Large model all-in-one machine for mine intelligent agent decision reasoning

Abstract

This application relates to the fields of artificial intelligence and edge computing technology, and discloses a large-scale integrated model for decision-making and inference in mining intelligent agents, including a data acquisition and preprocessing module, a scene recognition and SLA generation module, a model scheduling and inference module, and a result output and execution module. The method involves: first, processing the acquired data to generate a scene profile containing task type and service level constraints; then, the model scheduling and inference module, based on this scene profile and combined with a pre-established model profile library storing performance parameters of each model, calculates a comprehensive score for candidate large-scale models through a multi-dimensional matching degree scoring mechanism, thereby selecting the optimal target large-scale model for inference. This invention solves the problem of how to adaptively schedule large-scale models according to dynamic scenes under resource-constrained conditions at the edge of a mine, improving the efficiency, applicability, and security of decision-making.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and edge computing technology, specifically to a large-scale integrated machine for decision-making and reasoning of intelligent agents in mining. Background Technology

[0002] With the development of artificial intelligence technology, technologies represented by large-scale models are gradually being applied to traditional industrial fields such as mining, power, and manufacturing to improve the intelligence and automation level of production processes. In the mining operation environment, using large-scale models to comprehensively analyze massive amounts of sensor data, equipment operating status, and environmental parameters can provide intelligent decision support for key aspects such as safety monitoring, production scheduling, and equipment maintenance.

[0003] Currently, large-scale model deployments are primarily centered around cloud servers. This deployment model relies on stable, high-bandwidth network connections to transmit data collected on-site to the cloud for computation, and then return the decision results to the on-site environment. However, mining environments often suffer from incomplete network coverage, unstable signals, or limited communication bandwidth. This leads to higher network latency and reliability risks for cloud-based decision-making models, making it difficult to meet the stringent real-time requirements of safety-critical tasks such as emergency response to gas exceedances and immediate equipment failures.

[0004] To address these issues, deploying models at the edge, closer to the data source, has become a technological trend. By deploying edge computing devices at the mine site, data processing and model inference can be completed locally, reducing reliance on network connectivity. However, existing edge deployment solutions are often rudimentary, typically embedding a single, lightweight model onto the edge device. This "one-size-fits-all" static deployment approach cannot adapt to the diversity of mining operations. For example, simple equipment status monitoring and complex production process optimization have drastically different requirements for model accuracy, response speed, and computing power. Using a single model makes it difficult to balance efficiency and performance across all scenarios, easily leading to resource waste or insufficient accuracy.

[0005] Furthermore, existing solutions have shortcomings in ensuring the reliability of decision-making. For high-risk decision-making tasks directly related to personnel safety or significant property security, relying solely on the output of a single model carries the risk of misjudgment. Current technologies lack a mechanism to dynamically adjust inference strategies based on the risk level of the task and to effectively verify the decision results, which limits the in-depth application of large models in core automated control processes in mines. Therefore, how to adaptively schedule and collaboratively infer multiple large models at resource-constrained edge environments based on dynamically changing business scenarios and risk levels is a pressing issue in the current technological field. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a large-scale integrated model for decision-making and reasoning in mining intelligent agents. It solves the problem of how to adaptively schedule large models to achieve efficient and reliable decision-making under the condition of limited computing resources at the edge of the mine, based on dynamically changing business scenarios and risk levels.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a large-scale integrated model for decision-making and reasoning of intelligent agents in mining, comprising: The data acquisition and preprocessing module is used to acquire raw data from data sources in the mining operation environment and process the raw data to generate a standardized data stream. The scene recognition and SLA generation module is used to receive the standardized data stream and generate a scene profile containing task scenario, task type and service level constraints based on the data stream. The model scheduling and inference module is used to receive the scene profile and select a target large model from at least one large model deployed in the all-in-one machine for inference based on the scene profile in order to generate a decision result. The result output and execution module is used to receive the decision result and output the decision result to the execution unit or the human-computer interaction interface.

[0008] In one embodiment, the service level constraints include at least one of the following: maximum tolerable response latency, minimum acceptable accuracy, task risk level, and context length required for inference.

[0009] In one embodiment, the model scheduling and inference module is specifically used for: maintaining a model profile library, which stores model profiles corresponding to the at least one large model; calculating a comprehensive score for each large model based on the scene profile and the model profiles in the model profile library; and sorting the at least one large model based on the comprehensive score to select the top-ranked target large model.

[0010] In one embodiment, the overall score is determined by a weighted sum of a delay matching function, a precision matching function, a resource matching function, and a task bias function.

[0011] In one embodiment, the model scheduling and inference module is further configured to: select multiple large models to perform inference in parallel based on the task type or task risk level in the scene profile.

[0012] In one embodiment, when multiple large models are selected to perform inference in parallel, the model scheduling and inference module is further configured to: perform secure arbitration on multiple decision results generated by the multiple large models. The secure arbitration evaluates the consistency and confidence of the multiple decision results; when the evaluation results meet preset conditions, a consistent decision result is output; when the evaluation results do not meet preset conditions, the multiple decision results are fused.

[0013] In one embodiment, the model scheduling and inference module is further configured to: dynamically update the model profiles in the model profile library based on the running logs recorded by the result output and execution module. The model profiles include parameters such as model parameter size, average inference latency, accuracy, and scene-specific subvectors.

[0014] In one embodiment, the result output and execution module is specifically used to: format the decision result into a preset protocol message or structured data format, and send it through a communication interface.

[0015] A second aspect of the present invention provides a method for operating a large-scale integrated model for decision-making and reasoning in mining intelligent agents, comprising the following steps: S1: Collect raw data from the mining operation environment and process the raw data to generate a standardized data stream; S2: Receive the standardized data stream and generate a scenario profile containing task scenarios, task types, and service level constraints based on the data stream; S3: Receive the scene profile and select a target large model from at least one large model deployed in the all-in-one machine for reasoning based on the scene profile to generate a decision result; S4: Receive the decision result and output the decision result to the execution unit or the human-computer interaction interface.

[0016] This invention provides a large-scale integrated model for decision-making and reasoning in mining intelligent agents. It offers the following advantages: 1. This invention, by constructing a scenario profile that includes task scenarios, task types, and service level constraints, can transform abstract mining business requirements into precise and calculable performance indicators. Based on this, model selection is performed, enabling the integrated machine to adaptively process different types of tasks, improving the relevance of decision-making and the effectiveness of resource utilization.

[0017] 2. This invention establishes a comprehensive model scoring mechanism based on multi-dimensional matching degree, which comprehensively weighs multiple dimensions of model characteristics such as inference latency, accuracy, resource consumption, and scenario expertise during model scheduling. This allows for the matching of the most suitable large model to the current task, avoiding resource waste by high-performance models on simple tasks or the inability of low-performance models to meet the requirements of critical tasks, thus achieving optimal scheduling under limited edge resources.

[0018] 3. This invention introduces a risk-driven multi-model collaborative reasoning and security arbitration mechanism. For high-risk decision-making tasks, it employs multiple large models for parallel reasoning and verifies the consistency and confidence of the results. This mechanism can promptly detect anomalies by comparing results when a single model makes a misjudgment, and effectively avoid the automatic execution of errors through arbitration or degradation processing, thereby improving the reliability and security of the system's decision-making in critical security scenarios. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; The module consists of: 10. Data acquisition and preprocessing module; 20. Scene recognition and SLA generation module; 30. Model scheduling and inference module; and 40. Result output and execution module. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: Please see the appendix Figure 1 - Appendix Figure 2 The present invention provides a large model integrated machine for decision-making and reasoning of intelligent agents in mines, including: a data acquisition and preprocessing module 10, a scene recognition and SLA generation module 20, a model scheduling and reasoning module 30, and a result output and execution module 40.

[0022] Reference Figure 1 , Figure 1 The present invention relates to a method for operating a large-scale integrated model for decision-making and reasoning of intelligent mining agents, comprising the following steps: S1, the data acquisition and preprocessing module 10 receives real-time data from the mining operation environment and preprocesses the real-time data to generate a standardized data stream.

[0023] S2, the scene recognition and SLA generation module 20 receives the standardized data stream, analyzes it to identify the task scene and task type, and generates a structured scene profile containing service level constraints.

[0024] S3, the model scheduling and inference module 30 receives the scene profile and, based on the service level constraints therein, filters and selects one or more models from the local model profile library; then, it executes the corresponding inference strategy according to the task type to generate decision results.

[0025] S4, the result output and execution module 40 receives the decision result and converts it into control instructions or structured information for output, so as to complete the closed-loop execution of the decision or human-computer interaction.

[0026] The data acquisition and preprocessing module 10 is designed to establish communication links with various external data sources in the mining environment. To this end, the all-in-one machine is equipped with multiple standard data interfaces, such as an RJ45 interface for Ethernet communication, an RS-485 / RS-232 interface for fieldbus or serial communication, and a fiber optic interface for long-distance or high-bandwidth data transmission. Through these physical interfaces, the all-in-one machine can establish physical connections with different networks and devices.

[0027] To handle heterogeneous communication protocols used by different devices and systems, this module integrates a protocol adaptation layer. This adaptation layer contains a set of configurable protocol drivers or parsers, enabling it to parse various industrial and network communication protocols.

[0028] When a new data source is connected, the corresponding protocol driver and communication parameters, such as device address, port number, and data point table, can be specified for that data source through the system configuration interface. For non-standard or proprietary protocols used by some devices, this module provides a secondary development interface, allowing customized parsing plugins to be written according to the communication specifications provided by the device manufacturer and dynamically loaded into the protocol adaptation layer to achieve compatibility with specific devices. In this way, the module can deconstruct the original sensor values ​​or device status information from data packets carried by different protocols, providing a unified data foundation for subsequent processing.

[0029] After acquiring raw data from the protocol adaptation layer, the data acquisition and preprocessing module 10 immediately performs data cleaning and normalization processing. This aims to transform input data of varying quality into a high-quality, structurally consistent data stream for use by subsequent modules. This processing mainly includes the following steps: Data cleaning. This module performs cleaning operations to eliminate noise, anomalies, and missing values ​​in the raw data. For abnormal data points introduced by momentary sensor malfunctions or electromagnetic interference, filtering algorithms such as moving average filtering, median filtering, or statistical methods (such as the 3-sigma criterion) can be used to mark or remove data that exceeds the normal range, based on preset rules. For missing data caused by communication interruptions or equipment offline, this module fills in the missing data according to the data type and preset strategies. For example, linear interpolation can be used to fill in continuously changing physical quantities such as gas concentration and temperature; for status data such as switch quantities, the status value of the previous valid time step can be used.

[0030] Data normalization. The cleaned data will undergo a normalization process to ensure consistent format and units. This module converts data records from diverse sources into a predefined structured data format. Specifically, each valid data record is constructed as a key-value pair set containing a unified timestamp, a unique data source identifier, a metric name, and a metric value. Simultaneously, this module performs unit conversions, such as converting pressure units reported by different devices to international standard units (OSUs).

[0031] Timestamp alignment. To ensure the time-series comparability of multi-source data and the accuracy of fusion analysis, this module performs timestamp alignment. The system uses the integrated machine's own system clock as a unified time base, which can be synchronized with the mine dispatch center's master clock via a network time protocol. The timestamps of all incoming data are calibrated to this unified time baseline, thereby eliminating time deviations introduced by network latency or differences in device clocks.

[0032] After the above processing, the data acquisition and preprocessing module 10 outputs a stable, reliable and consistent standardized data stream, which is then transmitted to the scene recognition and SLA generation module 20.

[0033] After receiving a standardized data stream, the scene recognition and SLA generation module 20 performs automatic recognition of task scenarios and types. This process aims to parse events or requests with clear business meaning from continuous data and classify them.

[0034] This recognition process can be implemented based on a pre-configured rule engine, which stores a series of association rules. Each rule maps specific data features to a scene identifier and task type. The specific recognition steps are as follows: Scene identification and recognition. The scene recognition and SLA generation module 20 compares the features in the input data stream with the rule base through pattern matching. In a specific implementation, the rule base defines the triggering conditions for a specific scene.

[0035] Task type identification, in determining Subsequently, the scene recognition and SLA generation module 20 further determines its... The criteria for determining the task type may include the urgency of the event or the semantics of the task request. For example, if If there is an abnormality in the gas at the mining face, and the gas concentration exceeds the emergency threshold requiring immediate action, then... It is classified as an automatic decision-making category. If the concentration value is only at the warning level, then... The request is classified as a decision-making suggestion, requesting ventilation adjustment recommendations from the system. For task requests input by the user via natural language, this module can determine its type through keyword matching; for example, requests containing words like "analysis" or "statistics" will be identified. It was classified as an analysis class.

[0036] As another implementation of the rule engine, the aforementioned recognition process can also be accomplished using a lightweight, pre-trained classification model. This model takes key features from the data stream as input and directly outputs the corresponding scene identifier and task type. The training and deployment of this classification model are well-known technologies in this field and will not be elaborated upon here.

[0037] After recognition is completed, the generated and It will be used as part of the scene profile for subsequent processing.

[0038] After determining the task scenario and type, the scenario identification and SLA generation module 20 performs a dynamic quantification generation process for service level constraints. This process, based on a pre-established quantification rule library, will identify... and This is mapped to a set of precise, computable performance metrics. These metrics collectively constitute the service level constraints portion of the scenario profile. The quantification process specifically includes the following steps: Maximum tolerable response delay Quantification. This module is based on task type. Based on the nature of the condition, the corresponding latency limit is retrieved from the rule base and set. For example, for... For emergency tasks involving automated decision-making, such as emergency equipment shutdown, the rule base will assign a very short latency value, such as 200 milliseconds. For decision-making tasks, this value can be relaxed to several seconds; while for analytical tasks, this value can be set to the minute level.

[0039] Minimum acceptable precision Quantification. This module is based on the task scenario. The importance of the associated business determines the minimum accuracy requirement for the decision-making outcome. For example, in safety-related scenarios, such as abnormal gas levels at a mining face, the rule base will set an extremely high accuracy requirement, such as... Not less than 0.99; for scenarios involving production efficiency, such as load optimization of the main transportation system, the accuracy requirement can be appropriately reduced, such as... Not less than 0.95.

[0040] Task risk level Quantification. To dynamically assess the overall risk of a task, this module calculates using a quantitative model. One specific implementation is that the rule base is for each Its potential impact was predefined. and time urgency The baseline value. Risk level. Calculated using the following formula: ; in: This is the final calculated task risk level, normalized to the [0,1] interval.

[0041] The normalized impact level represents the degree of safety or production loss that a mission failure may cause.

[0042] The normalized urgency level represents the degree of urgency at which the task needs to be addressed.

[0043] and Let be the weighting coefficient, satisfying + =1, used to adjust the relative importance of impact and urgency in risk assessment.

[0044] Context length required for reasoning Quantification. This module is based on the task scenario. The required historical information determines the length of context data needed for inference. For example, for scenarios requiring trend prediction, the rule base will specify a longer context length to include enough historical data points; while for events that only need to respond to the current state, a shorter context length will be specified.

[0045] Through the above steps, the scene recognition and SLA generation module 20 has completed the process of... , , , Dynamic quantization of SLA parameters, and with , Together, they form a complete structured scene profile. The data is then transmitted to the model scheduling and inference module 30.

[0046] The model scheduling and inference module 30 internally maintains a model profile library. This library stores the key feature parameters of all available large models deployed within the all-in-one machine in a structured manner, providing a data foundation for subsequent model scheduling decisions. Each model in the library has its own independent model profile file.

[0047] A specific model portrait It can be defined by a set of parameters, and its data structure is as follows: ; in: This represents the model parameter scale, an inherent property of the model.

[0048] The maximum context length supported by the model; this value is an inherent property of the model.

[0049] This represents the average inference latency of the model under standard load.

[0050] This represents the model's accuracy on the benchmark set.

[0051] This represents the average power consumption when the model is running at full load.

[0052] The model is assigned a scenario-specific vector, where each dimension of the vector corresponds to a specific mining operation scenario, and its value represents the degree of specialization of the model in that scenario.

[0053] The process of building and maintaining this model profile library is as follows: Initial model profiling. When the model is deployed to the appliance, the system performs a one-time offline benchmark test. During this phase, the appliance tests the model's inference capabilities using a predefined benchmark dataset within a standardized hardware and software environment to measure and record its initial average inference latency. Precision and average power consumption Scenario-specific vectors The initial value can be preset based on the model's performance on a specific scenario dataset or based on the domain distribution of its training data.

[0054] Dynamic maintenance of model profiles. To ensure that model profiles reflect performance changes during actual operation, the system establishes a dynamic maintenance mechanism. After each decision task is completed, the result output and execution module 40 generates an operation log containing information such as task ID, model used, actual inference time, and final decision confidence. The model scheduling and inference module 30 periodically reads these logs and updates the parameters in the model profile library according to the following method.

[0055] One specific way to implement the update is to use the exponentially weighted moving average method to update the model's performance parameters.

[0056] Similarly, the model's expertise in specific scenarios is also considered. It is also dynamically updated.

[0057] Through this mechanism, the parameters in the model profile library can adaptively reflect the evolution of model performance, ensuring that the basis for scheduling decisions always remains timely.

[0058] In the model scheduling and inference module 30, one of the core steps is to implement a comprehensive model scoring mechanism based on multi-dimensional matching degree. This mechanism aims to perform a refined evaluation of candidate models that meet the basic conditions in order to select the model most suitable for the current task. The scoring process is as follows: Candidate model selection and comprehensive score calculation. First, the model scheduling and inference module 30 is based on the scene profile. The context length required for inference as defined in [the relevant document] A preliminary screening is performed on all models in the model profiling library, selecting those with the maximum supported context length. Less than The models that are eliminated are used to form a set of candidate models.

[0059] Subsequently, this module processes each model in the candidate model set. Calculate a comprehensive score These weighting coefficients are not fixed, but rather vary depending on the task type. Perform adaptive adjustments. For example, for For high-risk tasks involving automated decision-making, the system will assign precision weights. And task emphasis weight Higher values; For analytical tasks, the focus is more on latency weights. and resource weight One specific implementation is that this adaptive adjustment can be achieved through a pre-configured lookup table that maps task types to specific weight vectors.

[0060] Calculate the delayed matching function This function measures how well the model's response speed matches the task's latency requirements. Its calculation formula is: ; in, For the model Average inference delay, This represents the maximum tolerable response latency for the current task. The function's range is [0,1]. When the model latency is much smaller than the latency limit, the function value is close to 1; when the latency is equal to or exceeds the latency limit, the function value is 0.

[0061] Calculate the risk-adaptive precision matching function This function evaluates whether the model's accuracy meets the dynamic requirements under the current task risk. Its calculation formula is: ; in, For the model The reference accuracy, This represents the minimum acceptable accuracy for the task. The risk level is determined by the task. This is a preset risk adjustment coefficient. The significance of this formula is that the higher the risk level of the task, the higher the actual threshold for the accuracy requirements of the model.

[0062] Calculate resource matching function This function evaluates the adaptability of the power consumption and parameter scale required for model operation to the all-in-one machine's hardware resources. The average power consumption when the model is running at full load is... Exceeding the current available power budget for all-in-one PCs If the model does not meet the resource constraints, then the model does not satisfy the resource constraints. The value is set directly to 0. Otherwise, the calculation formula is: ; in, For model parameter scale, As the baseline model parameter scale, , , These are all empirical coefficients used to adjust the function's decay rate. The function value decreases as the model's power consumption and size increase.

[0063] Computation task bias function This function enhances the selection preference for scenario-specific models, particularly for mission-critical tasks. Its calculation formula is as follows: ; in, To the model Scene-specific expertise vector The extracted expertise score corresponding to the current scenario. This is a task type amplification factor, and its value is based on... Settings. For example, for tasks involving automated decision-making, Choose a value greater than 1 to amplify the strength of the expertise model in this score.

[0064] After comprehensively evaluating the candidate models, the model scheduling and inference module 30 executes a risk-driven multi-model collaborative inference and security arbitration mechanism. This mechanism dynamically selects the inference mode based on the inherent risk of the task and verifies the results to ensure the reliability and security of the decision. This process specifically includes the following steps: Execute a differentiated inference strategy based on task type. The model scheduling and inference module 30 determines the task type based on the scene profile. Select a model from the sorted list of candidate models and perform inference: For analytical tasks, the single model with the highest overall score is selected to perform inference.

[0065] For decision-making and suggestion tasks, the two models with the top two overall scores are selected to perform inference in parallel.

[0066] For automated decision-making tasks, to achieve redundant verification, multiple models with the top two or three comprehensive scores are selected to perform inference in parallel.

[0067] Generate a multi-model inference result set. For inference tasks requiring multiple models, the module schedules k selected models to perform inference in parallel and encapsulates the output of each model into a structured result. The results of all models together constitute a result set. Its structure is as follows: ; in, For the model The original output result, The confidence level attached to the model's output of this result. For the model Specialty score in the current scenario.

[0068] Security arbitration for high-risk tasks. For high-risk tasks categorized as "automatic decision-making," the module initiates a security arbitration mechanism. This mechanism determines whether to execute automatic decision-making by evaluating the consistency and confidence levels of multiple model outputs. First, the average confidence level is calculated. Consistency coefficient : ; ; in, It is a difference measure function used to calculate the difference between two output results. and The difference between them has a range of [0,1]. The specific implementation of this function is task-dependent. For example, for classification tasks, the function value is 0 when the output categories are the same and 1 when they are different; for numerical prediction tasks, it can be the absolute value of the normalized difference between the two.

[0069] Subsequently, the module follows the preset consistency threshold. and confidence threshold Perform a judgment. If the condition is met... ≥ and ≥ If the current decision is deemed highly reliable, the system will directly transmit the consistent decision result to the result output and execution module 40. If any condition is not met, it indicates that there is a significant divergence between the models or that the overall confidence level is insufficient. In this case, to mitigate risk, the system will downgrade the execution method of the task and instead perform result fusion.

[0070] Execution results fusion. For decision-making suggestion tasks or tasks that have been downgraded by automated decision-making, the module's execution results are fused to generate decision suggestions for human reference. First, the fusion weights of each model's results are calculated. : ; This weight combines the model's own confidence level with its historical performance in this scenario. Subsequently, the module uses this weight to synthesize multiple results, form a final decision recommendation, and pass it to the result output and execution module 40.

[0071] After receiving the decision results from the model scheduling and inference module 30, the result output and execution module 40 executes the formatted output of the decision results. This process aims to convert internal, standardized decision data objects into a specific format that can be directly received and parsed by specific execution units or human-machine interfaces in the mine.

[0072] This module maintains an output format mapping table. This table establishes the correspondence between task scenarios, decision result types, and target output formats. The output formatting process may include the following steps: The decision result type is parsed and the output format is matched. The result output and execution module 40 first identifies the type of the incoming decision result, which can be a control instruction to be executed directly, a decision suggestion for human reference, or an analysis report for review. Based on this type, the module looks up the corresponding target format from the output format mapping table.

[0073] Perform format conversion. Based on the matched target format, the module converts and encapsulates the decision results.

[0074] If the decision result is a control command and the target is an industrial controller such as a PLC, the module will encapsulate the command (such as device ID, target status value) into a message that conforms to a specific industrial protocol.

[0075] If the decision result is a recommendation or analysis report, and the target is a host computer monitoring software or a mobile terminal, the module converts it into a structured text format, such as JSON or XML. One specific implementation involves filling fields such as the recommendation content, confidence level, and reasons into a predefined JSON template for parsing and visualization by the user interface.

[0076] If the decision result is an analysis conclusion in text form, the module can directly format it as a plain text string and attach metadata such as timestamps and task sources.

[0077] In this way, the output and execution module 40 ensures that its output can be seamlessly integrated into the mine's diverse information and control systems, providing a standardized data interface for the final implementation of decisions.

[0078] After formatting the decision results, the result output and execution module 40 sends the formatted data to the designated external target through the physical data interface configured on the all-in-one machine, so as to complete the interaction with the execution unit and the human-machine interface.

[0079] For control instructions targeting an execution unit, the module directly sends the pre-encapsulated protocol message through the corresponding communication interface. For example, a control instruction for adjusting the frequency of a frequency converter will be encapsulated as a Modbus write register message and sent to a specific address of the PLC unit connected to the frequency converter through the RS-485 interface. After receiving and parsing the message, the PLC will execute the corresponding frequency adjustment action.

[0080] For decision-making suggestions or analysis reports targeting a human-computer interaction interface, this module sends formatted structured data via an Ethernet interface using the TCP / IP protocol to the host computer system or designated mobile terminal application at the mine dispatch center. The host computer or terminal application receiving this data is responsible for parsing the structured data and visualizing it on its user interface, such as displaying the decision content to operators in the form of pop-up alerts, trend charts, or text reports. In this way, system decision-making is combined with human monitoring and intervention.

[0081] During each decision-making task execution, the results output and execution module 40 is responsible for recording key information throughout the entire task process to generate a decision-making process log. This log provides data support for post-event traceability, system performance analysis, and dynamic maintenance of model profiles.

[0082] The system generates structured log entries. For each decision event, a unique event ID is generated. The module then integrates key data from each stage of the decision-making process into a structured log entry around this ID. One specific implementation is that this log entry is a JSON object, which may contain fields such as: Event timestamp; The original data characteristics that triggered the event; A complete scene profile generated by the scene recognition and SLA generation module 20, including the task scene, task type and various SLA constraints; The ID of the final model or model combination selected by the model scheduling and inference module 30; The actual inference time, raw output results, and confidence level of each invoked model; If security arbitration is carried out, record the results of the arbitration process, such as the consistency and confidence assessment values; The final decision or recommendation output by the system; Decision completion timestamp.

[0083] Log storage involves writing generated log entries to non-volatile storage media on the integrated machine, forming persistent log files. These log files are archived and managed by time or size, and can be periodically uploaded to the mine's central data server for backup, depending on the configuration. The recorded performance data, such as actual inference time and decision confidence, serves as input for the model scheduling and inference module 30 to dynamically maintain the model profile library, thus constituting a self-optimization mechanism for system performance.

Claims

1. A large-scale integrated model for decision-making and reasoning in mining intelligent agents, characterized in that: include: The data acquisition and preprocessing module is used to acquire raw data from data sources in the mining operation environment and process the raw data to generate a standardized data stream. The scene recognition and SLA generation module is used to receive the standardized data stream and generate a scene profile containing task scenario, task type and service level constraints based on the data stream. The model scheduling and inference module is used to receive the scene profile and select a target large model from at least one large model deployed in the all-in-one machine for inference based on the scene profile in order to generate a decision result. The result output and execution module is used to receive the decision result and output the decision result to the execution unit or the human-computer interaction interface.

2. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 1, is characterized in that... The service level constraints include at least one of the following: maximum tolerable response latency, minimum acceptable accuracy, task risk level, and context length required for inference.

3. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 1, is characterized in that... The model scheduling and inference module is specifically used for: Maintain a model image library, which stores model images corresponding to the at least one large model; Based on the scene profile and the model profiles in the model profile library, a comprehensive score is calculated for each large model; The at least one large model is ranked based on the comprehensive score, and the target large model with the highest ranking is selected.

4. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 3, is characterized in that... The overall score is determined by a weighted sum of at least one of the following matching functions: delay matching function, precision matching function, resource matching function, and task-biased function.

5. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 1, is characterized in that... The model scheduling and inference module is also used to: select multiple large models to perform inference in parallel based on the task type or task risk level in the scene profile.

6. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 5, is characterized in that... The model scheduling and inference module is also used for: A secure arbitration is performed on the multiple decision results generated by the parallel inference of the multiple large models. The secure arbitration includes evaluating the consistency and confidence of the multiple decision results. When the evaluation results meet the preset conditions, a consistent decision result is output; When the evaluation results do not meet the preset conditions, the results of the multiple decision results are merged.

7. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 3, is characterized in that... The model scheduling and inference module is also used for: Based on the output results and the running logs recorded by the execution module, the model profiles in the model profile library are dynamically updated.

8. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 7, is characterized in that... The model profile includes at least one of the following: Model parameter size, average inference latency, accuracy, and scenario-specific subvectors.

9. The large-scale integrated model for decision-making and reasoning of intelligent agents in mining, as described in claim 1, is characterized in that... The result output and execution module is specifically used to: format the decision result into a preset protocol message or structured data format, and send it through a communication interface.

10. A method for operating a large-scale integrated model for decision-making and reasoning of intelligent mining agents, comprising the large-scale integrated model for decision-making and reasoning of intelligent mining agents as described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Collect raw data from the mining operation environment and process the raw data to generate a standardized data stream; S2: Receive the standardized data stream and generate a scenario profile containing task scenarios, task types, and service level constraints based on the data stream; S3: Receive the scene profile and select a target large model from at least one large model deployed in the all-in-one machine for reasoning based on the scene profile to generate a decision result; S4: Receive the decision result and output the decision result to the execution unit or the human-computer interaction interface.

Images