Large-model-driven business index anomaly causal attribution method and system

Abstract

The invention relates to the technical field of supply chain business data analysis, and discloses a large model driven business index anomaly causal attribution method and system, and the method comprises the steps: collecting heterogeneous index data; constructing a unified semantic index set; identifying a potential abnormal index set; an optimized cause and effect graph is generated through time lag relevancy clipping; executing attribution analysis based on the structural equation model and an anti-factual reasoning mechanism; and generating a multi-granularity attribution interpretation result in combination with the large model. The technical problem that in the prior art, a scheme lacking unified index normalization and multi-scale causal interpretation is relied on, and particularly under the condition of crossing multiple business systems such as an ERP system, a CRM system and an MES system, accurate anomaly recognition and causal attribution cannot be achieved is solved. Due to the fact that the large model embedder, the causal graph pruning mechanism and the structured anti-fact modeling method are introduced, structural traceability and hierarchical interpretation of index abnormity are achieved, and the intelligent level of supply chain service monitoring is improved.

Description

Technical Field

[0001] This invention relates to the field of supply chain business data analysis technology, and in particular to a large model-driven method and system for attributing the causes of abnormal business indicators. Background Technology

[0002] Currently, with the continuous advancement of enterprise digital transformation, information systems such as Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM) are widely deployed in various business processes, including procurement, production, sales, logistics, and customer service. These systems collect and manage large amounts of structured indicator data from different business dimensions, playing a crucial role in supporting enterprise business decisions, process optimization, and risk management. However, current systems commonly suffer from data silos, semantic inconsistencies, and differences in indicator definitions, severely hindering cross-system data fusion and global collaborative analysis capabilities. In real-world business scenarios, when a key business indicator (such as inventory turnover days, customer order fulfillment rate, or energy consumption per unit of output) experiences abnormal fluctuations, enterprises typically need to trace the root cause of the anomaly, identify whether it is caused by other abnormal operations in upstream systems (such as procurement delays, production scheduling imbalances, or abnormal customer behavior), and further assess the impact of this anomaly on other downstream indicators or overall business objectives. Therefore, it is necessary to construct causal paths for indicators that encompass multiple systems to support root cause analysis and decision support based on anomaly signals.

[0003] The existing technologies suffer from the following main shortcomings: First, most current business systems employ their own independent data models and indicator definitions, lacking a unified indicator semantic mapping mechanism, making it impossible to establish stable and reusable indicator correlation links between different systems. Second, traditional indicator anomaly detection methods are mostly based on static threshold judgments or statistical deviation calculations of isolated indicators, making it difficult to identify interconnected anomalies and weak signal disturbances in the context of multi-indicator coupling. Third, in terms of attribution analysis, existing solutions typically rely on rule engines or coarse-grained graph structures based on statistical correlations, lacking the ability to accurately model causal direction, time lag, and intervention response, making it difficult to achieve multi-level explanations and quantitative attributions of abnormal behavior in complex systems. Furthermore, current methods generally lack explanatory models with large-scale semantic understanding capabilities, making it difficult to map attribution results to highly readable explanatory views.

[0004] Therefore, there is an urgent need for a large-scale model-driven causal attribution method for abnormal business indicators, which can accurately identify, locate the root cause, and explain the multi-level causal relationship of key indicator anomalies in complex enterprise information system environments, thereby improving the enterprise's intelligent decision-making and anomaly intervention capabilities. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a large-model-driven causal attribution method for abnormal business indicators. This method aims to solve the technical problem that existing technologies rely on solutions lacking unified indicator normalization and multi-scale causal explanations, especially when dealing with multiple business systems such as ERP, CRM, and MES systems, where accurate anomaly identification and causal attribution are impossible.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a large model-driven causal attribution method for abnormal business indicators. The large model-driven causal attribution method for abnormal business metrics includes: Step S10: Collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM). Based on the heterogeneous indicator data, construct a unified semantic indicator set at time t using a dynamic semantic and temporal mapping mechanism. ; Step S20: Based on the unified semantic index set A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. ; Step S30: Set up the abnormal indicators A knowledge mapping mechanism is used to project the knowledge onto a pre-defined set of business knowledge graph nodes, thus constructing an initial causal directed graph. For the initial causal directed graph Directed edges in The time-lag correlation degree was calculated using Pearson correlation analysis based on time lag. Based on time lag correlation For the initial causal directed graph Perform a pruning operation and output an optimized causal directed graph. ; Step S40: Based on the optimized causal directed graph A counterfactual attribution task is performed using a counterfactual reasoning-based attribution mechanism, and the root cause contribution set C is output. Step S50: Generate multi-granularity attribution display results based on the root cause contribution set C and a large model based on the Transformer structure.

[0007] Preferably, in step S10, ,in, Represents a vector of supply availability metrics; Represents a vector of inventory level indicators; Represents a vector of sales revenue metrics; This represents a vector of demand indicators.

[0008] Preferably, in step S10, heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM) are collected, and a unified semantic indicator set for time t is constructed based on the heterogeneous indicator data using a dynamic semantic and temporal mapping mechanism. The steps specifically include: Step S101: Collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM). Input the heterogeneous indicator data into a preset large model embedding device. The large model embedding device outputs the semantic vector of the i-th indicator data. The semantic vector of the j-th indicator data ; and based on the semantic vector of the i-th indicator data The semantic vector of the j-th indicator data The first semantic similarity is calculated using cosine similarity analysis. The large model embedder is built based on the Sentence-BERT framework. Step S102: Introduce semantic drift penalty coefficient And obtain the semantic change magnitude of the first semantic similarity within the preset semantic change collection period. Based on semantic drift penalty coefficient and semantic change range For the first semantic similarity Make corrections and output the second semantic similarity. Based on second semantic similarity Generate semantic similarity matrix ; Step S103: Perform time-scale indicator alignment operations on heterogeneous indicator data using the pandas and statsmodels libraries in Python, and output the aligned indicator series; Step S104: Based on the semantic similarity matrix Combined with the aligned index sequence and a pre-defined causal prior constraint graph The semantic vector of the i-th indicator data The semantic vector of the j-th indicator data Perform a structure normalization inference task, output the causal chain node attribution, and generate a unified semantic index set based on the causal chain node attribution. .

[0009] Preferably, in step S20, based on the unified semantic index set A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. The steps specifically include: Step S201: Based on the unified semantic index set The sliding window prediction method is used to calculate the full-dimensional standardized residual set; Step S202: Construct a large-scale model-driven semantic importance weighting model. Input the full-dimensional standardized residual set into the semantic importance weighting model, and the semantic importance weighting model outputs a business importance weight set; and based on the business importance weight set and the unified semantic index set... A comprehensive anomaly scoring function is constructed using the weighted L1 norm principle; Step S203: Set the comprehensive anomaly threshold When the comprehensive anomaly scoring function is greater than or equal to the comprehensive anomaly threshold When an overall business anomaly is detected, a single-dimensional standardized residual is obtained from the full-dimensional standardized residual set, and an anomaly indicator set is generated and output based on the single-dimensional standardized residual. .

[0010] Preferably, in step S202, the semantic importance weighted model is constructed using the BERT attention mechanism; the semantic importance weighted model includes a semantic embedding layer for receiving input from a full-dimensional standardized residual set; a multi-head attention layer for learning the semantic interaction weights between different indicators; and a normalized Softmax layer for outputting a set of business importance weights.

[0011] Preferably, in step S40, the causal directed graph is optimized. The steps for performing a counterfactual attribution task using a counterfactual reasoning-based attribution mechanism and outputting the root cause contribution set C specifically include: Step S401: Optimize the causal directed graph The set of upstream nodes that have causal path connections with the preset sales target Y is denoted as the candidate root cause indicator set S; based on the candidate root cause indicator set S, the structural equation model (SEM) of the sales target Y is constructed using the structural causal modeling principle; ,in, This represents a multivariate function learned through a predefined structural causal modeling function. Indicates the structural error term; Step S402: Construct counterfactual conditional expectation based on structural equation modeling (SEM) using the counterfactual conditional expectation modeling method. ; Step S403: Based on counterfactual conditional expectations The do-means algorithm based on expected difference is used to calculate the average causal effect set corresponding to the candidate root cause index set S. Further based on the average causal effect set The root cause contribution set C is constructed and output using the Softmax normalization principle.

[0012] Preferably, in step S50, the multi-granularity attribution display results include a technical indicator layer, which is used to display the standardized residual change trend, average causal effect contribution rate and lag correlation characteristics of each indicator; a business operation layer, which is used to display the causal path from abnormal indicators to key business KPIs, the contribution of nodes in the path and the cumulative amplification effect; and a strategic decision layer, which is used to display the business logic in natural language text format.

[0013] This invention also provides a large model-driven causal attribution system for abnormal business metrics, comprising: The unified data preprocessing module is used to collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM) systems. Based on the heterogeneous indicator data, a unified semantic indicator set at time t is constructed using a dynamic semantic and temporal mapping mechanism. ; Anomaly detection and scoring module, used for anomaly detection based on a unified semantic index set. A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. ; The cause-effect graph construction and pruning module is used to construct and prune sets of abnormal indicators. A knowledge mapping mechanism is used to project the knowledge onto a pre-defined set of business knowledge graph nodes, thus constructing an initial causal directed graph. For the initial causal directed graph Directed edges in The time-lag correlation degree was calculated using Pearson correlation analysis based on time lag. Based on time lag correlation For the initial causal directed graph Perform a pruning operation and output an optimized causal directed graph. ; The counterfactual causal attribution module is used to optimize the causal directed graph. A counterfactual attribution task is performed using a counterfactual reasoning-based attribution mechanism, and the root cause contribution set C is output. The Attribution Explanation and Multi-Granularity Visualization module is used to generate multi-granularity attribution display results based on the root cause contribution set C combined with a large model based on the Transformer structure.

[0014] The present invention also provides a large model-driven business indicator anomaly causal attribution device, comprising: a memory, a processor, and a large model-driven business indicator anomaly causal attribution program stored in the memory and executable on the processor. When the large model-driven business indicator anomaly causal attribution program is executed by the processor, it implements a large model-driven business indicator anomaly causal attribution method.

[0015] The present invention also provides a computer program product, including a large model-driven business indicator anomaly causal attribution program, which, when executed by a processor, implements the large model-driven business indicator anomaly causal attribution method.

[0016] The beneficial effects of this invention are as follows: By introducing a large model embedding mechanism, this invention performs semantic alignment and temporal standardization on heterogeneous indicators from systems such as ERP, CRM, and MES, and constructs a unified indicator space. This effectively solves the problem of misjudgment of anomaly identification caused by inconsistent indicator semantics and inconsistent data formats in the prior art, and fundamentally improves the accuracy and interpretability of indicator-level anomaly monitoring.

[0017] This invention integrates causal graph pruning, structural equation modeling, and counterfactual reasoning mechanisms. While identifying anomalies in key indicators, it can also uncover their potential upstream root causes. Through large-scale model generation technology and multi-level explanations of business and strategy, it significantly improves the efficiency of handling abnormal events and decision support capabilities in supply chain scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the first embodiment of a large model-driven causal attribution method for abnormal business metrics according to the present invention.

[0020] Figure 2 This is a schematic diagram of a device for a large model-driven causal attribution method for abnormal business metrics according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the large-model-driven causal attribution method for abnormal business metrics of the present invention, which presents the first embodiment of the large-model-driven causal attribution method for abnormal business metrics of the present invention.

[0023] In the first embodiment, the large model-driven causal attribution method for abnormal business metrics includes: Step S10: Collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM). Based on the heterogeneous indicator data, construct a unified semantic indicator set at time t using a dynamic semantic and temporal mapping mechanism. ; It should be noted that the "dynamic semantic and temporal mapping mechanism" refers to a unified processing flow for indicator fields in multi-source heterogeneous business systems. This mechanism consists of three parts: a semantic standardization module, an indicator mapping engine, and a temporal alignment module. The semantic standardization module parses the field names, descriptive metadata, and historical value patterns of indicators in ERP, MES, and CRM systems, extracting their underlying business meaning and semantic nesting hierarchy. The indicator mapping engine calculates the semantic similarity between indicators using structured word vector representations (such as BERT embedding) and performs semantic normalization based on a preset business ontology or mapping rule base. The temporal alignment module, through sampling window normalization, data resampling, and timestamp standardization, achieves temporal consistency alignment of indicators across different systems, providing a temporal consistency guarantee for building a unified semantic indicator set.

[0024] Understandably, this mechanism allows indicator data that were originally scattered across different systems with different naming conventions and sampling frequencies to be mapped into an indicator space with unified naming, unified business semantics, and unified time granularity. This results in a unified semantic indicator set with strong cross-system comparability and consistent structure. This indicator set not only significantly improves the semantic accuracy in subsequent causal path modeling and anomaly tracing analysis, but also provides a reliable semantic input foundation for models such as structural equation modeling and counterfactual reasoning, avoiding attribution misjudgments caused by semantic bias or time misalignment.

[0025] For example, in a typical pilot application at a manufacturing enterprise, the original field names for "Purchase Order Time" from the ERP system, the "Raw Material Receiving Time" field from the MES system, and the "Customer Complaint Response Time" field from the CRM system were inconsistent, with time precisions of daily, hourly, and second-level, respectively. After applying the dynamic semantic and temporal mapping mechanism of this invention, these fields were successfully mapped uniformly to the "Supply Response Delay Index," aligned at the daily granularity. Subsequently, in an inventory turnover anomaly analysis, this unified semantic index was identified as the primary source of the anomaly, significantly improving the accuracy of causal attribution analysis and the clarity of business explanations. Specific data shows that after adopting the method of this invention, the anomaly identification accuracy and attribution path coverage improved, verifying the effectiveness and promotional value of this mechanism in real-world business environments.

[0026] Step S20: Based on the unified semantic index set A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. ; It should be noted that the "residual-driven multi-index anomaly detection mechanism" refers to a multi-dimensional anomaly identification method that utilizes the fusion of prediction residuals and semantic weights. This mechanism first constructs a time-series prediction model for each index in a unified semantic index set, and uses the sliding window method to calculate the deviation between the predicted value and the actual value at the current moment, forming a standardized residual set. Then, based on a pre-set large model embedding unit, it calculates the importance weight of each index in terms of business semantics, and weights and fused the residual values ​​and weight values ​​to form a comprehensive anomaly score across multiple index dimensions.

[0027] It should be understood that, compared to traditional detection methods based on static thresholds or isolated indicator biases, this invention features innovative improvements in its algorithm. First, by introducing a sliding window adaptive prediction model (such as the ARIMA-LSTM hybrid model), the ability to fit non-stationary time series is enhanced. Second, by introducing a semantic importance weighting mechanism, explanatory weights are assigned to indicators within the business logic, thereby overcoming the deficiency of "equal weighting of detection results for each indicator" in traditional methods. Finally, through a weighted L1 norm aggregation strategy, the anomaly scoring process becomes more robust, effectively suppressing interference from single-point extrema.

[0028] For example, in a supply chain monitoring scenario, the system performs residual-driven detection on three indicators: "procurement and warehousing cycle" in the ERP system, "production takt time" in the MES system, and "order fulfillment time" in the CRM system. When "production takt time" fluctuates little in the short term but "procurement and warehousing cycle" continues to rise, the overall anomaly score calculated by weighting the residuals shows a significant increase, indicating "supply chain input anomaly risk." Subsequent verification confirmed that there were indeed delays in raw material delivery during this period, leading to a short-term imbalance in the inventory structure. This validates the significant technical advantages and feasibility of the residual-driven and weighted aggregation mechanism in complex business environments.

[0029] Step S30: Set up the abnormal indicators A knowledge mapping mechanism is used to project the knowledge onto a pre-defined set of business knowledge graph nodes, thus constructing an initial causal directed graph. For the initial causal directed graph Directed edges in The time-lag correlation degree was calculated using Pearson correlation analysis based on time lag. Based on time lag correlation For the initial causal directed graph Perform a pruning operation and output an optimized causal directed graph. ; It should be noted that the "knowledge mapping mechanism" refers to projecting the set of abnormal indicators output in step S20 onto a pre-defined set of business knowledge graph nodes through semantic matching, system affiliation, and business process dependency rules, thereby locating the nodes and paths related to the abnormal indicators in the graph structure. This graph can include key business activity nodes, data indicator nodes, and their causal dependencies across multiple system dimensions (such as ERP, MES, CRM, etc.), forming a unified cross-system business causal graph. The subgraph structure formed based on the mapping results in the graph is defined as an "initial causal directed graph," where the nodes are the semantic matching results of the abnormal indicators, and the directed edges represent the initial causal assumptions between different indicators based on business processes or data flows.

[0030] Understandably, by introducing a pruning mechanism based on time-lag correlation, the edge relationships of the initial causal directed graph can be effectively filtered and optimized at the data level, improving the accuracy and sparsity of the causal graph structure. Specifically, a lag sliding window is constructed based on the time series of indicators, and the "time-lag Pearson correlation analysis" is used to evaluate the linear dependence strength between indicator pairs at different lag times. Edge relationships with strong correlation and clear direction are retained based on the maximum correlation lag and statistical significance. This operation not only ensures the dynamic validity of edges in the causal graph but also suppresses spurious causal paths caused by error propagation or synchronization fluctuations.

[0031] It should be understood that, compared to traditional solutions that use static business process diagrams or manually defined causal paths, this invention introduces a data-driven time-lag correlation pruning mechanism, making the generation process of the causal graph both semantically reasonable and data-supported, thus enhancing the reliability of the graph structure in subsequent causal analysis. Especially in real-world environments where there are differences in sampling frequencies or propagation delays among cross-system and multi-scale indicators, traditional methods easily construct non-realistic synchronous edge relationships. This method, through time-off search and maximum lag correlation control, effectively avoids the retention of spurious correlation paths, improving the semantic accuracy and causal credibility of the overall causal graph.

[0032] For example, in actual enterprise operational data analysis, after identifying an anomaly in "inventory turnover days," the initial causal graph constructed causal connections between it and nodes such as "procurement receipt cycle," "order processing cycle," and "customer complaint frequency." By performing a lagged Pearson analysis on the time series data from the past three months, it was found that the maximum correlation lag between "procurement receipt cycle" and "inventory turnover days" was 5 days, with a Pearson coefficient reaching 0.71, indicating a significant causal shift signal. While "customer complaint frequency" also fluctuated, its correlation lag was unstable and its correlation coefficient fluctuated significantly, leading to automatic pruning. The resulting optimized causal directed graph clearly revealed that the main cause of the "inventory anomaly" was "delays in the procurement process," providing a high-confidence structural basis for subsequent root cause attribution and operational intervention.

[0033] Step S40: Based on the optimized causal directed graph A counterfactual attribution task is performed using a counterfactual reasoning-based attribution mechanism, and the root cause contribution set C is output. It should be noted that the "counterfactual attribution mechanism" refers to establishing a structural equation model for a target anomalous indicator (such as sales revenue, inventory turnover days, or order fulfillment rate) based on an optimized causal directed graph, and combining this with a counterfactual scenario generation method to simulate interventions and evaluate the effects on a set of candidate root cause indicators. This mechanism first constructs a structural equation model based on upstream nodes in the causal graph that have path relationships with the target indicator. Using the observed values, causal weights, and structural error terms of each node as inputs, an explicit functional relationship is established between the target variable and the causal nodes. Subsequently, a do-operation intervention is applied to the candidate root cause variable in a virtual scenario, and the expected difference in the target indicator before and after the intervention is calculated, thereby achieving quantitative inference of causal effects in a counterfactual sense.

[0034] Understandably, by using counterfactual reasoning mechanisms, it is possible to simulate states under several "assumptions" while maintaining data consistency, thereby quantitatively assessing the contribution of each candidate indicator to the target anomaly.

[0035] It should be understood that, compared to traditional attribution schemes based on regression residuals, gradient attribution, or simple correlation analysis, the counterfactual reasoning mechanism of this invention differs fundamentally in model structure and explanatory depth. Traditional methods can only identify "strongly correlated" characteristic variables, while this invention, through a combination of causal modeling and counterfactual intervention, explicitly removes the influence of confounding factors and covariates, thereby ensuring that the attribution results are directional and causally independent. Furthermore, this method employs a joint calculation framework of Average Causal Effect (ACE) and Normalized Causal Contribution Score (C-score), enabling the causal influence between different root cause indicators to be compared within a unified numerical space, effectively avoiding attribution bias caused by differences in dimensions.

[0036] For example, in a supply chain operations monitoring scenario, when an anomaly of "declining sales" is detected, the optimized causal graph reveals three candidate causal paths: "inventory turnover days," "purchase receipt cycle," and "customer return rate." After structural equation modeling, it is calculated that shortening the "purchase receipt cycle" to a normal level would increase the counterfactual expected sales by approximately 8.7%; while keeping the "inventory turnover days" unchanged and only adjusting the "customer return rate," the counterfactual expected sales would only increase by approximately 2.4%. After normalization, the output root cause contribution set C = {purchase receipt cycle: 0.67, inventory turnover days: 0.23, customer return rate: 0.10} indicates that "delayed purchase receipt cycle" is the primary root cause of the sales anomaly. This result is consistent with the company's subsequent actual verification data.

[0037] Step S50: Generate multi-granularity attribution display results based on the root cause contribution set C and a large model based on the Transformer structure.

[0038] It should be noted that the "large model based on Transformer architecture" refers to a pre-trained language model as its core. By introducing attribution context vectors, indicator semantic labels, and causal path hint templates, it performs linguistic modeling of the causal paths, weight distributions, and business background information of each indicator in the root cause contribution set C, and automatically generates readable attribution results covering different granularities (e.g., summary level, indicator level, path level). In practical implementation, this large model can be embedded into the attribution result interpretation module using a parameter-frozen LLM (such as BERT, T5, ChatGLM, etc.) or a lightweight, fine-tuned version of the Transformer architecture (such as LoRA, Adapter, etc.) to ensure its response timeliness and semantic consistency.

[0039] Understandably, by introducing the language generation capabilities of large models, highly structured but difficult-to-understand numerical information in the original causal contribution set C can be transformed into natural language text with logical explanations, business context, and actionable suggestions. This not only preserves the causal rigor of counterfactual attribution results but also gives the results stronger readability in the business context, lowering the threshold for enterprise operators, decision-makers, or engineers to understand and use causal attribution results.

[0040] For example, for the root cause contribution set C = {Purchase receipt cycle: 0.67, Inventory turnover days: 0.23, Customer return rate: 0.10} output in step S40, the large model first generates an attribution summary: "The decline in sales is mainly driven by anomalies in the procurement process"; then it further outputs an indicator-level explanation: "The purchase receipt cycle was extended by 2.8 days, contributing 67% to the decline in sales"; finally, it generates an explanation at the path level: "Delayed order approval in the ERP system (purchase order #) → Extended receipt cycle → Delayed inventory replenishment → Reduced inventory turnover → Reduced sales." In addition, the large model can also generate recommended strategies based on needs, such as: "It is recommended to set automatic reminders and overdue intervention rules for the purchase approval process to reduce the occurrence of similar anomalies in the future."

[0041] Example 2: Furthermore, the present invention provides a large-model-driven causal attribution system for abnormal business metrics, employing a large-model-driven causal attribution method for abnormal business metrics in the above embodiments, which can solve the technical problem of causal attribution of abnormal business metrics driven by a large model. Compared with the prior art, the beneficial effects of the large-model-driven causal attribution system for abnormal business metrics provided by the present invention are the same as the beneficial effects of the large-model-driven causal attribution method for abnormal business metrics provided in the above embodiments, and other technical features of the large-model-driven causal attribution system for abnormal business metrics are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0042] Example 3: This invention provides a large-model-driven business indicator anomaly causal attribution device. Please refer to... Figure 2A large-model-driven business metric anomaly causal attribution device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to execute the large-model-driven business metric anomaly causal attribution method described in Embodiment 1 above. The large-model-driven business metric anomaly causal attribution device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This large-model-driven business metric anomaly causal attribution device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this invention. A large-model-driven business indicator anomaly causal attribution device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the large-model-driven business indicator anomaly causal attribution device. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows a large-model-driven business indicator anomaly causal attribution device to communicate wirelessly or wiredly with other devices to exchange data. While the figure illustrates a large-model-driven business indicator anomaly causal attribution device with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0043] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described large-model-driven causal attribution method for abnormal business metrics. The computer program product provided by this invention can solve the technical problem of causal attribution of abnormal business metrics driven by a large model. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the causal attribution method for abnormal business metrics driven by a large model provided in the above embodiments, and will not be repeated here.

[0044] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0045] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0046] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A large-model-driven causal attribution method for abnormal business metrics, characterized in that, The methods include: Step S10: Collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM). Based on the heterogeneous indicator data, construct a unified semantic indicator set at time t using a dynamic semantic and temporal mapping mechanism. ; Step S20: Based on the unified semantic index set A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. ; Step S30: Set up the abnormal indicators A knowledge mapping mechanism is used to project the knowledge onto a pre-defined set of business knowledge graph nodes, thus constructing an initial causal directed graph. For the initial causal directed graph Directed edges in The time-lag correlation degree was calculated using Pearson correlation analysis based on time lag. Based on time lag correlation For the initial causal directed graph Perform a pruning operation and output an optimized causal directed graph. ; Step S40: Based on the optimized causal directed graph A counterfactual attribution task is performed using a counterfactual reasoning-based attribution mechanism, and the root cause contribution set C is output. Step S50: Generate multi-granularity attribution display results based on the root cause contribution set C and a large model based on the Transformer structure.

2. The large-model-driven causal attribution method for abnormal business metrics as described in claim 1, characterized in that, In step S10, ,in, Represents a vector of supply availability metrics; Represents a vector of inventory level indicators; Represents a vector of sales revenue metrics; This represents a vector of demand indicators.

3. The large-model-driven causal attribution method for abnormal business metrics as described in claim 1, characterized in that, In step S10, heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM) are collected. Based on the heterogeneous indicator data, a unified semantic indicator set for time t is constructed using a dynamic semantic and temporal mapping mechanism. The steps specifically include: Step S101: Collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM). Input the heterogeneous indicator data into a preset large model embedding device. The large model embedding device outputs the semantic vector of the i-th indicator data. The semantic vector of the j-th indicator data ; and based on the semantic vector of the i-th indicator data The semantic vector of the j-th indicator data The first semantic similarity is calculated using cosine similarity analysis. The large model embedder is built based on the Sentence-BERT framework. Step S102: Introduce semantic drift penalty coefficient And obtain the semantic change magnitude of the first semantic similarity within the preset semantic change collection period. Based on semantic drift penalty coefficient and semantic change range For the first semantic similarity Make corrections and output the second semantic similarity. Based on second semantic similarity Generate semantic similarity matrix ; Step S103: Perform time-scale indicator alignment operations on heterogeneous indicator data using the pandas and statsmodels libraries in Python, and output the aligned indicator series; Step S104: Based on the semantic similarity matrix Combined with the aligned index sequence and a pre-defined causal prior constraint graph The semantic vector of the i-th indicator data The semantic vector of the j-th indicator data Perform a structure normalization inference task, output the causal chain node attribution, and generate a unified semantic index set based on the causal chain node attribution. .

4. The large-model-driven causal attribution method for abnormal business metrics as described in claim 1, characterized in that, In step S20, based on the unified semantic index set A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. The steps specifically include: Step S201: Based on the unified semantic index set The sliding window prediction method is used to calculate the full-dimensional standardized residual set; Step S202: Construct a large-scale model-driven semantic importance weighting model. Input the full-dimensional standardized residual set into the semantic importance weighting model, and the semantic importance weighting model outputs a business importance weight set; and based on the business importance weight set and the unified semantic index set... A comprehensive anomaly scoring function is constructed using the weighted L1 norm principle; Step S203: Set the comprehensive anomaly threshold When the comprehensive anomaly scoring function is greater than or equal to the comprehensive anomaly threshold When an overall business anomaly is detected, a single-dimensional standardized residual is obtained from the full-dimensional standardized residual set, and an anomaly indicator set is generated and output based on the single-dimensional standardized residual. .

5. The large-model-driven causal attribution method for abnormal business metrics as described in claim 4, characterized in that, In step S202, the semantic importance weighted model is constructed using the BERT attention mechanism; the semantic importance weighted model includes a semantic embedding layer, which is used to receive the input of the full-dimensional standardized residual set; A multi-head attention layer is used to learn the semantic interaction weights between different metrics; The normalized Softmax layer is used to output the set of business importance weights.

6. The large-model-driven causal attribution method for abnormal business metrics as described in claim 1, characterized in that, In step S40, based on the optimized causal directed graph... The steps for performing a counterfactual attribution task using a counterfactual reasoning-based attribution mechanism and outputting the root cause contribution set C specifically include: Step S401: Optimize the causal directed graph The set of upstream nodes that have causal path connections with the preset sales target Y is denoted as the candidate root cause indicator set S; based on the candidate root cause indicator set S, the structural equation model (SEM) of the sales target Y is constructed using the structural causal modeling principle; ,in, This represents a multivariate function learned through a predefined structural causal modeling function. Indicates the structural error term; Step S402: Construct counterfactual conditional expectation based on structural equation modeling (SEM) using the counterfactual conditional expectation modeling method. ; Step S403: Based on counterfactual conditional expectations The do-means algorithm based on expected difference is used to calculate the average causal effect set corresponding to the candidate root cause index set S. Further based on the average causal effect set The root cause contribution set C is constructed and output using the Softmax normalization principle.

7. The large-model-driven causal attribution method for abnormal business metrics as described in claim 1, characterized in that, In step S50, the multi-granularity attribution display results include a technical indicator layer, which displays the standardized residual change trend, average causal effect contribution rate and lag correlation characteristics of each indicator; and a business operation layer, which displays the causal path from abnormal indicators to key business KPIs, the contribution of nodes in the path and the cumulative amplification effect. The strategic decision-making layer is used to present the business logic in natural language text format.

8. A large-model-driven business indicator anomaly causal attribution system, applied to the large-model-driven business indicator anomaly causal attribution method according to any one of claims 1 to 7, characterized in that, The large model-driven business metric anomaly attribution system includes: The unified data preprocessing module is used to collect heterogeneous indicator data from Enterprise Resource Planning (ERP), Manufacturing Execution System (MES), and Customer Relationship Management (CRM) systems. Based on the heterogeneous indicator data, a unified semantic indicator set at time t is constructed using a dynamic semantic and temporal mapping mechanism. ; Anomaly detection and scoring module, used for anomaly detection based on a unified semantic index set. A residual-driven multi-index anomaly detection mechanism is used to perform residual calculation and weighted aggregation operations, outputting a set of anomaly indicators. ; The cause-effect graph construction and pruning module is used to construct and prune sets of abnormal indicators. A knowledge mapping mechanism is used to project the knowledge onto a pre-defined set of business knowledge graph nodes, thus constructing an initial causal directed graph. For the initial causal directed graph Directed edges in The time-lag correlation degree was calculated using Pearson correlation analysis based on time lag. Based on time lag correlation For the initial causal directed graph Perform a pruning operation and output an optimized causal directed graph. ; The counterfactual causal attribution module is used to optimize the causal directed graph. A counterfactual attribution task is performed using a counterfactual reasoning-based attribution mechanism, and the root cause contribution set C is output. The Attribution Explanation and Multi-Granularity Visualization module is used to generate multi-granularity attribution display results based on the root cause contribution set C combined with a large model based on the Transformer structure.

9. A large-model-driven business indicator anomaly causal attribution device, characterized in that, The large model-driven business indicator anomaly causal attribution device includes: a memory, a processor, and a large model-driven business indicator anomaly causal attribution program stored in the memory and executable on the processor. When the large model-driven business indicator anomaly causal attribution program is executed by the processor, it implements a large model-driven business indicator anomaly causal attribution method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a large model-driven business indicator anomaly causal attribution program, which, when executed by a processor, implements a large model-driven business indicator anomaly causal attribution method according to any one of claims 1 to 7.

Images