AI-driven intelligent work reporting method, system, equipment and medium

Abstract

The invention relates to an AI-driven intelligent work reporting method, system and device and a medium. The method comprises the steps of collecting original operation data of production equipment and performing anomaly detection to generate a standardized anomaly event stream; carrying out unified representation on the abnormal event and the production context thereof, and generating a space-time aligned feature sequence; calculating causal intensity between abnormal events based on a preset model and constructing a dynamic time sequence causal graph; back diffusion iteration is carried out by simulating abnormal influence to trace root cause nodes, and a root cause score sorting vector is generated; positioning a key root cause and tracing a propagation path of the key root cause, analyzing backlog data for root cause types such as network interruption, and performing compensation work reporting; and finally, optimizing causal intensity model parameters by utilizing artificial feedback. The method overcomes the defects that in the prior art, abnormal events are analyzed in an isolated mode, and causes and effects cannot be correlated, intelligent root cause diagnosis and work report data self-correction of a complex abnormal chain can be achieved, and therefore production data reliability and operation and maintenance efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent manufacturing and production management information technology, specifically to an AI-driven intelligent work reporting method, system, equipment, and medium. Background Technology

[0002] In the digital transformation of modern manufacturing, intelligent work reporting methods serve as a crucial link connecting physical production and digital management. Their core lies in automatically and accurately converting real-time operational data from production equipment into task completion quantities that can be recognized and processed by the Manufacturing Execution System (MES). This method aims to replace traditional work reporting methods that rely on manual recording or simple rule-based judgments. By integrating data acquisition, anomaly detection, and logical processing, it achieves real-time feedback on production progress, working hours, and material consumption, thereby providing a precise data foundation for production scheduling, quality traceability, and performance analysis. A typical intelligent work reporting system needs to process data streams of varying types and from different devices, such as weight, length, and counts. Based on preset business rules such as planned weight and standard equipment capacity, it determines the production status and automatically triggers work reporting records to improve production efficiency and management transparency.

[0003] However, existing intelligent work reporting technologies have significant limitations in dealing with multi-source heterogeneous anomalies in complex production environments. When production data exhibits various anomalies such as weight jumps, communication interruptions, and failure to reset accumulated values, current systems typically only record and alarm for isolated events based on fixed thresholds, generating discrete anomaly reports. This approach fails to understand the temporal correlation and potential causal logic between different anomalies. For example, it cannot identify the continuous causal chain of "data surge" caused by temporary network outages leading to data backlog and subsequent centralized reporting after connection restoration. Due to the lack of ability to mine and reason about the causal relationships between anomalies, the system cannot automatically diagnose the root causes of a series of anomalies such as specific equipment failures, network node instability, or operational process oversights, nor can it provide intelligent data repair or compensation work reporting decisions based on root cause diagnosis. Maintenance personnel must spend a significant amount of time manually checking massive amounts of isolated alarm information, resulting in low fault location efficiency. Furthermore, the distortion of work reporting data caused by anomalies cannot be corrected in a timely and automatic manner, affecting the integrity and reliability of production data. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide an AI-driven intelligent work reporting method, system, device and medium that can automatically analyze the causal relationship between multiple abnormal events, realize root cause intelligent diagnosis and drive accurate compensation work reporting.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides an AI-driven intelligent work reporting method, comprising the following steps:

[0007] S1: Collect and process the raw operating data reported in real time by the production equipment and the work order task data of the manufacturing system. Based on the preset reporting logic rules, identify abnormal events that deviate from the normal production logic and generate a standardized raw abnormal event stream.

[0008] S2: Perform unified characterization processing on the original abnormal event stream and the context data of the manufacturing system. By encoding event attributes and fusing production work orders and equipment status information, generate an abnormal event feature sequence with spatiotemporal alignment.

[0009] S3: Calculate the causal strength between abnormal events in the abnormal event feature sequence based on the preset causal strength calculation model and construct a weighted directed graph to generate a dynamic temporal causal graph;

[0010] S4: Tracing root cause nodes in the dynamic temporal causal graph, simulating the process of anomalous influence spreading backward along the causal edge, iteratively calculating the probability of each node as the starting point of propagation, and generating a root cause score sorting vector for all nodes.

[0011] S5: Make diagnostic decisions based on root cause score sorting vector and dynamic temporal cause-effect graph. Locate key root cause nodes according to the score and trace their propagation path. Perform timestamp parsing and compensation reporting on the production data backed up during network outages according to the root cause type. Generate and execute intelligent reporting instructions.

[0012] S6: Based on manually verified real root cause feedback data, perform closed-loop optimization of the causal strength calculation model parameters and update the learnable parameters of the causal strength calculation model.

[0013] In one embodiment, S1 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0014] S11: Real-time analysis and work order association processing of the raw operating data reported by the production equipment, extracting equipment identification, production weight, length and timestamp, and matching it with the task requirements of the current production work order to generate production data points with work order tags;

[0015] S12: Perform rule-based anomaly identification processing on production data points with work order tags. Based on the weight jump threshold, continuous interruption duration and planned roll-up weight in the preset work reporting logic rules, determine whether the data point constitutes an isolated abnormal event and generate an initial abnormal event marker.

[0016] S13: Aggregate and standardize the initial abnormal event markers, merge the same type of abnormal markers triggered by the same device within a short time window, and uniformly record the event type, key values, device identifier and timestamp to generate a standardized original abnormal event stream.

[0017] In one embodiment, step S2 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0018] S21: Perform multi-dimensional attribute encoding on a single event in the original abnormal event stream, convert the event type and device identifier into a one-hot encoded vector, and standardize the associated abnormal values ​​to generate a basic event feature vector.

[0019] S22: Perform production context embedding processing on the basic event feature vector, query the manufacturing system based on the event timestamp and equipment identifier to obtain the corresponding material specifications, upper and lower limits of coil weight and equipment operating mode, and transform them into fixed-dimensional context feature vectors through a learnable embedding layer.

[0020] S23: Concatenate the basic event feature vector and the context feature vector into a unified high-dimensional feature vector, and sort the feature vectors of all events according to the timestamp to generate an abnormal event feature sequence with spatiotemporal alignment.

[0021] In one embodiment, S3 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0022] S31: Perform graph structure initialization processing on the abnormal event feature sequence, instantiate each feature vector in the abnormal event feature sequence into a graph node, and create preliminary temporal links between nodes in time order to generate the initial time sequence graph skeleton;

[0023] S32: Perform causal strength quantification on the node pairs with time-series links in the initial time-series graph skeleton, input the feature vectors of the two nodes into the causal strength calculation model, and calculate the causal strength score that combines feature correlation, temporal proximity and device consistency.

[0024] S33: Perform directed edge filtering and weighting on the causal intensity score. Based on the preset causal intensity threshold, connect the nodes with causal intensity scores exceeding the threshold with directed edges, and use the causal intensity score as the edge weight to synthesize a dynamic temporal causal graph.

[0025] In one embodiment, step S4 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0026] S41: Extract edge weight information from the weighted causal edges in the dynamic temporal causal graph, summarize the weight values ​​of all outgoing causal edges of each node, and generate the initial anomaly influence value of each node.

[0027] S42: Based on the initial abnormal influence value of each node and the weighted causal edge, perform iterative back propagation to simulate the reverse diffusion of abnormal influence along the causal edge. In each iteration, update the probability score of each node to the sum of the node's own influence value and the propagation scores of all downstream nodes from the node, and generate the updated node probability score set.

[0028] S43: Based on the node probability score set of the current iteration, perform convergence judgment and sorting processing, calculate the score change between adjacent iterations and determine whether it is less than the preset convergence threshold, sort the final converged scores in descending order of value, and generate the root cause score sorting vector of all nodes.

[0029] In one embodiment, S5 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0030] S51: Perform key root cause node localization processing on the root cause score sorting vector, select the top-ranked nodes as candidate key root causes according to the score, and generate a set of candidate root cause nodes.

[0031] S52: Perform anomaly propagation path tracing processing on the candidate root cause node set in the dynamic temporal causal graph. Starting from each candidate node, traverse downstream nodes along its outgoing causal edge, extract the complete anomaly propagation directed path, and generate the root cause propagation path set.

[0032] S53: Perform data compensation reporting on the paths related to the root cause of network outage in the root cause propagation path set, parse the lost data period corresponding to the time abnormal node in the path, extract the backlog of production data during the lost data period from the cache, and report the data to the work order in a compensatory manner according to the timestamp, and generate and execute intelligent reporting instructions.

[0033] In one embodiment, step S6 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0034] S61: Perform alignment and matching processing on the manually confirmed true root cause feedback data, associate the true root cause events identified in the feedback with the dynamic time-series cause-effect graph and root cause score sorting vector generated in the historical analysis, and determine the corresponding verified root cause nodes in the graph.

[0035] S62: Using verified root cause nodes as supervision labels, the training sample construction process is performed on the historical abnormal event feature sequence and dynamic temporal cause-effect graph that generated verified root cause nodes to generate supervised learning samples for model tuning.

[0036] S63: Based on supervised learning samples, the parameters of the causal strength calculation model are optimized. By minimizing the ranking loss of the model for the verified root cause nodes, the internal weight parameters of the model are adjusted to generate the learnable parameters of the updated causal strength calculation model.

[0037] Secondly, the present invention provides an AI-driven intelligent work reporting system, which is configured with the following modules:

[0038] The abnormal event detection module is used to collect and detect abnormalities in the raw operating data reported in real time by the production equipment and the work order task data of the manufacturing system. Based on the preset reporting logic rules, it identifies abnormal events that deviate from the normal production logic and generates a standardized raw abnormal event stream.

[0039] The anomaly feature characterization module is used to perform unified characterization processing on the original anomaly event stream and the context data of the manufacturing system. By encoding event attributes and fusing production work orders and equipment status information, it generates anomaly event feature sequences with spatiotemporal alignment.

[0040] The dynamic causal graph construction module is used to calculate the causal strength between abnormal events in the feature sequence of abnormal events based on a preset causal strength calculation model and construct a weighted directed graph to generate a dynamic time-series causal graph.

[0041] The root cause node tracing module is used to trace the root cause nodes of a dynamic time-series causal graph, simulate the process of the reverse diffusion of abnormal influence along the causal edge, iteratively calculate the probability of each node as the starting point of propagation, and generate a root cause score sorting vector for all nodes.

[0042] The intelligent work reporting instruction generation module is used to make diagnostic decisions based on the root cause score sorting vector and dynamic temporal cause-effect graph. It locates key root cause nodes according to the score and traces their propagation path. Based on the root cause type, it performs timestamp parsing and compensation work reporting processing on the production data backed up during the network outage, and generates and executes intelligent work reporting instructions.

[0043] The causal model optimization module is used to perform closed-loop optimization of the causal strength calculation model based on manually confirmed real root cause feedback data, and to update the learnable parameters of the causal strength calculation model.

[0044] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned AI-driven intelligent reporting methods.

[0045] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned AI-driven intelligent work reporting methods.

[0046] In summary, the AI-driven intelligent work reporting method provided in this application lays a unified data foundation for the correlation analysis of multi-source heterogeneous production anomalies by establishing a standardized raw anomaly event stream; furthermore, it integrates multi-dimensional contextual information such as production work orders and equipment status to construct an anomaly feature sequence with spatiotemporal alignment, thereby overcoming the data silo problem; based on this, by calculating the causal strength between anomaly events and constructing a dynamic temporal causal graph, it can fundamentally reveal the potential temporal correlation and causal propagation path between isolated anomaly events, thus solving the deficiency of existing technologies that can only record surface phenomena and cannot uncover deep connections; furthermore, through By tracing root cause nodes and generating sorting vectors through an iterative algorithm that simulates the back-diffusion of the impact of anomalies, the system can automatically and accurately locate the source of complex anomaly chains such as equipment failures, network interruptions, or operational errors, replacing inefficient manual investigation. Ultimately, based on accurate root cause diagnosis results, the system can intelligently trigger targeted data compensation and correction reporting logic, such as parsing and compensating for backlogged data due to network interruptions, to achieve self-correction of production data and ensure the accuracy of reporting results. At the same time, it utilizes human feedback to perform closed-loop optimization of the causal model, enabling the system to have continuous learning and adaptive capabilities, and comprehensively improving the intelligence level and reliability of production management.

[0047] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0048] Figure 1 A flowchart illustrating an AI-driven intelligent work reporting method provided in an embodiment of this application;

[0049] Figure 2 This is a schematic diagram of the structure of an AI-driven intelligent work reporting device provided in another embodiment of this application. Detailed Implementation

[0050] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0052] In one embodiment, such as Figure 1 As shown, an AI-driven intelligent work reporting method is provided. This embodiment illustrates the method's application to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0053] S1: Collects and processes the raw operating data reported in real time by the production equipment and the work order task data of the manufacturing system. Based on the preset work reporting logic rules, it identifies abnormal events that deviate from the normal production logic and generates a standardized raw abnormal event stream.

[0054] Specifically, the system collects raw operating data from production equipment and work order data from the manufacturing system. Raw operating data includes core quantitative output indicators, equipment operating status parameters, and communication link status data. The system acquires this data through data acquisition devices adapted to the production characteristics of different equipment. The deployment of these devices is determined based on the equipment's output type and data acquisition requirements. The system supports multiple industrial communication protocols and uses a built-in protocol parsing mechanism to convert raw data transmitted via different protocols into a unified data format, ensuring data transmission compatibility. Work order data is synchronously acquired from the manufacturing system and includes core business data such as work order number, material code, product specifications, work order type, reel-related parameters, planned production volume, and production process parameters. The synchronization process is triggered in real-time with work order status updates, and the system also executes a periodic verification process to ensure consistency between synchronized data and manufacturing system data.

[0055] Furthermore, the system performs anomaly detection based on preset reporting logic rules. These rules include production data anomaly judgment rules and auxiliary status anomaly judgment rules. The production data anomaly judgment rules are set based on the time-series change patterns and numerical fluctuation characteristics of the data, while the auxiliary status anomaly judgment rules are related to the normal range of equipment operating status parameters and communication link status data. The system performs periodic analysis on the collected raw data, marking data that meets the anomaly judgment rules as anomalies, and extracting key information corresponding to the anomalies, including the unique event identifier, anomaly type, occurrence time, associated equipment identifier, associated work order number, data change characteristics, and event duration. The system organizes this key information in chronological order to generate a standardized raw anomaly event stream.

[0056] S2: Perform unified characterization processing on the original abnormal event stream and the context data of the manufacturing system. By encoding event attributes and integrating production work orders and equipment status information, an abnormal event feature sequence with spatiotemporal alignment is generated.

[0057] Specifically, the system integrates the original abnormal event stream with the manufacturing system's context data, which includes equipment context, production context, and environmental context. Equipment context covers equipment model, historical operation records, maintenance records, and data acquisition device calibration status; production context includes the process requirements corresponding to the work order, current production progress, and material characteristics; and environmental context includes key parameters of the production environment and network topology information. The system establishes an association index between the original abnormal event stream and the context data using work order numbers and equipment identifiers, ensuring accurate association of various data types. The system encodes abnormal event attributes: discrete attributes are mapped to a unified code using a standardized vocabulary; textual attributes are converted into low-dimensional vectors through embedding encoding; and numerical attributes undergo standardization to eliminate dimensional differences, enabling different types of abnormal event attributes to be represented in the same feature space.

[0058] Preferably, the system employs a feature fusion mechanism to fuse the encoded abnormal event attributes with contextual data. The fusion process assigns weights based on the correlation between different types of data and the abnormal events. This weight allocation is determined based on the actual correlation logic of industrial production, ensuring that the fused features comprehensively reflect the essential characteristics and associated background of the abnormal events. The system constructs an abnormal event feature sequence, aligns the fused data temporally at a uniform time granularity, and fills time windows without abnormal events with uniform labeling information to ensure the temporal continuity of the sequence. Simultaneously, spatial alignment is performed based on equipment identifiers and work order numbers, ensuring that each feature sequence corresponds to a unique combination of equipment and work order. The feature sequence contains time window data of a fixed length, and each time window corresponds to a set of fused feature vectors.

[0059] S3: Based on the preset causal strength calculation model, calculate the causal strength between abnormal events in the abnormal event feature sequence and construct a weighted directed graph to generate a dynamic temporal causal graph.

[0060] Specifically, the system employs a hybrid causal strength calculation model to calculate the causal strength between anomalous events in an anomalous event feature sequence. This model combines statistical methods with prior knowledge of industrial production. Statistical methods quantify the information transmission relationships between anomalous events and capture potential correlations. Prior knowledge of industrial production is incorporated into the model in the form of rules to correct the results obtained by the statistical methods, ensuring that the causal strength calculation conforms to the actual logic of industrial production. The system trains the causal strength calculation model based on historical production data. The training data contains a large number of complete anomalous event chains and corresponding root cause information. During model training, the accuracy of the causal strength calculation is optimized by adjusting internal parameters, ensuring that the model can effectively identify causal relationships between different anomalous events. The system performs pairwise analysis on each anomalous event in the anomalous event feature sequence, calculating the causal strength between any two anomalous events to clarify the direction and magnitude of the causal relationship between events.

[0061] The system constructs a weighted directed graph, treating each anomalous event as a node. Node attributes contain complete information about the anomalous event. When the causal strength of two anomalous events reaches a set association threshold, the system establishes a directed edge between the corresponding nodes. The direction of the directed edge aligns with the causal association direction, and the edge weight corresponds to the causal strength. The system establishes a dynamic update mechanism: as new anomalous events occur, the causal strength between the new and historical anomalous events is calculated in real time. Based on the calculation results, the nodes and edges of the weighted directed graph are updated, generating a dynamic temporal causal graph. The dynamic temporal causal graph is stored in the form of an adjacency matrix and a node attribute table. The adjacency matrix records the causal strength between nodes, and the node attribute table records the details of the anomalous events corresponding to each node.

[0062] S4: Tracing root cause nodes in the dynamic temporal causal graph, simulating the process of anomalous influence spreading backward along the causal edge, iteratively calculating the probability of each node as the starting point of propagation, and generating a root cause score sorting vector for all nodes.

[0063] Specifically, the system designs a backdiffusion algorithm based on a dynamic temporal causal graph to trace the root cause nodes of anomalous events. The backdiffusion algorithm treats the directed edges in the dynamic temporal causal graph as influence propagation channels. The influence of an anomalous event propagates backward along the causal edges. The efficiency of influence transmission during propagation is related to the causal strength of the edge; the higher the causal strength, the higher the efficiency of influence transmission. The system assigns an initial influence value to each node, determined based on the scope and degree of interference of the anomalous event on the production process, reflecting the severity of the anomalous event. The system sets a diffusion decay parameter and an upper limit on the number of iterations. The upper limit on the number of iterations ensures algorithm convergence, with the convergence condition being that the change in root cause score obtained from two consecutive iterations meets the set requirements. The system iteratively calculates the root cause score of each node. During each iteration, each node receives the diffusion influence transmitted from subsequent nodes through the reverse edge, adds this diffusion influence to its own initial influence value, and obtains the root cause score for the current iteration. The iteration process continues until the convergence condition is met or the upper limit on the number of iterations is reached. After completing iterative calculations, the system sorts all nodes from highest to lowest root cause score, extracts key information for each node, including node identifier, root cause score, anomaly type, and occurrence time, and organizes this information according to the sorting results to generate a root cause score sorting vector. This root cause score sorting vector allows for filtering candidate root cause nodes based on set conditions, focusing on key analysis objects.

[0064] S5: Based on the root cause score sorting vector and dynamic temporal cause-effect graph, diagnostic decisions are made. Key root cause nodes are located according to their scores and their propagation paths are traced. Based on the root cause type, the backlog of production data during network outages is timestamped and compensated for, and intelligent work reporting instructions are generated and executed.

[0065] Specifically, the system analyzes the differences in root cause scores among nodes in the root cause score ranking vector. Combining the anomaly type corresponding to each node with the associated context data, it initially identifies candidate nodes for critical root causes. When the difference between the highest root cause score and other candidate nodes meets a set condition, the system directly identifies the node corresponding to the highest root cause score as the critical root cause node. When the score difference does not meet the set condition, the system calls relevant data such as historical fault records, operation logs, and production environment parameters for further verification to clarify the critical root cause node. Starting from the critical root cause node, the system traverses forward along the directed edges of the dynamic temporal cause-effect graph, recording all abnormal events affected by that root cause node, and sorting out the correlation and temporal sequence of these events to form a complete anomaly propagation path.

[0066] To address the backlog of production data caused by network outages, the system performs timestamp parsing on the backlogged data. Based on the start and recovery times of the communication interruption, combined with the output patterns during normal equipment production, the system breaks down the backlogged data into corresponding production time periods and completes accurate timestamps for each backlogged data entry. The system determines a compensation reporting scheme based on the root cause type and work order type. For sample line work orders, the actual output corresponding to the backlogged data is directly included in the total reporting quantity. For non-sample line work orders, the reporting method is determined based on the relationship between the weight difference corresponding to the backlogged data and the relevant parameters of the production line spool. If the reporting conditions are met, the work order is reported directly; otherwise, it is merged with subsequent normal production data before reporting. The system generates structured intelligent reporting instructions, including core information such as work order number, equipment identifier, reporting type, reporting quantity, timestamp range, root cause diagnosis results, and compensation strategy description. These instructions are synchronized to relevant systems such as the manufacturing system, warehouse management system, and customer relationship management system through standardized interfaces. During synchronization, the system receives execution feedback from each system and initiates a retry mechanism for cases of failed execution feedback to ensure the effective execution of the intelligent reporting instructions.

[0067] S6: Based on manually verified real root cause feedback data, perform closed-loop optimization of the causal strength calculation model parameters and update the learnable parameters of the causal strength calculation model.

[0068] Specifically, the system collects feedback data on genuine root causes confirmed by manual verification. This feedback data includes the true identifiers of key root causes, deviations in root cause determination, and relevant optimization suggestions. The feedback data is entered into the system through the production management platform. The system assigns a unique identifier to each piece of feedback data, linking it to the corresponding abnormal event chain and root cause diagnosis results to ensure the correspondence between the feedback data and the original analysis data. The system uses the feedback data as supervisory data, pairing it with the corresponding original abnormal event feature sequences to construct a model optimization dataset. This dataset is divided into subsets for model training and performance evaluation according to a set ratio. The system uses a supervised learning iterative optimization approach to update the learnable parameters of the causal strength calculation model. The optimization process aims to minimize the deviation between the model's predicted root cause score and the manually confirmed genuine root cause. A specified loss function is used to calculate the deviation value, and the model's learnable parameters, including statistical calculation parameters and rule coefficients incorporated from prior industrial production knowledge, are adjusted using a gradient descent algorithm.

[0069] Furthermore, the system establishes a model optimization trigger mechanism, including real-time optimization and batch optimization modes. Real-time optimization triggers incremental training after receiving a set amount of valid feedback data, while batch optimization triggers full training at fixed intervals. The system records the model version after each optimization, storing model parameters, training data information, and performance evaluation metrics, including key indicators such as accuracy and recall. The system monitors the performance of the new model version in real time. When a performance degradation reaches a set threshold, the system automatically initiates a version rollback mechanism, restoring the model to its historical best version. This ensures the continuous stability of the causal strength calculation model's performance, providing reliable support for subsequent causal analysis of abnormal events.

[0070] In summary, the AI-driven intelligent work reporting method provided in this application lays a unified data foundation for the correlation analysis of multi-source heterogeneous production anomalies by establishing a standardized raw anomaly event stream; furthermore, it integrates multi-dimensional contextual information such as production work orders and equipment status to construct an anomaly feature sequence with spatiotemporal alignment, thereby overcoming the data silo problem; based on this, by calculating the causal strength between anomaly events and constructing a dynamic temporal causal graph, it can fundamentally reveal the potential temporal correlation and causal propagation path between isolated anomaly events, thus solving the deficiency of existing technologies that can only record surface phenomena and cannot uncover deep connections; furthermore, through By tracing root cause nodes and generating sorting vectors through an iterative algorithm that simulates the back-diffusion of the impact of anomalies, the system can automatically and accurately locate the source of complex anomaly chains such as equipment failures, network interruptions, or operational errors, replacing inefficient manual investigation. Ultimately, based on accurate root cause diagnosis results, the system can intelligently trigger targeted data compensation and correction reporting logic, such as parsing and compensating for backlogged data due to network interruptions, to achieve self-correction of production data and ensure the accuracy of reporting results. At the same time, it utilizes human feedback to perform closed-loop optimization of the causal model, enabling the system to have continuous learning and adaptive capabilities, and comprehensively improving the intelligence level and reliability of production management.

[0071] In one embodiment, S1 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0072] S11: Real-time analysis and work order association processing of raw operating data reported by production equipment, extracting equipment identifier, production weight, length and timestamp, and matching it with the task requirements of the current production work order to generate production data points with work order tags.

[0073] Specifically, the system receives raw operating data reported in real time from production equipment and processes the data instantly through a built-in real-time parsing module. This parsing module is compatible with various communication protocols commonly used in industrial production scenarios and can automatically identify the raw data formats output by different devices, converting scattered data streams into a unified, parsable structure to ensure that data from all types of equipment can be effectively processed by the system. During the parsing process, the system accurately extracts three types of core information: equipment identifiers to clearly identify the source equipment and corresponding machine head; production weight and length as core quantitative production indicators reflecting equipment output; and timestamps to record the precise moment the data was generated to ensure temporal integrity.

[0074] Simultaneously, the system retrieves the task requirements of the current production work orders from the Manufacturing Execution System in real time, covering key information such as work order number, material-related parameters, and production specifications. Through the mapping relationship between equipment identifiers and production tasks, the extracted production data is matched with the corresponding current production work orders. The matching process, based on the equipment's production scheduling and the work order's production sequence, confirms the specific production task corresponding to each piece of production data. After matching, the system adds a unique work order tag to each piece of production data. The tag contains a unique work order number and core work order attribute information, ultimately generating production data points with work order tags. Each data point integrates equipment identifier, production weight, length, timestamp, and work order tag, achieving precise binding between production data and work order tasks.

[0075] S12: Perform rule-based anomaly identification processing on production data points with work order tags. Based on the weight jump threshold, continuous interruption duration and planned roll-up weight in the preset work reporting logic rules, determine whether the data point constitutes an isolated abnormal event and generate an initial abnormal event marker.

[0076] Specifically, the system invokes preset reporting logic rules to perform rule-based anomaly identification and processing on production data points tagged with work orders. These reporting logic rules include three core judgment criteria, corresponding to weight jumps, continuity interruptions, and planned roll-up weight-related anomalies. The system processes each tagged data point sequentially according to the time sequence, comparing the production weight of the current data point with the production data points of the same equipment at adjacent time nodes. Based on the weight jump judgment criteria, it determines whether the weight change conforms to normal production patterns. Simultaneously, it continuously monitors the data reception status, judging whether there are data transmission interruptions or missing data based on the continuity interruption judgment criteria. Furthermore, the system extracts the planned roll-up weight-related requirements corresponding to the current work order and compares the actual production weight in the data point with the judgment criteria corresponding to the planned roll-up weight. Through this multi-dimensional comprehensive judgment, the system determines whether each data point constitutes an isolated anomaly event. An isolated anomaly event specifically refers to an event that has no direct connection to other anomalies and is caused by a single data point deviating from the normal range. For data points identified as isolated anomalies, the system generates an initial anomaly event marker. The marker includes the work order label corresponding to the data point, the equipment identifier, the type of judgment criteria on which the anomaly is based, key numerical information, and the timestamp of the anomaly occurrence, ensuring that the initial marker can completely retain the core characteristics of the anomaly event.

[0077] S13: Aggregate and standardize the initial abnormal event markers, merge the same type of abnormal markers triggered by the same device within a short time window, and uniformly record the event type, key values, device identifier and timestamp to generate a standardized original abnormal event stream.

[0078] Specifically, the system aggregates and standardizes all initial abnormal event tags. First, a unified time window is set to filter abnormal tags generated within a short period. The time window is set based on the operating characteristics of industrial production equipment and the common temporal patterns of abnormal occurrences, ensuring effective identification of similar abnormalities occurring in clusters within the same equipment within a short timeframe. The system categorizes and aggregates the initial abnormal event tags by equipment identifier. Within the tag set corresponding to each equipment, it further identifies similar abnormal tags triggered within the same time window. Similar abnormal tags refer to tags generated based on the same judgment criteria in the reporting logic rules.

[0079] Furthermore, the system merges these initial anomaly event tags that meet the aggregation conditions, eliminating duplicate records and integrating key information about the anomalies within that time period, clarifying the start and end times of the anomalies and the range of change of key values. After merging, the system standardizes the recording of anomalies, unifying and standardizing record fields, including event type, key values ​​and their range of change, device identifier, anomaly start timestamp, anomaly end timestamp, and corresponding work order label, ensuring that each anomaly event record has a consistent format and complete information. Finally, the system organizes and sorts the standardized anomalies according to the chronological order of their occurrence timestamps, forming a standardized raw anomaly event stream, which is managed through a stable storage and transmission mechanism.

[0080] In one embodiment, step S2 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0081] S21: Perform multi-dimensional attribute encoding on a single event in the original abnormal event stream, convert the event type and device identifier into a one-hot encoded vector, and standardize the associated abnormal values ​​to generate a basic event feature vector.

[0082] Specifically, the system performs multi-dimensional attribute encoding processing on each single event in the original abnormal event stream. The event attributes include two categories: discrete attributes and continuous attributes. The system encodes and transforms discrete attributes. The event type covers all types of abnormal situations defined in the original abnormal event stream. The device identifier clearly identifies the specific device and corresponding machine head from which the event originates. The system constructs standardized vocabularies for these two types of discrete attributes respectively, mapping each specific event type and device identifier to a unique one-hot encoded vector. The one-hot encoded vector uniquely represents the corresponding attribute through the activation state of binary bits, ensuring that different discrete attributes are independent and distinguishable from each other in the feature space.

[0083] Furthermore, the system standardizes the abnormal values ​​associated with the events. Since abnormal values ​​are continuous attributes, and different types of abnormalities have different numerical dimensions, standardization eliminates the influence of dimensions through a unified transformation logic, enabling subsequent analysis of the numerical features of different types of abnormalities at the same scale. The system integrates the encoded one-hot encoded vectors with the standardized abnormal numerical features in a preset order. This integration process ensures that the positions of various attribute features are fixed, forming a structurally unified basic event feature vector. Each basic event feature vector fully retains the core attribute information of a single abnormal event.

[0084] S22: Perform production context embedding processing on the basic event feature vector, query the manufacturing system based on the event timestamp and equipment identifier to obtain the corresponding material specifications, upper and lower limits of coil weight and equipment operating mode, and transform them into fixed-dimensional context feature vectors through a learnable embedding layer.

[0085] Specifically, the system performs production context embedding processing on the basic event feature vector, with the event timestamp and equipment identifier as the core query criteria. The system establishes real-time data interaction with the manufacturing system through a standardized interface. Based on the event timestamp, it determines the production period when the abnormal event occurred. Combined with the equipment identifier, it locates the corresponding production equipment and accurately obtains three types of key production context information from the manufacturing system: material specifications associated with the specific parameters of the production material corresponding to the current work order; upper and lower limits of the reel weight providing quantitative benchmarks for anomaly identification and work reporting; and equipment operating mode reflecting the equipment's production operation status during that period.

[0086] The system structures and organizes the three types of contextual information it acquires, removing redundant data and retaining core, effective information. The organized contextual information is then input into a learnable embedding layer. Through training, the embedding layer learns the correlation between contextual information and anomalous event features, transforming the non-vector-based contextual information into fixed-dimensional contextual feature vectors. The vector dimension is set and maintained uniformly according to feature representation requirements, ensuring that the contextual information of each anomalous event is presented in a standardized vector form, thus achieving effective connection between production context and basic event features.

[0087] S23: Concatenate the basic event feature vector and the context feature vector into a unified high-dimensional feature vector, and sort the feature vectors of all events according to the timestamp to generate an abnormal event feature sequence with spatiotemporal alignment.

[0088] Specifically, the system performs a concatenation operation on the basic event feature vector and the context feature vector. The concatenation process follows a fixed order of feature dimensions, integrating the feature dimensions of the two types of vectors to form a unified high-dimensional feature vector. The high-dimensional feature vector simultaneously contains the core attribute features of the abnormal event itself and the corresponding production context features, realizing the comprehensive fusion of multi-dimensional information of a single event and ensuring that the feature vector can fully represent the essential characteristics of the abnormal event and its associated production background.

[0089] After concatenation, the system extracts the event timestamp corresponding to each high-dimensional feature vector and sorts all event high-dimensional feature vectors according to the chronological order of the timestamps. The sorting process strictly follows temporal logic to ensure that the abnormal events corresponding to adjacent feature vectors are continuous in the temporal dimension. Through temporal sorting, the system constructs an abnormal event feature sequence with spatiotemporal alignment. The high-dimensional feature vector at each position in the sequence retains its own complete features and reflects the temporal correlation between events through temporal arrangement. In the spatial dimension, the aggregation of events in the same production scenario is ensured through equipment identification and work order association. This feature sequence provides structured and temporally ordered input data for subsequent causal strength calculation and root cause tracing, ensuring that subsequent analysis can be carried out based on complete event features and spatiotemporal correlation.

[0090] In one embodiment, S3 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0091] S31: Perform graph structure initialization processing on the abnormal event feature sequence, instantiate each feature vector in the abnormal event feature sequence into a graph node, and create preliminary temporal links between nodes in time order to generate the initial time sequence graph skeleton.

[0092] Specifically, the system initializes the graph structure of the abnormal event feature sequence, traversing each high-dimensional feature vector in the sequence and instantiating it into an independent graph node of a dynamic temporal causal graph. Each graph node carries all the information of the corresponding feature vector and is associated with core information such as the event's unique identifier, device identifier, timestamp, and anomaly type, achieving a precise mapping between graph nodes and the original abnormal events. Preferably, the system re-verifies and sorts all instantiated nodes based on time to ensure that the node time sequence is consistent with the actual occurrence time sequence of the abnormal events. Based on the sorting result, the system completes the structured representation of node features through a feature mapping formula, which is:

[0093]

[0094] in, This represents the structured feature vector of the i-th graph node. This represents the i-th high-dimensional feature vector in the feature sequence of anomalies. This represents the feature mapping function, used to convert the original high-dimensional feature vectors into a feature representation form that adapts to the graph structure. Subsequently, the system creates preliminary temporal links between nodes based on temporal relationships, with the link direction pointing from the node corresponding to the earlier event to the node corresponding to the later event, generating an initial temporal graph skeleton. The initial temporal graph skeleton is managed by partitioning according to device identifiers, with nodes and links of the same device belonging to the same partition, and different partitions being independent of each other.

[0095] S32: Perform causal strength quantification on the node pairs with time-linked connections in the initial time series graph skeleton, input the feature vectors of the two nodes into the causal strength calculation model, and calculate the causal strength score that combines feature correlation, temporal proximity and device consistency.

[0096] Specifically, the system traverses all preliminary time-series links in the initial time-series graph skeleton, locates the node pairs corresponding to each link, and extracts the high-dimensional feature vectors of the two nodes. These vectors are then input into the causal strength calculation model for quantification. The model calculates the causal strength score by integrating three dimensions: feature relevance, temporal proximity, and device consistency. Feature relevance is obtained through vector similarity calculation, temporal proximity is determined based on event time intervals, and device consistency is verified through device identifier matching. To achieve effective fusion of multi-dimensional information, the model uses a weighted summation formula to calculate the final causal strength score. The formula is:

[0097]

[0098] in, Let represent the causal strength score between the i-th node and the j-th node. , , These are the weighting coefficients for feature correlation, temporal proximity, and device consistency, respectively, and they satisfy... , This represents the similarity between the feature vectors of two nodes. This represents the quantized value indicating the temporal proximity of events between two nodes. This represents the device consistency flag, which is 1 when two nodes correspond to the same device and 0 otherwise.

[0099] S33: Perform directed edge filtering and weighting on the causal intensity score. Based on the preset causal intensity threshold, connect the nodes with causal intensity scores exceeding the threshold with directed edges, and use the causal intensity score as the edge weight to synthesize a dynamic temporal causal graph.

[0100] Specifically, the system invokes a preset causal strength threshold, which is determined based on the causal relationship verification results of historical anomalous events. The system iterates through all node pairs with calculated causal strength scores and uses a filtering formula to determine whether a valid causal relationship exists between the node pairs. The filtering formula is as follows:

[0101]

[0102] in, For indicator functions, The causal strength score for the node pair. The system uses a preset causal strength threshold. When the indicator function is 1, the node pair is considered to have a valid causal relationship, and the system creates a directed edge connecting the node pair, with the direction of the directed edge consistent with the initial temporal link. When the indicator function is 0, the initial temporal link corresponding to the node pair is considered to have no valid causal relationship, is retained, but no directed edge is created. The system directly assigns the causal strength score to the corresponding directed edge as the edge weight, completing the directed edge filtering and weighting process. Based on the above processing, the system synthesizes a dynamic temporal causal graph on the initial temporal graph skeleton. This graph is stored in partitions according to device identifiers, supports multi-dimensional fast retrieval and traversal, and has dynamic update capabilities. It can supplement nodes and edges as new abnormal event feature sequences are input, providing structured causal relationship graph data support for subsequent root cause node tracing.

[0103] In one embodiment, step S4 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0104] S41: Extract edge weight information from the weighted causal edges in the dynamic temporal causal graph, summarize the weight values ​​of all outgoing causal edges of each node, and generate the initial anomaly influence value of each node.

[0105] Specifically, the system extracts weight information from the weighted causal edges in the dynamic temporal causal graph, traversing all nodes in the graph and locating all outgoing causal edges corresponding to each node. An outgoing causal edge is a directed edge from the current node to other nodes, and its weight is the previously calculated causal strength score. The system summarizes the weight values ​​of all outgoing causal edges for each node and calculates the initial anomaly influence value for each node using a summation formula:

[0106]

[0107] in, This represents the initial anomaly influence value of the i-th node. Let represent the set of outgoing neighbor nodes of the i-th node, that is, the set of nodes connected to the i-th node through a directed edge and located at the end of the edge. This represents the weight of the weighted causal edge pointing from the i-th node to the j-th node. The system associates and stores the initial anomaly influence value of each node with its node identifier, generating a set of initial anomaly influence values.

[0108] S42: Based on the initial abnormal influence value of each node and the weighted causal edge, perform iterative backpropagation to simulate the reverse diffusion of abnormal influence along the causal edge. In each iteration, update the probability score of each node to the sum of the node's own influence value and the propagation scores of all downstream nodes from the node, and generate the updated node probability score set.

[0109] Specifically, the system initiates an iterative backpropagation process of anomaly influence based on the initial anomaly influence values ​​of each node and the weighted causal edges of the dynamic temporal causal graph. The backpropagation process simulates the diffusion of anomaly influence along causal edges from downstream nodes to upstream nodes. In each iteration, the system updates the probability score of each node using the following formula:

[0110]

[0111] in, Let represent the probability score of the i-th node in the k-th iteration. This is the damping coefficient, used to balance the initial influence of a node itself with the propagation influence of downstream nodes. Let be the initial anomaly influence value of the i-th node. Let represent the set of incident neighbor nodes of the i-th node, that is, the set of nodes connected to the i-th node through a directed edge and located at the starting end of the edge. This represents the weight of the weighted causal edge pointing from the j-th node to the i-th node. This represents the sum of the weights of all outgoing causal edges from the j-th node. This represents the probability score of the j-th node in the (k-1)-th iteration. The system updates the probability scores of all nodes sequentially according to a preset iteration order, generating an updated set of node probability scores after one iteration.

[0112] S43: Based on the node probability score set of the current iteration, perform convergence judgment and sorting processing, calculate the score change between adjacent iterations and determine whether it is less than the preset convergence threshold, sort the final converged scores in descending order of value, and generate the root cause score sorting vector of all nodes.

[0113] Specifically, the system performs convergence testing based on the set of node probability scores in the current iteration and the set of node probability scores in the previous iteration. Convergence testing is achieved by calculating the overall change in the probability scores of all nodes between two adjacent iterations; the formula is as follows:

[0114]

[0115] in, This represents the maximum change in node probability scores between the k-th iteration and the (k-1)-th iteration. and These are the probability scores of the i-th node in the k-th and (k-1)-th iterations, respectively. The system will calculate the... Compare with the preset convergence threshold; if If the probability score is less than the preset convergence threshold, the iteration process is considered converged, the iteration stops, and the final set of node probability scores is output. If the convergence condition is not met, the next iteration continues until the convergence condition is met or the preset maximum number of iterations is reached. After convergence, the system sorts the final node probability scores in descending order of value, associates them with the corresponding node identifiers, abnormal event types, and other information, and generates a root cause score ranking vector for all nodes. This vector intuitively reflects the probability of each node being an abnormal root cause.

[0116] In one embodiment, S5 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0117] S51: Perform key root cause node localization processing on the root cause score sorting vector, select the top-ranked nodes as candidate key root causes based on their scores, and generate a set of candidate root cause nodes.

[0118] Specifically, the system receives the root cause score ranking vector of all nodes. This vector contains node identifiers, abnormal event types, and other related information, sorted in descending order of probability scores. The system performs key root cause node localization processing on the root cause score ranking vector. First, it extracts the root cause score and corresponding ranking information of each node in the ranking vector, and then selects candidate key root cause nodes according to a preset ranking filtering rule. To quantify the filtering criteria, the system uses a ranking indicator function to achieve precise filtering of candidate nodes. The formula is:

[0119]

[0120] in, For ranking indicator functions, This represents the rank of the i-th node in the root cause score sorting vector. A preset ranking threshold for candidate nodes. When the indicator function is 1, the node is included in the candidate key root cause node category; when the value is 0, it is excluded. The system integrates all nodes that meet the conditions to generate a set of candidate root cause nodes. Each element in the set is associated with complete node attribute information, including root cause score, abnormal event identifier, device identifier, and occurrence timestamp.

[0121] S52: Perform anomaly propagation path tracing processing on the candidate root cause node set in the dynamic temporal causal graph. Starting from each candidate node, traverse downstream nodes along its outgoing causal edge, extract the complete anomaly propagation directed path, and generate the root cause propagation path set.

[0122] Specifically, the system calls dynamic temporal cause-effect graph data and sequentially performs anomaly propagation path tracing processing on each node in the candidate root cause node set. The tracing process starts with a single candidate root cause node and traverses downstream nodes along the weighted causal edges emanating from that node. The traversal rule is to only follow the direction of the directed edges, sequentially visiting the emanating neighbor nodes of each downstream node until a terminal node without emanating causal edges is reached or a path loop is detected. To ensure path integrity, the system records the core information of each tracing path using a path representation formula, which is:

[0123]

[0124] in, This represents the t-th abnormal propagation path. The candidate root cause node at the start of the path. This represents a weighted causal edge from the k-th node to the (k+1)-th node. Let m be the terminal node of the path, and m be the number of nodes on the path minus one. The system performs deduplication on each tracing path, eliminating duplicate path structures, and integrates all complete anomaly propagation paths into a root cause propagation path set. Each path in the set is associated with a corresponding candidate root cause node identifier, edge weight sequence, and node anomaly type sequence.

[0125] S53: Perform data compensation reporting on the paths related to the root cause of network outage in the root cause propagation path set, parse the lost data period corresponding to the time abnormal node in the path, extract the backlog of production data during the lost data period from the cache, and report the data to the work order in a compensatory manner according to the timestamp, and generate and execute intelligent reporting instructions.

[0126] Specifically, the system performs classification and filtering on the root cause propagation path set, extracting propagation paths where the root cause node is a network outage type. These paths contain time-anomaly nodes, which are associated with information related to data continuity interruptions. The system parses the attribute information of the time-anomaly nodes to determine the start and end times of the data loss period, and calculates the duration of the lost data using the time difference. The system accesses the local cache module of the edge computing gateway to extract the backlogged production data during the data loss period. This data includes equipment operating parameters, production quantification indicators, and local timestamps. Preferably, the system determines the amount of compensation data allocated to each time segment using a compensation data amount calculation formula, which is:

[0127]

[0128] in, This represents the amount of compensation data for the s-th time segment. This represents the total amount of backlogged production data during the period when the data was lost. This represents the duration of the s-th time segment, where S represents the total number of time segments divided into periods of lost data. The system matches the amount of compensation data for each time segment with the corresponding unified timestamp and the current production work order information, and encapsulates it into a smart work reporting instruction according to the MES system interface specification. The instruction includes details of the compensation data, work order identifier, equipment identifier, root cause type, and propagation path information. The system uploads the work reporting instruction to the MES system through a RESTful API interface. After legality verification, the compensation work reporting is executed, synchronized to the relevant manufacturing systems, and the execution results are recorded to ensure that the compensation data is consistent with the actual production progress and work order requirements.

[0129] In one embodiment, step S6 of the AI-driven intelligent work reporting method provided by the present invention specifically includes the following steps:

[0130] S61: Perform alignment and matching processing on the manually confirmed true root cause feedback data, associate the true root cause events identified in the feedback with the dynamic time-series cause-effect graph and root cause score sorting vector generated in the historical analysis, and determine the corresponding verified root cause nodes in the graph.

[0131] Specifically, the system receives manually confirmed root cause feedback data, which includes core fields such as the identifier of the root cause event, its occurrence time, associated device information, and anomaly type. The system performs alignment and matching processing on the feedback data. First, it extracts key matching fields from the feedback data, including the occurrence time of the root cause event, the associated device identifier, and the anomaly type. These fields are used as matching indexes, and the system correlates them with the dynamic time-series cause-effect graph and root cause score sorting vector generated during historical analysis. To quantify matching accuracy, the system uses a matching degree calculation function to achieve precise correlation. The formula is:

[0132]

[0133] in, This represents the matching degree between the i-th historical node and the f-th piece of human feedback data. , , The weighting coefficients for device identifier matching, time range matching, and anomaly type matching, respectively, and satisfying the following conditions: , and These are device identifiers for historical nodes and feedback data, respectively. and These represent the event occurrence time corresponding to the historical node and the root cause occurrence time of the feedback data, respectively. For the preset time matching tolerance range, and These are the anomaly types for historical nodes and feedback data, respectively. This is an indicator function; it takes a value of 1 if a match is successful, and 0 otherwise. When the preset matching threshold is reached, the system determines that the historical node is the verified root cause node corresponding to the feedback data, and integrates all verified nodes to generate a set of verified root cause nodes.

[0134] S62: Using verified root cause nodes as supervision labels, the training sample construction process is performed on the historical abnormal event feature sequence and dynamic temporal cause-effect graph that generated verified root cause nodes, generating supervised learning samples for model tuning.

[0135] Specifically, the system uses nodes from the verified root cause node set as supervision labels, with the label value set as the root cause confirmation result identifier corresponding to that node. The system retrospectively generates historical data for verified root cause nodes, including the anomalous event feature sequence to which the node belongs and the corresponding dynamic temporal causal graph structure data. The system performs training sample construction processing on this historical data. Each sample contains two parts: input features and supervision labels. The input features consist of a high-dimensional feature vector of the anomalous event feature sequence, node association information of the dynamic temporal causal graph, and the initial calculation result of causal strength. The supervision label is the identifier of the verified root cause node and the corresponding root cause confidence level label. To standardize the sample format, the system integrates the input features through a sample representation formula, which is:

[0136]

[0137] in, Let k represent the k-th supervised learning sample. This is the sequence of anomalous event characteristics to which the root cause node has been verified. This is the corresponding dynamic temporal cause-effect graph structure data. This is the initial causal strength matrix of the causal graph. The label is the supervisory label, and its value is the index of the verified root cause node and the root cause confirmation identifier. The system performs deduplication and normalization processing on the constructed samples, removes duplicate samples and invalid samples with missing features, and divides the valid samples into training subsets and validation subsets according to a preset ratio, which together form a supervised learning sample set for model tuning.

[0138] S63: Based on supervised learning samples, the parameters of the causal strength calculation model are optimized. By minimizing the ranking loss of the model for the verified root cause nodes, the internal weight parameters of the model are adjusted to generate the learnable parameters of the updated causal strength calculation model.

[0139] Specifically, the system inputs the supervised learning sample set into the causal strength calculation model and initiates the model parameter optimization process. The optimization objective is to minimize the model's ranking loss for verified root cause nodes, thereby ensuring that verified root cause nodes rank higher in the model's output root cause score ranking. The system uses a ranking loss function to quantify the deviation between the model's prediction results and the supervised labels. The loss function formula is as follows:

[0140]

[0141] in This represents the total sorting loss. To supervise the number of learning samples, The set of supervised labels for the k-th sample (root cause node index verified). and These are the root cause scores for the validated root cause node i and the non-root cause node j in the k-th sample output by the model, respectively. A preset interval threshold is used to differentiate the scores of root cause nodes from non-root cause nodes. The system employs a gradient descent optimization algorithm, calculating the gradient of the loss function with respect to the model's internal weight parameters through backpropagation. Parameter values ​​are adjusted according to the gradient direction, iteratively updated until the loss function converges to a preset minimum or reaches a preset maximum number of iterations. After optimization, the system outputs the updated learnable parameters of the causal strength calculation model, replacing the original model parameters and deploying them to the inference server for subsequent causal strength calculations of anomalous events, achieving a closed-loop improvement in model performance.

[0142] Preferably, in the production site of a smart wire and cable factory, single-head large drawing machines often experience issues such as unreset meter weights leading to concentrated weight reporting from multiple work orders and abnormally high data spikes during specific periods. Double-head large drawing machines suffer from issues like uncollected weight data, deviations in length-to-weight conversion, and incorrect wire diameter selection. Medium drawing machines face delays in weight reset after wire breakage, repeated jumps in data acquisition, and errors in weight calculation for machine #2. Small drawing, micro drawing, and horizontal flattening machines commonly experience delays / inaccuracies in reporting due to unreset current weights, as well as sudden data jumps. These scenarios severely impact the accuracy of reporting data and production management efficiency. The technical solution of this application specifically addresses the above problems: The system first receives the raw operating data reported in real time by each device, parses and extracts the device identifier, production weight, length, and timestamp, and then accesses the manufacturing execution system's work order database to match the data with the task requirements of the current production work order, adding work order identification tags to generate production data points with work order tags. Based on preset reporting logic rules, the system performs rule-based anomaly identification on data points. It determines whether a data point constitutes an isolated anomaly event by using weight jump threshold, continuous interruption duration, and planned rollover weight. Then, it merges the anomaly tags of the same type within a short time window of the same device, uniformly records the event type, key values, device identifier, and timestamp, and generates a standardized original anomaly event stream.

[0143] Furthermore, the system performs multi-dimensional attribute encoding on single abnormal events, transforming event type and equipment identifier into one-hot encoded vectors, and standardizing abnormal values ​​to generate basic event feature vectors. Combining event timestamps and equipment identifiers, the system queries the manufacturing system to obtain material specifications, upper and lower limits of coil weight, and equipment operating modes. These are then transformed into fixed-dimensional context feature vectors through a learnable embedding layer. The two types of vectors are concatenated into a unified high-dimensional feature vector and sorted by timestamp to form a spatiotemporally aligned abnormal event feature sequence.

[0144] The system instantiates each feature vector in the feature sequence as a graph node, creates preliminary temporal links in chronological order to generate an initial temporal graph skeleton, and inputs the causal strength calculation model for node pairs with temporal links. It then calculates causal strength scores by integrating feature correlation, temporal proximity, and device consistency. Based on a preset threshold, it filters and creates directed edges, assigns edge weights, and synthesizes a dynamic temporal causal graph. Using the sum of the weights of all outgoing causal edges of each node as the initial anomaly influence value, it initiates iterative backpropagation to simulate the back diffusion of anomaly influence. Each iteration updates the node probability score to the sum of its own influence value and the propagation score of downstream nodes, until the score change between adjacent iterations is less than the convergence threshold. The final scores are then sorted in descending order to generate a root cause score ranking vector.

[0145] Based on the sorting vector, the top-ranked nodes are selected as candidate key root causes. Starting from each candidate node, downstream nodes are traversed along the outgoing causal edge to extract the complete directed path of anomaly propagation. Paths related to root causes such as network interruption and equipment reset failure are filtered out. The time period of lost data corresponding to the time anomaly node is analyzed. Backlogged production data is extracted from the cache, and the compensation data volume is split according to time segment and matched with work orders and timestamps. The intelligent work reporting instruction is encapsulated according to the MES system interface specification. After legality verification, the compensation work reporting is executed, synchronized to the relevant manufacturing system, and the execution result is recorded.

[0146] In this application scenario, the proposed solution accurately addresses the reporting anomalies of various types of equipment. Its compatibility with multiple protocols, devices, and industries allows for further expansion to more production scenarios, providing solid support for seamless data integration throughout the intelligent manufacturing process. The above examples, citing core production equipment in the wire and cable industry (such as large-diameter, medium-diameter, small-diameter, micro-diameter, and flat-bed wire and cable) mentioned in the technical briefing, illustrate the application scenario of the AI-driven intelligent reporting method. In practice, this method can be adapted to discrete / continuous production equipment with quantitative output in industries such as pipe manufacturing, textiles, and packaging, covering a wider range of intelligent manufacturing scenarios.

[0147] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0148] Based on the same inventive concept, this application also provides an AI-driven intelligent work reporting device for implementing the AI-driven intelligent work reporting method described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations of the one or more AI-driven intelligent work reporting device embodiments provided below can be found in the limitations of the AI-driven intelligent work reporting method described above, and will not be repeated here.

[0149] Preferably, such as Figure 2 As shown, the present invention provides an AI-driven intelligent work reporting system 700, which is configured with the following modules:

[0150] The abnormal event detection module 710 is used to collect and detect abnormalities in the raw operating data reported in real time by the production equipment and the work order task data of the manufacturing system. Based on the preset reporting logic rules, it identifies abnormal events that deviate from the normal production logic and generates a standardized raw abnormal event stream.

[0151] The anomaly feature characterization module 720 is used to perform unified characterization processing on the original anomaly event stream and the context data of the manufacturing system. By encoding event attributes and fusing production work orders and equipment status information, it generates anomaly event feature sequences with spatiotemporal alignment.

[0152] The dynamic causal graph construction module 730 is used to calculate the causal strength between abnormal events in the abnormal event feature sequence based on the preset causal strength calculation model and construct a weighted directed graph to generate a dynamic time-series causal graph.

[0153] The root cause node tracing module 740 is used to trace the root cause nodes of a dynamic time-series causal graph, simulate the process of the abnormal influence spreading backward along the causal edge, iteratively calculate the probability of each node as the starting point of propagation, and generate the root cause score sorting vector of all nodes.

[0154] The intelligent work reporting instruction generation module 750 is used to make diagnostic decisions based on the root cause score sorting vector and dynamic temporal cause-effect graph, locate key root cause nodes according to the score and trace their propagation path, perform timestamp parsing and compensation work reporting processing on the production data backed up during the network interruption according to the root cause type, and generate and execute intelligent work reporting instructions.

[0155] The causal model optimization module 760 is used to perform closed-loop optimization of the causal strength calculation model based on manually confirmed real root cause feedback data, and to update the learnable parameters of the causal strength calculation model.

[0156] Preferably, the abnormal event detection module 710 provided in this application is configured with the following units:

[0157] The data parsing and association unit is used to perform real-time parsing and work order association processing on the raw operating data reported by the production equipment, extract the equipment identifier, production weight, length and timestamp, and match them with the task requirements of the current production work order to generate production data points with work order tags.

[0158] The anomaly identification and marking unit is used to perform rule-based anomaly identification processing on production data points with work order tags. Based on the weight jump threshold, continuous interruption duration and planned roll-up weight in the preset work reporting logic rules, it determines whether the data point constitutes an isolated anomaly event and generates an initial anomaly event marker.

[0159] The anomaly aggregation and standardization unit is used to aggregate and standardize the initial anomaly event tags. It merges the same type of anomaly tags triggered by the same device within a short time window, and uniformly records the event type, key values, device identifier and timestamp to generate a standardized original anomaly event stream.

[0160] Preferably, the anomaly feature characterization module 720 provided in this application is configured with the following units:

[0161] The event attribute encoding unit is used to perform multi-dimensional attribute encoding processing on a single event in the original abnormal event stream, converting the event type and device identifier into a one-hot encoded vector, and standardizing the associated abnormal values ​​to generate a basic event feature vector.

[0162] The context feature embedding unit is used to perform production context embedding processing on the basic event feature vector. Based on the event timestamp and equipment identifier, it queries the manufacturing system to obtain the corresponding material specifications, upper and lower limits of coil weight and equipment operating mode, and transforms them into a fixed-dimensional context feature vector through a learnable embedding layer.

[0163] The feature sequence generation unit is used to concatenate the basic event feature vector and the context feature vector into a unified high-dimensional feature vector, and sort the feature vectors of all events according to the timestamp to generate an abnormal event feature sequence with spatiotemporal alignment.

[0164] Preferably, the dynamic cause-effect graph construction module 730 provided in this application is configured with the following units:

[0165] The graph structure initialization unit is used to perform graph structure initialization processing on the abnormal event feature sequence. It instantiates each feature vector in the abnormal event feature sequence into a graph node, and creates preliminary temporal links between nodes in time order to generate the initial time sequence graph skeleton.

[0166] The causal intensity quantization unit is used to quantify the causal intensity of node pairs with time links in the initial time series graph skeleton. The feature vectors of the two nodes are input into the causal intensity calculation model to calculate the causal intensity score that combines feature correlation, temporal proximity and device consistency.

[0167] The causal graph weighting and synthesis unit is used to perform directed edge filtering and weighting on the causal intensity score. Based on the preset causal intensity threshold, it connects the nodes with causal intensity scores exceeding the threshold with directed edges, and uses the causal intensity score as the edge weight to synthesize a dynamic temporal causal graph.

[0168] Preferably, the root cause tracing module 740 provided in this application is configured with the following units:

[0169] The initial influence calculation unit is used to extract the edge weight information of the weighted causal edges in the dynamic temporal causal graph, summarize the weight values ​​of all outgoing causal edges of each node, and generate the initial abnormal influence value of each node.

[0170] The influence iterative propagation unit is used to perform iterative back propagation based on the initial abnormal influence value of each node and the weighted causal edge, simulating the back diffusion of abnormal influence along the causal edge. In each iteration, the probability score of each node is updated to the sum of the node's own influence value and the propagation scores of all downstream nodes from the node, generating an updated set of node probability scores.

[0171] The root cause score sorting unit is used to perform convergence judgment and sorting based on the node probability score set of the current iteration, calculate the score change between adjacent iterations and determine whether it is less than the preset convergence threshold, sort the final converged scores in descending order of value, and generate the root cause score sorting vector of all nodes.

[0172] Preferably, the intelligent work reporting instruction generation module 750 provided in this application is configured with the following units:

[0173] The key root cause localization unit is used to perform key root cause node localization processing on the root cause score sorting vector, select the top-ranked nodes as candidate key root causes according to the score, and generate a set of candidate root cause nodes.

[0174] The propagation path tracing unit is used to perform abnormal propagation path tracing processing on the candidate root cause node set in the dynamic temporal causal graph. Starting from each candidate node, it traverses downstream nodes along its outgoing causal edge, extracts the complete abnormal propagation directed path, and generates the root cause propagation path set.

[0175] The data compensation reporting unit is used to process data compensation reports for paths related to the root cause of network outage in the root cause propagation path set. It parses the lost data period corresponding to the time anomaly node in the path, extracts the backlog of production data during the lost data period from the cache, and reports the data to the work order in a compensatory manner according to the timestamp. It also generates and executes intelligent reporting instructions.

[0176] Preferably, the causal model optimization module 760 provided in this application is configured with the following units:

[0177] The root cause feedback alignment unit is used to perform alignment and matching processing on the real root cause feedback data confirmed by humans. It associates the real root cause events identified in the feedback with the dynamic time-series cause-effect graph and root cause score sorting vector generated in the historical analysis to determine the corresponding verified root cause nodes in the graph.

[0178] The supervised sample construction unit is used to construct training samples from the historical anomalous event feature sequence and dynamic temporal causal graph that generated the verified root cause nodes, using the verified root cause nodes as supervision labels, and to generate supervised learning samples for model tuning.

[0179] The model parameter optimization unit is used to optimize the parameters of the causal strength calculation model based on supervised learning samples. It adjusts the internal weight parameters of the model by minimizing the ranking loss of the verified root cause nodes, and generates the learnable parameters of the updated causal strength calculation model.

[0180] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described AI-driven intelligent reporting method.

[0181] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned AI-driven intelligent reporting method.

[0182] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0183] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0184] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI-driven intelligent work reporting method, characterized in that, Includes the following steps: S1: Collect and process the raw operating data reported in real time by the production equipment and the work order task data of the manufacturing system. Based on the preset reporting logic rules, identify abnormal events that deviate from the normal production logic and generate a standardized raw abnormal event stream. S2: Perform unified characterization processing on the original abnormal event stream and the context data of the manufacturing system, and generate an abnormal event feature sequence with spatiotemporal alignment by encoding event attributes and fusing production work orders and equipment status information; S3: Calculate the causal strength between abnormal events in the abnormal event feature sequence based on the preset causal strength calculation model and construct a weighted directed graph to generate a dynamic temporal causal graph; S4: Perform root cause node tracing on the dynamic temporal causal graph, simulate the process of abnormal influence spreading backward along the causal edge, and iteratively calculate the probability of each node as the starting point of propagation, generating a root cause score sorting vector for all nodes. S5: Make diagnostic decisions based on the root cause score sorting vector and the dynamic temporal cause-effect graph, locate key root cause nodes according to the score and trace their propagation path, perform timestamp parsing and compensation reporting processing on the production data accumulated during the network interruption according to the root cause type, and generate and execute intelligent reporting instructions. S6: Based on manually confirmed real root cause feedback data, perform closed-loop optimization of the model parameters of the causal strength calculation model, and update the learnable parameters of the causal strength calculation model.

2. The method according to claim 1, characterized in that, S1 includes: S11: Real-time analysis and work order association processing of the raw operating data reported by the production equipment, extracting equipment identification, production weight, length and timestamp, and matching it with the task requirements of the current production work order to generate production data points with work order tags; S12: Perform rule-based anomaly identification processing on the production data points with work order tags. Based on the weight jump threshold, continuous interruption duration and planned roll-up weight in the preset work reporting logic rules, determine whether the data points constitute isolated abnormal events and generate initial abnormal event markers. S13: Aggregate and standardize the initial abnormal event markers, merge the same type of abnormal markers triggered by the same device within a short time window, and uniformly record the event type, key values, device identifier and timestamp to generate a standardized original abnormal event stream.

3. The method according to claim 1, characterized in that, S2 includes: S21: Perform multi-dimensional attribute encoding processing on a single event in the original abnormal event stream, convert the event type and device identifier into a one-hot encoded vector, and standardize the associated abnormal values ​​to generate a basic event feature vector. S22: Perform production context embedding processing on the basic event feature vector, query the manufacturing system according to the event timestamp and equipment identifier to obtain the corresponding material specifications, upper and lower limits of coil weight and equipment operation mode, and transform it into a fixed-dimensional context feature vector through a learnable embedding layer. S23: Concatenate the basic event feature vector and the context feature vector into a unified high-dimensional feature vector, and sort the feature vectors of all events according to the timestamp to generate an abnormal event feature sequence with spatiotemporal alignment.

4. The method according to claim 1, characterized in that, S3 includes: S31: Perform graph structure initialization processing on the abnormal event feature sequence, instantiate each feature vector in the abnormal event feature sequence into a graph node, and create preliminary temporal links between nodes in time order to generate an initial temporal graph skeleton. S32: Perform causal strength quantification on the node pairs with time-series links in the initial time-series graph skeleton, input the feature vectors of the two nodes into the causal strength calculation model, and calculate the causal strength score that combines feature correlation, temporal proximity and device consistency. S33: Perform directed edge filtering and weighting on the causal intensity score, connect the node pairs whose causal intensity scores exceed the preset causal intensity threshold with directed edges based on the preset causal intensity threshold, and use the causal intensity score as the edge weight to synthesize a dynamic temporal causal graph.

5. The method according to claim 1, characterized in that, S4 includes: S41: Extract edge weight information from the weighted causal edges in the dynamic temporal causal graph, summarize the weight values ​​of all outgoing causal edges of each node, and generate the initial abnormal influence value of each node. S42: Based on the initial abnormal influence value of each node and the weighted causal edge, perform iterative back propagation to simulate the reverse diffusion of abnormal influence along the causal edge. In each iteration, update the probability score of each node to the sum of the node's own influence value and the propagation scores from all downstream nodes of the node, and generate an updated set of node probability scores. S43: Based on the node probability score set of the current iteration, perform convergence judgment and sorting processing, calculate the score change between adjacent iterations and determine whether it is less than the preset convergence threshold, sort the final converged scores in descending order of value, and generate the root cause score sorting vector of all nodes.

6. The method according to claim 1, characterized in that, S5 includes: S51: Perform key root cause node localization processing on the root cause score sorting vector, select the top-ranked nodes as candidate key root causes according to the score, and generate a candidate root cause node set. S52: The set of candidate root cause nodes is subjected to anomaly propagation path tracing processing in the dynamic temporal causal graph. Starting from each candidate node, the downstream nodes are traversed along the causal edge it emanates from, and the complete anomaly propagation directed path is extracted to generate a set of root cause propagation paths. S53: Perform data compensation reporting on the paths related to the root cause of network interruption in the root cause propagation path set, parse the lost data period corresponding to the time abnormal node in the path, extract the backlogged production data of the lost data period from the cache and perform compensatory reporting according to the timestamp and work order, and generate and execute intelligent reporting instructions.

7. The method according to any one of claims 1-6, characterized in that, S6 includes: S61: Perform alignment and matching processing on the manually confirmed true root cause feedback data, associate the true root cause events identified in the feedback with the dynamic time-series cause-effect graph and the root cause score sorting vector generated in the historical analysis, and determine the corresponding verified root cause nodes in the graph. S62: Using the verified root cause node as a supervision label, the historical abnormal event feature sequence and dynamic temporal cause-effect graph that generated the verified root cause node are processed to construct training samples, generating supervised learning samples for model tuning. S63: Based on the supervised learning samples, the parameters of the causal strength calculation model are optimized by minimizing the ranking loss of the model for the verified root cause nodes to adjust the internal weight parameters of the model, thereby generating the updated learnable parameters of the causal strength calculation model.

8. An AI-driven intelligent work reporting system, characterized in that, The system includes: The abnormal event detection module is used to collect and detect abnormalities in the raw operating data reported in real time by the production equipment and the work order task data of the manufacturing system. Based on the preset reporting logic rules, it identifies abnormal events that deviate from the normal production logic and generates a standardized raw abnormal event stream. The abnormal feature characterization module is used to perform unified characterization processing on the original abnormal event stream and the context data of the manufacturing system. By encoding event attributes and fusing production work orders and equipment status information, it generates an abnormal event feature sequence with spatiotemporal alignment. The dynamic causal graph construction module is used to calculate the causal strength between abnormal events in the abnormal event feature sequence based on a preset causal strength calculation model and construct a weighted directed graph to generate a dynamic temporal causal graph. The root cause node tracing module is used to trace the root cause nodes of the dynamic temporal causal graph, simulate the process of the abnormal influence spreading backward along the causal edge and iteratively calculate the probability of each node as the starting point of propagation, and generate the root cause score sorting vector of all nodes. The intelligent work reporting instruction generation module is used to make diagnostic decisions based on the root cause score sorting vector and the dynamic temporal cause-effect graph, locate key root cause nodes according to the score and trace their propagation path, perform timestamp parsing and compensation work reporting processing on the production data backed up during the network interruption according to the root cause type, and generate and execute intelligent work reporting instructions. The causal model optimization module is used to perform closed-loop optimization of the model parameters of the causal strength calculation model based on manually confirmed real root cause feedback data, and to update the learnable parameters of the causal strength calculation model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images