An agent-driven differentiated report generation method, system and storage medium

Abstract

The application discloses an intelligent agent driven differentiated report generation method and system and a storage medium. The method comprises the following steps: monitoring key index data flow in real time, and analyzing the characteristics of the key index data flow based on a statistical model to generate a state switching instruction; receiving the instruction and switching among multiple working states such as a normal state, a warning state and an emergency state. In response to the state switching, the report generation processing flow is dynamically adjusted, and the adjustment at least includes changing a data sampling frequency, switching a data preprocessing model and selecting a report generation logic matched with the current state. By introducing the event driven polymorphic working mode, the application realizes the transformation from passive report generation to active and quasi-real-time decision support, significantly improves the response speed of the system to emergencies and the intelligent level of decision support, and balances the system energy efficiency and performance.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, specifically, it relates to artificial intelligence, knowledge graph and big data processing technology, and in particular to an agent-driven differentiated report generation method, system and computer-readable storage medium. Background Technology

[0002] In modern enterprise management, especially for large conglomerates with businesses spanning multiple industries such as coal, power, chemicals, and new energy, data-driven decision support is crucial. As the core carrier of data analysis results, the efficiency, accuracy, and decision support capabilities of reports directly impact the group's operational efficiency and risk management level.

[0003] Existing report generation technologies mainly fall into the following categories: First, data filling methods based on fixed templates. These methods pre-define report chapters, charts, and indicators, automatically retrieving data from a database interface and filling it into the corresponding positions. Their advantages are simplicity and speed, but their disadvantages are also significant. Firstly, the rigid structure of the template cannot adapt to the differentiated needs of different management levels (such as group headquarters, second-tier subsidiaries, and grassroots production units). Group headquarters focus on cross-industry macro-level compliance and strategic risks, while grassroots units require real-time operational data specific to a particular production line or workshop; fixed templates struggle to accommodate the analytical requirements of different granularities. Secondly, these methods are typically designed for a single business area. For multi-industry groups, the data standards, indicator naming, and business logic differ greatly between industries (e.g., "raw coal production" in the coal industry versus "grid-connected electricity" in the power industry). Simple template filling cannot achieve effective cross-industry integration and comparative analysis.

[0004] Secondly, there are self-service analytical reports based on business intelligence (BI) tools. These tools offer business personnel a degree of flexibility, allowing them to customize reports through drag-and-drop. However, this requires users to possess strong data analysis skills and a deep understanding of the business. For senior management, what they often need are direct, clear conclusions and recommendations, rather than complex interfaces. More importantly, BI tools themselves lack domain knowledge and cannot automatically perform compliance checks (such as determining whether pollutant emissions meet national standards) or provide decision-making suggestions (such as recommending appropriate remediation solutions when emissions exceed standards).

[0005] Third, preliminary intelligent solutions have been introduced using knowledge graph technology. Some advanced systems have begun to attempt to build domain-specific knowledge graphs, structuring information such as data, standards, and cases. However, this is usually limited to a single domain. For example, a knowledge graph specifically for pollution control may not be able to effectively link a company's production data, financial data, etc., resulting in a one-sided analytical dimension and failing to reveal the deep causal relationship between "business activities" and "environmental impact," thus significantly reducing the decision-making value of the report.

[0006] In summary, existing technologies generally face the following challenges when dealing with the complex reporting needs of multi-industry conglomerates:

[0007] 1. Data silos and poor adaptability: It is difficult to effectively integrate business data and environmental data from multiple industries (such as coal, power and chemicals) under the group, and there is a lack of a unified indicator alignment and correlation analysis mechanism, resulting in reports that are single-dimensional and fragmented.

[0008] 2. Lack of knowledge-based decision support: The reports mostly remain at the level of data listing, lacking functions such as automatic compliance verification, risk warning and intelligent decision suggestions based on industry standards, regulations and policies and historical success cases.

[0009] 3. Mismatch between hierarchical requirements: The report generation logic is rigid and cannot dynamically adjust the data granularity, analysis focus, and presentation format of the report according to the specific needs of different management levels such as the group, second-level enterprises, and grassroots units, resulting in weak practicality and relevance of the report.

[0010] 4. Delayed and passive response: Most systems adopt a periodic or request-based report generation mode. For sudden environmental warnings or emergencies during the production process, they lack near real-time intelligent perception and rapid response capabilities, and cannot generate emergency reports containing root cause analysis and disposal suggestions in the first time, which may lead to missing the best time for handling.

[0011] Therefore, there is an urgent need for a new intelligent report generation method that can overcome the above limitations, achieve deep cross-industry data integration, possess knowledge-based decision support capabilities, accurately match multi-level differentiated needs, and proactively and rapidly respond to emergencies. Summary of the Invention

[0012] The main objective of this invention is to provide an agent-driven method, system, and computer-readable storage medium for generating differentiated reports, aiming to solve technical problems in the prior art such as difficulties in cross-industry data fusion, lack of knowledge-based decision support, inability to match multi-level needs, and delayed emergency response.

[0013] To achieve the above objectives, a first aspect of the present invention provides an agent-driven differentiated report generation method, characterized by comprising: a state determination module that monitors key indicator data streams from one or more heterogeneous data sources in real time; the state determination module analyzing the statistical characteristics of the key indicator data streams based on a predefined statistical model to generate a state switching instruction; a report generation module receiving the state switching instruction and switching between multiple preset working states, including at least a normal state, an early warning state, and an emergency state, according to the instruction; the report generation module dynamically adjusting its report generation processing flow in response to the switching of working states, the adjustment including at least: changing the data sampling frequency, switching the data preprocessing model, and selecting a report generation logic matching the current working state; wherein, in the emergency state, the report generation logic is configured to bypass the conventional user requirement parsing steps and forcibly call a preset emergency report template to generate an emergency report.

[0014] Furthermore, the state determination module analyzes the statistical characteristics of the key indicator data stream, including: using a sliding window-based algorithm to continuously calculate and evaluate the key indicator data stream.

[0015] Furthermore, the sliding window-based algorithm is an exponentially weighted moving average (EWMA) control chart algorithm, which identifies trend changes or abnormal fluctuations in the data stream by calculating the weighted average of the key indicator data within the sliding window.

[0016] Furthermore, the EWMA control chart algorithm is configured as follows: a weighting factor λ between 0 and 1 is defined, where a smaller λ value assigns a higher weight to recent data points to enhance sensitivity to changes in the data stream; an upper control limit (UCL) and a lower control limit (LCL) are established based on the statistical standard deviation of the historical data of the key indicator, the upper control limit and the lower control limit are respectively located at K times the standard deviation of a center line, where K is a configurable risk coefficient with a preset value of 3; and when the EWMA calculated value of M consecutive data points in the key indicator data stream exceeds the upper control limit or the lower control limit, a state switching instruction for switching to the warning state or the emergency state is generated, where M is a preset integer with a value greater than or equal to 2.

[0017] Furthermore, under the emergency state, the report generation logic also includes: based on a pre-built multi-industry knowledge graph, executing one or more Structured Query Language (SPARQL) queries in parallel to automatically analyze the potential root causes of the events that trigger the emergency state; and according to the query results, associating and extracting corresponding emergency plan information from the knowledge graph, and filling the potential root causes and emergency plan information into the emergency report.

[0018] Furthermore, the data preprocessing model is an adaptive anomaly detection model, which is configured to periodically and automatically update a set of industry weight coefficients for identifying data anomalies based on the distribution changes of recently collected data, in order to adapt to data concept drift caused by changes in production processes or external environment.

[0019] Furthermore, the report generation process also includes: in the normal state, using a natural language understanding model to parse the received user report generation request to obtain a parsing intent; calculating a confidence score of the parsing intent based on the sequence probability output by a final classification layer of the natural language understanding model; and, when the confidence score is lower than a preset threshold, triggering a user-facing human-computer interaction clarification process, or generating a default report in a preset format.

[0020] A second aspect of the present invention provides an agent-driven differentiated report generation system, characterized in that it comprises: a state determination module configured to monitor key indicator data streams from one or more heterogeneous data sources in real time, and analyze the statistical characteristics of the key indicator data streams based on a predefined statistical model to generate a state switching instruction; and a report generation module configured to receive the state switching instruction, switch between multiple preset working states including at least a normal state, an early warning state, and an emergency state according to the instruction, and dynamically adjust its report generation processing flow in response to the switching of working states, wherein the adjustment includes at least changing the data sampling frequency, switching the data preprocessing model, and selecting a report generation logic that matches the current working state; wherein, in the emergency state, the report generation module is configured to bypass the conventional user requirement parsing steps and forcibly call a preset emergency report template to generate an emergency report.

[0021] Furthermore, the report generation module employs an adaptive anomaly detection model as the data preprocessing model. The adaptive anomaly detection model is configured to periodically and automatically update a set of industry weight coefficients for identifying data anomalies based on the distribution changes of recently collected data, in order to adapt to data concept drift.

[0022] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect of the present invention.

[0023] This invention introduces an event-driven, multi-mode switching architecture, enabling the system to autonomously switch between various states such as normal, early warning, and emergency based on the characteristics of real-time data streams, and dynamically schedule matching computing resources and processing logic. This design achieves "on-demand allocation" of system capabilities, bringing significant benefits:

[0024] 1. Balancing energy efficiency and performance: During more than 99% of stable operation, the system is in a low-power routine inspection state, which can reduce resource consumption by at least 80%. When anomalies such as sudden pollution events are detected, it can instantly switch to a high-performance emergency response state, reducing the generation time of emergency reports from the traditional hours to within minutes, achieving a qualitative leap from "post-event collection" to "near real-time response".

[0025] 2. Enhanced Robustness and Reliability: A data preprocessing model with adaptive concept drift capability is introduced, which can automatically adjust model parameters according to changes in data distribution. This effectively addresses data pattern changes caused by adjustments in production processes or changes in the external environment, reducing the false positive rate of abnormal data by 5-10%. Simultaneously, a confidence-aware demand parsing mechanism can proactively initiate clarification or generate default reports when handling ambiguous user commands, significantly reducing the generation of invalid or erroneous reports and improving the reliability of human-computer interaction and user experience.

[0026] 3. Intelligent and proactive decision support: By deeply integrating knowledge graphs and polymorphic working mechanisms, this invention can not only generate routine and analytical management reports on demand, but also proactively discover potential risks, provide real-time warnings, and automatically generate decision support reports containing potential root cause analysis and standardized emergency response suggestions in emergency situations. This allows the system to evolve from a passive reporting tool into an intelligent decision support platform with autonomous perception, dynamic adaptation, and hierarchical response capabilities. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of an enhanced system architecture based on polymorphic switching according to an embodiment of the present invention; Figure 2 This is a general flowchart of an agent-driven differential report generation method according to an embodiment of the present invention; Figure 3This is a flowchart illustrating a multi-source data access and indicator alignment method according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating a method for dynamically constructing a pollution control knowledge graph according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of a multi-level, multi-industry demand analysis model provided according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the state transition logic of a polymorphic intelligent agent core according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the workflow of an event flow triage module according to an embodiment of the present invention; Figure 8 This is an example diagram of the content structure of an emergency report provided according to an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0030] Example 1: System Architecture

[0031] Reference Figure 1 This embodiment provides an agent-driven differentiated report generation system. The system can be deployed on a group's central server or a private cloud platform, and its hardware infrastructure includes, but is not limited to, processors, memory, hard drives, and other storage devices, as well as network communication interfaces. The core innovation of the system lies in its software architecture, which introduces an event-driven polymorphic working mode switching mechanism.

[0032] like Figure 1 As shown, the system logically includes a data access layer, an event triage module (ETM) (101), a multi-state agent core (MSAC) (102), a pollution control knowledge graph subsystem, and a report output layer.

[0033] The data access layer is responsible for collecting structured and unstructured data from business systems (such as ERP and MES) of different industries under the group, such as coal, power, chemical and new energy, environmental monitoring systems (such as CEMS), and external data sources (such as national policy and regulation websites) through multi-source interfaces.

[0034] The Event Triage Module (ETM) (101), serving as the "nerve center" of this system, is configured as a state determination module. Located between the data access layer and the polymorphic agent core (102), it performs real-time, lightweight monitoring of key indicator data streams. For example, it continuously monitors the emission concentrations of key pollutants (such as COD and VOCs) and the operating parameters of key equipment (such as pressure and temperature). The ETM (101) analyzes the statistical characteristics of the data stream (such as mean, volatility, and rate of change) based on predefined statistical models, such as the Exponentially Weighted Moving Average (EWMA) control chart algorithm, which will be detailed later. The analysis result is not direct data content, but rather a State Transition Signal (103). This signal is sent to the polymorphic agent core (102).

[0035] The Multi-State Intelligent Agent Core (MSAC) (102), serving as the "execution brain" of this system, is configured as a report generation module. Internally, it maintains a finite state machine with at least three preset working states: routine inspection state (S_routine), early warning monitoring state (S_warning), and emergency response state (S_emergency). MSAC (102) receives state switching instructions (103) from the ETM (101) and switches its own working state accordingly. Crucially, each state corresponds to a completely different set of processing links and resource scheduling strategies.

[0036] • In the routine inspection state (S_routine) link (104), MSAC (102) samples and processes data at a low frequency (e.g., once per hour), performs routine data preprocessing, responds to user report generation requests, performs complex requirement analysis, and calls a multi-objective optimized template selection algorithm to generate routine and analytical reports. System resource consumption is minimal in this state.

[0037] • In the early warning monitoring state (S_warning) link (105), when ETM (101) detects a continuous, small-amplitude abnormal trend in the data, MSAC (102) switches to this state. It increases the data sampling frequency (e.g., once per minute), calls a more complex adaptive data preprocessing model, and may preload knowledge graph subgraphs related to potential problems for continuous monitoring and data caching, but usually does not actively generate reports. This state is designed to "prevent problems before they occur" and prepare for potentially more serious events.

[0038] • In the emergency response state (S_emergency) link (106), when the ETM (101) detects a momentary, significant, and drastic anomaly in the data (such as a severe exceedance of pollutant concentration), the MSAC (102) immediately switches to this state. It collects data at the highest frequency (e.g., in real time or once per second), bypasses the conventional demand analysis steps, forcibly invokes the preset emergency report template, quickly generates an emergency source tracing report, and may proactively push it to relevant personnel through various channels (such as SMS, system pop-ups). In this state, the system will mobilize maximum computing resources to ensure the fastest response.

[0039] The pollution control knowledge graph subsystem adopts a three-layer architecture of "ontology layer - data layer - reasoning layer" to provide knowledge support for MSAC (102) decision-making in any state. The report output layer is responsible for displaying and exporting various reports generated by MSAC (102) in a way that is suitable for different terminals such as PC and mobile devices.

[0040] Through this architecture, the system has transformed from a passive reporting tool into a proactive, context-aware decision support platform.

[0041] Example 2: Overall Method Flow

[0042] Reference Figure 2 This embodiment provides an agent-driven differential report generation method, which can be executed by the system described in Embodiment 1. The method includes the following steps:

[0043] Step S201: Real-time monitoring and status determination. After the system starts, the event flow triage module (101) continuously monitors the key indicator data flow from the data access layer, which originates from heterogeneous data sources across multiple industries. As mentioned earlier, this module uses a lightweight statistical model (such as EWMA) to analyze the dynamic characteristics of the data flow.

[0044] Step S202: Generate and transmit state switching instructions. Based on the analysis results of step S201, the event flow triage module (101) determines the current working state of the system. For example, if the data is stable, the normal state is maintained; if the data shows an abnormal trend, an instruction to switch to the warning state is generated; if the data changes abruptly and exceeds the danger threshold, an instruction to switch to the emergency state is generated. This state switching instruction (103) is sent to the multi-state agent core (102).

[0045] Step S203: Switch working state and adjust processing flow. After receiving the state switching instruction (103), the multi-state intelligent agent core (102) immediately switches its internal state machine to the specified state (normal, early warning, or emergency). This switch will trigger a series of dynamic adjustments to the processing flow.

[0046] Step S204: Execute the report generation logic that matches the current state. This is the core execution step of the method. The polymorphic agent core (102) executes a differentiated report generation process based on the current working state.

[0047] Under normal circumstances, a complete workflow designed to generate in-depth analysis reports is executed. This workflow includes detailed steps such as multi-source data access and indicator alignment, dynamic construction of a pollution control knowledge graph, multi-industry data preprocessing, multi-level and multi-industry requirement analysis, differentiated report generation, and report output and dynamic optimization. These steps will be detailed in subsequent embodiments.

[0048] • In an alert state, the processing flow is adjusted to focus on monitoring and data accumulation, and the report generation service is suspended. However, a more sensitive data preprocessing model is invoked to prepare for potential emergency responses.

[0049] • In emergency situations, the processing flow is greatly simplified and accelerated. The system will bypass the time-consuming user requirement analysis process, forcibly call the preset emergency report template, and use knowledge graphs to perform rapid root cause analysis and link it with emergency plans, generating and outputting emergency reports as quickly as possible.

[0050] Through the above steps, this method achieves intelligent scheduling of system resources and processing logic, ensuring optimal performance in different business scenarios.

[0051] Example 3: Detailed Processing Flow under Normal Conditions

[0052] In its normal state (S_routine), the system aims to respond to user requests and generate high-quality, in-depth, and personalized analysis reports. Its processing flow can be divided into the following stages:

[0053] Phase 1: Multi-source data access and indicator alignment

[0054] Reference Figure 3 This phase aims to address the issue of data heterogeneity across multiple industry groups.

[0055] Step S301: Multi-source data access. Through the data access layer interface, access environmental data such as "mine water COD value" from coal enterprises, "dust emission" from power enterprises, and "VOCs concentration" from chemical enterprises, as well as business data such as "raw coal production" and "power generation".

[0056] Step S302: Indicator Segmentation and Vectorization. For the accessed indicator names, a word segmentation tool is first used for word segmentation and part-of-speech tagging. For example, "sulfur dioxide emissions" can be processed into "sulfur dioxide [indicator name] - emissions [attribute] - mg / m³ [unit]". Subsequently, based on a Word2Vec model pre-trained using industry environmental standard text, the segmented indicator names are converted into high-dimensional (e.g., 128-dimensional) semantic vectors.

[0057] Step S303: Similarity Calculation and Indicator Alignment. The cosine similarity algorithm is used to calculate the similarity between the semantic vectors of indicators from different sources. A similarity threshold (e.g., 0.85) is set. When the vector similarity of two indicators (such as "sulfur dioxide emissions" and "SO2 emissions") is greater than or equal to this threshold, the system determines that they are the same entity, thereby achieving automatic indicator alignment across systems and industries.

[0058] Step S3D4: Tagging and Storage. After alignment, the system will automatically label the indicator data with industry tags (coal / electricity / chemical / new) and hierarchical tags (group / secondary enterprise / grassroots unit) based on its source path (e.g., "group database → power sector") or enterprise registration information, laying the foundation for subsequent hierarchical and differentiated analysis.

[0059] Phase Two: Dynamic Construction of a Knowledge Graph for Pollution Control

[0060] Reference Figure 4 This stage aims to transform fragmented data into structured, reasonable knowledge.

[0061] Step S401: Triple Generation. Using the RDF triple conversion algorithm, the aligned structured data is converted into triples in the format `<subject, predicate, object>`. For example, the data "Power Plant - Emissions - Sulfur Dioxide 30mg / m³" and the standard "Sulfur Dioxide - Limit - 35mg / m³ (GB 13223-2011)" are converted into corresponding RDF triples and stored in the data layer of the knowledge graph.

[0062] Step S402: Knowledge Reasoning and Knowledge Graph Enhancement. Based on the SWRL (Semantic Web Rule Language) rule reasoning algorithm, a series of industry rules are defined. For example, rule R1: "If the emission concentration of pollutant B by company A is greater than the standard limit, then A is determined to be an enterprise exceeding the standard." When new data enters, the reasoning engine automatically triggers these rules, generates new knowledge (such as "a certain chemical company is an enterprise exceeding the standard"), and adds it back to the knowledge graph.

[0063] Step S403: Dynamic Update of the Knowledge Graph. The system is equipped with a web crawler module that automatically crawls the latest policies, standards, and industry rectification cases published on the official websites of the Ministry of Ecology and Environment and local environmental protection bureaus every month. The rules of the data layer and inference layer of the knowledge graph are updated through the above algorithm to ensure the timeliness of the knowledge.

[0064] Phase 3: Multi-industry Data Preprocessing

[0065] This stage aims to improve the quality of the data entering the model. The system employs two core algorithms:

[0066] 1. Adaptive Isolation Forest Algorithm with Dynamic Weight Adjustment: Used for anomaly cleaning. Traditional isolation forest algorithms treat all data equally, but this invention improves upon this approach. First, based on the historical fluctuation characteristics of environmental data from various industries, initial weight coefficients are assigned to different industries (e.g., chemical industry data fluctuates greatly, weight = 1.2; new energy industry data fluctuates less, weight = 0.8). When constructing the isolation tree for splitting, industry weight coefficients are introduced to adjust the splitting threshold. More importantly, this model possesses adaptive capabilities. The system periodically (e.g., weekly) analyzes the distribution changes of recently collected data. If the data distribution of a certain industry changes significantly (concept drift), the model automatically adjusts the weight coefficient for that industry. This mechanism enables anomaly detection to adapt to external changes such as adjustments in production processes, significantly improving robustness.

[0067] 2. Industry-Adaptive Z-Score Standardization Algorithm: Used for data format standardization. To avoid the impact of differences in the units of measurement of different indicators, data standardization is necessary. Traditional Z-Score uses the mean and standard deviation of the sample itself, while this invention retrieves the more representative historical long-term mean μ and standard deviation σ of the corresponding industry indicator from the knowledge graph, and then calculates according to the formula:

[0068] x std = (x - μ) / σ

[0069] For example, the "smoke and dust emission of 25 mg / m³" for power companies is standardized to (25-20) / 5 = 1, and the "VOCs concentration of 90 mg / m³" for chemical companies is standardized to (90-80) / 10 = 1. This ensures that cross-industry indicators are numerically comparable.

[0070] Phase Four: Multi-level and Multi-industry Demand Analysis

[0071] Reference Figure 5 This stage aims to accurately understand the user's fuzzy query intent.

[0072] Step S501: Model Construction. This invention employs a BERT-BiLSTM-CRF hybrid model (500).

[0073] • Word embedding layer (501): A pre-trained Chinese BERT model (such as bert-base-chinese) is used to convert the user's request text (such as "secondary enterprise chemical environmental protection control") into a 768-dimensional vector containing deep semantic information.

[0074] • Feature extraction layer (502): The vector output by BERT is input into a bidirectional long short-term memory network (BiLSTM). BiLSTM effectively captures the contextual dependencies of the text through forward and backward propagation, such as identifying "secondary enterprise" as hierarchical information and "chemical industry" as industry information.

[0075] • Label prediction layer (503): A conditional random field (CRF) layer is introduced on top of BiLSTM. The CRF layer can learn the constraints between labels (such as "hierarchy" label cannot be directly followed by "index" label), thereby ensuring that the output label sequence is legal and optimal.

[0076] Step S502: Confidence Perception and Interaction. A key improvement of this invention is that after the model performs intent parsing, it does not directly adopt the result. The system uses the sequence probability output by the CRF layer to calculate the confidence score of the entire parsing result. A confidence threshold is set (e.g., 0.85). If the score is lower than this threshold, it indicates that the model is not sufficiently confident in the parsing result. At this time, the system will not blindly generate a report, but will trigger one of two subsequent behaviors: a) proactively initiating a clarifying question, such as "Do you mean you need an environmental compliance management report at the secondary enterprise level in the chemical industry?"; b) generating a preset, relatively overview-based default report. This mechanism significantly improves the reliability of human-computer interaction.

[0077] Step S503: Requirement Completion. For high-confidence analysis results, the system will also utilize knowledge graphs to complete the requirements. For example, if a user only mentions "environmental compliance," the system will automatically retrieve core environmental compliance indicators from the knowledge graph, such as "comparison data of pollutant emissions with national standards" and "statistics on the number of times exceeding standards," and add them to the report generation task.

[0078] Phase 5: Generation of Differentiated Reports

[0079] Based on the analyzed requirements, this stage generates the final report.

[0080] Step S504: Optimal Template Selection. The system pre-sets multiple report templates (such as group-wide coordination type, secondary control type, and grassroots practical type). A multi-objective optimization template selection algorithm is used, based on three dimensions: "template matching degree," "data integrity," and "generation efficiency," to calculate the comprehensive score of each template using a weighted summation method. The formula is:

[0081] S = w1 × M + w2 × I + w3 × E

[0082] Where S is the overall score, M, I, and E are the scores for the three dimensions, and w1, w2, and w3 are weighting coefficients (e.g., 0.5, 0.3, 0.2). The system selects the template with the highest score.

[0083] Step S505: Intelligent Content Filling. After the template is determined, the system retrieves data from the preprocessed database for filling. More importantly, it performs knowledge association based on a knowledge graph. For example, when filling in VOCs emission data for a chemical company, the system automatically executes a SPARQL query, retrieves the corresponding national standard limits and relevant historical rectification cases from the knowledge graph, and generates analytical text, such as "A chemical company's VOCs emission concentration is 90 mg / m³, exceeding the GB standard (80 mg / m³) by 12.5%. It is recommended to refer to the case and adopt activated carbon adsorption technology for rectification."

[0084] Phase Six: Report Output and Dynamic Optimization

[0085] The system supports outputting reports in multiple formats such as PDF and Word, and is adapted for display on different terminals. Simultaneously, the system collects user feedback on report modifications. This feedback is used to fine-tune the parameters of the requirement analysis model (BERT-BiLSTM-CRF) using the gradient descent algorithm, thus forming a closed loop of "data-knowledge-decision-feedback," enabling the report generation capability to continuously evolve.

[0086] Example 4: Multi-state switching and emergency response mechanism

[0087] This embodiment focuses on explaining the polymorphic working mechanism of the system, which is the core difference from the existing technology.

[0088] Reference Figure 6 The diagram illustrates the state transition logic of the Multi-State Agent Core (MSAC) (102). MSAC (102) can transition between a normal state (S_routine) (601), a warning state (S_warning) (602), and an emergency state (S_emergency) (603). The transition is triggered by instructions sent by the Event Triage Module (ETM) (101).

[0089] Reference Figure 7 This demonstrates the workflow of ETM (101).

[0090] Step S701: Parameter Setting and Baseline Update. The ETM (101) employs an Exponentially Weighted Moving Average (EWMA) control chart algorithm based on a sliding window. The algorithm is configured to define a weighting factor λ between 0 and 1. Furthermore, based on the statistical standard deviation of historical data for key indicators, an upper control limit (UCL) and a lower control limit (LCL) are established, each located at K times the standard deviation of a centerline, where K is a configurable risk coefficient (preset to 3). The system periodically (e.g., every 24 hours) dynamically updates the centerline and standard deviation of the control chart using historical stable data.

[0091] Step S702: Calculate the EWMA value in real time. At each time step t, ETM (101) is calculated based on the real-time data point p. t The EWMA value z from the previous moment t-1 Calculate the current EWMA value z t :

[0092] z t = λ · p t + (1 - λ) · z t-1

[0093] Smaller λ values ​​(such as 0.3) make the model more sensitive to recent data changes.

[0094] Step S703: State trigger logic judgment. ETM (101) determines whether to send a state transition command based on a series of rules:

[0095] • Triggering a warning (SIGNAL_TO_WARNING) or an emergency (SIGNAL_TO_EMERGENCY): When the EWMA calculated values ​​of M consecutive data points (e.g., M is a preset integer with a value greater than or equal to 2, such as 5) in the key indicator data stream exceed the upper or lower control limit, it indicates that the data has experienced a continuous, non-random offset, and ETM (101) generates the state switching instruction for switching to the warning state or the emergency state. Furthermore, if a single data point p... t Exceeding a preset hard threshold T_hard, representing a physical or regulatory limit (such as 150% of a standard limit), or z t If the rate of change dz / dt exceeds the mutation threshold T_slope, it indicates that a violent sudden event has occurred, and ETM (101) will immediately send an instruction to switch to emergency state (603).

[0096] • Restore to normal (SIGNAL_TO_ROUTINE): If the system is in a warning (602) or emergency (603) state, and the data of a preset number of consecutive windows has recovered to within the control limits, then send a command to restore to normal state (601).

[0097] Emergency response procedures

[0098] When MSAC (102) switches to emergency mode (603), its execution flow is as follows:

[0099] 1. Resource escalation: Immediately raise the priority of its own process to the highest level, preempting CPU and I / O resources.

[0100] 2. High-frequency data acquisition: Increase the sampling frequency of relevant data sources to the highest level (e.g., once per second).

[0101] 3. Bypass complex preprocessing: Skip time-consuming cleaning steps such as adaptive isolated forest and perform only the most basic formatting processing.

[0102] 4. Bypass requirements parsing: Completely skip the call to the BERT-BiLSTM-CRF model.

[0103] 5. Force the use of the emergency template: Directly load the preset "Emergency Source Tracing Report Template (T-EM-01)".

[0104] 6. Parallel knowledge retrieval: Refer to Figure 8 The content population logic of the report template is highly dependent on the knowledge graph. MSAC(102) executes multiple SPARQL queries in parallel to obtain key information at an extremely fast speed:

[0105] oEvent Snapshot(801): Query to obtain pollutant peak value, occurrence time, and duration.

[0106] o Impact range estimation (802): Based on the "pollutant-diffusion model-geographical location" relationship in the knowledge graph, query the downstream units and sensitive areas that may be affected.

[0107] Potential Root Cause Analysis (803): This is the most critical step. Parallel queries are performed on the "equipment operating data" (such as valve opening and pressure) and "personnel operation logs" of relevant production units within a short time window (such as 10 minutes) before and after the time of exceeding the standard. Through correlation analysis, the top 3 most likely sources of the anomaly are listed.

[0108] o Emergency Plan Association (804): Directly extract the document number and key handling measures of the Emergency Plan associated with this type of event from the knowledge graph.

[0109] 7. Report Push: Generate an emergency report containing the above information within a very short time (e.g., 45 seconds) and proactively push it to pre-set emergency response personnel through various means such as the system, SMS, and email.

[0110] Example 5: Variations and Alternatives

[0111] To further expand the scope of protection, the present invention also covers the following technical variations.

[0112] In the Event Flow Triage Module (ETM) (101), in addition to the EWMA algorithm, other statistical process control (SPC) algorithms can be used, such as the Cumulative Sum Control Chart (CUSUM) algorithm, which is more sensitive to small, persistent drifts in the process mean.

[0113] During the requirements analysis phase, the BERT part in the BERT-BiLSTM-CRF model can be replaced with other more advanced pre-trained language models, such as RoBERTa, XLNet, or models fine-tuned for specific industrial domains, in order to achieve higher semantic understanding accuracy.

[0114] For knowledge graph construction, in addition to RDF and SWRL, property graph databases (such as Neo4j) and their query languages ​​(such as Cypher) can also be used, which may have performance advantages when dealing with complex path queries and multi-hop association analysis.

[0115] The system of this invention is not limited to internal enterprise reporting; its methods and architecture are also applicable to a wider range of fields, such as intelligent early warning in regional environmental monitoring networks and real-time risk monitoring and reporting in the financial sector.

[0116] It should be noted that the present invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the methods described in any of the above embodiments. Simultaneously, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor, when executing the program, implements the methods described in any of the above embodiments.

[0117] In summary, this invention, through its innovative multi-mode switching architecture, combined with an adaptive algorithm model and deeply integrated knowledge graph technology, successfully addresses the multiple challenges faced by multi-industry group enterprises in report generation and management decision-making. It achieves a leap from passive and lagging data processing to proactive and real-time intelligent decision support, demonstrating high practical value and broad application prospects.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating differentiated reports driven by an intelligent agent, characterized in that, include: A state determination module is used to monitor key indicator data streams from one or more heterogeneous data sources in real time. The state determination module analyzes the statistical characteristics of the key indicator data stream based on a statistical model to generate a state switching instruction. The statistical model is an exponentially weighted moving average control chart algorithm based on a sliding window. The exponentially weighted moving average control chart algorithm identifies trend changes or abnormal fluctuations in the data stream by calculating the weighted average of the key indicator data within the sliding window. A report generation module receives the state switching instruction and switches between multiple preset working states, including at least normal state, early warning state and emergency state, according to the instruction; The report generation module responds to the switching of working status by dynamically adjusting its report generation process. The adjustment includes at least: changing the data sampling frequency, switching the data preprocessing model, and selecting the report generation logic that matches the current working status. The data preprocessing model is an adaptive anomaly detection model, which is configured to periodically and automatically update a set of industry weight coefficients for identifying data anomalies based on the distribution changes of recently collected data, in order to adapt to data concept drift caused by changes in production processes or external environment. In the emergency state, the report generation logic is configured to bypass the conventional user requirement parsing steps and forcibly invoke a preset emergency report template to generate an emergency report.

2. The method as described in claim 1, characterized in that, The exponentially weighted moving average control chart algorithm is configured as follows: a weighting factor λ between 0 and 1 is defined; an upper control limit and a lower control limit are established based on the statistical standard deviation of the historical data of the key indicator, the upper control limit and the lower control limit are respectively located at K times the standard deviation of a center line, where K is a configurable risk coefficient with a preset value of 3; and when the calculated value of the exponentially weighted moving average of M consecutive data points in the key indicator data stream exceeds the upper control limit or the lower control limit, a state switching instruction for switching to the warning state or the emergency state is generated, where M is a preset integer with a value greater than or equal to 2, which aims to ensure high confidence in the state switching decision and filter false alarms caused by transient noise.

3. The method as described in claim 1, characterized in that, In the emergency state, the report generation logic further includes: based on a pre-built multi-industry knowledge graph, executing one or more structured query language queries in parallel to obtain query results, so as to automatically analyze the potential root causes of the events that trigger the emergency state; and according to the query results, associating and extracting the corresponding emergency plan information from the knowledge graph, and filling the potential root causes and emergency plan information into the emergency report.

4. The method as described in claim 1, characterized in that, The report generation process also includes: In the aforementioned normal state, a natural language understanding model is used to parse the received user report generation request in order to obtain a parsing intent; Based on the sequence probability output by a final classification layer of the natural language understanding model, a confidence score for the parsed intent is calculated. In addition, when the confidence score is lower than a preset threshold, a user-facing human-computer interaction clarification process is triggered, or a default report in a preset format is generated.

5. A differential report generation system driven by an intelligent agent, characterized in that, include: A state determination module is configured to monitor key indicator data streams from one or more heterogeneous data sources in real time, and analyze the statistical characteristics of the key indicator data streams based on a statistical model to generate a state switching instruction. The statistical model is an exponentially weighted moving average control chart algorithm based on a sliding window. The exponentially weighted moving average control chart algorithm identifies trend changes or abnormal fluctuations in the data stream by calculating the weighted average of the key indicator data within the sliding window. A report generation module is configured to receive the state switching instruction, switch between multiple preset working states including at least a normal state, an early warning state, and an emergency state according to the instruction, and dynamically adjust its report generation processing flow in response to the switching of working states. The adjustment includes at least changing the data sampling frequency, switching the data preprocessing model, and selecting the report generation logic that matches the current working state. The data preprocessing model is an adaptive anomaly detection model, which is configured to periodically and automatically update a set of industry weight coefficients for identifying data anomalies based on the distribution changes of recently collected data, in order to adapt to data concept drift caused by changes in production processes or external environment. In the emergency state, the report generation module is configured to bypass the conventional user requirement parsing steps and forcibly invoke a preset emergency report template to generate an emergency report.

6. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images