An ai agent driven cross-department coordination cost dynamic management and control method and system

Abstract

The application provides an AI Agent driven cross-department collaboration cost dynamic management and control method and system, belonging to the technical field of cost management and control. The method comprises the following steps: receiving standardized data streams obtained by converting preset semantic models based on edge AI Agents deployed in department business systems; inputting the standardized data streams into a global cost calculation model for accounting to obtain cost estimation results; using an unsupervised learning algorithm to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines; when the cost estimation results reach a warning threshold rule, triggering a corresponding warning signal; calling a causal analysis model to perform attribution analysis on the cost anomaly triggering the warning to obtain an attribution result; classifying the cost anomaly into controllable cost anomaly or uncontrollable cost anomaly according to the attribution result; selecting and executing a corresponding hierarchical regulation mechanism for regulation according to the result of the cost anomaly classification to obtain a target management and control result.

Description

Technical Field

[0001] This application relates to the field of cost control technology, and in particular to an AI Agent-driven cross-departmental collaborative dynamic cost control method and system. Background Technology

[0002] In fast-paced business scenarios such as new retail and internet services, the limitations of existing centralized, manually operated, and periodic data processing systems are becoming increasingly apparent, making it difficult to meet the core business needs for real-time cost control. Data managers manually extract and organize heterogeneous data daily or weekly. The data preprocessing stage alone takes several hours or even half a day. Coupled with the time spent on data uploading and central accounting, the entire cost data processing flow has a delay of at least 24 hours. This causes the cost accounting results output by the central accounting platform to always lag behind the actual dynamics of business cost consumption, failing to accurately reflect the current cost operation status. When unexpected cost anomalies occur during business operations, the anomalies can only be detected and alerted in the next data extraction and accounting cycle. During the delayed period, abnormal cost consumption continues to occur, not only leading to a continuous increase in cost losses, but also causing abnormal data to intertwine with normal business data due to the long time span. Subsequent management personnel need to spend a lot of time sorting out the source of the anomalies, further extending the control response cycle. This post-event discovery and post-event handling control model completely deviates from the core needs of fast-paced business scenarios for pre-event warnings and in-event intervention of costs. The root cause of this problem is that the existing system relies on manual periodic data processing, which makes it impossible to identify cost anomalies in real time in the early stages, becoming a core bottleneck restricting the efficiency and effectiveness of enterprise cost control. Summary of the Invention

[0003] To address the aforementioned technical issues, this application provides an AI Agent-driven method and system for dynamic cost control of cross-departmental collaboration.

[0004] A first aspect of this application provides an AI Agent-driven method for dynamic cost control of cross-departmental collaboration, comprising:

[0005] The system receives standardized data streams generated by edge AI agents deployed in various departmental business systems based on a preset semantic model. These standardized data streams originate from local heterogeneous business data and include resource consumption records, job events, and time-related data.

[0006] The standardized data stream is input into the global cost calculation model for verification to obtain the cost estimation result;

[0007] Unsupervised learning algorithms are used to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines.

[0008] Based on the aforementioned cost behavior baseline, statistical process control methods are applied to generate multi-level early warning threshold rules.

[0009] When the cost estimation result reaches the warning threshold rule, the corresponding warning signal is triggered and associated with the cost data segment that triggered the warning signal and the corresponding cost anomaly pattern;

[0010] The causal analysis model is invoked to perform attribution analysis on the cost anomalies that triggered the early warning, and the attribution results are obtained.

[0011] Based on whether the attribution results point to internally controllable business processes or resource factors, cost anomalies are classified as controllable cost anomalies or uncontrollable cost anomalies.

[0012] Based on the results of cost anomaly classification, the corresponding hierarchical control mechanism is selected and implemented for control to obtain the target control results.

[0013] A second aspect of this application provides an AI Agent-driven cross-departmental collaborative cost dynamic management and control system, comprising:

[0014] The data acquisition module is used to receive standardized data streams converted by edge AI Agents deployed in the business systems of various departments based on a preset semantic model. The standardized data streams originate from local heterogeneous business data. The standardized data streams include resource consumption records, job events, and time-related data.

[0015] The cost accounting module is used to input the standardized data stream into the global cost calculation model for accounting and to obtain cost estimation results.

[0016] Anomaly detection module is used to perform cluster analysis and anomaly detection on the cost estimation results using an unsupervised learning algorithm to obtain potential cost anomaly patterns and cost behavior baselines;

[0017] The early warning generation module is used to generate multi-level early warning threshold rules based on the cost behavior baseline and by applying statistical process control methods.

[0018] The early warning execution module is used to trigger a corresponding early warning signal when the cost estimation result reaches the early warning threshold rule, and associate it with the cost data segment that triggered the early warning signal and the corresponding cost anomaly mode;

[0019] The attribution analysis module is used to call the causal analysis model to perform attribution analysis on the cost anomalies that trigger the warning and obtain the attribution results.

[0020] The attribution classification module is used to classify cost anomalies into controllable cost anomalies or uncontrollable cost anomalies based on whether the attribution results point to internally controllable business processes or resource factors.

[0021] The control execution module is used to select and execute the corresponding hierarchical control mechanism based on the results of cost anomaly classification to obtain the target control results.

[0022] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described AI Agent-driven cross-departmental collaborative cost dynamic management method.

[0023] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the AI ​​Agent-driven cross-departmental collaborative cost dynamic management method described above.

[0024] The beneficial effects of the AI ​​Agent-driven cross-departmental collaborative cost dynamic management method and system provided in this application are as follows: This application achieves standardized conversion of heterogeneous business data through an edge AI Agent, ensuring cross-departmental data interoperability. The federated learning framework allows sub-models from various departments to securely aggregate to form a global benchmark, protecting data privacy while ensuring the uniformity of cost accounting. Unsupervised clustering and anomaly detection accurately locate abnormal patterns, statistical process control generates multi-level thresholds, causal analysis and classification mechanisms clarify the controllability of anomalies, and hierarchical regulation addresses problems in a targeted manner. Each technical feature is progressively enhanced, forming a closed loop from data collection, accounting, detection, early warning, attribution to regulation, achieving precise and efficient dynamic cost management, improving the real-time performance and reliability of cross-departmental collaborative management, and avoiding cost waste and management lag. Attached Figure Description

[0025] Figure 1 A flowchart illustrating an AI Agent-driven method for dynamic cost control across departments, provided in an embodiment of this application;

[0026] Figure 2 This is a structural block diagram of an AI Agent-driven cross-departmental collaborative cost dynamic control system provided in an embodiment of this application;

[0027] Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0029] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0030] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an AI Agent-driven method for dynamic cost control across departments, as provided in an embodiment of this application. The method includes:

[0031] S101: Receives standardized data streams transformed by edge AI Agents deployed in various departmental business systems based on preset semantic models. The standardized data streams originate from local heterogeneous business data. The standardized data streams include resource consumption records, job events, and time-related data.

[0032] In this embodiment, a standardized data stream is received, transformed by an edge AI Agent deployed in various departmental business systems based on a preset semantic model. This standardized data stream originates from heterogeneous local business data within each department. The edge AI Agent achieves real-time collection and preprocessing of local data through lightweight deployment, avoiding cross-departmental data transmission delays and privacy risks. These departmental business systems include, for example, financial systems, procurement systems, production systems, human resources systems, and operations and maintenance systems.

[0033] The edge AI Agent has a built-in data acquisition adapter and semantic conversion module. The data acquisition adapter supports the acquisition of multiple formats of structured data (financial vouchers, purchase orders, work hour records), semi-structured data (process approval documents, log files), and unstructured data (contract documents, voice command records).

[0034] The edge AI Agent captures all business data by monitoring changes in business system interfaces or local databases in real time. The semantic conversion module, based on a pre-trained industry-specific semantic model (integrating enterprise business terminology and cost control dimensions), normalizes the collected heterogeneous data, including data field mapping, unit unification, redundant data removal, missing value completion (using interpolation algorithms based on business scenarios), and preliminary filtering of outliers (removing obviously invalid data, such as negative values ​​or extreme values ​​outside the reasonable range), ultimately generating a standardized data stream.

[0035] Standardized data flow specifically includes resource consumption records, operational events, and time-related data; among which, resource consumption records include quantitative resource consumption data such as labor costs (working hours, salary standards, social security and housing provident fund), material costs (purchase unit price, consumption quantity, loss rate), equipment costs (depreciation, operation and maintenance costs, energy consumption), and expense costs (office expenses, travel expenses, service fees).

[0036] The task event data includes the entire lifecycle of cross-departmental collaborative tasks, such as initiation, execution, transfer, and completion, as well as information on the departments, personnel, and resource scheduling associated with the tasks.

[0037] Time-driven data includes the start and end times of each task event, the time distribution of resource consumption, and the turnaround time of cross-departmental processes, providing a time dimension for cost accounting and behavioral analysis.

[0038] S102: Input the standardized data stream into the global cost calculation model for calculation to obtain the cost estimation result.

[0039] In this embodiment, standardized data streams are input into a global cost calculation model for multi-dimensional accounting, resulting in real-time and accurate cost estimation results. The cost estimation results refer to the multi-dimensional, full-cycle cost quantification results output after heterogeneous business data collected by edge AI agents from various departments is converted into standardized data streams, input into a global cost calculation model built on a federated learning framework, and undergoes cross-departmental cost allocation, time-driven correlation accounting, and accuracy verification. The cost estimation results are not a single-dimensional total cost value, but rather include hierarchical accounting data at the departmental, operational, and resource levels, along with auxiliary information such as accounting trajectory, data source, and deviation range, balancing accounting accuracy and traceability. Its generation process uses resource consumption records, operational events, and time-driven data from the standardized data stream as core inputs, relying on the cross-departmental time-driven rate benchmark aggregated by the global cost calculation model to accurately allocate the common costs generated by cross-departmental collaborative operations, solving the problems of ambiguous departmental boundaries and difficulty in tracing drivers in traditional accounting. This result not only characterizes the real-time cost consumption status of each department and operation, but also provides high-quality data samples for subsequent unsupervised learning anomaly detection. At the same time, it serves as the core basis for comparing early warning thresholds and evaluating the control effect, and runs through the entire cost control process.

[0040] The global cost calculation model is trained collaboratively with various edge AI agents through a federated learning framework. Since the business data of each department exhibits non-independent and identically distributed (Non-IID) characteristics, this application adopts a personalized federated learning framework. This framework allows each department's sub-model to be individually adjusted based on the global baseline model, while ensuring data privacy, to adapt to its respective business characteristics. The specific training process is as follows: Each edge AI Agent uses local data to train and maintain its department's operation time driver rate sub-model, and securely aggregates it through federated learning to form a globally unified operation time driver rate benchmark; Each edge AI Agent calculates the parameter gradient of its local operation time driver rate sub-model and performs homomorphic encryption on the parameter gradient; The encrypted parameter gradient is uploaded to the federated learning coordination node, which uses a clustered federated learning algorithm to group departments according to gradient similarity and aggregate different global models for different groups to cope with data heterogeneity; The updated global gradient values ​​after grouping and aggregation are fed back to each edge AI Agent, and each edge AI Agent adjusts its local sub-model parameters based on the updated global gradient values ​​of its group. The above process is repeated until the model converges, and the personalized operation time driver rate benchmark applicable to each department is determined.

[0041] In this embodiment, the core accounting logic of the global cost calculation model is built upon Time-Driven Activity-Based Costing (TD-ABC). Through a federated learning framework, it coordinates the training of various edge AI Agents to obtain a dynamic activity time driver rate benchmark, achieving real-time and accurate tracking of activity-level costs. Its calculation formula can be expressed as:

[0042]

[0043] in, Show resource pool The unit time rate (dynamically updated via federated learning); Indicates task Consume resources The actual time (collected in real time by the edge AI Agent); Indicates material The unit cost; Indicates task Consumable materials The quantity.

[0044] This embodiment utilizes this model to accurately allocate shared costs generated during cross-departmental collaboration to specific tasks and departments based on actual operation time and resource consumption, resolving the issue of unclear cost attribution caused by ambiguous departmental boundaries and difficulties in tracing cost drivers in traditional cost accounting. Compared to traditional post-event centralized accounting, this method leverages an edge AI agent to collect operational events and time-related data in real time, achieving real-time cost awareness—collecting and calculating costs as they occur—providing a highly timely and accurate cost data foundation for subsequent anomaly detection and dynamic control.

[0045] S103: Unsupervised learning algorithms are used to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines.

[0046] In this embodiment, an unsupervised learning algorithm is used to perform cluster analysis and anomaly detection on the cost estimation results, obtaining potential cost anomaly patterns and cost behavior baselines. This eliminates the need for manually labeled samples and can automatically identify unknown anomaly types. The cost behavior baselines are constructed based on normal cost patterns obtained from cluster analysis, combined with the statistical characteristics (mean, variance, median, and trend threshold) of historical cost data. They are divided into department-level baselines, operation-level baselines, and resource-type-level baselines, and these baselines have dynamic update capabilities. Baseline parameters are periodically adjusted based on new cost data and business changes to ensure the baselines are adaptable to actual business scenarios and prevent baseline failure due to business iterations.

[0047] S104: Based on the cost behavior baseline, apply statistical process control methods to generate multi-level early warning threshold rules.

[0048] In this embodiment, based on the cost behavior baseline, a statistical process control method (control chart) is applied, and combined with the enterprise's cost control objectives and risk tolerance, multi-level early warning threshold rules are generated to achieve hierarchical early warning for cost anomalies.

[0049] First, the fluctuation range of normal cost data is analyzed using statistical process control methods. Considering that cost data may not follow a normal distribution, this application uses kernel density estimation (KDE) to fit the probability distribution of cost data, thereby determining the upper control limit, lower control limit, and warning line of cost data. Subsequently, warning levels are divided according to the severity of anomalies into three levels: Level 1 warning (indicative level) indicates that cost data is close to the baseline fluctuation limit, with no obvious anomaly risk, and only the trend needs to be monitored; Level 2 warning (alert level) indicates that cost data exceeds the baseline fluctuation limit but does not break through the control limit, with potential anomaly risk, and preliminary verification needs to be initiated; Level 3 warning (emergency level) indicates that cost data breaks through the control limit, with a high anomaly risk, and intervention procedures need to be initiated immediately.

[0050] As a possible implementation, based on a cost behavior baseline, statistical process control methods are applied to generate multi-level early warning threshold rules. Combining the enterprise's cost control objectives, an adaptive threshold setting method is adopted: for normally distributed cost data, the 3σ principle is used to set the basic threshold; for non-normally distributed cost data, quantile regression is used to estimate different quantiles (e.g., 0.05, 0.95) based on historical data as threshold boundaries. A deviation exceeding the mean ±1σ (or 5% quantile) is set as a Level 1 early warning, exceeding the mean ±2σ (or 2.5% quantile) as a Level 2 early warning, and exceeding the mean ±3σ (or 0.5% quantile) as a Level 3 early warning. Simultaneously, considering the business characteristics of each department (e.g., the cost fluctuation differences between production and administrative departments), the early warning thresholds are dynamically calibrated to form personalized multi-level early warning threshold rules, ensuring the relevance and effectiveness of the early warning signals.

[0051] In the early warning generation mechanism of this embodiment, the 3σ principle is introduced to construct a multi-level cost anomaly early warning model. Based on the historical cost behavior baseline, the standard deviation σ of the cost data is calculated, and the following early warning triggering conditions are set:

[0052]

[0053] in, This is the estimated cost at the current moment; This represents the center value (mean) of the corresponding cost behavior baseline. This is the early warning level coefficient, which can be set to [value] according to control requirements. (Level 1 prompt) (Level 2 Warning) (Level 3 Emergency).

[0054] When real-time cost data deviates from the benchmark value by more than 3σ, the system automatically triggers the highest level warning and links with the Statistical Process Control (SPC) control chart to visualize the cost fluctuation trend, accurately locate the process link, responsible department and related operation where abnormal fluctuations occur, and achieve a rapid closed loop from anomaly detection to responsibility determination.

[0055] S105: When the cost estimation result reaches the early warning threshold rule, the corresponding early warning signal is triggered and associated with the cost data segment that triggered the early warning signal and the corresponding cost anomaly mode.

[0056] In this embodiment, a real-time monitoring mechanism is established to compare each output cost estimation result with multi-level early warning threshold rules in real time. When the cost estimation result reaches the early warning threshold rule, the corresponding early warning signal is triggered. The early warning signal is pushed to the relevant department heads, financial control personnel and enterprise management through multiple channels such as visualization platform, email, SMS, and WeChat, to ensure that relevant personnel respond in a timely manner.

[0057] Simultaneously, it links to the cost data segment that triggered the warning signal, including specific resource consumption details, operational event information, time nodes, departments involved, and related processes, and maps it to the corresponding cost anomaly mode (such as material consumption excess, cross-departmental cost allocation deviation, and time waste). This generates a warning tracing report, clearly presenting the source, composition, and correlation of the abnormal data, providing accurate data support for subsequent attribution analysis, and avoiding blind verification work.

[0058] S106: Call the causal analysis model to perform attribution analysis on the cost anomalies that triggered the warning, and obtain the attribution results.

[0059] In this embodiment, a causal analysis model is invoked to perform in-depth analysis of cost anomalies that trigger warnings. The causal analysis model is built upon a structural causal model (SCM), combining historical business data and domain expert knowledge to establish a causal graph between variables. When a cost anomaly occurs, the model identifies the root cause of the anomaly using causal inference algorithms (such as the backdoor criterion and the frontdoor criterion) and quantifies the impact strength of each cause. Simultaneously, the model outputs an attribution confidence score, indicating the reliability of the attribution result. To improve attribution accuracy, the system encourages business departments to conduct A / B testing or implement targeted interventions after a cost anomaly occurs, and uses the intervention results as new causal data input into the model to continuously iterate and optimize the causal graph. The attribution results are output in the form of triples (causal entity, impact strength, attribution confidence score), and combined with a controllability knowledge graph to label the controllable attributes (controllable / uncontrollable) and responsible parties (corresponding departments / positions) of each causal entity, providing a basis for subsequent classification.

[0060] S107: Based on whether the attribution results point to internally controllable business processes or resource factors, classify cost anomalies into controllable cost anomalies or uncontrollable cost anomalies.

[0061] In this embodiment, cost anomalies are classified into controllable cost anomalies or uncontrollable cost anomalies based on whether the attribution results point to internally controllable business processes or resource factors, providing a basis for the formulation of subsequent control strategies. Controllable cost anomalies refer to anomalies caused by factors within the enterprise that can be improved through process optimization, resource allocation adjustment, strengthened management, and standardized operations, such as wasted working hours, excessive material losses, redundant cross-departmental processes, and idle resources. Uncontrollable cost anomalies refer to anomalies caused by external factors that the enterprise cannot independently intervene in, such as soaring market raw material prices, industry policy adjustments, and supply chain disruptions caused by natural disasters.

[0062] In this embodiment, if the cost anomaly is caused by both controllable and uncontrollable factors during the classification process, the influence ratio of each factor will be clearly defined, and the controllable and uncontrollable parts will be marked separately. At the same time, the core responsible department and the cooperating department will be defined to ensure that the classification results are clear and the responsibilities are well-defined.

[0063] S108: Based on the results of cost anomaly classification, select and implement the corresponding hierarchical control mechanism to achieve the target control results.

[0064] In this embodiment, based on the results of cost anomaly classification, warning level, and attribution analysis conclusions, a corresponding hierarchical control mechanism is selected and executed for precise control, effectively resolving cost anomalies and achieving the target control results. The system incorporates a cross-departmental collaborative workflow engine for initiating, transferring, and responding to control tasks. Once a control task is generated, the workflow engine breaks it down into sub-tasks based on task type, involved departments, and urgency, and assigns them to the corresponding department's business system. Each department's edge AI Agent receives the sub-tasks, drives the local business system to execute them, and provides real-time feedback on the execution status. The workflow engine monitors the entire task execution process. If task delays or execution conflicts occur, it dynamically adjusts the task execution order or resource allocation based on a preset priority scheduling algorithm (e.g., based on urgency, impact scope, resource dependencies, etc.) to ensure efficient completion of the control task. The hierarchical control mechanism is designed with two dimensions: anomaly type and warning level, ensuring the targeted and effective control measures.

[0065] For controllable cost anomalies, control measures are implemented based on the warning level: Level 1 warning (indicative level) requires the corresponding department to independently carry out optimization and adjustments, such as streamlining resource consumption processes, strengthening personnel training, and optimizing local resource allocation, and submitting a rectification report after control measures are implemented; Level 2 warning (alert level) requires the financial control department to work with the corresponding department to develop a special control plan, such as adjusting cost budgets, optimizing cross-departmental collaboration processes, establishing special control indicators, and regularly tracking the progress of control measures; Level 3 warning (emergency level) requires the company's management to take the lead in coordinating resources from multiple departments to activate an emergency control mechanism, such as suspending high-consumption operations, re-planning resource allocation, revising management systems, and establishing a special team to supervise the implementation of control measures to ensure that the anomaly is resolved quickly.

[0066] For uncontrollable cost anomalies, the focus is on risk response and buffer control: Level 1 warning (advice level): establish a dynamic tracking mechanism to monitor the trend of abnormal changes and reserve resources in advance; Level 2 warning (alert level): adjust relevant business plans and budget allocations, optimize cost structure, and reduce the scope of the impact of the anomaly, such as finding alternative suppliers and adjusting product pricing strategies; Level 3 warning (emergency level): activate the enterprise-level risk response plan, coordinate cross-departmental resources to build a risk buffer system, and assess the impact of the anomaly on the overall cost target of the enterprise, and adjust the annual cost control target if necessary.

[0067] After the control measures are implemented, the system tracks and evaluates their effectiveness in real time, comparing cost data, operational efficiency, and business indicators before and after the control measures to determine whether the expected control objectives have been achieved. If the control effects are met, the cost behavior baseline and early warning threshold rules are updated, forming a closed-loop control system. If the control effects are not met, the system returns to the attribution analysis stage to re-examine the causes, optimize the control plan, and execute it again until the target control results are achieved. Simultaneously, the control process, results, and lessons learned are incorporated into the enterprise's cost control knowledge base, providing a reference for handling similar anomalies in the future and continuously improving cost control capabilities.

[0068] As can be seen from the above, this application achieves standardized transformation of heterogeneous business data through an edge AI agent, ensuring cross-departmental data interoperability. The federated learning framework allows sub-models from various departments to securely aggregate into a global benchmark, protecting data privacy while ensuring uniformity in cost accounting. Unsupervised clustering and anomaly detection accurately locate abnormal patterns, statistical process control generates multi-level thresholds, causal analysis and classification mechanisms clarify the controllability of anomalies, and hierarchical regulation addresses problems in a targeted manner. Each technical feature progresses layer by layer, forming a closed loop from data collection, accounting, detection, early warning, attribution to regulation, achieving precise and efficient dynamic cost control, improving the real-time performance and reliability of cross-departmental collaborative management, and avoiding cost waste and management lag.

[0069] In one embodiment of this application, the time-driven data in the standardized data stream generated by the edge AI Agent is obtained by the edge AI Agent monitoring or receiving raw logs from the department's business system, identifying cost-related job events, and estimating the standard man-hours or machine-hours consumed in executing a single job event.

[0070] In this embodiment, the edge AI Agent monitors and receives various types of log data, such as raw operation logs, equipment operation logs, and process flow logs, from various departmental business systems in real time based on preset log collection rules. Through the built-in job event recognition module, combined with natural language processing and rule engine technology, it accurately extracts and identifies job events directly related to cost consumption from massive amounts of raw logs, including but not limited to material procurement and warehousing events, equipment start-up and shutdown events, personnel job operation events, and cross-departmental process approval events.

[0071] For each type of identified work event, the edge AI Agent uses a composite algorithm combining historical data fitting and scenario-based correction to estimate the standard working hours or machine hours consumed by a single work event. First, it retrieves historical working hour data for the same type of work from the local cache and generates basic working hour estimates using algorithms such as linear regression and neural network fitting. Then, it dynamically corrects the estimates by combining the actual conditions of the current work scenario (such as equipment performance status, operator skill level, material characteristics, and process complexity), eliminating outliers in the historical data to ensure the accuracy of the working hour estimates. Simultaneously, the edge AI Agent collects and correlates supplementary information such as the actual start and end times of each work event, the time distribution characteristics of resource consumption, the turnaround time of cross-departmental processes, the duration and cause of work interruptions, etc. It then performs structured integration and redundancy removal on the aforementioned working hour data and time information to form complete and standardized time driver data. This provides accurate and comprehensive time dimension support for subsequent cost accounting by driver-based cost allocation and for identifying time dimension anomalies in behavioral analysis.

[0072] The tiered control mechanism in this embodiment implements differentiated control strategies for different types of cost anomalies:

[0073] For controllable cost anomalies (such as resource waste within the process, excessive procurement, etc.), based on a pre-set business causal knowledge graph, the Pareto optimal resource control algorithm is used to seek the optimal balance among multiple objectives (cost minimization, efficiency maximization, and process interference minimization), generate multiple optimized alternative solutions, and automatically push them to the business system of the responsible department for execution.

[0074] In response to uncontrollable cost anomalies (such as sudden manpower shortages or external price fluctuations), the system initiates a cross-departmental resource allocation mechanism based on real-time resource maps. By intelligently matching resource surpluses with demanding departments, it generates resource allocation plans and initiates inter-departmental negotiation processes with the help of edge AI agents. If the negotiation fails, it follows preset arbitration rules or escalates to management decision-making to ensure rapid resource allocation and avoid cost out-of-control due to external changes.

[0075] As can be seen from the above, this application, by adopting the aforementioned technical solution, enables the edge AI Agent to directly monitor or receive raw logs from departmental operations, accurately identifying cost-related operational events. The core lies in the precise estimation of standard working hours and machine hours for a single operational event, providing quantitative support for time-driven data. This data, as a key component of standardized data flow, accurately reflects the correlation between operations and costs, providing precise input for the global cost calculation model. The technical feature focuses on the quantification of working hours, compensating for the shortcomings of ambiguous time drivers in traditional cost control, enabling cost accounting to be based on actual operational consumption, improving the accuracy of cost estimation results, and laying a precise data foundation for subsequent anomaly detection and control decisions. This application constructs a distributed data platform through the AI ​​Agent, achieving real-time operational cost accounting based on Time-Driven Activity-Based Costing (TD-ABC), establishing a multi-level early warning mechanism based on the 3σ principle and SPC control charts, and combining Pareto optimal control algorithms to achieve intelligent allocation of cross-departmental resources. Each technical step is progressively advanced, forming a closed-loop management system covering cost incurrence, data collection, intelligent analysis, and real-time control. This fundamentally solves the pain points of traditional cost control, such as data lag, ambiguous attribution, and passive control, and significantly improves enterprises' cost control capabilities and operational efficiency in fast-paced, highly collaborative business scenarios.

[0076] In one embodiment of this application, an unsupervised learning algorithm is used to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines, including:

[0077] Construct an enhanced feature space that includes cost values, rates of change, and cost correlations between departments;

[0078] The distribution characteristics of data points in the enhanced feature space are evaluated to obtain evaluation results. Based on the evaluation results, the corresponding analysis strategy is selected and executed to obtain potential cost anomaly patterns and cost behavior baselines. If the data points show periodicity or trend in the time series dimension, and the variance of the cost correlation between departments is less than or equal to the preset first variance threshold, then the first analysis strategy is executed.

[0079] If the data points exhibit high dimensionality and sparsity, or if the variance of cost correlation between departments exceeds the preset first variance threshold, then the second analysis strategy will be executed.

[0080] In this embodiment, an enhanced feature space is first constructed. The enhanced feature space clearly includes three core features: cost value, cost change rate, and cost correlation between departments, which fully covers the cost's own attributes and cross-departmental linkage characteristics.

[0081] Subsequently, the distribution characteristics of data points in the enhanced feature space are evaluated, generating assessment results including temporal distribution characteristics, dimensional sparsity, and variance of cost correlation between departments. Based on the assessment results, corresponding analysis strategies are selected and executed to accurately obtain potential cost anomaly patterns and cost behavior baselines. The preset first variance threshold is an empirical threshold derived from historical cross-departmental cost data statistics, used to define the fluctuation range of cost correlation between departments.

[0082] If the data points exhibit periodicity (monthly or quarterly business cycles) or trend (increasing or decreasing trends) in the time series dimension, and the variance of the cost correlation between departments is less than or equal to the preset first variance threshold, it indicates that the cost linkage between departments is stable and the data distribution pattern is traceable, then the first analysis strategy will be executed.

[0083] If the data points exhibit high dimensional sparsity (i.e., the proportion of most dimensions in the feature space is less than the preset sparsity threshold), or the variance of the cost correlation between departments is greater than the preset first variance threshold, it indicates that the cost linkage between departments fluctuates drastically and the data distribution is complex and without obvious patterns. In this case, the second analysis strategy will be executed.

[0084] As can be seen from the above, this application, by adopting the aforementioned technical solutions, enhances the numerical value of feature space integration cost, rate of change, and departmental correlation. The two correlation calculation methods are adapted to different business scenarios, improving the completeness of feature dimensions. Based on data distribution characteristics, corresponding analysis strategies are selected to adapt to periodic, trend-based, and highly sparse data scenarios.

[0085] In one embodiment of this application, executing a first analysis strategy includes:

[0086] Density clustering algorithm is used to cluster data points in the enhanced feature space. By optimizing the neighborhood radius and minimum sample number parameters, one or more core clusters that represent cost behavior patterns are identified, and data points that do not belong to any core cluster and whose number of points in the neighborhood is less than a preset first point number threshold are identified as isolated outliers.

[0087] For each core cluster, a time-series decomposition model is used to fit the cost behavior baseline and normal fluctuation range of the core cluster.

[0088] Data points that deviate from the cost behavior baseline of their respective core clusters and exceed their normal fluctuation range, along with identified isolated outliers, are collectively identified as potential cost anomaly patterns.

[0089] In this embodiment, the first analysis strategy is executed, specifically including:

[0090] The first step involves using a density-based clustering algorithm to cluster data points in the enhanced feature space. A combination of grid search and cross-validation is used to optimize the neighborhood radius (Eps) and minimum sample size (MinPts) parameters, selecting the optimal parameter combination that best suits the current data distribution. Based on this combination, one or more core clusters representing typical cost behavior patterns are identified. Data points within a core cluster must satisfy the condition that the number of samples in their neighborhood is not less than the optimized minimum sample size and conforms to a temporal distribution pattern. Simultaneously, a preset first threshold (derived from historical normal data neighborhood distribution statistics) is set. Data points that do not belong to any core cluster and whose neighborhood contains fewer than this preset first threshold are directly identified as isolated outliers.

[0091] The second step involves introducing a time series decomposition model (such as the STL decomposition model) for each identified core cluster. This model decomposes the cost time series data within the core cluster into trend terms, periodic terms, and residual terms. By fitting the coordinated change patterns of the trend terms and periodic terms, a cost behavior baseline specific to each core cluster is obtained. At the same time, the normal fluctuation range (such as ±2 standard deviations) corresponding to the baseline is calculated based on the normal distribution characteristics of the residual terms.

[0092] The third step is to verify all data points in the enhanced feature space one by one. Data points that deviate from the cost behavior baseline of their respective core clusters and exceed the corresponding normal fluctuation range are summarized and integrated with the isolated anomalies identified in the first step to jointly determine potential cost anomaly patterns. The anomaly type (intra-cluster deviation / isolated) and the corresponding core cluster association information of each anomaly point are simultaneously labeled.

[0093] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses a density clustering algorithm with optimized parameters to accurately divide core clusters and lock in normal cost behavior patterns, while the rules for identifying isolated outliers quickly filter out extreme anomalies. The time-series decomposition model fits a baseline and fluctuation range for each core cluster, integrating baseline deviation anomalies with isolated outliers to form a complete set of anomaly patterns. Parameter optimization ensures the rationality of clustering, and time-series decomposition aligns with the time-series characteristics of costs; the two work together to accurately identify explicit anomalies and capture implicit baseline deviations. Compared to traditional single-anomaly judgment, this solution improves the comprehensiveness of anomaly identification, avoids missed or false judgments, and makes the cost behavior baseline more closely aligned with actual business operations.

[0094] In one embodiment of this application, executing a second analysis strategy includes:

[0095] A deep anomaly detection model based on reconstruction error is used to detect data points in the enhanced feature space and obtain an anomaly score for each data point.

[0096] Data points with anomaly scores greater than a preset anomaly score threshold are clustered, and the different clusters formed are identified as different potential cost anomaly patterns.

[0097] Data points with outlier scores less than or equal to a preset outlier score threshold are clustered. For each data cluster, a Gaussian process kernel function is selected or combined based on the temporal structure characteristics of the data cluster, and Gaussian process regression is used to fit and obtain the cost behavior baseline.

[0098] In this embodiment, the second analysis strategy is executed, specifically including:

[0099] The first step involves using a deep anomaly detection model based on reconstruction error (preferably an autoencoder series model, such as a stacked autoencoder (SAE) or a variational autoencoder (VAE)) to perform unsupervised anomaly detection on high-dimensional sparse data points in the enhanced feature space. The model first maps high-dimensional features to a low-dimensional latent space using an encoder, and then reconstructs the original features using a decoder. The reconstruction error quantifies the degree of anomaly of the data points, and an anomaly score is calculated for each data point. The higher the anomaly score, the greater the probability that the data point deviates from the normal pattern.

[0100] The second step is to set a preset anomaly score threshold (determined based on historical normal data reconstruction error statistics and enterprise cost control accuracy requirements, which can be optimized and calibrated through ROC curves), filter out data points with anomaly scores greater than the threshold, and use density clustering algorithm to perform cluster analysis on them, aggregating data points with similar anomaly characteristics, identifying the different clusters formed as different potential cost anomaly patterns, and labeling the core anomaly characteristics, involved departments, and data distribution characteristics of each cluster.

[0101] The third step is to cluster the data points whose abnormal scores are less than or equal to the preset abnormal score threshold (i.e., the data points that are judged to be within the normal range) to obtain multiple normal data clusters.

[0102] For each normal data cluster, we first analyze its temporal structure characteristics: if the data cluster exhibits stationary temporal characteristics, we select the quadratic exponential kernel function; if it exhibits periodic temporal characteristics, we select the periodic kernel function; if it has both stationary and periodic characteristics, we combine the quadratic exponential kernel function and the periodic kernel function to construct a Gaussian process kernel function that suits the data characteristics.

[0103] Subsequently, Gaussian process regression is performed based on the kernel function to fit the temporal variation pattern and distribution characteristics of cost data within the data cluster, thereby obtaining the cost behavior baseline corresponding to the data cluster. The baseline includes both the mean trajectory and the confidence interval boundary, adapting to the complex distribution characteristics of high-dimensional sparse data and providing accurate support for the subsequent generation of multi-level early warning threshold rules.

[0104] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses a deep anomaly detection model to generate anomaly scores based on reconstruction errors, accurately quantifying the degree of data anomalies. Cluster analysis categorizes high-anomaly-score data into different anomaly patterns, achieving differentiation of anomaly types. Low-anomaly-score data is fitted to a baseline via Gaussian process regression, with the selection and combination of kernel functions appropriately matching the temporal structure characteristics. The synergy between deep model detection and Gaussian regression fitting not only adapts to high-dimensional sparse data scenarios but also ensures the accuracy of baseline fitting. These technical features solve the challenges of anomaly detection and baseline construction under sparse data, improve the applicability of cost control in complex data scenarios, and make anomaly pattern classification clearer and baselines more valuable for reference.

[0105] In one embodiment of this application, an enhanced feature space is constructed, including cost values, rates of change, and inter-departmental cost correlations, comprising:

[0106] Extract the cost sequence of each department within a preset time window from the standardized data stream;

[0107] The cost sequence is subjected to first-order difference calculation to obtain the instantaneous rate of change sequence;

[0108] The instantaneous rate of change sequence is exponentially smoothed based on a preset time window to generate a smooth rate of change;

[0109] Based on the operational chain described by the global cost calculation model, the impact coefficient of changes in the operational efficiency of upstream departments on the costs of downstream departments is calculated as the cost correlation degree between departments.

[0110] For each time point in the cost series:

[0111] The cost series is synchronously smoothed based on a preset time window to obtain a processed cost value series. The cost value corresponding to the time point is obtained from the processed cost value series.

[0112] Obtain the rate of change value corresponding to that time point from the smoothed rate of change sequence;

[0113] Obtain the inter-departmental cost correlation vector corresponding to a given time point;

[0114] The cost values, smoothed rate of change, and inter-departmental cost correlation vectors are concatenated to form an enhanced feature vector at a given time point.

[0115] The enhanced feature space is constructed based on the enhanced feature vectors at all time points.

[0116] In this embodiment, an enhanced feature space is constructed, which includes cost values, rates of change, and cost correlations between departments. Specifically, this includes:

[0117] The first step is data extraction and time window setting: From the standardized data stream generated in step 1, extract the cost sequence of each department within the preset time window by department dimension. The preset time window is set based on the characteristics of the enterprise's business cycle (such as daily, weekly or monthly, which can be dynamically adjusted according to the control precision requirements). The core of the cost sequence includes the resource consumption summary data at each moment within the corresponding time window to ensure the integrity of the data time sequence.

[0118] The second step is the generation of instantaneous rate of change sequences: First-order difference calculations are performed on the extracted cost sequences of each department. The instantaneous cost changes are quantified by the difference between cost values ​​at adjacent time points, resulting in the instantaneous rate of change sequences for each department, accurately capturing the short-term fluctuation characteristics of costs. The third step is the generation of smoothed rate of change sequences: To reduce the impact of instantaneous fluctuation noise on feature quality, exponential smoothing processing (preferably exponentially weighted moving average, EWMA) is performed on the instantaneous rate of change sequences based on the aforementioned preset time window. By setting a smoothing coefficient (calibrated based on historical data fluctuation amplitude, with a value range of 0-1), the interference of abnormal instantaneous values ​​is weakened, generating smoothed rates of change that better reflect the true trend of cost changes.

[0119] The fourth step is to calculate the inter-departmental cost correlation: Based on the cross-departmental operational links described by the global cost calculation model, the upstream and downstream linkages of each department's operations are broken down, and a departmental operational influence matrix is ​​constructed. Regression analysis is used to quantify the impact of changes in upstream departmental operational efficiency on downstream departmental costs, and the corresponding influence coefficient is calculated. This influence coefficient is used as the inter-departmental cost correlation, accurately representing the transmission effect of cross-departmental operational linkages on costs, and forming an inter-departmental cost correlation vector at each time point.

[0120] Step 5, Time-by-Time Enhancement Feature Vector Construction: For each time point in the cost sequence, perform the following operations: First, perform synchronous smoothing on the cost sequence based on the aforementioned preset time window (consistent with the rate smoothing method to ensure unified data processing logic), obtaining a processed cost value sequence, and extracting the cost value corresponding to that time point; second, match and obtain the rate of change value corresponding to that time point from the smoothed rate of change sequence generated in Step 3; third, retrieve the inter-department cost correlation vector corresponding to that time point generated in Step 4; fourth, concatenate the obtained cost value, smoothed rate of change, and inter-department cost correlation vector according to a preset dimension order to form the enhanced feature vector for that time point.

[0121] Step 6: Enhanced feature space construction: Summarize the enhanced feature vectors from all time points to construct a complete enhanced feature space. This feature space simultaneously covers the static cost values, dynamic trends, and cross-departmental correlation characteristics, providing high-quality feature support for subsequent unsupervised learning analysis.

[0122] As can be seen from the above, by adopting the aforementioned technical solutions, this application eliminates data noise interference and improves the stability of the rate of change by differentiating and smoothing the cost sequence. The inter-departmental correlation is calculated based on the operational chain, accurately representing the cross-departmental cost impact relationship. The concatenation of feature vectors at each time point integrates multi-dimensional information to construct a comprehensive and accurate enhanced feature space. Data preprocessing ensures feature quality, correlation calculation aligns with business logic, and feature concatenation achieves multi-dimensional information fusion. These three aspects work together to solve the problem of the one-sidedness of traditional feature construction, providing high-quality input for subsequent anomaly detection algorithms, improving the overall accuracy of anomaly identification and baseline fitting, and strengthening the data support capability for cost control.

[0123] In one embodiment of this application, based on the temporal structure characteristics of the data cluster, a Gaussian process kernel function is selected or combined, and Gaussian process regression is used to fit a cost behavior baseline, including:

[0124] Select or combine Gaussian process kernel functions based on the temporal structure characteristics of the data cluster;

[0125] Using the selected kernel function, a Gaussian process model is trained based on historical data of the data cluster to obtain the posterior distribution of the predicted cost value;

[0126] The mean function of the posterior distribution is used as the central trajectory of the cost behavior baseline, and the envelope formed by adding or subtracting twice the standard deviation function to the mean function is used as the fluctuation boundary of the cost behavior baseline.

[0127] In this embodiment, for each normal data cluster, the cost behavior baseline is fitted according to the following process:

[0128] First, based on the temporal structure characteristics of the data cluster, select or combine Gaussian process kernel functions: if the data cluster exhibits stationary temporal characteristics (no obvious trend or period, stable fluctuation amplitude), select the quadratic exponential kernel function to adapt to the smooth and continuous characteristics of the data; if it exhibits periodic temporal characteristics (repeated fluctuations at fixed time intervals), select the periodic kernel function to accurately capture the periodic change pattern; if it has both stationary and periodic characteristics, linearly combine the quadratic exponential kernel function and the periodic kernel function to take into account both the smoothness and periodicity adaptation requirements of the data.

[0129] Secondly, using the selected Gaussian process kernel function, a Gaussian process model is trained based on the historical cost time series data of the data cluster. The covariance relationship between data points is defined through the kernel function, and a complete Gaussian process regression framework is constructed. The posterior distribution of the predicted cost value is obtained by solving the problem. This posterior distribution can fully characterize the uncertainty range of the predicted value.

[0130] Finally, a cost behavior baseline is constructed based on the posterior distribution: the mean function of the posterior distribution is used as the central trajectory of the cost behavior baseline, which accurately reflects the temporal trend of normal cost changes; the envelope formed by adding and subtracting two standard deviations from the mean function is used as the fluctuation boundary of the cost behavior baseline. This boundary can cover most normal cost fluctuation scenarios (confidence level of about 95.45%), taking into account both the stability and flexibility of the baseline, and providing accurate baseline support for the subsequent generation of multi-level early warning threshold rules.

[0131] As can be seen from the above, this application, by adopting the aforementioned technical solution, ensures that the selection and combination of kernel functions are well-suited to the temporal characteristics of the data cluster, thus guaranteeing that the Gaussian process model closely reflects the cost change patterns. The trained model outputs a predicted posterior distribution of costs, with the mean function serving as the baseline center trajectory, and the mean plus or minus twice the standard deviation forming the fluctuation boundary. Kernel function adaptation ensures model fitting accuracy, and the posterior distribution provides both the baseline core trajectory and a clear normal fluctuation range. These technical features allow the cost behavior baseline to not only have a central reference but also a dynamic fluctuation range, adapting to the temporal characteristics of cost changes, avoiding the limitations of a fixed baseline, improving the baseline's adaptability to actual cost behavior, and providing a scientific basis for setting early warning thresholds.

[0132] In one embodiment of this application, selecting or combining Gaussian process kernel functions based on the temporal structure characteristics of the data cluster includes:

[0133] The temporal patterns of the data clusters are encoded by a variational autoencoder to obtain the latent spatial feature vectors.

[0134] The latent space feature vectors are input into the kernel function selection network to obtain the combined weights:

[0135] Based on the latent space feature vectors, multiple preset Gaussian process kernel functions are selected or combined to obtain a combined kernel function;

[0136] Gaussian process kernel functions include: radial basis function kernel, periodic kernel, Matern kernel, and rational quadratic kernel;

[0137] The kernel function is obtained by weighted summation of the combined weights and the combined kernel function.

[0138] In this embodiment, a Gaussian process kernel function is selected or combined based on the temporal structure characteristics of the data cluster, as follows:

[0139] The first step is to use a variational autoencoder (VAE) to encode the temporal cost data of the data cluster. The VAE first performs serialization reconstruction preprocessing on the temporal data to eliminate data noise interference. Then, the encoder network maps the temporal pattern to a low-dimensional latent space and outputs a latent space feature vector that represents the core features of the temporal structure, accurately capturing key characteristics such as the stationarity, periodicity, and fluctuation intensity of the data cluster.

[0140] The second step is to input the latent space feature vector into a preset kernel function selection network (preferably a lightweight fully connected neural network). The network learns the mapping relationship through training and outputs the combined weights of each Gaussian process kernel function. The sum of the combined weights is normalized to 1 to ensure that the weight allocation is reasonable.

[0141] The third step is to select a single kernel function or combine multiple kernel functions from the preset Gaussian process kernel function set based on the temporal structure characteristics represented by the latent space feature vectors, so as to obtain the initial combined kernel function. The preset Gaussian process kernel function set includes radial basis function kernel (RBF kernel, which is suitable for smooth and stable temporal features), periodic kernel (suitable for periodic temporal features), Matern kernel (suitable for non-smooth and noisy temporal features), and rational quadratic kernel (suitable for temporal features with gradually changing fluctuation amplitude).

[0142] The fourth step is to perform a weighted summation of the combined weights of the network output and the corresponding initial combined kernel function to obtain the final kernel function that adapts to the temporal structure characteristics of the current data cluster, taking into account both the adaptability of temporal features and the model fitting accuracy.

[0143] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses a variational autoencoder to encode temporal patterns and extract latent features from data clusters, providing a precise basis for kernel function selection. The kernel function selection network outputs combined weights to achieve a reasonable fusion of multiple kernel functions, adapting to complex temporal features. Multiple preset kernel functions cover different temporal scenarios, and weighted summation forms the optimal kernel function. Encoding extracts core features, the network optimizes weight allocation, and multi-kernel fusion adapts to complex scenarios; these three aspects work together to solve the problem of insufficient adaptation by a single kernel function. The optimized kernel function improves the fitting accuracy of the Gaussian process model, making the cost baseline and prediction more accurate, while enhancing the model's adaptability to different temporal patterns and strengthening the flexibility of cost control.

[0144] In one embodiment of this application, an AI Agent-driven method for dynamic management and control of cross-departmental collaborative costs further includes:

[0145] Based on the posterior distribution, the prediction uncertainty for each prediction point is calculated, and the temporal correlation of the prediction uncertainty is evaluated.

[0146] When the prediction uncertainty at multiple consecutive time points exceeds a preset first threshold, a baseline reconstruction signal is triggered.

[0147] In response to the baseline reconstruction signal, the sensitivity of the warning threshold rule is adjusted, wherein the warning threshold fluctuates within the range of the mean function plus or minus K times the standard deviation function, where the K value is adjusted according to the level of prediction uncertainty, the higher the uncertainty, the larger the K value.

[0148] In this embodiment, based on the posterior distribution output by the Gaussian process model, the standard deviation corresponding to each prediction point is extracted and used as a quantitative indicator of the prediction uncertainty for that prediction point. The larger the standard deviation, the higher the uncertainty and the lower the reliability of the prediction result. Simultaneously, autocorrelation analysis is used to assess the temporal correlation of the prediction uncertainty. By calculating the autocorrelation coefficient of the prediction uncertainty at consecutive time points, it is determined whether the uncertainty fluctuations exhibit temporal clustering characteristics, providing a basis for subsequent threshold determination.

[0149] Set a preset first threshold (based on the enterprise cost control fault tolerance rate and historical forecast data statistical calibration), and monitor the forecast uncertainty of consecutive time points in real time. When the forecast uncertainty of multiple consecutive time points (the consecutive number threshold is set according to the business cycle, such as 3 daily time points or 1 weekly time point) is greater than the preset first threshold, it indicates that the current cost behavior baseline is no longer able to adapt to the data fluctuation characteristics, and immediately trigger the baseline reconstruction signal.

[0150] In response to baseline reconstruction signals, the sensitivity of the early warning threshold rules is dynamically adjusted to avoid false alarms or missed alarms due to excessive prediction uncertainty. The adjustment method involves floating the early warning threshold within the range of the mean function plus or minus K times the standard deviation function, where K is a dynamic adjustment coefficient. Its value is strictly adapted to the level of prediction uncertainty; the higher the prediction uncertainty, the larger the K value, expanding the early warning threshold range to reduce the risk of false alarms; the lower the prediction uncertainty, the smaller the K value, narrowing the early warning threshold range to improve early warning sensitivity. The specific range of K values ​​is iteratively calibrated based on historical early warning accuracy, ensuring that the adjusted early warning rules adapt to the current data fluctuation state.

[0151] As can be seen from the above, this application, by adopting the aforementioned technical solution, accurately captures the model's predicted risks through uncertainty calculation and time correlation assessment. The baseline reconstruction signal triggering mechanism promptly addresses high-uncertainty scenarios, and the K value is dynamically adjusted based on uncertainty to optimize the sensitivity of the warning threshold. Increasing the K value under high uncertainty reduces the false alarm rate; decreasing the K value under low uncertainty improves the timeliness of warnings. The synergy between uncertainty assessment and dynamic K value adjustment ensures that the warning threshold adapts to the predicted risk level, avoiding false alarms and missed alarms caused by fixed thresholds in high-uncertainty scenarios, improving the flexibility and accuracy of the warning mechanism, and ensuring the reliability of cost anomaly warnings.

[0152] In one embodiment of this application, the K value is adjusted according to the level of prediction uncertainty, including:

[0153] Obtain the standard deviation of the posterior distribution of the Gaussian process at the current prediction time point as a measure of prediction uncertainty;

[0154] The uncertainty level is obtained by normalizing the forecast uncertainty measure for all time points within the current forecast window.

[0155] Based on the departmental business attribute weights and cost type sensitivity coefficients associated with the cost behavior baseline, the uncertainty level is weighted and adjusted to obtain the adjusted uncertainty level.

[0156] The modified uncertainty level is mapped to the base K value using a pre-defined monotonically decreasing function. The higher the uncertainty level, the smaller the mapped base K value.

[0157] Multiply the base K value by the preset base sensitivity coefficient and truncate the upper and lower limits to obtain the final K value used to calculate the warning threshold.

[0158] In this embodiment, the K value is adjusted according to the level of prediction uncertainty, specifically including:

[0159] The first step is to obtain the standard deviation of the posterior distribution of the Gaussian process at the current prediction time point, and directly use it as the core metric for the prediction uncertainty at that time point, thus preserving the original uncertainty quantification characteristics.

[0160] The second step is to normalize the prediction uncertainty metric values ​​for all time points within the current prediction window (using the min-max normalization method to map the values ​​to the [0,1] interval) to eliminate the influence of dimensional differences and obtain the standardized uncertainty level, where 0 corresponds to the lowest uncertainty and 1 corresponds to the highest uncertainty.

[0161] The third step involves a weighted adjustment of the normalized uncertainty level based on the departmental business attribute weights and cost type sensitivity coefficients associated with the current cost behavior baseline. Departmental business attribute weights are preset according to departmental coreity and business impact (core business departments have higher weights than auxiliary departments), and cost type sensitivity coefficients are set according to cost control priorities (e.g., R&D costs and core production consumable costs have higher sensitivity coefficients than general administrative costs). A correction coefficient is obtained by weighted summing of these two factors. The normalized uncertainty level is then multiplied by this correction coefficient to obtain the adjusted uncertainty level, ensuring that the K-value adjustment aligns with actual business needs.

[0162] The fourth step involves mapping the corrected uncertainty level to a base K value using a preset monotonically decreasing function (preferably an inverse proportional function or an exponential decreasing function, with parameters calibrated using historical data). The higher the corrected uncertainty level, the smaller the mapped base K value, thus initially achieving the inverse adaptation between uncertainty and K value.

[0163] The fifth step involves multiplying the base K value by a preset base sensitivity coefficient (set based on the enterprise's early warning accuracy requirements, with a default value of 1.0) to obtain the initial K value. Simultaneously, upper and lower limits for the K value are set (lower limit 1.5, upper limit 3.0, covering most control scenarios), truncating the initial K value to ensure it remains within a reasonable and effective range. This results in the final K value used to calculate the early warning threshold. The overall adjustment logic for the K value can be iteratively calibrated using historical early warning data to further improve adaptability.

[0164] As can be seen from the above, this application, by adopting the aforementioned technical solutions, eliminates the influence of dimensions through uncertainty measurement normalization, and weighted correction combines business attributes and cost sensitivity, making the uncertainty level more aligned with actual control needs. A monotonically decreasing function maps the basic K value, and upper and lower limit truncation ensures the rationality of the K value. Normalization ensures a unified evaluation standard, weighted correction improves the accuracy of uncertainty assessment, and function mapping and truncation optimize the adaptability of the K value. These technical features enable the scientific and dynamic adjustment of the K value, deeply binding the early warning threshold with predicted risks and business needs, avoiding false alarms and missed alarms, ensuring the timeliness of early warning response, and further optimizing the adaptability and reliability of the early warning mechanism.

[0165] In one embodiment of this application, based on a cost behavior baseline, a statistical process control method is applied to generate multi-level early warning threshold rules, including:

[0166] For each cost behavior baseline, a hidden Markov model is constructed to simulate the transition probability of cost states between normal, slightly abnormal, and severely abnormal.

[0167] Based on the temporal characteristics of the hidden Markov model and cost estimation results, the probability distribution of each hidden state is calculated.

[0168] Using the probability distribution as input, a pre-trained deep Q-network outputs the optimal combination of multi-level warning thresholds for the current situation. The deep Q-network is trained with the reward function being the weighted combined cost of minimizing false alarms, false alarms, and warning response delay.

[0169] The generated multi-level early warning threshold rules are correlated and verified with the historical attribution results of the causal analysis model. If a new cost anomaly pattern is found to trigger an early warning, but there is no similar attribution in the historical attribution database, the early warning rule is automatically marked as an observation rule and its trigger level is temporarily reduced. At the same time, an active learning process is started to collect data under this pattern.

[0170] In this embodiment, a separate Hidden Markov Model (HMM) is constructed for each cost behavior baseline. This HMM defines the cost operation state as three hidden states: normal, slightly abnormal, and severely abnormal. Based on the central trajectory and fluctuation boundaries of the cost behavior baseline, the initial probabilities, state transition probability matrices, and observation probability matrices for each state are initialized. The state transition probabilities are obtained by analyzing the switching frequency statistics of different states in historical cost data, and are used to accurately simulate the dynamic transition patterns of cost states between normal, slightly abnormal, and severely abnormal, adapting to the time-series fluctuation characteristics of costs.

[0171] The second step involves organizing the cost estimation results by time series, extracting their temporal features, and inputting them into the hidden Markov model corresponding to the cost behavior baseline. A forward-backward algorithm is then used to calculate the probability distribution of the cost estimation results at different hidden states at each time point, quantifying the uncertainty and anomaly tendency of the current cost operation status.

[0172] The hidden state probability distribution calculated above is used as the core input and fed into a pre-trained deep Q-network. The network outputs the optimal combination of multi-level warning thresholds for the current situation. The deep Q-network is trained with the reward function of minimizing the weighted comprehensive cost of false alarm rate, false negative rate, and warning response delay. The weights are set according to the enterprise's cost control priority (false alarm weight is higher than response delay weight in core business scenarios). Through iterative training, the network learns the optimal threshold configuration strategy under different probability distribution scenarios, so that the multi-level warning thresholds take into account both warning accuracy and timeliness.

[0173] The generated multi-level early warning threshold rules are correlated and verified with the historical attribution results of the causal analysis model to construct a mapping relationship between threshold rules, abnormal patterns, and attribution results. If the verification finds a new cost anomaly pattern that triggers an early warning, but there are no similar attribution cases in the historical attribution database, it indicates that the attribution experience for the corresponding abnormal pattern of the rule is insufficient. The early warning rule is automatically marked as an observation rule, and its trigger level is temporarily reduced (e.g., mild anomaly is reduced to warning, moderate anomaly is reduced to mild) to avoid misjudgment interfering with normal business operations. At the same time, an active learning process is initiated to continuously collect cost data, business scenario information, and subsequent handling results under this abnormal pattern, supplementing the attribution database and providing data support for rule optimization.

[0174] The resulting multi-level early warning threshold rule system is divided into three levels according to the severity of the anomaly: mild anomaly, moderate anomaly, and severe anomaly. Each level corresponds to a clear threshold range, triggering conditions, and response time requirements. Furthermore, it can be continuously and dynamically iterated based on proactive learning data and control feedback results to continuously improve its adaptability.

[0175] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses a Hidden Markov Model to simulate cost state transitions, accurately capturing the patterns of cost state changes. A deep Q-network outputs the optimal threshold combination, balancing false positives, false negatives, and response delays. An association verification mechanism identifies new anomaly patterns and actively learns and collects data to optimize rules. Multiple models collaboratively generate thresholds, and the verification and learning mechanism optimizes rules, ensuring both the rationality of the current thresholds and enabling dynamic rule iteration. These technical features address the limitations of traditional fixed thresholds, enabling dynamic optimization of multi-level early warning thresholds, improving the adaptability and accuracy of the early warning mechanism, and simultaneously providing data accumulation for the management of new anomaly patterns, strengthening the long-term effectiveness of early warning rules.

[0176] In one embodiment of this application, when the cost estimation result reaches a warning threshold rule, a corresponding warning signal is triggered, and this is associated with the cost data segment that triggered the warning signal and the corresponding cost anomaly pattern, including:

[0177] At the moment the warning signal is triggered, a dynamic data capture instruction is generated, which propagates backward along the data tracing chain of the edge AI Agent;

[0178] Each relevant edge AI Agent uploads, according to instructions, the current standardized data stream segment that triggers the alert, as well as the preceding event and status data stream segments within a preset causal time window. The causal time window is determined based on the operation link topology.

[0179] The central system will reconstruct the collected cross-departmental and cross-time data stream fragments into snapshots of the business status before and after the warning time, based on business process logic and timestamps.

[0180] Extract the feature vector from the business status snapshot image and perform similarity matching with the features of attributed abnormal cases in the historical case library. Associate the historical cost anomaly pattern with the highest matching degree and its root cause chain with this warning.

[0181] In this embodiment, at the moment the warning signal is triggered, the global management platform automatically generates a dynamic data capture instruction. This instruction carries core information such as the warning level, trigger threshold, and associated cost baseline, and propagates backward along the data traceability chain preset by the edge AI Agent to accurately reach the edge AI Agents of various departments related to the warning, ensuring the targeting and comprehensiveness of data collection.

[0182] Each edge AI Agent receiving the instruction synchronously uploads two types of data stream fragments as required by the instruction: the current standardized data stream fragment that triggers the alert and the preceding event and status data stream fragments within a preset causal time window. The current standardized data stream fragment that triggers the alert includes key details such as resource consumption records, work events, and time-related data at the corresponding time point. The preceding event and status data stream fragments within the preset causal time window are determined based on the work link topology described by the global cost calculation model, and the duration is calibrated by factors such as work execution cycle and cross-departmental linkage delay, fully covering the business pre-state and event link before the alert occurs.

[0183] After collecting all relevant cross-departmental and cross-time data stream fragments uploaded by edge AI agents, the data fragments are first timestamped and their integrity verified to remove invalid and duplicate data. Then, based on the enterprise's preset business process logic and data timestamp order, time-series splicing and association mapping algorithms are used to reconstruct the discrete data stream fragments into a complete snapshot of the business status before and after the warning time, clearly restoring the cross-departmental operational linkage status, resource allocation, and cost change trajectory at and before the warning occurs.

[0184] Through feature extraction algorithms, core feature vectors are extracted from the reconstructed business status snapshot. Feature dimensions include key indicators such as cost fluctuation amplitude, cross-departmental correlation strength, operational efficiency, and resource consumption ratio. These feature vectors are then compared with the features of attributable anomaly cases in the historical case library using cosine similarity algorithm to calculate the matching degree, selecting the historical cost anomaly patterns with the highest matching degree. Simultaneously, the root cause chain (core triggers, impact paths, responsible parties, etc.) corresponding to this historical pattern is synchronously associated with this alert, providing preliminary reference for subsequent in-depth attribution analysis and improving attribution efficiency.

[0185] As can be seen from the above, this application, by adopting the aforementioned technical solution, dynamically captures data commands for reverse tracing, collects cross-departmental and cross-time data stream fragments, and reconstructs and restores the early warning scenario using business status snapshots. Feature matching associates historical anomaly patterns with root cause chains, quickly locating the source of the anomaly. The tracing mechanism ensures data integrity, snapshot reconstruction restores the real scenario, and feature matching accelerates root cause location. These three elements work together to achieve accurate association between early warnings and anomaly data and patterns, solving the problem that traditional early warnings only indicate anomalies and cannot quickly trace the source, improving the efficiency of anomaly handling, providing complete scenario support for subsequent attribution analysis, and strengthening the closed-loop capability of cost control.

[0186] In one embodiment of this application, cost anomalies are classified into controllable cost anomalies or uncontrollable cost anomalies based on whether the attribution result points to internally controllable business processes or resource factors, including:

[0187] Establish and maintain a dynamic controllability knowledge graph, which defines all business process entities, resource entities and their attributes within the enterprise in an ontological form, and labels them with controllability tags (such as fully controllable, conditionally controllable, and uncontrollable) and responsible entities; among them, entities related to operational efficiency and internal resource pricing are labeled as conditionally controllable by default.

[0188] The attribution results output by the causal analysis model are formalized as a series of triples (causal entity, influence strength, attribution confidence).

[0189] The causal entity in each triple is linked to the controllability knowledge graph, and reasoning is performed along the relational edges in the graph in a finite step length to determine whether the cause can be intervened or offset through internal enterprise actions (such as process adjustment and resource reallocation); the reasoning process needs to be comprehensively evaluated by combining the attribution confidence and the confidence of the nodes in the knowledge graph.

[0190] Calculate the weighted sum of the influence strength of all causal entities that can be linked to controllable or conditionally controllable nodes (the weights combine influence strength and confidence). If the weighted sum exceeds a preset threshold for the total influence strength, it is classified as an anomaly of controllable costs; otherwise, it is classified as an anomaly of uncontrollable costs.

[0191] For nodes with controllable conditions, the system will simultaneously generate a list of conditions that must be met to make them controllable, and assess the feasibility of meeting the conditions.

[0192] After each classification decision, the attribution result and classification result are used as feedback to update the labels and confidence of relevant nodes in the controllability knowledge graph.

[0193] In this embodiment, a controllability knowledge graph is established and dynamically maintained. This knowledge graph uses an ontological approach to structurally define all business process entities, resource entities, and their core attributes within the enterprise, covering dimensions such as process stages, resource types, responsibility boundaries, and control permissions. Each entity is labeled with a controllability tag, specifically categorized into three types: fully controllable, conditionally controllable, and uncontrollable. Entities related to operational efficiency and internal resource pricing are labeled as conditionally controllable by default, clarifying their control prerequisites. Each entity is synchronously associated with its corresponding responsible entity (department / position), forming a complete association link between entity, attribute, controllability, and responsible entity. The knowledge graph employs an incremental update mechanism, periodically optimizing entity definitions and associations based on business adjustments and control experience to ensure adaptation to dynamic changes in the enterprise's business.

[0194] This embodiment formalizes the attribution results output by the causal analysis model into a series of triple structures (causal entity, influence strength, attribution confidence). The causal entity corresponds to the standardized entity name in the controllability knowledge graph, the influence strength quantifies the proportion of the entity's contribution to cost anomalies (range 0-1), and the attribution confidence characterizes the reliability of the attribution conclusion (range 0-1), providing standardized input for subsequent reasoning and weighted calculation.

[0195] This embodiment precisely links and matches the causal entity in each triple with the controllability knowledge graph. Finite-step reasoning (preset to 2-3 steps to balance reasoning depth and efficiency) is performed along the semantic relationship edges (including relationships, associations, and influence relationships) in the graph to determine whether the causal entity can be intervened or offset through internal enterprise actions (process adjustments, resource reallocation, operational standard optimization, enhanced access control, etc.). During the reasoning process, a weighted fusion algorithm is used to comprehensively evaluate the attribution confidence and the confidence of the corresponding node in the knowledge graph, eliminating low-confidence associations to ensure the accuracy of the controllability judgment.

[0196] Calculate the weighted sum of the influence strength of all causal entities that can be linked to controllable or conditionally controllable nodes. The weight is composed of the product of the entity's influence strength and attribution confidence, ensuring that entities with high influence and high confidence dominate the classification. Set a preset percentage threshold (calibrated based on enterprise cost control targets and historical data, with a default value of 70%). If the above weighted sum exceeds the preset percentage threshold of total influence strength, it indicates that the core cost anomaly is driven by internal controllable factors, and it is classified as a controllable cost anomaly; if the weighted sum is lower than or equal to the preset percentage threshold, it is classified as an uncontrollable cost anomaly.

[0197] For the cause entity linked to the conditionally controllable node, the system automatically generates a list of conditions that must be met to transform it from conditionally controllable to fully controllable. The conditions include resource allocation requirements, process optimization directions, replacement of responsible entities, and improvement of control mechanisms. At the same time, the system assesses the feasibility of meeting these conditions by combining factors such as the enterprise's existing resource capabilities, business cycle, and implementation costs, and outputs a feasibility score and implementation priority suggestions to support the formulation of subsequent control measures.

[0198] After each classification decision is completed, the attribution results, classification results, and preliminary feedback on the handling are used as incremental data and fed back to the controllability knowledge graph. The controllability labels (e.g., nodes deemed conditionally controllable are updated to "fully controllable" after verification of stable management) and confidence levels are dynamically updated, continuously optimizing the accuracy of the knowledge graph. Simultaneously, for controllable cost anomalies, the weighted sum of the impact strength of controllable cause entities is calculated. Based on the anomaly's spread speed and its impact on the company's overall cost objectives, three handling priorities—high, medium, and low—are assigned. Uncontrollable cost anomalies are prioritized according to their impact scope and risk level, ensuring that core anomalies receive priority response and handling.

[0199] As can be seen from the above, this application, by adopting the aforementioned technical solution, clarifies the controllability of business entities and the responsible parties through a controllability knowledge graph. The attribution result triples are linked to the graph, enabling the judgment of the controllability of causal entities. Weighted sum calculations and proportional thresholds scientifically classify controllable and uncontrollable anomalies, and a feedback mechanism updates the graph node information. The graph provides the basis for judgment, linking reasoning clarifies controllability, and the feedback mechanism optimizes the graph's accuracy. These technical features solve the ambiguity in anomaly controllability determination, achieving precise anomaly classification and providing a clear basis for subsequent tiered control. Simultaneously, through iterative optimization of the graph, classification accuracy is improved, strengthening the targeted nature of cost control.

[0200] In one embodiment of this application, based on the anomaly classification result, a corresponding hierarchical control mechanism is selected and executed for control to obtain the target control result, including:

[0201] If the abnormality is a controllable cost, an optimized alternative solution, including cost and efficiency trade-offs, is generated using a multi-objective optimization algorithm based on a pre-set business causal knowledge graph. This alternative solution is then pushed to the corresponding responsible department's business system to drive its execution. The system monitors the execution status of the solution and feeds back the execution results and cost changes to the global cost calculation model and causal analysis model for learning.

[0202] If the cost is an uncontrollable anomaly, the resource map is used to match the resource needs and surplus status of different departments, generate a resource allocation plan, and initiate a cross-departmental negotiation and confirmation process through the edge AI Agent of the relevant departments. After reaching a consensus within a preset time, the allocation is executed.

[0203] If no consensus is reached, the matter will be escalated to the pre-established arbitration rules or management decision-making process.

[0204] For all implemented control actions, the system continuously tracks changes in cost estimation results, evaluates the control effects, and stores the control cases and their effects in the historical case database for use in optimizing future early warning, attribution, and control decisions.

[0205] In this embodiment, firstly, the controllable cost anomaly control process: If a cost anomaly is classified as a controllable cost anomaly, the system, based on a preset business causal knowledge graph, accurately locates the business processes, resource allocation nodes, and responsible entities associated with the anomaly, clarifying the core direction of control and optimization. Subsequently, a multi-objective optimization algorithm (preferably the non-dominated sorting genetic algorithm NSGA-Ⅲ) is used, with cost minimization, business efficiency maximization, and process interference minimization as core objectives, to generate multiple optimized alternatives that include cost and efficiency trade-offs. The solutions include streamlined process paths, resource reallocation schemes, and operational specification optimization details—all actionable measures. The optimized alternatives are pushed to the business systems of the corresponding responsible departments, simultaneously generating execution guidelines and timeline requirements to drive implementation. During execution, the edge AI Agent monitors the execution status of the solutions in real time (including execution progress, resource input, process compliance, etc.), periodically collects execution data and cost changes, and feeds this data back to the global cost calculation model and causal analysis model for incremental learning, correcting model parameters to improve subsequent accounting and attribution accuracy.

[0206] If cost anomalies are classified as uncontrollable cost anomalies, the system first correlates with the enterprise resource map to comprehensively analyze the resource balance status, resource allocation scope, and resource usage constraints of each department, accurately matching cross-departmental resource needs with available resources. Based on the matching results, multiple resource allocation plans are generated, clearly defining the type, quantity, flow path, and responsibilities of each department for the allocated resources. Subsequently, a cross-departmental negotiation and confirmation process is initiated through the edge AI Agent of the relevant departments, pushing the resource allocation plans to all related departments and allowing departments to propose adjustments based on actual business needs. A negotiation timeframe is set (preset based on the urgency of the resources, such as 24 hours / 3 working days). Once all departments reach a consensus within the preset time, the resource allocation execution process is immediately initiated, with the edge AI Agent tracking the resource flow status throughout to ensure the plan is implemented.

[0207] If the relevant departments fail to reach a consensus on the resource allocation plan within the preset negotiation period, an escalation mechanism will be triggered. The disputed matters and the opinions of each department will be summarized and organized, and a preliminary arbitration recommendation will be generated according to preset arbitration rules (such as arbitration logic based on departmental coreness, resource ownership, and business priority). Alternatively, it may be directly submitted to the management decision-making process, whereby management will issue a final decision. Once the decision is finalized, it will be synchronized with the relevant departments to drive the implementation of resource allocation or alternative control measures, preventing the risk from escalating due to stalled negotiations.

[0208] For all implemented control measures, we continuously track the dynamic changes in subsequent cost estimation results. We construct a quantitative evaluation system from multiple dimensions, including cost reduction, efficiency improvement, risk control effectiveness, and business impact, to comprehensively assess the control effects. Simultaneously, we store complete information about each control case (including anomaly details, classification results, control plan, execution process, and evaluation conclusions) in a historical case database. This information is then added to the data sources for early warning rule optimization, attribution model training, and control strategy iteration, forming a closed loop of control-evaluation-reflection-optimization. This continuously improves the accuracy and efficiency of future early warning, attribution, and control decisions.

[0209] As can be seen from the above, this application, by adopting the aforementioned technical solutions, achieves a multi-objective optimization scheme for controllable anomalies that balances cost and efficiency, forming a closed loop between implementation and effect feedback; for uncontrollable anomalies, resource allocation plans and negotiation mechanisms quickly resolve cross-departmental resource conflicts, and arbitration processes ensure the implementation of these plans. Tracking control effects and storing cases optimize subsequent decision-making. Tiered control tailored to anomaly types, along with a closed-loop feedback and case accumulation optimization mechanism, achieves both precise anomaly handling and enhances the long-term effectiveness of control plans. These technical features address the problems of traditional control plans being singular and difficult to implement, improving the efficiency and effectiveness of cost-effective anomaly handling, and strengthening cross-departmental collaborative control capabilities.

[0210] In one embodiment of this application, an AI Agent-driven method for dynamic management and control of cross-departmental collaborative costs further includes:

[0211] Incremental learning and parameter updates are performed on the global cost calculation model and causal analysis model using the results of target control.

[0212] Based on the actual implementation effect of the optimization plan or the results of resource allocation, the causal association weights in the business causal knowledge graph and the resource status information in the real-time resource graph are dynamically corrected.

[0213] In this embodiment, the obtained target control results are used to carry out targeted incremental learning and parameter optimization on the global cost calculation model and the causal analysis model.

[0214] For the global cost calculation model, the actual cost data and resource consumption details collected during the management and control process are compared with the original estimated results to calculate the deviation value. With the goal of minimizing the deviation, the core parameters in the model, such as the time driver rate benchmark and the cross-department cost allocation coefficient, are corrected. At the same time, accounting samples under new business scenarios are added to improve the model's adaptability to complex collaborative scenarios.

[0215] For causal analysis models, the attribution conclusions after the verification of target control results (including accurate attribution cases and biased attribution cases) are used as incremental training data to optimize the causal reasoning chain of the model, adjust the correlation weight between the causal entity and the abnormal result, improve the attribution confidence and the accuracy of core cause location, and reduce attribution bias.

[0216] Based on the actual implementation effect of the optimization plan or the results of resource allocation, the business causal knowledge graph and the real-time resource graph are dynamically updated and corrected. For the business causal knowledge graph, based on cost and efficiency changes after the optimization plan is implemented, the weights of causal relationships between business processes, resource allocation, and cost anomalies in the graph are verified and adjusted. For causal relationships that have been verified as effective in practice, their weights are increased; for relationships that are invalid or have significant deviations after verification, their weights are reduced and marked for optimization. Simultaneously, newly discovered causal relationships are added to improve the graph's reasoning capabilities. For the real-time resource graph, based on the actual implementation of resource allocation (including the quantity of resources transferred, arrival time, and surplus updates), the resource status information (such as availability, occupancy status, and allocation permissions) in the graph is dynamically corrected. The attributes of cross-departmental resource association nodes are updated synchronously to ensure that the graph accurately reflects the current resource allocation status of the enterprise, providing reliable data support for the generation of contingency plans for resource allocation in the event of uncontrollable cost anomalies.

[0217] Based on the updated model parameters and corrected knowledge graph results, the multi-level early warning threshold rules and hierarchical control strategies were further optimized. The corrected model output and its association with the knowledge graph were integrated into the early warning rule verification system, and the threshold combinations and triggering conditions were adjusted. Combined with feedback on control effectiveness, the logic for generating control schemes was optimized, and strategy templates for efficient control cases were added. This continuously improved the entire process control system—from data collection and calculation to detection, early warning, attribution, control, and iteration—and enhanced the enterprise's cross-departmental collaborative cost dynamic control capabilities.

[0218] As can be seen from the above, this application, by adopting the aforementioned technical solution, uses control results to drive incremental model learning, updating parameters to improve model calculation and attribution accuracy, and aligning with business change patterns. Dynamic correction of the business causal knowledge graph and resource graph optimizes causal association weights and resource status information, improving the quality of decision-making data support. The synergy between model iteration and graph correction allows the control system to continuously adapt to business changes, solving the rigidity problem of traditional control systems. These technical features enable the self-optimization and upgrading of the control system, improving the accuracy, adaptability, and long-term effectiveness of cost control, forming a virtuous cycle of "control-feedback-optimization," and strengthening the sustainability of the technical solution.

[0219] Corresponding to the AI ​​Agent-driven cross-departmental collaborative cost dynamic management method in the above embodiment, Figure 2This is a structural block diagram of an AI Agent-driven cross-departmental collaborative cost dynamic control system provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The AI ​​Agent-driven cross-departmental collaborative cost dynamic management and control system 20 includes: a data acquisition module 21, a cost accounting module 22, an anomaly detection module 23, an early warning generation module 24, an early warning execution module 25, an attribution analysis module 26, an attribution classification module 27, and a control execution module 28.

[0220] The data acquisition module 21 is used to receive standardized data streams converted by edge AI Agents deployed in the business systems of various departments based on preset semantic models. The standardized data streams originate from local heterogeneous business data. The standardized data streams include resource consumption records, job events, and time-related data.

[0221] Cost accounting module 22 is used to input standardized data streams into the global cost calculation model for accounting and to obtain cost estimation results;

[0222] Anomaly detection module 23 is used to perform cluster analysis and anomaly detection on cost estimation results using unsupervised learning algorithms to obtain potential cost anomaly patterns and cost behavior baselines;

[0223] The early warning generation module 24 is used to generate multi-level early warning threshold rules based on the cost behavior baseline and by applying statistical process control methods.

[0224] The early warning execution module 25 is used to trigger the corresponding early warning signal when the cost estimation result reaches the early warning threshold rule, and associate it with the cost data segment that triggered the early warning signal and the corresponding cost anomaly mode;

[0225] Attribution analysis module 26 is used to call the causal analysis model to perform attribution analysis on the cost anomalies that trigger the early warning and obtain the attribution results;

[0226] The attribution classification module 27 is used to classify cost anomalies into controllable cost anomalies or uncontrollable cost anomalies based on whether the attribution results point to internally controllable business processes or resource factors.

[0227] The control execution module 28 is used to select and execute the corresponding hierarchical control mechanism based on the results of cost anomaly classification to obtain the target control results.

[0228] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the data acquisition module 21, cost accounting module 22, anomaly detection module 23, early warning generation module 24, early warning execution module 25, attribution analysis module 26, attribution classification module 27, and control execution module 28 are shown.

[0229] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0230] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0231] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0232] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the AI ​​Agent-driven cross-departmental collaborative cost dynamic control method provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0233] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0234] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0235] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0236] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0237] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0238] The units described as separate components may or may not be physically separate. 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 units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0239] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0240] The above are merely specific embodiments 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 equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered 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. A method for dynamic cost control of cross-departmental collaboration driven by AI Agent, characterized in that, include: Receive standardized data streams transformed by edge AI Agents deployed in various departmental business systems based on preset semantic models; The standardized data stream is input into the global cost calculation model for verification to obtain the cost estimation result; Unsupervised learning algorithms are used to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines. Based on the aforementioned cost behavior baseline, statistical process control methods are applied to generate multi-level early warning threshold rules. When the cost estimation result reaches the warning threshold rule, the corresponding warning signal is triggered and associated with the cost data segment that triggered the warning signal and the corresponding cost anomaly pattern; The causal analysis model is invoked to perform attribution analysis on the cost anomalies that triggered the early warning, and the attribution results are obtained. Based on whether the attribution results point to internally controllable business processes or resource factors, cost anomalies are classified as controllable cost anomalies or uncontrollable cost anomalies. Based on the results of cost anomaly classification, select and implement the corresponding hierarchical control mechanism to achieve the target control results; The unsupervised learning algorithm is used to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines, including: Construct an enhanced feature space that includes cost values, rates of change, and cost correlations between departments; The distribution characteristics of data points in the enhanced feature space are evaluated to obtain evaluation results. Based on the evaluation results, the corresponding analysis strategy is selected and executed to obtain the potential cost anomaly patterns and cost behavior baselines. If the data points show periodicity or trend in the time series dimension, and the variance of the cost correlation between departments is less than or equal to a preset first variance threshold, then the first analysis strategy is executed. If the data points exhibit high dimensionality and sparsity, or if the variance of the cost correlation between departments exceeds the preset first variance threshold, then the second analysis strategy will be executed. The execution of the first analysis strategy includes: Density clustering algorithm is used to cluster data points in the enhanced feature space. By optimizing the neighborhood radius and minimum sample number parameters, one or more core clusters that represent cost behavior patterns are identified, and data points that do not belong to any core cluster and whose number of points in the neighborhood is less than a preset first point number threshold are identified as isolated outliers. For each core cluster, a time-series decomposition model is used to fit the cost behavior baseline and normal fluctuation range of the core cluster. Data points that deviate from the cost behavior baseline of their respective core clusters and exceed their normal fluctuation range, along with identified isolated anomalies, are collectively identified as potential cost anomaly patterns. The execution of the second analysis strategy includes: A deep anomaly detection model based on reconstruction error is used to detect data points in the enhanced feature space and obtain an anomaly score for each data point. Data points with anomaly scores greater than a preset anomaly score threshold are clustered, and the different clusters formed are identified as different potential cost anomaly patterns. Data points with anomaly scores less than or equal to a preset anomaly score threshold are clustered. For each data cluster, a Gaussian process kernel function is selected or combined based on the temporal structure characteristics of the data cluster, and Gaussian process regression is used to fit and obtain the cost behavior baseline.

2. The AI ​​Agent-driven cross-departmental collaborative cost dynamic management method according to claim 1, characterized in that, The construction includes an enhanced feature space comprising cost values, rates of change, and inter-departmental cost correlations, including: Extract the cost sequence of each department within a preset time window from the standardized data stream; The cost sequence is subjected to first-order difference calculation to obtain the instantaneous rate of change sequence; The instantaneous rate of change sequence is exponentially smoothed based on a preset time window to generate a smooth rate of change; Based on the operational chain described by the global cost calculation model, the impact coefficient of changes in the operational efficiency of upstream departments on the costs of downstream departments is calculated as the cost correlation degree between departments. For each time point in the cost sequence: The cost sequence is synchronously smoothed based on the preset time window to obtain a processed cost value sequence, and the cost value corresponding to the time point is obtained from the processed cost value sequence. Obtain the rate of change value corresponding to the time point from the smooth rate of change sequence; Obtain the inter-departmental cost correlation vector corresponding to a given time point; The cost value, the smooth change rate, and the inter-departmental cost correlation vector are concatenated to form an enhanced feature vector at a given time point. The enhanced feature space is constructed based on the enhanced feature vectors at all time points.

3. The AI ​​Agent-driven cross-departmental collaborative cost dynamic management method according to claim 1, characterized in that, The step of selecting or combining Gaussian process kernel functions based on the temporal structure characteristics of the data cluster, and using Gaussian process regression to fit and obtain the cost behavior baseline includes: Select or combine Gaussian process kernel functions based on the temporal structure characteristics of the data cluster; Using the selected kernel function, a Gaussian process model is trained based on the historical data of the data cluster to obtain the posterior distribution of the predicted cost value; The mean function of the posterior distribution is used as the central trajectory of the cost behavior baseline, and the envelope formed by adding and subtracting twice the standard deviation function to the mean function is used as the fluctuation boundary of the cost behavior baseline.

4. The AI ​​Agent-driven cross-departmental collaborative cost dynamic control method according to claim 3, characterized in that, The step of selecting or combining Gaussian process kernel functions based on the temporal structure characteristics of the data cluster includes: The temporal patterns of the data cluster are encoded by a variational autoencoder to obtain a latent spatial feature vector. The latent space feature vector is input into the kernel function selection network to obtain the combined weights: Based on the latent space feature vector, multiple preset Gaussian process kernel functions are selected or combined to obtain a combined kernel function; The Gaussian process kernel functions include: radial basis function kernel, periodic kernel, Matern kernel, and rational quadratic kernel; The combined weights and combined kernel function are weighted and summed to obtain the kernel function.

5. The AI ​​Agent-driven cross-departmental collaborative cost dynamic management method according to claim 4, characterized in that, Also includes: Based on the posterior distribution, the prediction uncertainty for each prediction point is calculated, and the temporal correlation of the prediction uncertainty is evaluated. When the prediction uncertainty at multiple consecutive time points exceeds a preset first threshold, a baseline reconstruction signal is triggered. In response to the baseline reconstruction signal, the sensitivity of the warning threshold rule is adjusted, wherein the warning threshold is fluctuated within the range of the mean function plus or minus K times the standard deviation function, wherein the K value is adjusted according to the level of the prediction uncertainty, and the higher the uncertainty, the larger the K value.

6. The AI ​​Agent-driven cross-departmental collaborative cost dynamic management method according to claim 5, characterized in that, The K value is adjusted according to the level of the prediction uncertainty, including: Obtain the standard deviation of the posterior distribution of the Gaussian process at the current prediction time point as a measure of prediction uncertainty; The uncertainty level is obtained by normalizing the forecast uncertainty measure for all time points within the current forecast window. Based on the departmental business attribute weights and cost type sensitivity coefficients associated with the cost behavior baseline, the uncertainty level is weighted and corrected to obtain the corrected uncertainty level. The modified uncertainty level is mapped to the base K value using a pre-defined monotonically decreasing function. The higher the uncertainty level, the smaller the mapped base K value. The base K value is multiplied by the preset base sensitivity coefficient, and then truncated to the upper and lower limits to obtain the final K value used to calculate the warning threshold.

7. An AI Agent-driven cross-departmental collaborative cost dynamic management and control system, characterized in that, include: The data acquisition module is used to receive standardized data streams converted by edge AI Agents deployed in the business systems of various departments based on a preset semantic model. The standardized data streams originate from local heterogeneous business data. The standardized data streams include resource consumption records, job events, and time-related data. The cost accounting module is used to input the standardized data stream into the global cost calculation model for accounting and to obtain cost estimation results. Anomaly detection module is used to perform cluster analysis and anomaly detection on the cost estimation results using an unsupervised learning algorithm to obtain potential cost anomaly patterns and cost behavior baselines; The unsupervised learning algorithm is used to perform cluster analysis and anomaly detection on the cost estimation results to obtain potential cost anomaly patterns and cost behavior baselines, including: Construct an enhanced feature space that includes cost values, rates of change, and cost correlations between departments; The distribution characteristics of data points in the enhanced feature space are evaluated to obtain evaluation results. Based on the evaluation results, the corresponding analysis strategy is selected and executed to obtain the potential cost anomaly patterns and cost behavior baselines. If the data points show periodicity or trend in the time series dimension, and the variance of the cost correlation between departments is less than or equal to a preset first variance threshold, then the first analysis strategy is executed. If the data points exhibit high dimensionality and sparsity, or if the variance of the cost correlation between departments exceeds the preset first variance threshold, then the second analysis strategy will be executed. The execution of the first analysis strategy includes: Density clustering algorithm is used to cluster data points in the enhanced feature space. By optimizing the neighborhood radius and minimum sample number parameters, one or more core clusters that represent cost behavior patterns are identified, and data points that do not belong to any core cluster and whose number of points in the neighborhood is less than a preset first point number threshold are identified as isolated outliers. For each core cluster, a time-series decomposition model is used to fit the cost behavior baseline and normal fluctuation range of the core cluster. Data points that deviate from the cost behavior baseline of their respective core clusters and exceed their normal fluctuation range, along with identified isolated anomalies, are collectively identified as potential cost anomaly patterns. The execution of the second analysis strategy includes: A deep anomaly detection model based on reconstruction error is used to detect data points in the enhanced feature space and obtain an anomaly score for each data point. Data points with anomaly scores greater than a preset anomaly score threshold are clustered, and the different clusters formed are identified as different potential cost anomaly patterns. Data points with anomaly scores less than or equal to a preset anomaly score threshold are clustered. For each data cluster, a Gaussian process kernel function is selected or combined based on the temporal structure characteristics of the data cluster, and Gaussian process regression is used to fit and obtain the cost behavior baseline. The early warning generation module is used to generate multi-level early warning threshold rules based on the cost behavior baseline and by applying statistical process control methods. The early warning execution module is used to trigger a corresponding early warning signal when the cost estimation result reaches the early warning threshold rule, and associate it with the cost data segment that triggered the early warning signal and the corresponding cost anomaly mode; The attribution analysis module is used to call the causal analysis model to perform attribution analysis on the cost anomalies that trigger the warning and obtain the attribution results. The attribution classification module is used to classify cost anomalies into controllable cost anomalies or uncontrollable cost anomalies based on whether the attribution results point to internally controllable business processes or resource factors. The control execution module is used to select and execute the corresponding hierarchical control mechanism based on the results of cost anomaly classification to obtain the target control results.

Images