Coal feeder DCS integrated control method and system

Abstract

The invention provides a coal feeder DCS integrated control method and system, and relates to the field of automatic control. The method comprises the following steps: acquiring multi-dimensional sensor data during operation of a coal feeder by utilizing a DCS (Distributed Control System), and carrying out data cleaning and abnormal value elimination to obtain pre-processed multi-source data; and generating a real-time state evaluation result and a health degree index of the coal feeder through the state evaluation model. The actual coal quantity of the coal feeder is calculated through an intelligent coal quantity calculation and calibration algorithm, one-key intelligent calibration is triggered, and a coal quantity calibration coefficient is obtained and updated. And dynamically adjusting control parameters by applying a self-adaptive PID control strategy, and generating and outputting a closed-loop control instruction of the coal feeder. Standard preprocessing data, control instructions and system states are monitored in real time, abnormity is recognized through a fault diagnosis and fault-tolerant processing mechanism, a standby control mode is automatically switched when faults are detected, and fault feature records are generated and stored at the same time. The technical problems that parameters in the calibration process cannot be monitored and failure reasons are difficult to analyze are solved.

Description

Technical Field

[0001] This application relates to the field of automation control, and in particular to a DCS integrated control method and system for a coal feeder. Background Technology

[0002] In the pulverizing system of a thermal power plant boiler, the coal feeder, as the core equipment for continuous and uniform coal feeding, directly affects the unit's combustion efficiency and safe operation due to its operational stability and control accuracy. Currently, the widely used weighing belt feeders (such as the CS2024 model) in the industry mostly employ independent microcomputer controllers for operation control. These controllers integrate dedicated control boards such as power boards, CPU boards, and motor speed controllers, and complete coal quantity calculation and control by collecting signals from weighing sensors and speed sensors, providing a fundamental guarantee for unit operation. However, in actual use, the existing coal feeder control system lacks effective monitoring methods for its overall calibration process, requiring manual operation. During manual calibration, key parameters such as overall sensor data and calculated coal quantity cannot be visualized in real time, making it difficult to trace the root cause after calibration failure. It is impossible to quickly determine whether the cause is abnormal sensor signals, unstable operating conditions, or coefficient calculation deviations, which not only increases the calibration workload but also reduces the calibration success rate and efficiency. Summary of the Invention

[0003] This application provides a DCS integrated control method and system for a coal feeder, which solves the technical problems of existing technology where calibration process parameters cannot be monitored and the causes of failure are difficult to analyze.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a DCS integrated control method for a coal feeder includes: collecting multi-dimensional sensor data from the coal feeder during operation using a DCS system, performing data cleaning and outlier removal to obtain preprocessed multi-source data; fusing and analyzing the preprocessed multi-source data with historical operating data in a memory database, and generating real-time status assessment results and a health index for the coal feeder through a status assessment model; calculating the actual coal quantity of the coal feeder using an intelligent coal quantity calculation and calibration algorithm based on the real-time status assessment results and the health index, triggering one-click intelligent calibration, and acquiring and updating the coal quantity calibration coefficient; dynamically adjusting control parameters using an adaptive PID control strategy based on preset coal quantity values, real-time coal quantity, and real-time status assessment results, generating and outputting closed-loop control commands for the coal feeder; and monitoring standardized preprocessed data, control commands, and system status in real time, identifying anomalies through fault diagnosis and fault-tolerant processing mechanisms, automatically switching to a backup control mode when a fault is detected, and simultaneously generating and storing fault characteristic records.

[0005] In conjunction with the first aspect mentioned above, one possible implementation involves generating real-time status assessment results and a health index for the coal feeder through a status assessment model. Specifically, this includes: constructing a multi-dimensional feature vector based on preprocessed multi-source data, where the multi-dimensional feature vector includes temporal, spatial, and frequency domain features; constructing a hierarchical status assessment model comprising a pulse coding layer, a feature fusion layer, a memory association layer, and a decision output layer; inputting the multi-dimensional feature vector into the hierarchical status assessment model, where the pulse coding layer converts the preprocessed multi-source data into a pulse sequence; the feature fusion layer extracting local correlations and temporal dependencies from the pulse sequence using spatiotemporal convolution kernels to generate a fused feature map; the memory association layer, with its built-in memory units, storing feature templates for historical normal operation modes and typical fault modes, and outputting the pattern matching probability and confidence level by calculating the similarity between the fused feature map and each memory template; and the decision output layer generating the real-time status assessment result and health index based on the pattern matching probability and confidence level, combined with preset fuzzy inference rules.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves calculating the actual coal quantity of the coal feeder using an intelligent coal quantity calculation and calibration algorithm, triggering one-click intelligent calibration, and obtaining and updating the coal quantity calibration coefficients. Specifically, this includes: determining the confidence level of the current coal quantity measurement system based on real-time status assessment results and health index; if the confidence level is high, entering the main mass flow calculation mode to calculate the initial real-time coal quantity; if the confidence level is low, entering the auxiliary volume calculation mode, calculating the volumetric estimated coal quantity based on the material cross-section model matched from historical data using belt speed and the material density characteristics inverted from vibration sensors. An adaptive data fusion strategy is established, assigning fusion weights to the initial real-time coal quantity or the volumetric estimated coal quantity according to the confidence level and fusion feature map, resulting in the final actual coal quantity output. Based on calibration commands, the coal feeder is controlled to enter calibration mode to form a calibration dataset. The calibration coefficient optimization module aims to minimize the error between the calculated coal quantity and the standard load value, using a fitting algorithm to iteratively solve for the optimal set of coal quantity calibration coefficients.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves using a calibration coefficient optimization module to minimize the error between the calculated coal quantity and the standard load value, and iteratively solving for the optimal set of coal quantity calibration coefficients using a fitting algorithm. Specifically, this includes: controlling the coal feeder to enter calibration mode based on calibration commands; synchronously collecting steady-state output data from each sensor to form a calibration dataset; inputting the calibration dataset into the built-in calibration coefficient optimization module, and using a fitting algorithm to iteratively solve for the optimal set of coal quantity calibration coefficients, with the goal of minimizing the error between the calculated coal quantity and the standard load value. The calibration coefficient set includes at least coefficient K for linear correction and coefficient B for zero-point calibration; updating the optimal set of coal quantity calibration coefficients to the parameter storage area in the DCS control logic, and simultaneously generating a calibration execution report. The report includes coefficient update records, error comparison information, and data quality assessment results.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the process of dynamically adjusting control parameters using an adaptive PID control strategy to generate and output closed-loop control commands for the coal feeder specifically includes: determining the current control condition category based on real-time status assessment results and a health index; the condition category includes at least stable, transient, abnormal, and fault-tolerant conditions. Based on the condition category, the corresponding basic PID parameter set is retrieved from a preset condition parameter mapping library, and the basic PID parameter set and confidence level are corrected according to the pattern matching probability to obtain a weighted PID parameter set. A PID controller is constructed based on the weighted PID parameter set, and the deviation between the preset coal quantity value and the actual coal quantity output is calculated to obtain the preliminary control quantity. The health index is used as a feedback gain adjustment factor to dynamically adjust the gain of the preliminary control quantity to obtain the final control quantity. The final control quantity is converted into a standard current signal through a DCS system and output to the coal feeder frequency converter to form a closed-loop control command.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves identifying anomalies through fault diagnosis and fault tolerance mechanisms. Specifically, this includes: constructing a multi-source feature collaborative diagnosis network, which comprises a time-domain and frequency-domain feature extraction layer, a dynamic correlation weight calculation layer, a fault mode matching layer, and a fault tolerance strategy mapping layer. The time-domain and frequency-domain feature extraction layer processes multi-dimensional sensor data in standardized preprocessed data in parallel, extracting multi-dimensional feature vectors including instantaneous deviation, trend slope, spectral peak value, and covariance matrix eigenvalues. The dynamic correlation weight calculation layer calculates the correlation weight coefficients of each sensor signal under the current operating condition based on real-time state assessment results and health index. The fault mode matching layer incorporates a fault mode memory matrix, calculates the weighted similarity between the current multi-dimensional feature vector and each fault template, and obtains the fault type with the highest matching degree and confidence level. The fault tolerance strategy mapping layer calls the corresponding fault tolerance scheme from a preset fault tolerance strategy library based on the matched fault type and confidence level.

[0010] In conjunction with the first aspect mentioned above, one possible implementation also includes a visualization module: based on real-time data streams and historical data in the memory database, a multimodal operation knowledge graph is constructed with the coal feeder as the core entity. The nodes of the multimodal operation knowledge graph include equipment entity nodes, sensor nodes, control command nodes, status nodes, fault nodes, and time series nodes, while the edges include data flow edges, causal relationship edges, temporal correlation edges, and spatial topology edges. Using a dynamic visualization rendering engine, the real-time coal quantity, control commands, real-time status assessment results, health index, and fault characteristic records in the multimodal operation knowledge graph are integrated and displayed in the form of multi-dimensional panels, trend curves, topology diagrams, heat maps, and fault trees, and interactive historical backtracking and comparative analysis based on time windows are supported.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the visualization module also includes an update mechanism: dynamically updating the node attributes and edge weights in the knowledge graph based on feedback of control command execution effects, fault handling records, and state transition events, and generating a graph analysis report to support predictive maintenance and optimization decisions.

[0012] In conjunction with the first aspect mentioned above, one possible implementation involves using a DCS system to collect multi-dimensional sensor data during the operation of the coal feeder, and then performing data cleaning and outlier removal to obtain preprocessed multi-source data. Specifically, this includes: the DCS system concurrently collecting multi-dimensional real-time data from multiple sensors during the operation of the coal feeder. This multi-dimensional real-time data includes weight signals from weighing sensors, rotational speed signals from speed sensors, vibration spectrum data from vibration sensors, real-time temperature values ​​from temperature sensors, and displacement signals from the driven wheel inspection device. A sliding window-based dynamic threshold filtering algorithm is used to perform primary cleaning of the multi-dimensional real-time data, resulting in pre-cleaned data. Based on the pre-cleaned data, outlier detection and removal are performed by constructing a multivariate covariance matrix to obtain standard data. The dimension of the covariance matrix is ​​equal to the number of sensor types. The outlier determination condition is that at a certain moment, the Mahalanobis distance between the data vector of each sensor and the center of the covariance matrix exceeds a set threshold. Data points exceeding the threshold are marked as outliers and interpolated for compensation. Standard data is aligned and merged according to time series to generate a standardized data stream with a unified timestamp, which is then pushed to the real-time data cache in the memory database in real time. At the same time, a data quality assessment report is generated, which includes data integrity rate, outlier ratio and cleaning efficiency indicators.

[0013] Secondly, a DCS integrated control system for a coal feeder is provided, comprising: a multi-source data acquisition and fusion module, a DCS module, and a control command issuance and execution monitoring module. The multi-source data acquisition and fusion module uses the DCS system to collect multi-dimensional sensor data during coal feeder operation, performs data cleaning and outlier removal to obtain preprocessed multi-source data. The DCS module integrates and analyzes the preprocessed multi-source data with historical operating data in a memory database, generating real-time status assessment results and a health index for the coal feeder through a status assessment model. Based on the real-time status assessment results and health index, the actual coal quantity of the coal feeder is calculated using an intelligent coal quantity calculation and calibration algorithm, triggering one-click intelligent calibration to obtain and update the coal quantity calibration coefficient. According to the preset coal quantity value, real-time coal quantity, and real-time status assessment results, an adaptive PID control strategy is applied to dynamically adjust control parameters, generating and outputting closed-loop control commands for the coal feeder. The control command issuance and execution monitoring module monitors standardized preprocessed data, control commands, and system status in real time, identifies anomalies through fault diagnosis and fault-tolerant processing mechanisms, automatically switches to standby control mode when a fault is detected, and simultaneously generates and stores fault characteristic records.

[0014] This application provides a DCS integrated control method and system for a coal feeder. By incorporating calibration logic into the DCS, it supports automatic calibration under no-load and chain code load conditions. During operation, it can display key parameters such as sensor data and coal quantity calculations in real time. Then, it generates correction and calibration coefficients through least squares fitting and simultaneously outputs an error comparison report. This completely solves the pain points of traditional coal feeders, such as "opaque calibration process and difficulty in analyzing failure causes." The entire calibration process requires no manual intervention, and key parameters can be monitored in real time. This avoids human error and allows for rapid identification of core causes such as abnormal sensor signals, unstable operating conditions, or coefficient calculation deviations when calibration fails, significantly improving calibration success rate and efficiency while significantly reducing calibration workload. Attached Figure Description

[0015] Figure 1 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 2 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 3 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 4 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 5 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 6 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 7 A flowchart illustrating an integrated DCS control method for a coal feeder provided in this application embodiment; Figure 8 This is a schematic diagram of the structure of an integrated DCS control system for a coal feeder provided in an embodiment of this application. Detailed Implementation

[0016] like Figure 1 As shown in the embodiment of this application, a DCS integrated control method for a coal feeder includes: Step 101: Use the DCS system to collect multi-dimensional sensor data during the operation of the coal feeder, and perform data cleaning and outlier removal to obtain preprocessed multi-source data.

[0017] Specifically, this includes: the DCS system acquiring multi-dimensional real-time data from multiple sensors during the operation of the coal feeder in parallel. This multi-dimensional real-time data includes weight signals from weighing sensors, rotational speed signals from speed sensors, vibration spectrum data from vibration sensors, real-time temperature values ​​from temperature sensors, and displacement signals from the driven wheel inspection device. A sliding window-based dynamic threshold filtering algorithm performs initial cleaning on the multi-dimensional real-time data, resulting in pre-cleaned data. Based on this pre-cleaned data, outlier detection and removal are performed by constructing a multivariate covariance matrix to obtain standard data. The dimension of the covariance matrix equals the number of sensor types. The outlier determination condition is that the Mahalanobis distance between the data vector of each sensor and the center of the covariance matrix exceeds a set threshold at a certain moment. Data points exceeding the threshold are marked as outliers and interpolated for compensation. The standard data is aligned and fused according to time series to generate a standardized data stream with a unified timestamp, which is pushed in real-time to the real-time data cache in the memory database. Simultaneously, a data quality assessment report is generated, including data integrity rate, outlier ratio, and cleaning efficiency indicators.

[0018] In some implementations, multiple sensors on the coal feeder are first activated using the DCS system, allowing them to collect data in parallel. The collected data is then immediately preprocessed. During preprocessing, the standard deviation of each signal is calculated based on historical data from the most recent N periods. A filtering threshold is dynamically set, and data exceeding the threshold is smoothed. Next, a multivariate covariance matrix is ​​constructed, and the Mahalanobis distance between each sensor data vector and the matrix center at each moment is calculated. If the distance exceeds a threshold set based on historical anomaly patterns, the data at that moment is marked as an anomaly, and linear interpolation of data from preceding and following moments is used for compensation to obtain standard data. This standard data is then aligned using a unified timestamp, allowing lossless compression algorithms to reduce data volume. Finally, the data is pushed to the real-time cache of the memory database. Simultaneously, a quality report is generated, including indicators such as data integrity rate and the proportion of anomalies.

[0019] The construction of the multivariate covariance matrix specifically includes: In the process of multidimensional sensor data cleaning, assuming that m types of sensor signals are collected (e.g., weighing, rotational speed, vibration, temperature, displacement), the data at each time t can be represented as a vector: ,in Let represent the measurement value of the i-th sensor at time t. Then, we can obtain a segment of historical data as a baseline, assuming we select data from the most recent n times to form a sample matrix: .

[0020] Thus directly through Calculate the mean vector of data from each sensor. Then through Calculate each element in the covariance matrix S ,at this time This represents the covariance between the data from the i-th sensor and the j-th sensor. It is the measurement value of the i-th sensor at time t. It is the mean of the data from the i-th sensor. It is the measurement value of the j-th sensor at time t. It is the mean of the data from the j-th sensor. Therefore, the covariance matrix can be directly expressed as... ,in It is a vector of length n consisting entirely of 1s.

[0021] Step 102: The preprocessed multi-source data is fused and analyzed with the historical operating data in the memory database, and the real-time status assessment results and health index of the coal feeder are generated through the status assessment model.

[0022] In some implementations, a memory database is built and maintained within the DCS to continuously store historical operating data from weighing, speed measurement, vibration, temperature, and driven wheel inspection devices, indexed by timestamp, equipment number, and operating condition type. When new preprocessed multi-source data is input in real time, historical data sequences of the same equipment under similar operating conditions can be retrieved directly from the memory database based on the current time window and equipment identifier, performing time alignment and feature matching. Through a fusion analysis step, real-time and historical data are compared and correlated at the feature level (e.g., calculating the feature distance between real-time features and historical normal patterns, consistency of change trends, and matching degree of abnormal patterns). The fused multidimensional feature vector is then input into a pre-trained state assessment model (using a machine learning model trained on historical data, such as a gradient boosting tree or deep neural network; the training process includes data cleaning, feature engineering, labeling historical states (e.g., normal, wear, minor anomaly, severe anomaly, etc.), and model tuning), allowing the model to generate real-time state classification results and continuous health indices.

[0023] Step 103: Based on the real-time status assessment results and health index, calculate the actual coal quantity of the coal feeder through the intelligent coal quantity calculation and calibration algorithm, and trigger one-click intelligent calibration to obtain and update the coal quantity calibration coefficient.

[0024] The intelligent coal quantity calculation and calibration algorithm refers to an algorithm integrated into the DCS (Distributed Control System). Based on the real-time operating status of the coal feeder and data from multiple sensors, it adaptively calculates a coal quantity closer to the actual value and can automatically or semi-automatically complete the calibration process to correct measurement errors. One-click intelligent calibration refers to triggering the system to automatically execute the entire calibration process through an integrated operation command on the DCS interface, without requiring manual intervention. The coal quantity calibration coefficient is one or a set of parameters used to correct the coal quantity calculation model, including correction coefficient k and accuracy coefficients, which can be obtained and updated during the calibration process to ensure calculation accuracy.

[0025] In some implementations, the reliability of the current measurement system is judged based on real-time status assessment results and health index. If the status assessment is normal and the health index is high, the system uses a classic mass flow calculation model based on weighing and speed sensor signals (the basic model is: real-time coal quantity = weight per unit length of belt × belt speed) as the basis to calculate the real-time coal quantity. Simultaneously, the displacement signal from the driven wheel inspection device is introduced as the basis for speed verification and slippage compensation. If the status assessment indicates a potential anomaly in the weighing sensor or a decrease in the health index, the confidence weight of the weighing signal will be dynamically reduced, and the auxiliary calculation based on the volumetric model will be enhanced. This means that estimations will rely more on belt speed and a preset material accumulation cross-section model to achieve seamless switching and ensure the continuity of coal quantity output.

[0026] Once the operator triggers the one-click intelligent calibration command, the system directly enters calibration mode: the DCS controls the coal feeder to operate stably under no-load and known load (e.g., using chain codes). It automatically collects the output values ​​of each sensor under this stable operating condition and compares them with the known standard load. Through a built-in optimization algorithm (e.g., least squares fitting), a new set of coal quantity calibration coefficients (including correction coefficient k and accuracy coefficient) is automatically calculated to minimize the error between the calculated coal quantity and the standard coal quantity. This optimization process uses a preset objective function (i.e., minimizing the sum of squared errors). After calculating the new coefficients, it automatically updates the relevant parameters in the DCS logic, replaces the old calibration coefficients, and generates a calibration report, recording the coefficient values ​​before and after calibration, error comparisons, and data quality assessments. The entire process is completed automatically under DCS logic control, and the calibration status and key parameters are displayed in real-time through the human-machine interface.

[0027] Step 104: Based on the preset coal quantity value, real-time coal quantity, and real-time status evaluation results, apply an adaptive PID control strategy to dynamically adjust the control parameters, generate and output closed-loop control commands for the coal feeder.

[0028] In some implementations, a preset coal quantity value is used as the setpoint, and the real-time coal quantity after intelligent calculation and calibration is obtained as a process variable feedback, allowing the controller to calculate the deviation between the setpoint and the real-time coal quantity.

[0029] The core logic of the adaptive PID control strategy is to closely integrate the current deviation magnitude and trend with the real-time status assessment results. If the status assessment indicates that the coal feeder is operating smoothly and the health index is high, the PID controller uses a set of preset control parameters optimized for the "ideal working condition" (such as a large proportional gain and a moderate integral time) to pursue fast response and accurate tracking. If the status assessment indicates a slight abnormality (such as slight noise in the weighing signal or a slight tendency for belt slippage) and the health index decreases, the adaptive algorithm will dynamically adjust the control parameters according to preset rules or online adjustment mechanisms.

[0030] The adjustments can be based on a pre-defined fuzzy rule table (constructed during the system design phase based on expert experience and simulation testing; its inputs are the state assessment level and deviation characteristics, and its outputs are the adjustment direction and magnitude of the PID parameters). For example, when the health index falls below a certain threshold, the adaptive logic may automatically reduce the proportional gain to decrease sensitivity to measurement noise, while appropriately increasing the integral time to prevent integral saturation. The controller then uses the adjusted parameters to perform PID calculations and determine a control increment.

[0031] Finally, the control increment is converted into a specific control command (such as a 4-20mA analog signal or a digital frequency command) and output to the speed control actuator of the coal feeder, thus forming a complete closed-loop control loop for coal quantity.

[0032] Step 105: Monitor standardized preprocessed data, control commands and system status in real time, identify anomalies through fault diagnosis and fault tolerance mechanisms, and automatically switch to standby control mode when a fault is detected, while generating and storing fault feature records.

[0033] System status refers to the hardware and software operating conditions monitored by the DCS integrated control system itself, including internal health information such as card communication status, controller load, power supply status, and network status. The fault diagnosis and fault-tolerant handling mechanism is a logical system integrating rule-based judgment, trend analysis, and pattern matching, used to automatically identify various anomalies and faults occurring in the coal feeder control system. Backup control mode refers to a simplified and robust control scheme automatically activated when the main control strategy (such as closed-loop PID control based on precise weighing) fails to operate normally due to a fault; examples include volumetric control mode or fixed speed mode.

[0034] In some implementations, during operation, standardized preprocessed data, currently issued control command values, and the DCS system's own status (such as I / O card communication status and CPU load rate) are continuously monitored in parallel in real time. A fault diagnosis and fault-tolerant processing mechanism is run based on a preset diagnostic rule base. This diagnostic mechanism not only monitors instantaneous values ​​but also analyzes data trends: for example, detecting a continuous deviation between the belt speed and the driven pulley inspection device speed to diagnose "belt slippage." When the diagnostic mechanism identifies any preset abnormal fault mode, the fault-tolerant processing flow is immediately triggered.

[0035] The fault-tolerant processing procedure will first automatically decide and switch to a preset backup control mode based on the fault type and severity level. For example, for "weighing sensor failure" or "speed sensor failure", it will automatically and seamlessly switch to "volume control mode". In this mode, the control commands are calculated based on a preset coal cross-section model and belt speed (or backup speed source), and the coal quantity is maintained stably through PID or simpler proportional control. At the same time, a fault characteristic record is immediately generated. This record automatically captures and packages key data snapshots within a certain period of time before, during, and after the fault, including all relevant sensor raw and pre-processed values, control command sequences, PID parameters, status assessment results, system alarm logs, etc., and attaches a timestamp, fault code, and diagnostic conclusion. This record is then stored in a dedicated fault history database.

[0036] The diagnostic rule base is built upon the analysis of common fault modes of coal feeders (such as sensor signal disconnection, over-range, jamming, belt slippage, motor overload, communication interruption, etc.), and pre-set corresponding judgment conditions and thresholds. For example, when the rate of change of the weighing signal exceeds the physical limit threshold for multiple consecutive cycles, or the signal value remains unchanged for a long time (jamming), or the sensor power supply status is abnormal, it is diagnosed as "weighing sensor fault".

[0037] Another possible implementation includes a visualization module: based on real-time data streams and historical data in a memory database, a multimodal operation knowledge graph is constructed with the coal feeder as the core entity. The nodes of the multimodal operation knowledge graph include equipment entity nodes, sensor nodes, control command nodes, status nodes, fault nodes, and time-series nodes. Edges include data flow edges, causal relationship edges, temporal correlation edges, and spatial topology edges. Using a dynamic visualization rendering engine, real-time coal quantity, control commands, real-time status assessment results, health index, and fault characteristic records from the multimodal operation knowledge graph are integrated and displayed in the form of multi-dimensional panels, trend curves, topology diagrams, heat maps, and fault trees, supporting interactive historical backtracking and comparative analysis based on time windows. Through a graph self-learning update mechanism, the node attributes and edge weights in the knowledge graph are dynamically updated based on the feedback of newly generated control command execution effects, fault handling records, and state transition events, generating a graph analysis report to support predictive maintenance and optimization decisions.

[0038] Among them, the dynamic visualization rendering engine refers to a software module that can receive data in real time and drive the dynamic update of the graphical interface, supporting the integrated display and interactive operation of various chart types (such as curves, panels, topology diagrams, heatmaps, and fault trees). The graph self-learning update mechanism refers to the algorithmic process of automatically adjusting the node attributes and edge weights in the knowledge graph based on newly generated operational data (such as control effect feedback, fault records, and state changes), which is used to achieve the continuous evolution and knowledge accumulation of the graph.

[0039] In some implementations, real-time coal quantity data, issued control commands, real-time status assessment results, health indices, and recorded fault characteristics are extracted from the memory database and associated with corresponding historical data sequences. This allows for the construction of a structured, multimodal operational knowledge graph with the coal feeder as the core entity. Equipment entity nodes are associated with all their sensor nodes; sensor nodes organize data into time-series nodes; control command nodes are associated with their trigger times and execution effects; status nodes record assessment results and health at different time points; and fault nodes are associated with their occurrence time, type, handling measures, and impact range. All types of nodes are connected through data flow edges (e.g., sensor data flowing to status assessment), causal relationship edges (e.g., command execution causing status changes), temporal correlation edges (e.g., data changes before and after a fault), and spatial topology edges (e.g., the physical connection between the coal feeder and the motor and belt).

[0040] This allows for direct access to the dynamic visualization rendering engine, which integrates and renders real-time and historical data from the graph, displaying it as an integrated interface on the DCS operator station. This interface includes a real-time coal quantity curve panel, a control command time sequence table, a status and health dashboard, an equipment topology diagram, a fault heat map distribution, and a fault tree analysis view, allowing operators to perform historical data backtracking and multi-dimensional comparative analysis by sliding the time axis or selecting nodes.

[0041] Finally, during use, the system continuously monitors the coal quantity feedback deviation after the execution of newly generated control commands, the characteristics of newly recorded faults and their processing results, as well as state transition events. Through the graph self-learning update mechanism, the system dynamically adjusts the relevant node attributes (such as updating the health decay coefficient) and edge weights (such as strengthening the correlation between certain faults and sensor anomalies) using feedback data. The system also regularly generates a knowledge graph analysis report that combines text and graph. This report summarizes recent operating modes, potential risk points, and optimization suggestions, supporting engineers in predictive maintenance and strategy optimization.

[0042] It should be noted that when constructing a knowledge graph, it is necessary to ensure that the definitions of nodes and edges conform to the actual process flow and physical connection relationship, so as to avoid introducing invalid or misleading associations.

[0043] Based on the above technical solutions, multi-dimensional sensor data is processed through dynamic threshold filtering and covariance matrix anomaly removal to ensure data accuracy, solving the problems of large interference and low effectiveness of traditional data, and providing a reliable foundation for subsequent analysis. Real-time and historical data can be fused, and a status assessment model can generate status results and health indices, enabling real-time control of equipment operation and solving the problem of unpredictable equipment status. This allows the system to calculate actual coal quantity through intelligent algorithms and trigger one-click calibration, simplifying the calibration process, improving measurement accuracy, and solving the pain points of complex and error-prone traditional calibration. Finally, an adaptive PID control strategy is used to dynamically adjust parameters, achieving precise closed-loop control of coal quantity and improving the poor adaptability of fixed parameters. Fault diagnosis and fault tolerance mechanisms can monitor anomalies in real time and switch to standby mode, while recording fault characteristics, significantly improving system stability and solving the practical problems of difficult fault diagnosis and significant downtime impact, comprehensively optimizing the operating efficiency and maintenance experience of the coal feeder. Simultaneously, by constructing a multimodal operational knowledge graph and achieving dynamic visualization, fusion, and self-learning updates, previously scattered real-time data, historical records, status assessments, and fault information are integrated into a knowledge network with semantic associations and temporal context. This effectively utilizes a graph structure model to organically organize multi-source heterogeneous data, providing an intuitive and comprehensive operational perspective through the fusion of multiple views, and introducing a self-learning mechanism to ensure the knowledge graph continuously evolves with accumulated operational experience. This solves the technical problems of isolated information in traditional DCS monitoring interfaces, reliance on manual correlation for historical data analysis, and difficulty in tracing the root causes of faults. It achieves deep visualization of the coal feeder's operational status, intelligent correlation analysis of historical data, and systematic accumulation and reuse of operational knowledge, thereby improving system transparency, fault analysis efficiency, and predictive maintenance capabilities. It effectively breaks through the bottleneck of "controller black box, difficult fault diagnosis," enabling direct causal tracing through fault trees and trend curves after a fault occurs without disassembling the equipment, quickly identifying the root causes of faults such as sensor disconnection, belt slippage, and control parameter drift, shortening fault handling time, significantly reducing the workload of maintenance personnel, and comprehensively improving the maintainability and operational reliability of the coal feeder control system.

[0044] In one possible implementation of the embodiments of this application, combined with Figure 1-2 As shown, the process of generating real-time status assessment results and health index of the coal feeder through the status assessment model can be achieved through the following steps 201 to 206, which are explained in detail below: Step 201: Construct a multidimensional feature vector based on preprocessed multi-source data. The multidimensional feature vector includes time features, spatial features, and frequency domain features.

[0045] In some implementations, sensor data sequences within a specified sliding time window are extracted from preprocessed multi-source data according to a unified timestamp. This allows for the direct concatenation of all time statistics (mean, variance, slope, etc.), spatial correlation indicators (principal component contribution rate, etc.), and frequency domain distribution parameters (peak frequency, frequency band energy percentage, etc.) in a predetermined order, forming a fixed-dimensional, comprehensive multi-dimensional feature vector, which serves as the unified input for subsequent state assessment models.

[0046] For the construction of time features, for numerical signals such as weighing sensors, speed sensors, and temperature sensors, the mean (reflecting the average level), variance (reflecting the degree of fluctuation), and trend slope (reflecting the trend of change) obtained by linear fitting are calculated within the window respectively; while for the raw spectrum data output by vibration sensors, the average spectrum is calculated within the window.

[0047] For constructing spatial features, real-time data points from different types of sensors, such as weighing, speed measurement, temperature, and driven wheel displacement, within the same time window are used to form a multi-dimensional data vector, and the covariance matrix between these vectors is calculated. Principal component analysis is performed on this covariance matrix to extract the first N principal components and their corresponding variance contribution rates. The magnitude of the variance contribution rate reflects the strength and direction of the overall correlation between data from different sensors.

[0048] The construction of frequency domain features specifically refers to the processing of vibration sensor signals: performing a fast Fourier transform on the vibration time-domain signal (such as acceleration waveform) within the window to obtain its power spectral density; identifying the peak frequency points with significant energy from the power spectrum; and calculating the percentage of energy in the predefined key frequency bands (such as bearing characteristic frequency bands and belt resonance frequency bands) to represent the specific state of the mechanical component.

[0049] Step 202: Construct a hierarchical state evaluation model that includes a pulse coding layer, a feature fusion layer, a memory association layer, and a decision output layer.

[0050] The network consists of several layers. The pulse coding layer, a neural network layer mimicking the pulse firing mechanism of biological neurons, primarily converts continuous input feature values ​​into discrete pulse sequences. The timing and frequency of pulse firing encode the intensity and variation of the input information. The feature fusion layer is a neural network layer that integrates and refines features from different sources or dimensions through operations such as convolution, thereby extracting more discriminative fused features. The memory association layer is a network layer with built-in accessible storage units (memory units) that store historically learned typical pattern features. The current input feature is matched and associated with these stored patterns based on similarity. The decision output layer is the last layer of the network, responsible for generating the final classification or regression output, such as a status label and health index, based on the processing results of the previous layers and pre-defined rules or mapping relationships.

[0051] In some implementations, the pulse coding layer is designed as a leak integral and fire (LIF) neuron model, where each neuron corresponds to one dimension or a set of related dimensions of the input multidimensional feature vector. Its parameters include membrane potential time constant, threshold potential, etc., and the initial values ​​are set according to the feature dynamic range. The function of this layer is to transform continuous feature values ​​into a series of pulse events on the time axis. Its pulse firing rate is proportional to the offset of the input feature value relative to its resting level.

[0052] The feature fusion layer is built on top of the pulse coding layer, and usually adopts the Spiking CNN structure. A set of trainable convolutional kernels is defined, which slide in the time and feature dimensions to perform convolution operations on the pulse sequence passed from the upper layer and extract local spatiotemporal correlation patterns. The initial weights of the convolutional kernels can be set by random initialization or based on small sample pre-training. During training, the kernels are optimized by algorithms adapted to the pulse network, such as backpropagation in time (BPTT). The goal is to make the fused features more discriminative for different device states.

[0053] The core of the memory association layer is to introduce an updatable memory matrix (i.e., memory unit), where each row vector of the matrix represents a stored historical state pattern feature (such as "normal operation", "slight wear", "belt slippage", etc.). These pattern features are typical feature vector centers learned from the training dataset during the model training phase through clustering (such as K-Means) or autoencoders. During the forward propagation of the model, the output feature vector of the current feature fusion layer is compared with each pattern vector in the memory matrix to calculate the cosine similarity, resulting in a set of similarity scores. An optional attention mechanism is used to weight the scores to focus on the most relevant historical patterns.

[0054] The decision output layer receives the similarity score vector (or attention-weighted vector) from the memory association layer. It primarily consists of a fully connected layer followed by a Softmax function for state classification (outputting the probability distribution of labels such as "normal," "slightly abnormal," and "severely abnormal"), and simultaneously maps the similarity score to a continuous health index between 0 and 100 through another fully connected layer or regressor. The weights and bias parameters of the decision layer are optimized along with other parts of the network during training. The loss function typically combines the cross-entropy loss for state classification and the mean squared error loss for health index prediction.

[0055] Step 203: Input the multidimensional feature vector into the hierarchical state evaluation model, and the pulse coding layer converts the preprocessed multi-source data into a pulse sequence.

[0056] In some implementations, the constructed multidimensional feature vector is used as input to a deployed and loaded hierarchical state evaluation model, at which point the first processing layer (pulse coding layer) of the model begins operation. This allows the feature values ​​to be converted into an input current injected into the neuron upon input of the multidimensional feature vector. The magnitude of this current is proportional to the deviation of the feature value from its preset baseline value (usually the average value under historical normal operating conditions). The neuron then operates according to its membrane potential dynamics equation (including charging, leakage, and threshold comparison processes): the input current causes the membrane potential to rise; without input, the membrane potential decays exponentially over time (leakage); when the membrane potential reaches or exceeds a preset firing threshold, the neuron immediately generates a pulse (output 1), and the membrane potential is then reset to a reset potential and enters a brief refractory period.

[0057] This process is repeated at continuous simulation time steps or discrete time points. Therefore, the change of each feature component over time is encoded by a sequence of pulses generated by its corresponding neuron. The higher the feature value relative to the baseline or the faster it changes, the higher the pulse firing frequency of its corresponding neuron. The feature's pattern information is encoded in the timing pattern of the pulse firing. Ultimately, the pulse coding layer transforms the entire multidimensional feature vector—a static or slowly varying numerical representation—into a dynamic, parallel sequence of pulses.

[0058] Step 204: The feature fusion layer extracts local correlations and temporal dependencies in the pulse sequence through spatiotemporal convolution kernels to generate a fused feature map.

[0059] In some implementations, the parallel pulse sequence output by the pulse coding layer is unfolded in the time dimension to form a three-dimensional tensor, which serves as the input to the feature fusion layer. Its three dimensions are typically represented as batch size, time step, and number of input channels (corresponding to different feature components or sensor types). The feature fusion layer internally defines a set of trainable spatiotemporal convolutional kernels. Each kernel is itself a three-dimensional tensor, its dimensions covering a small time window (time dimension) and some or all of the input channels (spatial dimension or channel dimension).

[0060] During the forward propagation of the model, each spatiotemporal convolutional kernel performs a sliding convolution operation on the temporal and channel dimensions of the input tensor (calculating the dot product between the input value (whether a pulse appears or its intensity) and the kernel weights within the local spatiotemporal region covered by the kernel, adding a bias term, and then typically transforming it through a nonlinear activation function (such as ReLU or its spiking neural network variants). This computation process is repeated at all sliding positions of the input tensor.

[0061] Therefore, for each convolution kernel, the results of its convolution calculation at all positions are aggregated to form a two-dimensional "feature map" (its two dimensions are usually the output channel and the time step). The value at each position reflects the degree of matching between the input pulse sequence and the specific spatiotemporal pattern encoded by the convolution kernel (e.g., "vibration pulses of a specific pattern that occur shortly after the occurrence of the weighing pulse") in the corresponding time segment and channel combination.

[0062] By performing the above operation on multiple convolutional kernels (i.e. multiple different spatiotemporal pattern detectors) simultaneously, multiple feature maps will be generated. These feature maps will then be stacked together to form the final fused feature map.

[0063] Step 205: The memory association layer has built-in memory units that store feature templates of historical normal operation modes and typical failure modes. By calculating the similarity between the fused feature map and each memory template, the pattern matching probability and confidence level are output.

[0064] In some implementations, after the feature fusion layer generates the fused feature map for the current time step, the memo-association layer begins its work: First, it aggregates the fused feature map (which may be a two-dimensional or three-dimensional tensor) within the current time window, for example, through global average pooling or a specific attention convergence mechanism, to obtain a fixed-length "current state feature vector". Then, it calculates the similarity between this current state feature vector and each feature template vector stored in the memory unit (common similarity measures include cosine similarity or the inverse of Euclidean distance). For each template, the original similarity scores of all templates are normalized using the Softmax function, making the sum of all scores equal to 1, thus transforming the original scores into a probability distribution, i.e., obtaining the "pattern matching probability" of the current input matching each historical state pattern. Simultaneously, the entropy of these probability distributions can be calculated, or the difference between the highest and second-highest matching probabilities can be calculated, or a "confidence" score can be comprehensively calculated based on the absolute magnitude of the original similarity scores, thereby measuring the reliability of the current matching result. For example, if the highest matching probability is much higher than other probabilities and its original similarity score is high, the confidence level is high; if all matching probabilities are close and not high, the confidence level is low. Finally, the memory association layer passes the output pattern matching probability vector (representing the probability that the current state belongs to various historical patterns) and a summarized confidence value to the next layer (the decision output layer) for final decision-making.

[0065] The memory units in the memory association layer are constructed as a learnable storage matrix. The construction process includes: using fused feature vectors (usually averaged or aggregated features within a time window) generated from training samples processed by the feature fusion layer and derived from historical normal operation data and labeled typical fault data as the learning object. Through clustering algorithms (such as K-Means) or unsupervised / self-supervised learning methods, several representative cluster centers or prototype vectors are extracted from these historical features. These prototype vectors then constitute the initial "feature templates" stored in the memory units. Each template corresponds to a learned state mode (e.g., "Mode 1: Stable heavy load," "Mode 2: Idle operation," "Mode 3: Moderate belt slippage," "Mode 4: Early bearing wear vibration characteristics," etc.). The number of templates is usually preset as a hyperparameter and can be adjusted according to clustering quality indicators.

[0066] Step 206: The decision output layer generates real-time status assessment results and health index based on pattern matching probability and confidence level, combined with preset fuzzy inference rules.

[0067] In some implementations, the system receives a pattern matching probability vector (denoted as P, containing K probability values ​​corresponding to K historical pattern templates) and a total confidence value (denoted as C). Simultaneously, it maintains a pre-defined fuzzy inference rule base. During the actual decision-making process, the key elements and confidence value C in the current input pattern matching probability P are fuzzified according to a pre-defined membership function to obtain their membership degrees on various linguistic variables. The fuzzy inference engine is then activated, matching the fuzzified input against all rules in the rule base. For each rule, the activation strength of its condition part is calculated (usually through T-norm operations such as minimization or product). Implication operations are then applied to generate a preliminary conclusion fuzzy set based on the activation strength of each rule and the output fuzzy set of its conclusion part. Finally, the conclusion fuzzy sets generated by all activated rules are aggregated (usually using S-norm operations such as maximization or summation) to form a comprehensive conclusion fuzzy set.

[0068] The comprehensive conclusion fuzzy set (for state assessment, this might be a fuzzy membership distribution of state labels; for health index, it's a fuzzy set over a numerical range) is processed. For state assessment results, the maximum membership principle or weighted average method is typically used to select one or a few state labels with the highest membership as the final output, such as "minor anomaly - belt slippage". For the health index, defuzzification methods such as the centroid method can be used to calculate a definite value, which is then combined with a preset health baseline value (e.g., 100 for new equipment) and rule-triggered adjustments to ultimately generate a health index between 0 and 100.

[0069] The fuzzy inference rule base is constructed based on domain expert knowledge, historical fault analysis, and an understanding of the meaning of each pattern template in the memory unit, through manual definition or automatic summarization by rule learning algorithms. Each fuzzy inference rule typically has the form "IF (condition) THEN (conclusion)". The "condition" part usually involves a fuzzy judgment of the input variables, for example: "IF the matching probability of pattern A is high AND the overall confidence is moderate or higher AND (optional: the matching probability of pattern B is low)". Here, "high", "moderate or higher", and "low" are linguistic variables, quantified into values ​​between [0,1] using a predefined membership function. The "conclusion" part defines the state assessment result and health index adjustment that should be output under this condition, for example: "THEN the state assessment result is 'slight anomaly - sensor A drift', and the health index is reduced by 15 points". A rule base contains multiple rules, potentially covering normal, multiple anomaly, and combinations thereof.

[0070] In another possible implementation, the health index can be directly calculated by weighting the pattern matching probability with the preset health influence weight of each pattern template, and then adjusting it in conjunction with the confidence level.

[0071] Based on the above technical solution, by constructing multi-dimensional feature vectors covering time, space, and frequency domains, the system comprehensively characterizes equipment operation data from multiple dimensions, solving the problems of one-sided feature representation and insufficient data validity in traditional single-dimensional features, and providing comprehensive data support for status assessment. A four-layer hierarchical status assessment model is then built. The pulse coding layer transforms preprocessed data into pulse sequences, reducing noise interference and transmission delay; the feature fusion layer extracts data correlations through spatiotemporal convolution kernels, overcoming the problem of isolated and uncoupled data; the memory association layer relies on historical pattern template matching to output probabilities and confidence levels, solving the problem of lack of historical references in status recognition; and the decision output layer combines fuzzy inference rules to generate quantitative results and health indices, filling the gaps in traditional assessments that are fuzzy and lack quantitative indicators. This enables real-time and accurate judgment of equipment operating status and quantitative assessment of health, solving the practical pain points of difficult equipment status prediction and difficulty in tracing the root cause of faults, and significantly improving equipment reliability and maintenance efficiency.

[0072] In one possible implementation of the embodiments of this application, combined with Figure 1-3 As shown, the process of calculating the actual coal quantity of the coal feeder through the intelligent coal quantity calculation and calibration algorithm, triggering one-click intelligent calibration, and obtaining and updating the coal quantity calibration coefficient can be achieved through the following steps 301 to 303, which are explained in detail below: Step 301: Based on the real-time status assessment results and health index, determine the confidence level of the current coal quantity measurement system. If the confidence level is high, enter the main mass flow calculation mode to calculate the initial real-time coal quantity; if the confidence level is low, enter the auxiliary volume calculation mode, and calculate the volume estimated coal quantity based on the material cross-section model matched from historical data by the belt speed and the material density characteristics inverted by the vibration sensor.

[0073] The primary mass flow calculation mode refers to the working mode that calculates the coal quantity in real time based on the signals from weighing and speed sensors and according to the classical mass flow formula (coal quantity = weight per unit length of belt × belt speed). The auxiliary volumetric calculation mode is a backup calculation mode that, when the confidence level of the measurement system is low, does not rely on weighing sensors but instead uses belt speed, material cross-sectional models from historical data, and material density characteristics retrieved from vibration sensors to calculate the volume of material per unit length of the belt and estimate its density to obtain the coal quantity. The material cross-sectional model is a mathematical model that describes the shape and cross-sectional area of ​​the material accumulation on the belt as a function of the coal feed rate, obtained through learning or calibration from historical operating data. The material density characteristics retrieved from vibration sensors refer to the characteristic information that indirectly reflects the density or accumulation state of the currently conveyed material by using the vibration signals of the belt or idler detected by vibration sensors through spectral analysis or feature extraction.

[0074] In some implementations, the confidence level of the current coal quantity measurement system can be determined directly based on the real-time status assessment results and health index, according to preset rules or mapping tables. If the confidence level is high, the system enters the main mass flow calculation mode. At this time, the system directly reads the unit length weight signal output by the weighing sensor and the belt speed signal output by the speed sensor, and calculates the initial real-time coal quantity in real time according to the mass flow calculation formula.

[0075] If the confidence level is low, the system enters the auxiliary volume calculation mode. In this mode, the current belt speed is first acquired. Then, from the historical operating data in the memory database, a material cross-sectional model similar to the current belt speed and coal feeder operating conditions is matched. This model, typically constructed during the initial system commissioning phase through chain code calibration under no-load and different load conditions and regression analysis of historical operating data, characterizes the functional relationship between belt speed, coal feeder output, and material cross-sectional area. Simultaneously, vibration spectrum data collected in real time by vibration sensors is used. A pre-trained inversion model (based on a large amount of historical data, using machine learning algorithms to establish a mapping relationship between vibration characteristic frequency band energy, dominant frequency, etc., and material density or packing state) is used to estimate the density characteristics of the current material. Finally, combining the matched material cross-sectional area, the current belt speed, and the inverted material density, the estimated coal quantity is calculated using the volume estimation formula (coal quantity = cross-sectional area × belt speed × estimated density).

[0076] Step 302: Establish an adaptive data fusion strategy. Based on the confidence level and fusion feature map, assign fusion weights to the initial real-time coal quantity or volumetric estimated coal quantity to obtain the final actual coal quantity output.

[0077] In some implementations, the system receives the confidence level determination result and the fused feature map at the current moment. It can then directly invoke a preset adaptive data fusion strategy (weight allocation function or model). Thus, it can use the confidence level and fused feature map as input to output the fusion weights for the initial real-time coal quantity and the volumetric estimated coal quantity. The assigned weights are then weighted and summed with the corresponding coal quantity values, resulting in: Final actual coal quantity = (Initial real-time coal quantity × First weight) + (Volume estimated coal quantity × Second weight), thereby obtaining a single, continuous, and more reliable coal quantity output.

[0078] The weight allocation model is constructed by collecting a large amount of sample data through historical operation records or specially designed test conditions. Each sample should include: the confidence level determined by the system at that time (which can be discretized into several levels such as high, medium, and low), the fusion feature map of the corresponding moment extracted from the state assessment model (usually a multi-dimensional feature vector), the initial real-time coal quantity value obtained through the main mass flow calculation mode, the volumetric estimated coal quantity value obtained through the auxiliary volume calculation mode, and the verified reference coal quantity value at that moment (i.e., the "true value", which can be obtained through a high-precision temporary measuring device, strict offline calibration, or statistical average under long-term stable operation). The fusion feature map can then be standardized or normalized, the confidence level can be uniquely encoded (if a neural network is used), and the initial real-time coal quantity and the volumetric estimated coal quantity can be uniformly measured as necessary. Then define the model structure: If a lightweight neural network is used, a network containing an input layer, several hidden layers, and an output layer can be designed. The input is the concatenated confidence level code and the fused feature map vector, and the output is two weight values ​​between 0 and 1 that sum to 1. If a rule-based fuzzy logic system is used, the membership function of the input variables (confidence level and some key fused feature map dimensions), the fuzzy rule base (e.g., "IF confidence level IS low AND feature map stability index IS low THEN main weight IS small, auxiliary weight IS large") and the defuzzification method need to be predefined.

[0079] The loss function is redefined: the mean squared error (MSE) between the fused coal quantity and the reference coal quantity is used as the loss function. For neural networks, backpropagation and gradient descent optimizers (such as Adam) are used for iterative training on the training dataset, updating the network weights by minimizing the loss function. For fuzzy logic systems, intelligent optimization algorithms such as genetic algorithms and particle swarm optimization may be used to optimize the parameters in the rule base (such as the center point and width of the membership function, or the weights of the rule conclusion part) to minimize the same defined error. Finally, during training, the dataset is divided into training, validation, and test sets. The validation set is used to monitor model performance and prevent overfitting, while the test set is used to evaluate the generalization ability of the final model. After the model training is completed, its performance needs to be evaluated on an independent test set to ensure that the assigned weights effectively reduce the fusion error and that the weight changes conform to physical intuition and operational experience. Finally, the trained model or optimized fuzzy system is integrated into the DCS control logic to achieve online dynamic weight allocation.

[0080] Step 303: Based on the calibration command, control the coal feeder to enter the calibration working condition to form a calibration dataset. Through the calibration coefficient optimization module, with the goal of minimizing the error between the calculated coal quantity and the standard load value, use the fitting algorithm to iteratively solve for the optimal set of coal quantity calibration coefficients.

[0081] The process of optimizing the calibration coefficients aims to minimize the error between the calculated coal quantity and the standard load value. It involves iteratively solving for the optimal set of coal quantity calibration coefficients using a fitting algorithm. Specifically, this includes: controlling the coal feeder to enter calibration mode based on calibration commands; synchronously collecting steady-state output data from various sensors to form a calibration dataset; inputting the calibration dataset into the built-in calibration coefficient optimization module; and using a fitting algorithm to iteratively solve for the optimal set of coal quantity calibration coefficients, which includes at least coefficient K for linear correction and coefficient B for zero-point calibration. Finally, updating the optimal set of coal quantity calibration coefficients to the parameter storage area in the DCS control logic and simultaneously generating a calibration execution report, which includes coefficient update records, error comparison information, and data quality assessment results.

[0082] In some implementations, upon receiving a one-click intelligent calibration command, the DCS controls the coal feeder to enter a preset calibration sequence. For example, it first controls the feeder to run at a certain reference speed under no-load conditions. Once the operation stabilizes and the fluctuations in sensor data are less than preset thresholds, the system automatically or prompts the operator to apply a standard load of known mass (such as a chain code), and then waits for the operation to reach a new state. Under each steady-state condition, the system synchronously collects the output data of all relevant sensors over a period of time (e.g., 30 seconds), including weight signals from weighing sensors and speed signals from tachometers, forming a calibration dataset containing multiple pairs of "sensor data and corresponding standard load values." A predefined coal quantity calculation model can then be used, typically in the form: Calculated coal quantity = K × (sensor measurement) + B. This calibration dataset is input into the coal quantity calculation model, and the values ​​of K and B are continuously adjusted to minimize the error until the optimal solution is found. After obtaining the optimal set of coal quantity calibration coefficients, these new coefficients are automatically updated to the storage area specifically for storing coal quantity calculation parameters in the DCS control logic, replacing the old calibration coefficients. At the same time, the report generation function is invoked to automatically generate a structured calibration execution report based on the calibration data, the comparison of old and new coefficients, the error analysis of the calculated coal quantity and standard load value (such as maximum error, average error, root mean square error), and the quality assessment results of the collected data (such as stability and signal-to-noise ratio), for operation and maintenance personnel to review and archive.

[0083] Based on the above technical solution, confidence levels are determined by real-time status assessment results and health index. At high confidence, the main mass flow calculation mode is used; at low confidence, it switches to an auxiliary volumetric mode based on belt speed, historical material cross-section models, and vibration-derived density. This solves the problem of inaccuracy in single measurement modes when sensors malfunction, ensuring the continuity of coal quantity measurement. Furthermore, an adaptive data fusion strategy dynamically allocates fusion weights for the two types of coal quantities based on confidence level and fusion feature map, further improving the accuracy of actual coal quantity output and overcoming the problem of poor adaptability between different measurement data. Finally, through multi-sensor steady-state data acquisition under calibration conditions, combined with an iterative solution of the linear correction coefficient K and zero-point calibration coefficient B using a fitting algorithm aimed at minimizing error, the calibration process is made precise and automated. This solves the pain points of uncontrollable parameters, large errors, and reliance on manual labor in traditional calibration, comprehensively improving the reliability and operational efficiency of the coal feeder control.

[0084] In one possible implementation of the embodiments of this application, combined with Figure 1-4 As shown, the process of dynamically adjusting control parameters and generating and outputting closed-loop control commands for the coal feeder using an adaptive PID control strategy can be achieved through the following steps 401 to 405, which are explained in detail below: Step 401: Based on the real-time status assessment results and health index, determine the current control condition category. The control condition category includes at least stable condition, transitional condition, abnormal condition and fault-tolerant condition.

[0085] In some implementations, the system receives real-time status assessment results and health indices from a hierarchical status assessment model. A pre-defined operating condition judgment rule base is then invoked. The current real-time status assessment results and health indices are input into this rule base for matching. Each rule is iterated through sequentially, and if the condition is met, the corresponding operating condition category is output. Simultaneously, the system checks whether a pre-defined specific fault alarm (such as sensor disconnection or communication interruption) has been triggered in the real-time data stream. If triggered, it is directly determined to be a fault-tolerant operating condition.

[0086] The construction of the operating condition judgment rule base includes: collecting a large amount of historical operating data and labeling it with corresponding operating condition categories based on historical operating data, expert experience, and simulation tests; extracting features such as status assessment labels, health index, and volatility of key sensor data; and using decision tree or rule induction algorithms to learn the mapping relationship between features and operating condition categories, forming a judgment rule set in the form of "IF-THEN", such as "IF Status Assessment = 'Normal' AND Health Index > 85 THEN Operating Condition Category = Stable Operating Condition".

[0087] Step 402: According to the operating condition category, call the corresponding basic PID parameter set from the preset operating condition parameter mapping library, and modify the basic PID parameter set and confidence level according to the pattern matching probability to obtain the weighted PID parameter set.

[0088] The operating condition parameter mapping library refers to a database or data table pre-stored in the DCS, which stores a set of basic PID parameters corresponding to different operating condition categories (such as stable operating condition, transient operating condition, abnormal operating condition, and fault-tolerant operating condition). Each parameter set includes three basic values: proportional coefficient K, integral time Ti0, and derivative time Td0. The construction process involves, at the beginning of the project, replaying and analyzing the control effect under different operating conditions based on the historical operating data of the coal feeder. Using expert experience combined with control theory (such as the critical proportional method and trial and error method), a set of relatively stable and high-performance PID parameters is tuned as initial basic values ​​for each type of operating condition. Then, through small-scale pilot tests on a simulation platform or in the field, the control effect under different loads and different coal qualities is tested and fine-tuned, and finally solidified into the mapping library.

[0089] In some implementations, based on the determined operating condition category, the corresponding basic PID parameter set (K, Ti0, Td0) is retrieved from a pre-defined operating condition parameter mapping library. The pattern matching probability vector P=[p1,p2,…,pn] (where n is the total number of historical patterns) and confidence level C from the memory association layer output of the hierarchical state evaluation model are then obtained. Based on a pre-defined table of pattern and parameter weight coefficients (predefined and associated with the influence weight Wi of each historical operating mode on PID parameter adjustment), the weighted correction value Kp for each basic PID parameter (e.g., K) is then calculated. weighted =Kp0×Σ(pi×Wi), where pi is the probability of the current state matching the i-th historical pattern, and Wi is the preset weight coefficient corresponding to the i-th pattern (this coefficient is usually set by experts based on the sensitivity of the historical pattern to the control parameters; for example, the "severe belt slippage" pattern may correspond to a lower Kp weight to reduce the system response speed). Perform the same weighting calculation on Ti0 and Td0 to obtain the weighted parameter set (Kp... weighted Ti weighted ,Td weighted Finally, the weighted parameter set is fine-tuned based on the confidence level C: if the confidence level C is higher than the preset confidence threshold (e.g., 0.8), the pattern matching result is considered reliable, and the weighted parameter set is directly used as the output; if the confidence level C is lower than the threshold, the current state judgment is considered to have uncertainty, and for the sake of system safety and stability, the weighted parameter set is conservatively processed, for example, Kp is reduced. weighted Scaling down proportionally, Ti weighted The weighted PID parameters are increased proportionally to reduce the response speed and overshoot risk of the control system. After this step, the final weighted PID parameter set is obtained.

[0090] Step 403: Construct a PID controller based on the weighted PID parameter set, calculate the deviation between the preset coal quantity value and the actual coal quantity output, and obtain the preliminary control quantity.

[0091] In some implementations, the obtained weighted PID parameter set is loaded into the real-time running PID controller algorithm module. In this case, the PID controller is typically implemented in the DCS as a function block or a custom algorithm module. That is, in each control cycle (e.g., 200ms), the preset coal quantity value (setpoint SP) and the actual coal quantity output (process variable PV) obtained through data fusion are synchronously read. Then, the controller calculates the deviation at the current moment e(t) = SP - PV. Subsequently, the controller executes the complete PID calculation: the proportional term output is Kp. weighted ×e(t); The integral term output is obtained by accumulating the historical deviations and multiplying by a coefficient, i.e., Kp. weighted / Ti weighted )× e(t)dt, the integral term is used to eliminate the steady-state error; the output of the differential term is obtained by calculating the difference between the current deviation and the deviation in the previous cycle and multiplying by a coefficient, that is, Kp weighted ×Td weighted ×[de(t) / dt]. The differential term is used to predict the change trend of the deviation and suppress overshoot. Adding the outputs of the three terms, the preliminary control quantity u(t)=Kp weighted ×[e(t)+(1 / Ti weighted )× e(t)dt+Td weighted ×de(t) / dt].

[0092] Step 404: Use the health index as the feedback gain adjustment factor to dynamically adjust the gain of the preliminary control quantity to obtain the final control quantity.

[0093] In some implementation manners, read the health index HI, and by analyzing historical data, study the change trend of the control gain required to maintain system stability and control accuracy at different health levels. For example, when the device is healthy (HI>80), the control system can withstand a stronger control action to pursue fast response; when the device health deteriorates (for example, 60 < HI ≤ 80), in order to reduce the disturbance to possible minor abnormalities (such as wear), the control intensity should be appropriately weakened; when the health is poor (HI ≤ 60), in order to protect the device and avoid instability caused by control actions, the control gain should be significantly reduced. Based on this analysis, a reference health value H base (for example, 80) and a gain adjustment coefficient α can be set. Then compare the current health index HI with the reference value H base and apply a preset calculation rule (the final control quantity u final = preliminary control u × [1 + α × (HI - H base )]) for dynamic gain adjustment. Where α is a preset gain adjustment sensitivity coefficient (which can be positive or negative), and its sign and magnitude determine the direction and amplitude of the impact of health changes on the control quantity. Usually, when HI > H_base, the adjustment coefficient is greater than 1, enhancing the control action; when HI < H_base, the adjustment coefficient is less than 1, weakening the control action.

[0094] Step 405: Convert the final control quantity into a standard current signal through the DCS system and output it to the coal feeder frequency converter to form a closed-loop control instruction.

[0095] In this context, "standard current signal" refers to a standard analog signal commonly used in industrial automation, typically a 4-20mA DC current signal, used to transmit continuous control commands between the controller and field actuators. "Feder frequency converter" refers to an AC variable frequency speed control device used to drive the feeder belt motor. It controls the motor speed by receiving externally input analog signals (such as 4-20mA), thereby adjusting the coal feed rate. "Closed-loop control command" refers to the final, complete control signal chain that automatically adjusts the feeder speed based on actual coal feed feedback to achieve a preset coal feed target.

[0096] In some implementations, the final control variable u final The signal is transmitted to the DCS system, and the pre-configured control current signal conversion relationship is used to calculate the output current I. out The DCS system then uses its analog output (AO) module in its I / O subsystem to convert the digital command into a stable 4-20mA DC current output. This physical current signal is transmitted to the analog input terminal of the coal feeder frequency converter via a shielded cable. Upon receiving this current signal, the frequency converter, according to its internally set current and frequency curves (usually linear, e.g., 4mA corresponds to 0Hz, and 20mA corresponds to the motor's rated frequency of 50Hz), converts the current signal I... out This is converted into a corresponding output frequency, thereby driving the coal feeder belt motor to operate at a specific speed. Changes in motor speed directly lead to changes in belt speed, which in turn alter the coal feed rate. Simultaneously, the coal feeder's weighing and speed sensors continuously measure the actual coal quantity and feed the signals back to the DCS system, forming a closed loop: the DCS compares the preset coal quantity with the actual coal quantity, identifies the deviation, and then generates new control commands through the aforementioned adaptive PID control strategy. This cycle repeats continuously, achieving closed-loop automatic control of the coal feed rate.

[0097] The conversion relationship between control current signals is a linear or nonlinear mapping preset during the DCS configuration phase based on the technical specifications of the field frequency converter and the control requirements of the coal feeder. It is typically a simple linear calibration process: if the final control quantity's calculation range is [0, 100%], and the corresponding standard current signal range received by the frequency converter is 4-20mA, then the conversion formula is: Output current I... out =4+(u final / 100%)×(20-4)mA.

[0098] Based on the above technical solution, by introducing the pattern matching probability and confidence level output by the hierarchical state evaluation model as the basis for parameter correction, the historical operating experience of the equipment and the credibility of the current state are directly integrated into the control parameter generation process, and dynamic gain adjustment is performed in conjunction with the health index. This realizes the transformation of control parameters from static preset to dynamic adaptation, and from single feedback to multi-source knowledge fusion. It effectively improves the adaptability and robustness of the PID controller to the complex and variable operating conditions of the coal feeder (such as belt slippage, sensor malfunction, and equipment aging), and solves the technical problems of traditional fixed-parameter PID controllers where control performance degrades and is prone to overshoot or oscillation when equipment state changes. Thus, while ensuring control accuracy, it enhances the stability and reliability of the system.

[0099] In one possible implementation of the embodiments of this application, combined with Figure 1-5 As shown, the process of identifying anomalies through fault diagnosis and fault tolerance mechanisms can be achieved through the following steps 501 to 505, which are explained in detail below: Step 501: Construct a multi-source feature collaborative diagnosis network, which includes a time-domain and frequency-domain feature extraction layer, a dynamic correlation weight calculation layer, a fault mode matching layer, and a fault tolerance strategy mapping layer.

[0100] During the construction process, the network architecture was first designed as a four-layer progressive structure, consisting of a time-domain and frequency-domain feature extraction layer, a dynamic correlation weight calculation layer, a fault mode matching layer, and a fault tolerance strategy mapping layer.

[0101] For the time-domain and frequency-domain feature extraction layer, an independent feature extraction channel is configured for each type of sensor. The weighing and speed measurement sensor channels use a sliding window mechanism to calculate instantaneous deviation and trend slope, while the vibration sensor channel extracts the spectral peak value and frequency band energy ratio through fast Fourier transform. All sensor data are sent to the covariance analysis module to calculate feature values, and finally spliced ​​together to form a unified multidimensional feature vector.

[0102] In the dynamic association weight calculation layer, the status assessment results and health index are received in real time. The corresponding templates in the historical normal operation mode library are called to calculate the Mahalanobis distance between the current sensor data and the template. The association weight coefficients of each sensor are dynamically allocated according to the preset index weight formula, so that the weight of sensors with low confidence is reduced and the weight of sensors with high confidence is increased.

[0103] The fault mode matching layer has a pre-set fault mode memory matrix (obtained through training with historical fault data). It stores feature templates and attribute information of different fault types. The system calculates the weighted cosine similarity between the weighted multidimensional feature vector and each template in the matrix, and outputs the fault type with the highest matching degree and the corresponding confidence level.

[0104] The fault tolerance strategy mapping layer retrieves the corresponding processing solution from the predefined fault tolerance strategy library based on the matching results. At this time, the processing solution will clearly specify the control mode switching, backup equipment activation, parameter adjustment range and smooth transition function parameters to ensure that the system output does not change during the execution process. At the same time, this layer will trigger fault data recording and visualization interface updates to complete the closed-loop diagnosis of the entire process from feature extraction to fault tolerance execution.

[0105] Step 502: The time-domain and frequency-domain feature extraction layer is used to perform parallel processing on the multi-dimensional sensor data in the standardized preprocessed data, and extract multi-dimensional feature vectors including instantaneous deviation, trend slope, spectral peak, and covariance matrix eigenvalues.

[0106] In some implementations, a standardized preprocessed data stream is received. For continuous numerical signals such as weighing sensors and speed sensors, an independent time-domain processing unit is allocated to each channel. Each unit uses a sliding window of length N. The arithmetic mean of the data within the window is calculated in real time as a local benchmark. The instantaneous deviation is obtained by subtracting the latest sampled value from this average value. At the same time, the least squares method is used to perform linear regression on all data points within the window to fit the trend slope. These two calculation results reflect the instantaneous anomaly and medium-term trend of the signal, respectively.

[0107] For vibration sensor signals, the vibration time-domain waveform within the sliding window is first windowed (e.g., using a Hanning window) and subjected to a Fast Fourier Transform to obtain its power spectral density map. The frequency point with the highest amplitude is identified as the spectral peak. Simultaneously, based on several key frequency bands predefined by the coal feeder's mechanical structure (e.g., the motor rotation fundamental frequency band, belt meshing frequency band, and bearing fault characteristic frequency band), the percentage of the power spectral integral value in each frequency band relative to the total frequency band integral value is calculated to obtain the energy proportion of each frequency band. Meanwhile, the data vectors from all sensors (including weighing, speed measurement, temperature, displacement, etc.) within the current sliding window are fed in parallel into the covariance analysis module. This module first standardizes the data from each channel to zero mean and unit variance, and then constructs an m×m covariance matrix (where m is the number of sensor types).

[0108] The real symmetric matrix is ​​decomposed into eigenvalues, and the top k largest eigenvalues ​​are extracted as feature vector components representing the joint state of the multi-sensor system. Finally, the instantaneous deviation, trend slope, spectral peak value, energy proportion of each frequency band, and the top k eigenvalues ​​of the covariance matrix calculated from each channel are concatenated in a predetermined order to form a fixed-dimensional multi-dimensional feature vector.

[0109] Step 503: The dynamic correlation weight calculation layer calculates the correlation weight coefficient of each sensor signal under the current operating condition based on the real-time status assessment results and health index.

[0110] In some implementations, the system receives real-time status assessment results and health indices, and retrieves the sensor data template corresponding to the operating condition most similar to the current status assessment result from the historical normal operating mode library in the memory database. Then, for each sensor channel, the mean vector of the sensor data within the current sliding window is calculated and compared with the baseline mean vector of the corresponding sensor in the retrieved historical template to calculate the Mahalanobis distance between the two.

[0111] The Mahalanobis distance is then initially corrected based on the real-time status assessment results. If the status assessment result is "normal operation" and the health index is higher than 85, the Mahalanobis distance of all sensors is reduced proportionally to enhance their weight. If the status assessment result is "sensor abnormal" or the health index is lower than 60, the Mahalanobis distance of the corresponding abnormal sensor will be amplified to reduce its weight.

[0112] The corrected Mahalanobis distance values ​​of each sensor are then substituted into the preset exponential weighting calculation formula. A set of normalized association weight coefficients is obtained, where Let be the weighting coefficient of the i-th sensor. The corrected Mahalanobis distance for the i-th sensor. This is an adjustment factor (range: 0.5~2.0). The total number of sensors, This is the sum of the calculated values ​​for all sensors.

[0113] Step 504: The fault mode matching layer has a built-in fault mode memory matrix. It calculates the weighted similarity between the current multidimensional feature vector and each fault template to obtain the fault type with the highest matching degree and confidence level.

[0114] The fault mode memory matrix is ​​a two-dimensional data structure in which rows represent different fault types and columns represent fault-related feature dimensions. Each element in the matrix stores the typical value or statistic of a specific fault type on the corresponding feature.

[0115] In some implementations, the weighted multidimensional feature vectors are compared one by one with the templates in the fault mode memory matrix. The weighted cosine similarity calculation method is used to find the fault template with the highest matching degree, and the fault type and confidence level are output.

[0116] It should be noted that the system receives a dynamically associated weight coefficient vector and a multi-dimensional feature vector. This allows for the calculation of a weighted cosine similarity between the current multi-dimensional feature vector and each row of the template vector in the fault mode memory matrix. After traversing all fault templates, a set of similarity values ​​is obtained. The fault type corresponding to the maximum value is then found as the preliminary matching result through comparison.

[0117] The construction of the fault mode memory matrix includes: collecting multi-dimensional feature data of various typical faults (such as load cell drift, belt slippage, bearing wear, motor overload, etc.) before and after their occurrence during historical operation; extracting feature data samples under stable abnormal states for each type of fault; calculating the mean or median of each feature dimension; and storing the mean vector as the feature template of this type of fault in the corresponding row of the matrix. At the same time, recording the additional attributes of this type of fault, such as fault level, trigger threshold, and range of affected equipment.

[0118] It should be noted that the confidence score is determined by two factors: first, the difference between the highest and second-highest similarity values ​​(a larger difference indicates higher confidence); and second, the absolute value of the highest similarity value (a higher value indicates higher confidence). The confidence score is specifically calculated using a preset linear or non-linear mapping function, which is determined during the training phase by fitting historical matching data and manual confirmation results. The final output is the fault type with the highest matching degree and its corresponding confidence score. If the highest similarity is lower than a preset matching threshold (e.g., 0.7) and the confidence score is lower than a confidence threshold (e.g., 0.6), then the "Unknown Anomaly" type is output.

[0119] Step 505: The fault tolerance strategy mapping layer calls the corresponding fault tolerance scheme from the preset fault tolerance strategy library based on the matched fault type and confidence level.

[0120] In some implementations, based on the received fault type, one or more candidate fault-tolerance schemes are quickly retrieved from the fault-tolerance strategy library. These schemes are then filtered or ranked based on confidence levels. If the confidence level is higher than a preset high-confidence threshold (e.g., 0.8), the main processing logic defined in that scheme is executed directly. If the confidence level is in the medium range (e.g., 0.5 to 0.8), a more conservative alternative logic may be executed, or enhanced monitoring and alarms may be implemented while executing the main logic. If the confidence level is lower than a low-confidence threshold (e.g., 0.5), only a general alarm may be triggered and recorded without performing complex control switching, awaiting manual confirmation. After selecting a specific scheme, the mapping layer converts the operation instruction sequence, parameter adjustment values, and switching timing information defined in the scheme into a set of instructions executable by the DCS and sends them to the downstream execution modules, simultaneously triggering the recording of the fault handling log.

[0121] The construction process of the fault-tolerant strategy library includes: first, systematically listing all known or possible fault types (such as load cell failure, speed signal loss, belt slippage, bearing overheating, communication interruption, etc.). Then, for each fault type, considering its impact on system safety, control accuracy, and equipment lifespan, designing corresponding fault-tolerant processing logic. For example, for a "load cell failure" fault, the solution might include a series of specific operations such as "switching to volumetric metering mode," "activating backup volumetric calculation parameters," "reducing PID control response speed," and "triggering an alarm and recording." Execution conditions are then defined for each solution, including fault type matching, confidence thresholds, and current operating condition constraints. Finally, all solutions are stored in the fault-tolerant strategy library in a structured format (such as XML or database tables), and an index relationship is established from fault type to solution entry.

[0122] In another possible implementation, the system will also automatically record data snapshots, fault codes, and processing logs for a period of time before and after the fault occurs, store them in the fault history area of ​​the memory database, and visualize them in the form of a fault tree on the DCS operation interface, supporting operators to conduct retrospective analysis.

[0123] Based on the above technical solution, the ability to identify complex faults is enhanced by introducing multi-dimensional feature extraction in the time and frequency domains. Combined with dynamic correlation weight calculation, the accuracy and adaptability of diagnosis are improved. Furthermore, precise fault type localization is achieved through fault mode memory matrix and weighted similarity matching. The fault tolerance strategy library and smooth transition mechanism effectively ensure smooth switching and continuous operation of the system under abnormal conditions. This solves the technical problems of traditional coal feeder control systems, such as reliance on manual fault diagnosis, poor fault tolerance, and easy fluctuations in coal quantity during switching processes, thus improving the system's reliability, maintainability, and automation level.

[0124] This application embodiment can divide a coal feeder DCS integrated control system into functional units based on the above method example. For example, each function can be divided into separate functional units, or two or more functions can be integrated into the same processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0125] When using integrated units, Figure 7 and 8The above-described embodiment illustrates an integrated DCS control system for a coal feeder, comprising a multi-source data acquisition and fusion module, a DCS module, and a control command issuance and execution monitoring module. The multi-source data acquisition and fusion module utilizes the DCS system to collect multi-dimensional sensor data during coal feeder operation, performing data cleaning and outlier removal to obtain preprocessed multi-source data. The DCS module integrates and analyzes the preprocessed multi-source data with historical operating data in a memory database, generating real-time status assessment results and a health index for the coal feeder through a status assessment model. Based on the real-time status assessment results and health index, it calculates the actual coal quantity of the coal feeder using an intelligent coal quantity calculation and calibration algorithm, triggering one-click intelligent calibration to obtain and update the coal quantity calibration coefficient. According to the preset coal quantity value, real-time coal quantity, and real-time status assessment results, it dynamically adjusts control parameters using an adaptive PID control strategy, generating and outputting closed-loop control commands for the coal feeder. The control command issuance and execution monitoring module monitors standardized preprocessed data, control commands, and system status in real time, identifies anomalies through fault diagnosis and fault-tolerant processing mechanisms, automatically switches to a backup control mode when a fault is detected, and simultaneously generates and stores fault characteristic records.

[0126] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A DCS integrated control method for a coal feeder, characterized in that, include: The DCS system is used to collect multi-dimensional sensor data during the operation of the coal feeder, and data cleaning and outlier removal are performed to obtain preprocessed multi-source data. The preprocessed multi-source data is fused and analyzed with historical operating data in the memory database, and the real-time status assessment results and health index of the coal feeder are generated through the status assessment model. Based on the real-time status assessment results and health index, the actual coal quantity of the coal feeder is calculated through the intelligent coal quantity calculation and calibration algorithm, and one-click intelligent calibration is triggered to obtain and update the coal quantity calibration coefficient. Based on the preset coal quantity value, the real-time coal quantity and the real-time status evaluation result, the adaptive PID control strategy is applied to dynamically adjust the control parameters, generate and output the closed-loop control command of the coal feeder. The system monitors the standardized preprocessed data, control commands, and system status in real time, identifies anomalies through fault diagnosis and fault tolerance mechanisms, and automatically switches to standby control mode when a fault is detected, while generating and storing fault feature records.

2. The integrated DCS control method for a coal feeder according to claim 1, characterized in that, The process of generating real-time status assessment results and health index of the coal feeder through the status assessment model specifically includes: A multidimensional feature vector is constructed based on the preprocessed multi-source data. The multidimensional feature vector includes temporal features, spatial features, and frequency domain features. A hierarchical state evaluation model is constructed, comprising a pulse coding layer, a feature fusion layer, a memory association layer, and a decision output layer. The multidimensional feature vector is input into the hierarchical state evaluation model, and the pulse coding layer converts the preprocessed multi-source data into a pulse sequence. The feature fusion layer extracts local correlations and temporal dependencies in the pulse sequence through spatiotemporal convolution kernels to generate a fused feature map. The memory association layer has a built-in memory unit that stores feature templates of historical normal operation modes and typical failure modes. By calculating the similarity between the fused feature map and each memory template, it outputs the pattern matching probability and confidence level. The decision output layer generates real-time status assessment results and health index based on the pattern matching probability and confidence level, combined with preset fuzzy inference rules.

3. The integrated DCS control method for a coal feeder according to claim 2, characterized in that, The process of calculating the actual coal quantity of the coal feeder through an intelligent coal quantity calculation and calibration algorithm, triggering one-click intelligent calibration, and obtaining and updating the coal quantity calibration coefficient specifically includes: Based on the real-time status assessment results and health index, the confidence level of the current coal quantity measurement system is determined. If the confidence level is high, the system enters the main mass flow calculation mode to calculate the initial real-time coal quantity. If the confidence level is low, the system enters the auxiliary volume calculation mode to calculate the volume estimated coal quantity based on the material cross-section model matched from historical data by the belt speed and the material density characteristics inverted by the vibration sensor. An adaptive data fusion strategy is established, and fusion weights are assigned to the initial real-time coal quantity or the volumetric estimated coal quantity based on the confidence level and fusion feature map to obtain the final actual coal quantity output. Based on the calibration command, the coal feeder is controlled to enter the calibration condition to form a calibration dataset. The calibration coefficient optimization module aims to minimize the error between the calculated coal quantity and the standard load value. The optimal set of coal quantity calibration coefficients is obtained by iteratively solving the fitting algorithm.

4. The integrated DCS control method for a coal feeder according to claim 3, characterized in that, The process of using a calibration coefficient optimization module to minimize the error between the calculated coal quantity and the standard load value, and iteratively solving for the optimal set of coal quantity calibration coefficients using a fitting algorithm, specifically includes: The coal feeder is controlled to enter the calibration condition based on the calibration command; the steady-state output data of each sensor is collected synchronously to form a calibration dataset; The calibration dataset is input into the built-in calibration coefficient optimization module. With the goal of minimizing the error between the calculated coal quantity and the standard load value, the optimal set of coal quantity calibration coefficients is obtained by iteratively solving the fitting algorithm. The set of calibration coefficients includes at least coefficient K for linear correction and coefficient B for zero-point calibration. The optimal coal quantity calibration coefficient set is updated to the parameter storage area in the DCS control logic, and a calibration execution report is generated simultaneously. The report includes coefficient update records, error comparison information, and data quality assessment results.

5. The integrated DCS control method for a coal feeder according to claim 4, characterized in that, The process of dynamically adjusting control parameters using an adaptive PID control strategy to generate and output closed-loop control commands for the coal feeder specifically includes: Based on the real-time status assessment results and health index, the current control condition category is determined. The control condition category includes at least stable condition, transitional condition, abnormal condition and fault-tolerant condition. According to the operating condition category, the corresponding basic PID parameter set is called from the preset operating condition parameter mapping library, and the basic PID parameter set and confidence level are corrected according to the pattern matching probability to obtain the weighted PID parameter set. A PID controller is constructed based on the weighted PID parameter set, and the deviation between the preset coal quantity value and the actual coal quantity output is calculated to obtain the preliminary control quantity. The health index is used as a feedback gain adjustment factor to dynamically adjust the initial control quantity to obtain the final control quantity. The final control quantity is converted into a standard current signal through the DCS system and output to the coal feeder frequency converter to form a closed-loop control command.

6. The integrated DCS control method for a coal feeder according to claim 5, characterized in that, The process of identifying anomalies through fault diagnosis and fault tolerance mechanisms specifically includes: A multi-source feature collaborative diagnosis network is constructed, which includes a time-domain and frequency-domain feature extraction layer, a dynamic correlation weight calculation layer, a fault mode matching layer, and a fault tolerance strategy mapping layer. The time-domain and frequency-domain feature extraction layer is used to perform parallel processing on the multi-dimensional sensor data in the standardized preprocessed data, and extract multi-dimensional feature vectors including instantaneous deviation, trend slope, spectral peak, and covariance matrix eigenvalues. The dynamic correlation weight calculation layer calculates the correlation weight coefficient of each sensor signal under the current operating condition based on the real-time status assessment result and health index. The fault mode matching layer has a built-in fault mode memory matrix, which calculates the weighted similarity between the current multidimensional feature vector and each fault template to obtain the fault type with the highest matching degree and confidence level. The fault tolerance strategy mapping layer calls the corresponding fault tolerance scheme from the preset fault tolerance strategy library according to the matched fault type and confidence level.

7. The integrated DCS control method for a coal feeder according to claim 6, characterized in that, It also includes a visualization module: Based on the real-time data stream and historical data in the memory database, a multimodal operation knowledge graph with the coal feeder as the core entity is constructed. The nodes of the multimodal operation knowledge graph include equipment entity nodes, sensor nodes, control command nodes, status nodes, fault nodes and time series nodes, and the edges include data flow edges, causal relationship edges, temporal correlation edges and spatial topology edges. Using a dynamic visualization rendering engine, the real-time coal quantity, control commands, real-time status assessment results, health index and fault characteristic records in the multimodal operation knowledge graph are integrated and displayed in the form of multi-dimensional panels, trend curves, topology diagrams, heat maps and fault trees, and support interactive historical retrospective and comparative analysis based on time windows.

8. The integrated DCS control method for a coal feeder according to claim 7, characterized in that, The visualization module also includes an update mechanism: based on the feedback of the control command execution effect, fault handling records and state transition events, it dynamically updates the node attributes and edge weights in the knowledge graph and generates a graph analysis report to support predictive maintenance and optimization decisions.

9. The integrated DCS control method for a coal feeder according to claim 8, characterized in that, The process of acquiring multi-dimensional sensor data from the coal feeder during operation using the DCS system, and then performing data cleaning and outlier removal to obtain preprocessed multi-source data specifically includes: The DCS system collects multi-dimensional real-time data from multiple sensors during the operation of the coal feeder in parallel. The multi-dimensional real-time data includes the weight signal from the weighing sensor, the rotational speed signal from the speed sensor, the vibration spectrum data from the vibration sensor, the real-time temperature value from the temperature sensor, and the displacement signal from the driven wheel inspection device. The dynamic threshold filtering algorithm based on the sliding window is used to perform primary cleaning on the multi-dimensional real-time data to obtain the data after primary cleaning. Based on the data after the initial cleaning, outlier detection and removal are performed by constructing a multivariate covariance matrix to obtain standard data. The dimension of the covariance matrix is ​​equal to the number of sensor types. The anomaly determination condition is that the Mahalanobis distance between each sensor data vector and the center of the covariance matrix exceeds a set threshold at a certain moment. Data points that exceed the threshold are marked as anomalies and interpolation compensation is performed. The standard data is aligned and fused according to time series to generate a standardized data stream with a unified timestamp, and pushed to the real-time data cache in the memory database in real time. At the same time, a data quality assessment report is generated, which includes data integrity rate, outlier ratio and cleaning efficiency indicators.

10. A DCS integrated control system for a coal feeder, characterized in that, To implement the integrated DCS control method for a coal feeder as described in any one of claims 1-9, the control system specifically includes: a multi-source data acquisition and fusion module, a DCS module, and a control command issuance and execution monitoring module; The multi-source data acquisition and fusion module is used to collect multi-dimensional sensor data during the operation of the coal feeder using the DCS system, and to perform data cleaning and outlier removal to obtain preprocessed multi-source data. The DCS module is used to fuse and analyze the preprocessed multi-source data with historical operating data in the memory database, generate real-time status assessment results and health index of the coal feeder through a status assessment model; based on the real-time status assessment results and health index, calculate the actual coal quantity of the coal feeder through an intelligent coal quantity calculation and calibration algorithm, trigger one-click intelligent calibration, obtain and update the coal quantity calibration coefficient; according to the preset coal quantity value, the real-time coal quantity and the real-time status assessment results, apply an adaptive PID control strategy to dynamically adjust the control parameters, generate and output closed-loop control commands for the coal feeder; The control command issuance and execution monitoring module is used to monitor the standardized preprocessed data, control commands and system status in real time, identify anomalies through fault diagnosis and fault tolerance mechanisms, and automatically switch to standby control mode when a fault is detected, while generating and storing fault feature records.

Images