decreasing the decay strength to maintain system stability when fluctuations are moderate. Finally, the dynamic decay factor finely controls the message propagation weights between nodes, ensuring that anomaly propagation calculations reflect the real-time dynamic characteristics of the supply chain environment.

[0170] As a specific implementation method, the solution of this application is implemented as follows: In the supply chain production and sales collaboration system, the edge weight matrix is ​​generated by fusing collaboration health score, spectrum anomaly offset, and production line cycle time disturbance coefficient. Normalization is performed using the maximum value normalization method, scaling the matrix elements to the [0,1] interval. The Hadamard product operation is executed through a GPU parallel computing unit, multiplying the standardized edge weight matrix element-wise with the original graph adjacency matrix representing the supplier-manufacturer-distributor hierarchical relationship. The dynamic decay factor is calculated based on the standard deviation of the edge weights within a 5-minute sliding window. When the standard deviation exceeds a threshold, the decay factor linearly increases to 0.8 to rapidly decay historical information; otherwise, the decay factor remains at 0.3 to preserve historical topological features. The updated graph structure represents the second graph neural layer input to the dynamic graph neural network, used for subsequent anomaly propagation calculations.

[0171] Through the above technical solution, this application can adaptively adjust the retention level of historical graph structure information according to the time series change rate of edge weights, enabling the graph neural network to distinguish between normal fluctuations and real abnormal events in the supply chain environment. The introduction of a dynamic decay factor effectively solves the problem of fixed historical information retention, ensuring that the message propagation weights between nodes are closely related to the current abnormal state. Ultimately, this achieves a simultaneous improvement in the timeliness and accuracy of abnormal propagation calculation, providing a reliable basis for outputting the comprehensive abnormal probability of supply chain entities.

[0172] Specifically, in some of the embodiments described above in this application, anomaly propagation calculation is proposed by inputting the abnormal state vector and the updated graph structure representation into the second graph neural layer of a dynamic graph neural network to output the comprehensive anomaly probability and contribution distribution. However, in its implementation, traditional methods lack the ability to fine-grainedly analyze cross-modal anomaly coupling relationships and cannot simultaneously realize the quantification of anomaly propagation and the contribution of each modality, resulting in the anomaly tracing process relying on human experience and being inefficient.

[0173] To address this, this application further proposes inputting the abnormal state vector and the updated graph structure representation into the second graph neural layer of a dynamic graph neural network. Anomaly propagation is calculated through a cross-modal graph attention mechanism, simultaneously outputting the comprehensive anomaly probability of each supply chain entity and the contribution distribution of each modality indicator, including: Based on the abnormal state vector and the updated graph structure representation, cross-modal attention coefficients between nodes are calculated through a multi-head attention network.

[0174] By utilizing cross-modal attention coefficients, anomalous state propagation is performed on the enhanced graph adjacency matrix, and the anomalous state representation of each node is updated through a gated graph neural network.

[0175] Based on the updated node abnormal state representation, the comprehensive abnormal probability of each supply chain entity is determined by the Sigmoid activation function of the output layer. At the same time, the contribution distribution of each modal indicator to the abnormal state is analyzed by backpropagation analysis of the cross-modal attention coefficient.

[0176] Among them, the cross-modal graph attention mechanism refers to a computational framework capable of dynamically modeling the correlation between multi-source heterogeneous data. It can be implemented using dynamic graph structure optimization techniques based on attention weights, aiming to capture the nonlinear interaction characteristics of physiological, collaborative, and spectral modalities in the frequency and semantic dimensions. Multi-head attention networks can be neural network structures that enhance feature parsing capabilities by computing multiple attention heads in parallel. They can be implemented using a variant of the multi-head self-attention mechanism in the Transformer architecture, aiming to extract modal coupling relationships from different feature subspaces and improve the accuracy of identifying complex anomaly propagation paths. The enhanced graph adjacency matrix refers to a normalized edge weight matrix generated by fusing normalization processing and Hadamard product operations. It can be implemented using matrix fusion techniques based on dynamic decay factors, aiming to integrate the original business topology with multimodal coupling strength to form an optimized graph structure reflecting anomaly propagation characteristics. Gated graph neural networks are variants of graph neural networks integrating gating mechanisms. They can be implemented by combining GRU structures with graph convolution operations, aiming to dynamically adjust the information flow intensity based on cross-modal attention coefficients, effectively suppressing false propagation caused by spectral noise or communication delays. The Sigmoid activation function is a non-linear transformation function that maps continuous state values ​​to probability intervals. It can be implemented using the standard Sigmoid function or its smoothed variants, aiming to transform the representation of anomalous node states into intuitively quantifiable anomalous probability values. Backpropagation analysis is a technique that calculates the contribution of analytical features using gradients. It can be implemented using gradient backpropagation techniques based on attention coefficients, aiming to establish an interpretable correlation between anomalous probabilities and various modal indicators, thereby quantifying the distribution of contribution.

[0177] Specifically, the proposed solution first generates cross-modal attention coefficients between nodes based on the abnormal state vector and the updated graph structure representation, using a multi-head attention network. These coefficients dynamically reflect the abnormal coupling strength of supply chain entities in physiological, collaborative, and spectral dimensions. Subsequently, abnormal state propagation is performed on the enhanced graph adjacency matrix using these attention coefficients. The gated graph neural network adaptively adjusts the information update intensity based on the attention coefficients, ensuring that the abnormal propagation process effectively filters noise interference and preserves key coupling relationships. Building upon this, the updated node abnormal state representation is converted into a comprehensive abnormal probability using a sigmoid activation function, achieving an intuitive quantification of the abnormality level. Furthermore, through backpropagation analysis of the cross-modal attention coefficients, the contribution distribution of each modality index to the abnormal state is directly derived. This allows abnormal propagation and contribution analysis to be completed synchronously within a unified computational framework, avoiding the complex process of constructing independent models in traditional methods, thus forming a complete closed loop for multimodal abnormal coupling analysis.

[0178] As a preferred embodiment, the solution of this application is implemented as follows: In the supply chain production and sales collaborative monitoring system, after the dynamic graph neural network receives the abnormal state vector and the updated graph structure representation, the multi-head attention network uses four parallel attention heads to process the feature interactions of the physiological modality, the collaborative modality, and the spectral modality respectively. Each attention head generates a corresponding cross-modal attention coefficient. These coefficients are used to perform message passing operations on the enhanced graph adjacency matrix. The gated graph neural network dynamically adjusts the information update intensity according to the attention coefficients. For example, when the attention coefficient is high, it enhances information passing to strengthen key coupling paths, and when the coefficient is low, it suppresses noise propagation. The updated node state is converted into an abnormal probability value between 0 and 1 through the Sigmoid function. At the same time, the system determines the relative contribution weights of the physiological modality, the collaborative modality, and the spectral modality to the abnormal state by calculating the gradient of the abnormal probability with respect to the attention coefficients. Finally, the contribution distribution of each modality is displayed in the visualization interface in the form of a heatmap.

[0179] Through the above scheme, this application achieves simultaneous analysis of anomaly propagation and the contribution of each modality, effectively overcoming the problems of low efficiency and reliance on manual experience in traditional methods for anomaly tracing. The cross-modal graph attention mechanism can capture the coupling relationships between multimodal data with fine granularity, making the anomaly propagation process more efficient. Backpropagation analysis directly derives the contribution distribution from the attention coefficients, avoiding additional computational overhead and improving the automation and timeliness of anomaly tracing, providing reliable technical support for real-time anomaly handling in the supply chain production and sales collaboration process.

[0180] Specifically, in some of the embodiments described above in this application, a cross-modal graph attention mechanism is proposed to output the contribution distribution of each modality indicator to quantify the impact of anomalies. However, in its implementation, the calculation of the contribution distribution lacks an analytical mechanism based on backpropagation, which makes it impossible to accurately quantify the relative contribution of each modality indicator (such as physiological modality, cooperative modality and spectral modality) to the abnormal state, thereby affecting the reliability and interpretability of anomaly tracing.

[0181] In response, this application further proposes to analyze the contribution distribution of each modal index to the abnormal state through backpropagation analysis of the cross-modal attention coefficients, including: In the output layer of the dynamic graph neural network, the partial derivative of the comprehensive anomaly probability with respect to the cross-modal attention coefficients is calculated to obtain the attention gradient matrix.

[0182] Based on the attention gradient matrix, gradient-weighted aggregation is performed on the attention channels corresponding to the physiological modality, collaborative modality, and spectral modality to obtain the initial contribution weights of each modality.

[0183] The initial contribution weights are subjected to softmax normalization to obtain a standardized contribution distribution, where the value of each dimension reflects the relative contribution of the corresponding modal index to the abnormal state.

[0184] The attention gradient matrix refers to the partial derivative matrix calculated through backpropagation. It can be implemented using an automatic differentiation framework (such as PyTorch or TensorFlow) to establish a differentiable mathematical relationship between the overall anomaly probability and cross-modal attention coefficients, avoiding quantization bias caused by relying on heuristic rules. Gradient weighted aggregation refers to the gradient aggregation operation performed on attention channels corresponding to different modalities. It can be implemented using a modality classifier or predefined channel mapping relationships to separate interference caused by multimodal data coupling, allowing the contributions of different dimensions such as physiological state, business collaboration, and device spectrum to be quantified independently. In practical applications, softmax normalization refers to the standardization process of converting the initial contribution weights into a probability distribution. It can be implemented using an exponential normalization function to give the contribution values ​​relative proportionality and normalization characteristics, supporting the intuitive identification of dominant anomaly modalities in visual tracing.

[0185] Specifically, the proposed solution calculates the partial derivative of the comprehensive anomaly probability with respect to the cross-modal attention coefficients at the output layer of a dynamic graph neural network to obtain the attention gradient matrix. This step utilizes the gradient information of the neural network to backtrack the sensitivity changes of the attention coefficients, providing a differentiable theoretical basis for contribution analysis. Subsequently, based on the attention gradient matrix, gradient-weighted aggregation is performed on the attention channels corresponding to physiological, collaborative, and spectral modalities. This step ensures that the contribution calculation reflects the unique role mechanism of each modality in anomaly propagation through modality classification and gradient weighting. Finally, the initial contribution weights are softmax normalized to generate a standardized contribution distribution, so that the value of each dimension directly represents the relative importance of the corresponding modality index, thereby supporting the tracing and interpretability analysis of anomaly events.

[0186] As a specific implementation method, the solution of this application is implemented as follows: In the output layer of the dynamic graph neural network, the automatic differentiation function of the PyTorch framework is used to calculate the partial derivative of the comprehensive anomaly probability with respect to the cross-modal attention coefficients, thereby obtaining the attention gradient matrix. Based on the predefined modal channel index mapping relationship, the channels belonging to the physiological modality, collaborative modality, and spectral modality in the attention gradient matrix are weighted and summed respectively to obtain the initial contribution weights of each modality. Subsequently, the standard softmax function is applied to normalize the initial contribution weights to generate a contribution distribution vector, where each element of the vector corresponds to the relative contribution ratio of the physiological modality, collaborative modality, or spectral modality.

[0187] Through the above scheme, this application realizes the quantification of the contribution of each modal indicator, solves the problem of inaccurate contribution analysis, improves the reliability and interpretability of anomaly tracing, and enables supply chain managers to accurately identify the dominant abnormal modalities and take targeted measures.

[0188] Specifically, in some of the embodiments described above in this application, an anomaly detection mechanism based on the comprehensive anomaly probability and the contribution distribution of each modality index is proposed. However, in its implementation, there is a lack of effective correlation mapping between the anomaly detection results and the original multimodal data fragments, which makes the anomaly tracing process of high-risk entities highly abstract and not intuitive. Operators find it difficult to quickly locate the root cause of the anomaly and understand the coupling relationship between multimodal data, thereby affecting the timeliness and accuracy of supply chain collaborative decision-making.

[0189] In response, this application further proposes a method based on the comprehensive anomaly probability and the contribution distribution of each modal indicator. Through time window matching and entity identification association, corresponding eye-tracking trajectory fragments, communication text fragments, and electromagnetic spectrum fragments are extracted from the multimodal data stream. This generates a visual interface that integrates business topology and multimodal source tracing evidence, including: High-risk supply chain entities are identified based on comprehensive anomaly probability, and dominant anomaly modes are determined based on contribution distribution. Corresponding time series data segments are extracted from multimodal data streams using a time window matching algorithm.

[0190] By using entity identifier resolution technology, the extracted eye-tracking trajectory fragments, communication text fragments, and electromagnetic spectrum fragments are associated and mapped with the corresponding supply chain entities to establish a multimodal evidence chain for abnormal events.

[0191] Based on the multimodal evidence chain and supply chain business topology, a visualization interface that integrates business topology and multimodal traceability evidence is generated through a visualization engine.

[0192] Among them, the time window matching algorithm refers to a technical solution that dynamically adjusts the data extraction window range based on the timestamp of the abnormal event. It can be implemented using an adaptive window expansion mechanism based on the rate of change of the anomaly probability or a predictive window positioning based on historical anomaly event patterns, aiming to ensure the spatiotemporal consistency of the extracted data fragments with the anomaly event. Entity identifier resolution technology refers to a technical means of achieving semantic alignment of multi-source heterogeneous data using unique identifiers of supply chain entities. It can be implemented using an entity identifier mapping table based on a distributed ledger or an entity relationship index structure based on a graph database, aiming to solve the problem of broken associations caused by heterogeneous formats of different modal data. Multimodal evidence chain refers to an anomaly data association structure across physiological, collaborative, and device dimensions. It can be constructed as an evidence chain list based on timestamp sequences or an evidence network based on a graph structure, aiming to clearly characterize the propagation path of anomalies across multiple dimensions. The visualization engine refers to a software component that generates integrated business topology and multimodal source tracing evidence. It can be implemented using a WebGL-based 3D visualization framework or a spreadsheet-based 2D interactive interface, aiming to provide an intuitive view of anomaly source tracing.

[0193] Specifically, this application's solution identifies high-risk supply chain entities based on comprehensive anomaly probability, focusing on actual risk points to avoid interference from irrelevant information. Simultaneously, it identifies dominant anomaly modalities based on contribution distribution, ensuring the targeted nature of extracted data segments. A time window matching algorithm dynamically adjusts the window range based on the timestamp of an anomaly occurrence, capturing raw data synchronized with the anomaly event. Entity identifier resolution technology utilizes unique entity identifiers to semantically align eye-tracking fragments, communication text fragments, and electromagnetic spectrum fragments with supply chain entities, establishing a cross-modal evidence chain. Finally, a visualization engine integrates the multimodal evidence chain with the supply chain business topology, generating an intuitive visualization interface that spatially maps the anomaly state of high-risk entities to their multimodal evidence, allowing users to interactively view traceability information from different perspectives.

[0194] As a specific implementation method, the solution of this application is implemented as follows: When the system detects that the overall anomaly probability of a certain electronic component supplier exceeds a threshold and the physiological modality contribution is the highest, the time window matching algorithm automatically extracts the eye movement trajectory fragments of the supplier's associated operators from the multimodal data stream 5 minutes before the anomaly occurred. Entity identification resolution technology maps the eye movement trajectory fragments to the supplier's entity identifier, and simultaneously extracts cross-departmental communication texts and equipment electromagnetic spectrum fragments from the same period. The visualization engine generates a business topology map containing supplier nodes, where the node is highlighted in red, and displays the corresponding eye movement heatmap, communication keyword cloud, and spectrum fluctuation curve through an expandable panel. Operators can click on the node to view detailed traceability evidence for each modality's data.

[0195] Through the above scheme, this application establishes an effective correlation mapping between anomaly detection results and original multimodal data fragments. The anomaly tracing process is transformed from an abstract description to an intuitive and visual evidence display. Operators can quickly locate the root cause of anomalies and understand the coupling relationship between multimodal data, thereby improving the timeliness and accuracy of supply chain collaborative decision-making.

[0196] In some of the embodiments described above in this application, a visual interface for generating integrated business topology and multimodal tracing evidence is proposed to integrate multidimensional evidence of supply chain anomalies. However, in its implementation, there is a lack of dynamic interaction mechanism between the main business view and the tracing evidence view. Users cannot trigger real-time updates of related views through the operation of a single view, which requires repeated manual switching of views for comparison during anomaly tracing. This not only increases the complexity of operation but also makes it difficult to intuitively track the causal relationship between multimodal data, reducing the efficiency of anomaly analysis and the speed of decision response.

[0197] In response, this application further proposes a visualization interface based on a multimodal evidence chain and supply chain business topology, which generates a visualization interface that integrates business topology and multimodal traceability evidence through a visualization engine, including: A main business view is generated based on the supply chain business topology, and multiple parallel traceability evidence views are generated based on the data types and relationships in the multimodal evidence chain.

[0198] By using entity identification mapping technology, a two-way association relationship is established between supply chain entities in the main business view and data fragments in each traceability evidence view, enabling interactive responses between views.

[0199] Deploy event listeners in the visual interface to respond to interactive operations on abnormal elements in any view and synchronously update the display content of all related views.

[0200] The main business view refers to a visual presentation interface built upon the supply chain business topology. It can be implemented using topology layout algorithms such as force-directed graphs or hierarchical tree diagrams, aiming to clearly display the business logic relationships between supply chain entities. The traceability evidence view can be understood as an independent display unit generated based on the data types and relationships in the multimodal evidence chain. It can be implemented using differentiated visualization forms such as time-series line charts, text semantic clouds, or spectral heatmaps, aiming to avoid heterogeneous data interference in the identification of key anomalies. Entity identification mapping technology refers to the technical mechanism for establishing data associations between views. It can be implemented using globally unique identifier hash mapping or graph database indexing technology, aiming to solve the problem of broken evidence chains caused by view fragmentation. Event listeners are software components used to capture user interaction behavior. They can be implemented using a front-end event listening framework or message queue subscription mechanism, aiming to trigger dynamic updates of associated views in real time.

[0201] Specifically, the solution in this application presents the business topology relationships between supply chain entities through a main business view. Simultaneously, it generates parallel traceability evidence views based on data types such as eye-tracking tracing, communication text, and electromagnetic spectrum in the multimodal evidence chain, forming an independent display layer for multi-dimensional evidence. Entity identification mapping technology constructs a bidirectional mapping channel between supply chain entities in the main business view and data fragments in each traceability evidence view. When a user selects a specific high-risk entity in the main business view, the system automatically locates the corresponding multimodal data fragment in all evidence views; conversely, when selecting a data fragment in an evidence view, it can also trace back to the related entities in the business topology. An event listener continuously monitors user interactions with abnormal elements in any view. Once a click or hover operation is detected, a synchronous update mechanism for related views is immediately triggered. For example, when an abnormal frequency band is selected in the electromagnetic spectrum view, the main business view automatically highlights the relevant equipment entity, and the communication text view synchronously displays the corresponding collaboration records for that time period, thereby achieving dynamic association of cross-modal evidence and forming a complete closed loop of the abnormal traceability evidence chain.

[0202] As a specific implementation manner, the solution of the present application is specifically implemented as follows: The supply chain business topology structure uses a force-directed graph to generate a main business view, where nodes represent entities such as suppliers, manufacturers, and logistics providers, and the connections represent business collaboration relationships. For the multi-modal evidence chain, an eye movement trajectory view (showing the attention distribution in the form of a fixation point heat map), a communication text view (showing key conversations in the form of a semantic network graph), and an electromagnetic spectrum view (showing abnormal offsets in the form of a dynamic spectrum graph) are respectively constructed. The entity identification mapping technology establishes a two-way association between the device nodes in the main business view and the frequency band data segments in the electromagnetic spectrum view through a globally unique identifier. The visualization interface deploys a JavaScript-based event listener. When the user clicks on the assembly station node in the main business view, the system automatically locates the spectrum anomaly segment corresponding to that time period in the electromagnetic spectrum view and highlights the collaboration record of the quality inspection department during that time period in the communication text view, realizing the联动更新 of the three views.

[0203] Through the above technical solution, the present application effectively solves the problem of the lack of dynamic interaction between the main business view and the traceability evidence view, reduces the operation steps for the user to manually switch views, enables the causal association between multi-modal data to be intuitively presented, and thus improves the efficiency of supply chain anomaly event analysis and the decision-making response speed.

[0204] All the above optional technical solutions can be combined arbitrarily to form optional embodiments of the present application, which will not be elaborated here one by one.

[0205] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including a computer program. The above computer program can be executed by a processor to complete the method for real-time detection and visualization of supply chain production and sales collaboration anomaly events in the above embodiments. For example, the computer-readable storage medium can be a Read-Only Memory (ROM), a Random Access Memory (RAM), a Compact Disc Read-Only Memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0206] In an exemplary embodiment, a computer program product or a computer program is also provided. The computer program product or the computer program includes program code, and the program code is stored in a computer-readable storage medium. The processor of the computer device reads the program code from the computer-readable storage medium, and the processor executes the program code, so that the computer device executes the method for real-time detection and visualization of supply chain production and sales collaboration anomaly events.

[0207] In some embodiments, the computer program involved in the present application embodiments may be deployed and executed on a computer device, or executed on multiple computer devices located in one location, or executed on multiple computer devices distributed in multiple locations and interconnected through a communication network. Multiple computer devices distributed in multiple locations and interconnected through a communication network may constitute a blockchain system.

[0208] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0209] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for real-time detection and visualization of abnormal events in supply chain production and sales coordination, characterized in that, The method includes: By deploying biosensors, enterprise communication system interfaces, and spectrum acquisition devices at key workstations on the production line, operator eye movement trajectories and heart rate variability sequences, cross-departmental business communication texts, and production line electromagnetic spectrum signals are acquired simultaneously to form a multimodal data stream. Physiological state analysis is performed on the eye movement trajectory and heart rate variability sequence to obtain the real-time attention distraction index and physiological stress level; sentiment semantic analysis is performed on the cross-departmental business communication text to obtain the collaborative health score for specific production work orders; spectral features are extracted from the electromagnetic spectrum signal of the production line to obtain the spectral anomaly offset and production line beat disturbance coefficient. Based on the real-time attention distraction index, the physiological stress level, the collaborative health score, the spectrum anomaly offset, and the production line beat disturbance coefficient, a multimodal anomaly coupling analysis is performed using a dynamic graph neural network to output the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modal indicator. Based on the comprehensive anomaly probability and the contribution distribution of each modal indicator, the corresponding eye-tracking trajectory fragments, communication text fragments, and electromagnetic spectrum fragments are extracted from the multimodal data stream through time window matching and entity identification association. A visualization interface that integrates business topology and multimodal source tracing evidence is generated, in which high-risk entities are synchronously associated with and displayed with the extracted data fragments.

2. The method according to claim 1, characterized in that, The process involves simultaneously acquiring operator eye-tracking trajectories and heart rate variability sequences, cross-departmental business communication text, and production line electromagnetic spectrum signals through biosensors, enterprise communication system interfaces, and spectrum acquisition devices deployed at key workstations on the production line, forming a multimodal data stream, including: Non-invasive eye trackers and heart rate sensors deployed at quality inspection and precision assembly stations are used to collect eye movement trajectory data and heart rate variability data of operators during the execution of specific production work orders. The system obtains cross-departmental email exchanges and instant messaging records associated with the target production work order through the enterprise communication system interface, and performs anonymization processing and topic clustering on the communication records. By deploying broadband radio frequency sensors around the production equipment, electromagnetic spectrum signals of the production line are collected during the standard production cycle, and a reference spectrum template for each production line equipment under normal operating conditions is established. The collected eye-tracking data, heart rate variability data, communication record text, and electromagnetic spectrum signals are timestamped and associated with work order identifiers to generate the multimodal data stream with spatiotemporal alignment markers.

3. The method according to claim 1, characterized in that, The physiological state analysis of the eye movement trajectory and heart rate variability sequence to obtain the real-time attention distraction index and physiological stress level includes: The eye movement trajectory data is subjected to sampling frequency normalization and noise filtering to extract the fixation point distribution entropy and pupil diameter variation coefficient per unit time. Frequency domain feature analysis was performed on the heart rate variability sequence to determine the ratio of low-frequency power to high-frequency power, and the standard deviation of heart rate interval between adjacent normal heart rate intervals was extracted. Based on the gaze point distribution entropy, the pupil diameter variation coefficient, the low-frequency to high-frequency power ratio, and the heart rate interval standard deviation, the data are processed by a pre-trained physiological state assessment model to output the real-time attention distraction index and the physiological stress level. The real-time attention distraction index comprehensively reflects the operator's visual attention concentration, while the physiological stress level quantitatively characterizes the operator's autonomic nervous system activation state.

4. The method according to claim 1, characterized in that, The process of performing sentiment and semantic analysis on the cross-departmental business communication text to obtain a collaborative health score for a specific production work order includes: The cross-departmental business communication text is segmented and named entity recognition is performed to extract key communication content associated with the production work order number; Determine the emotional polarity and emotional intensity values ​​of the key communication content, and statistically analyze the communication response time delay and communication interaction frequency of the key communication content; Based on preset weights, the emotional polarity value, the emotional intensity value, the communication response time delay, and the communication interaction frequency are weighted and calculated to obtain the emotional consistency index. A semantic association network is constructed based on the key communication content, the topological features of the semantic association network are determined, and the semantic conflict degree is determined based on the topological features. Based on the sentiment consistency index, the topological features of the semantic association network, and the semantic conflict degree, the collaborative health score is generated through normalization and scaling transformation.

5. The method according to claim 4, characterized in that, The process of determining the emotional polarity and emotional intensity values ​​of the key communication content, and statistically analyzing the communication response time delay and communication interaction frequency of the key communication content, includes: Based on the sentiment dictionary, sentiment words are matched and intensity is calculated for the key communication content to obtain the basic sentiment polarity value and basic sentiment intensity value of each sentence; Based on the communication timing relationship of the key communication content, the time interval between adjacent communication records is determined as the communication response time delay, and the number of communication interactions within a preset time window is counted as the communication interaction frequency. Extract the dialogue thread structure and semantic role labeling results from the key communication content to generate communication context features; Based on the basic emotional polarity value and the basic emotional intensity value, the emotional correction is performed by integrating the communication context features to obtain the emotional polarity value and the emotional intensity value.

6. The method according to claim 1, characterized in that, The step of extracting spectral features from the electromagnetic spectrum signal of the production line to obtain the spectral anomaly offset and the production line cycle time disturbance coefficient includes: Variational mode decomposition and frequency domain structure analysis were performed on the electromagnetic spectrum signal of the production line to obtain the spectrum anisotropy index and characteristic frequency band energy distribution; The autocorrelation coefficient sequence of the signal envelope is calculated based on the energy distribution of the characteristic frequency band, and the fluctuation variance and periodicity measure of the autocorrelation coefficient sequence are extracted. Based on the spectral anisotropy index, the fluctuation variance, and the periodicity metric, the spectral anomaly offset and the production line cycle disturbance coefficient are simultaneously calculated using a nonlinear mapping function. The spectral anomaly offset represents the degree of anomaly in the electromagnetic environment, and the production line cycle disturbance coefficient represents the stability of the production rhythm.

7. The method according to claim 1, characterized in that, The method, based on the real-time attention distraction index, physiological stress level, collaborative health score, spectral anomaly offset, and production line beat disturbance coefficient, performs multimodal anomaly coupling analysis through a dynamic graph neural network, outputting the comprehensive anomaly probability of key supply chain entities and the contribution distribution of each modality indicator, including: Based on the real-time attention distraction index and the physiological stress level, the abnormal state vector of the supply chain entity in the physiological dimension is generated through the first graph neural layer of the dynamic graph neural network. Based on the collaborative health score, the spectral anomaly offset, and the production line cycle time disturbance coefficient, a multimodal coupled edge weight matrix is ​​constructed, and the graph structure representation of the dynamic graph neural network is updated using the edge weight matrix. The abnormal state vector and the updated graph structure representation are input into the second graph neural layer of the dynamic graph neural network. Anomaly propagation is calculated through a cross-modal graph attention mechanism, and the comprehensive anomaly probability of each supply chain entity and the contribution distribution of each modality index are output synchronously.

8. The method according to claim 7, characterized in that, The process of generating an abnormal state vector of the supply chain entity in the physiological dimension through the first graph neural layer of the dynamic graph neural network, based on the real-time attention distraction index and the physiological stress level, includes: The real-time attention distraction index and the physiological stress level are subjected to time series standardization, and the statistical correlation coefficient and covariance matrix of the two are calculated. Based on the statistical correlation coefficient and the covariance matrix, an attention-stress coupling feature matrix is ​​constructed, and the principal component eigenvectors of the attention-stress coupling feature matrix are calculated. The principal component feature vector is concatenated with the basic attribute features of the supply chain entity and input into the graph convolutional network of the first graph neural layer. The abnormal state vector is generated by aggregating neighbor node information and transforming activation functions. The dimensions of the abnormal state vector are consistent with the number of supply chain entities, and the value of each dimension reflects the degree of abnormality of the corresponding entity in the physiological dimension.

9. The method according to claim 7, characterized in that, The construction of a multimodal coupling edge weight matrix based on the collaborative health score, the spectral anomaly offset, and the production line cycle time disturbance coefficient includes: Determine a first correlation factor between the collaborative health score and the spectrum anomaly offset, and simultaneously determine a second correlation factor between the production line cycle disturbance coefficient and the spectrum anomaly offset; Based on the first association factor and the second association factor, the relationship strength weights between supply chain entities are determined, and the business topology distances between entities are fused using the relationship strength weights to generate an initial edge weight matrix; The collaborative health score, the spectral anomaly offset, and the production line cycle time disturbance coefficient are fused to obtain the multimodal coupling coefficient; The initial edge weight matrix is ​​dynamically adjusted using the multimodal coupling coefficient to obtain the multimodal coupling edge weight matrix; The edge weight matrix contains each element that reflects the intensity of anomaly propagation between corresponding supply chain entities based on multimodal data.

10. The method according to claim 1, characterized in that, The step of inputting the abnormal state vector and the updated graph structure representation into the second graph neural layer of the dynamic graph neural network, performing anomaly propagation calculation through a cross-modal graph attention mechanism, and synchronously outputting the comprehensive anomaly probability of each supply chain entity and the contribution distribution of each modality indicator includes: Based on the abnormal state vector and the updated graph structure representation, cross-modal attention coefficients between nodes are calculated using a multi-head attention network. Using the cross-modal attention coefficients, anomaly state propagation is performed on the enhanced graph adjacency matrix, and the anomaly state representation of each node is updated through a gated graph neural network; Based on the updated node abnormal state representation, the comprehensive abnormal probability of each supply chain entity is determined by the Sigmoid activation function of the output layer. At the same time, the contribution distribution of each modal indicator to the abnormal state is analyzed by backpropagation analysis of the cross-modal attention coefficient.

Images