Coal-fired power plant safety monitoring system and method

Abstract

The invention relates to the technical field of computer programming languages, and particularly discloses a coal-fired power plant safety monitoring system and method. The system comprises a multi-modal data fusion platform, a federal learning agent network and a digital twinborn simulation engine. The edge computing node carries out unified acquisition and feature extraction on multi-source heterogeneous data through multi-protocol conversion, quantum noise suppression and a preprocessing chipset; the multi-modal data fusion platform realizes semantic mapping based on a knowledge graph, and fuses time-space correlation characteristics of data such as an infrared image and gas concentration by adopting a time-space encoder and a cross-modal attention mechanism. According to the method, the problems of early warning delay and high false alarm rate caused by low multi-source data decentralized processing efficiency and insufficient nonlinear correlation analysis in a traditional scheme are solved, efficient data fusion, complex risk accurate prediction and automatic safety response are realized, and the real-time performance and reliability of a coal-fired power plant monitoring system are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and particularly relates to a safety monitoring system and method for a coal-fired power plant. Background Art

[0002] The safety monitoring system of a coal-fired power plant is a core facility to ensure the safe production of the power plant, mainly used for real-time monitoring of coal yard environmental parameters (such as temperature, combustible gas concentration, dust concentration, etc.) to prevent safety accidents such as fires and explosions. The existing system usually consists of multiple types of sensors (such as infrared thermometers, gas detectors, dust monitors), data acquisition devices and a monitoring platform. By collecting real-time data of key areas in the coal yard and combining with early warning algorithms, it judges abnormal states and triggers alarms. Such a system relies on distributed hardware devices and a centralized data processing platform, needs to integrate multi-source heterogeneous data, and realizes visual monitoring and emergency response through a unified interface.

[0003] However, the existing technology has insufficient capabilities for real-time fusion and efficient analysis of multi-source heterogeneous data, resulting in delayed early warning responses and a high false alarm rate. Specifically, in the monitoring scenario of a coal-fired power plant, there are various types of sensors (such as temperature, gas, dust), high data acquisition frequencies and large scales. The existing system usually uses independent data processing modules to process different types of data segment by segment, resulting in inconsistent data formats and increased transmission delays. In addition, the existing early warning models rely on a single algorithm (such as threshold judgment or simple logistic regression), and it is difficult to dynamically adapt to the non-linear correlation characteristics in the complex environment of the coal yard. For example, when parameters such as temperature and gas concentration fluctuate simultaneously, the existing models are prone to misjudgment due to the lack of a multi-dimensional data collaborative analysis mechanism and require manual intervention for correction. The root cause of this problem lies in the dispersion of the data fusion architecture and the limitations of algorithm optimization capabilities, which restricts the real-time performance and accuracy of the monitoring system. Summary of the Invention

[0004] In view of the above technical pain points existing in the prior art, the present invention provides a safety monitoring system, method and device for a coal-fired power plant, which solves the pain points that the existing safety monitoring system for a coal-fired power plant has low real-time data fusion efficiency and insufficient analysis of non-linear correlation characteristics due to the decentralized processing of multi-source heterogeneous data, inconsistent formats and single algorithms, resulting in delayed early warning responses and increased false alarm rates.

[0005] In a first aspect, the present invention provides a safety monitoring system and method for a coal-fired power plant, including: Edge computing nodes, deployed in the pre-monitoring area of the coal yard, for real-time collection of infrared images, gas concentration time series signals and equipment log texts. The edge computing nodes include: Sensor interface module, configured with a multi-protocol conversion circuit and an AD sampling unit, which converts analog signals into digital signals. The AD sampling unit integrates an adaptive sampling frequency control module, dynamically adjusts the sampling frequency (0.1 kHz to 10 kHz) according to the change rate of the dust concentration gradient, and achieves frequency band matching through an NPU+FPGA heterogeneous architecture; Quantum noise suppression circuit, embedded in the sensor interface module, uses a quantum tunneling effect suppression unit to increase the power frequency noise signal-to-noise ratio to ≥35 dB; Preprocessing chipset, performs denoising, normalization, and feature extraction operations on the digital signals to obtain preprocessed data; Communication control unit, uploads the preprocessed data to the central processor through the 5G protocol; Multi-modal data fusion platform, connected to the edge computing node, including: Knowledge graph module, performs semantic mapping on heterogeneous data based on the coal yard safety ontology to generate a standardized data stream; Dynamic fusion analysis module, uses a spatio-temporal encoder and a cross-modal attention mechanism, calculates the cross-attention weights of the infrared image features and the gas concentration signal through a preset Transformer architecture in the dynamic fusion analysis module, and generates a feature vector that fuses thermodynamic correlations; Dynamic causal inference engine, constructs a causal graph of coal yard accidents based on a Bayesian network. When the wind speed > 8 m / s and the coal pile stacking angle > 38°, it activates the dust explosion probability inference path; Federated learning agent network, communicates with the multi-modal data fusion platform, collaboratively generates a risk assessment result based on a consensus algorithm. The federated learning agent network includes a gradient confusion mechanism, injects a random orthogonal matrix before gradient compression, and makes the gradient parameter distribution entropy value ≥6.2; Digital twin simulation engine, receives the risk assessment result, constructs a three-dimensional virtual coal yard to simulate the risk evolution path, and optimizes the early warning model parameters. The digital twin simulation engine includes a physical-data hybrid drive model, embeds an LSTM prediction unit in the discrete element method to correct the simulation error of the particle friction coefficient μ in real time; an actuator control unit, triggers a spray valve or a fan to perform a safety response action according to the output instruction of the digital twin simulation engine.

[0006] Furthermore, in the coal-fired power plant safety monitoring system of the present invention, the preprocessing chipset includes: an image feature extraction module, configured to extract infrared image features through convolutional operations to generate a 128-dimensional feature vector; a data processing module, configured to run an FIR filter to eliminate power frequency interference, and the cut-off frequency of the FIR filter is set to 1.2 times the sensor signal frequency band; wherein, the image feature extraction module and the data processing module process data in parallel through an NPU and an FPGA chip respectively, and transmit the processing results to the communication control unit through a data bus; the communication control unit transmits the data to the mirror module of the coal-fired power plant safety monitoring system, and the mirror module performs: performing simulation operation verification on the received data, and if error data is detected, marking the error data and generating simulated correction data; recombining the simulated correction data with the correct data to generate an available data set.

[0007] Furthermore, in the coal-fired power plant safety monitoring system of the present invention, the dynamic fusion analysis module includes: a spatio-temporal encoder, which encodes the sensor position into a spatial vector using a three-dimensional spherical coordinate system and splices it with the timestamp information to generate spatio-temporal features; a graph neural network, which constructs an adjacency matrix based on the sensor topological relationship, analyzes the implicit associations between multiple parameters, and obtains the implicit association analysis results between multiple parameters; a reinforcement learning agent, which configures weight parameters according to the implicit association analysis results between multiple parameters, and adjusts the real-time adjustment sensor weights of the coal pile volume and wind speed with the configured weight parameters to obtain the adjusted real-time adjustment sensor weights of the coal pile volume and wind speed; The digital twin simulation engine further includes: a multi-objective game optimizer, which coordinates the conflict between fire suppression efficiency and energy consumption through the Nash equilibrium algorithm to generate a Pareto optimal solution set.

[0008] Furthermore, in the coal-fired power plant safety monitoring system of the present invention, the federated learning agent network performs: Local model training: Each agent updates the model parameters on the edge computing node and adds Laplace noise (ε = 0.5) to achieve differential privacy; Gradient compression: Adopt Top-k sparsification to retain 10% important gradients, and combine Huffman coding to reduce the communication data volume; Consensus verification: When the failure rate of regional sensors exceeds 30%, trigger the view switching mechanism of the PBFT protocol to eliminate Byzantine nodes; The federated learning agent network further includes: a dynamic privacy budget allocation module, which allocates differential Laplace noise parameters according to the sensor data type, where the temperature data δ = 1e-4 and the gas concentration data δ = 1e-6; The Byzantine node elimination module triggers the view switching mechanism of the PBFT protocol when the node response timeout is 500 ms and the gradient deviation > 3σ.

[0009] Furthermore, in the coal-fired power plant safety monitoring system of the present invention, the digital twin simulation engine includes: A discrete element method modeling unit that simulates the morphological changes of the coal pile and dynamically adjusts the particle friction coefficient μ ∈ [0.3, 0.7]; A computational fluid dynamics module that predicts the fire spread path based on the Navier-Stokes equation, with a grid division accuracy of 0.5 m³; An automated machine learning component that screens the feature subset through a genetic algorithm and adjusts the CC and γ parameters of the SVM model using a Bayesian optimizer.

[0010] In a second aspect, the present invention provides a coal-fired power plant safety monitoring method, which is applied to the coal-fired power plant safety monitoring system described above, and includes the following steps: Step S1: Real-time collect infrared images, gas concentration time series signals, and equipment log texts through edge computing nodes, and preprocess the collected data to generate preprocessed data; Step S2: Perform semantic mapping on the preprocessed data based on the knowledge graph to generate a standardized data stream; Step S3: Use a spatio-temporal encoder to extract the spatio-temporal features of the standardized data stream, and calculate the correlation weight between the infrared image features and the gas concentration signal through a cross-modal attention mechanism to generate a feature vector that fuses thermodynamic correlations; Step S4: Construct a causal graph of coal yard accidents based on a Bayesian network, and activate the dust explosion probability inference path when abnormal wind speed and coal pile stacking angle are detected; Step S5: Aggregate the local model parameters of multiple edge nodes through a federated learning framework, inject a random orthogonal matrix before gradient compression, and generate a global risk assessment result based on a consensus algorithm; Step S6: Construct a three-dimensional virtual coal yard in the digital twin simulation engine, simulate the risk evolution path, and correct the simulation error through a physical-data hybrid drive model; Step S7: Optimize the early warning model parameters according to the simulation results, and trigger the spray valve or fan to execute safety response actions.

[0011] Furthermore, in the coal-fired power plant safety monitoring method of the present invention, the step S1 includes: Step S11: Receive sensor signals through a multi - protocol conversion circuit and dynamically adjust the AD sampling frequency according to the change rate of the dust concentration gradient; Step S12: Suppress power - frequency noise through a quantum noise suppression circuit; Step S13: Extract infrared image features through convolution operation to generate feature vectors; Step S14: Eliminate power - frequency interference through a FIR filter and upload the processed data through a 5G protocol.

[0012] Further, in the coal - fired power plant safety monitoring method of the present invention, the step S3 includes: Step S31: Encode the sensor position information using a three - dimensional spherical coordinate system and splice it with a timestamp to generate spatio - temporal features; Step S32: Analyze the topological relationship of the sensor network based on a graph neural network to mine the implicit associations between multi - parameters; Step S33: Dynamically adjust the sensor weight parameters through a reinforcement learning algorithm.

[0013] Further, in the coal - fired power plant safety monitoring method of the present invention, the step S5 includes: Step S51: Add Laplace noise to each edge node to achieve differential privacy; Step S52: Adopt Top - k gradient sparsification to compress the communication data volume; Step S53: When a node failure or abnormal gradient deviation is detected, trigger a consensus protocol to eliminate abnormal nodes.

[0014] Further, in the coal - fired power plant safety monitoring method of the present invention, the step S6 includes: Step S61: Simulate the morphological changes of the coal pile through the discrete element method and embed an LSTM prediction unit to correct the friction coefficient in real - time; Step S62: Simulate the fire spread path based on the Navier - Stokes equation; Step S63: Screen the feature subset through a genetic algorithm and optimize the classification model parameters.

[0015] Advantages of the present invention: Through the collaborative work of the multi - protocol conversion, quantum noise suppression of the edge computing node and the pre - processing chipset, the present invention converts heterogeneous data such as infrared images, gas concentration signals, and equipment logs into a unified format, eliminating format conflicts and transmission redundancy caused by decentralized data processing. The semantic mapping of the knowledge graph further realizes the semantic alignment of multi - modal data, breaks the data silos in traditional solutions, significantly improves the data fusion efficiency, and provides a structured input for subsequent analysis.

[0016] The spatio - temporal encoder combines a three - dimensional spherical coordinate system and timestamp encoding to accurately characterize the spatio - temporal characteristics of sensor data; The cross - modal attention mechanism and the graph neural network collaborate to mine the implicit associations between multi - parameters such as infrared images and gas concentrations (such as the non - linear coupling relationship between high - temperature areas and dust concentrations). The reinforcement learning agent dynamically adjusts the sensor weights, enabling the model to focus on key risk factors and solving the problem of insufficient complex association analysis in traditional single algorithms.

[0017] The federated learning agent network reduces the data transmission volume and shortens the aggregation delay of the central processor through distributed computing and gradient compression technologies at the edge nodes; the dynamic causal inference engine avoids misjudgment caused by fluctuations in a single parameter based on the multi-parameter joint probability analysis of the Bayesian network. The digital twin simulation engine corrects simulation errors in real time through a physical-data hybrid model, optimizes the warning threshold, and triggers precise execution actions (such as directional spraying for explosion suppression), significantly improving the response speed and accuracy.

[0018] The automated machine learning component (genetic algorithm + Bayesian optimization) dynamically filters the feature subset and optimizes the model parameters according to environmental changes, reducing the cost of manual parameter tuning; the multi-objective game optimizer coordinates the conflict between safety and energy consumption through the Nash equilibrium algorithm to generate a Pareto optimal solution set (such as a low-energy dust suppression plan). The collaboration between the discrete element method and computational fluid dynamics simulation enables accurate prediction of the coal pile shape and the fire spread path, optimizes resource allocation (such as spraying range, fan power), and reduces operating costs.

[0019] The federated learning framework prevents sensitive data leakage and improves the anti-interference ability of the model through differential privacy (such as Laplace noise injection) and Byzantine node elimination mechanisms; the data verification and correction function of the mirror module solves the problem of data distortion caused by sensor failures or environmental interference, ensuring the reliability of the input data and providing guarantee for long-term stable operation.

[0020] The present invention systematically solves the core problems in the safety monitoring of coal-fired power plants, such as low data fusion efficiency, insufficient mining of non-linear associations, high false alarm rate, and response delay, through a full-chain technical solution of multi-source data unified processing, dynamic correlation analysis, distributed collaborative computing, and digital twin simulation. Its beneficial effects are reflected in improved data fusion ability, accurate early warning of complex risks, optimized resource utilization, and enhanced system robustness, significantly superior to traditional solutions and meeting the comprehensive requirements of industrial scenarios for real-time performance, reliability, and economy. Brief Description of the Drawings

[0021] Figure 1 It is a timing diagram of the safety monitoring system for coal-fired power plants provided by the technical solution of the present invention. Detailed Embodiments

[0022] The following further describes an embodiment of the present invention with reference to the drawings.

[0023] In a first aspect, the present invention provides a safety monitoring system and method for coal-fired power plants, including: Edge computing nodes are deployed in the pre-monitoring area of the coal yard and are used to collect infrared images, gas concentration time series signals, and equipment log texts in real time. The edge computing nodes include: Sensor interface module, configured with a multi-protocol conversion circuit and an AD sampling unit, which converts analog signals into digital signals. The AD sampling unit integrates an adaptive sampling frequency control module, dynamically adjusts the sampling frequency (0.1 kHz - 10 kHz) according to the change rate of the dust concentration gradient, and realizes frequency band matching through an NPU+FPGA heterogeneous architecture; Quantum noise suppression circuit, embedded in the sensor interface module, uses a quantum tunneling effect suppression unit to increase the power frequency noise signal-to-noise ratio to ≥35 dB; Preprocessing chipset, performs denoising, normalization, and feature extraction operations on the digital signal to obtain preprocessed data; Communication control unit, uploads the preprocessed data to the central processor through the 5G protocol; Multi-modal data fusion platform, connected to the edge computing node, including: Knowledge graph module, performs semantic mapping on heterogeneous data based on the coal yard safety ontology to generate a standardized data stream; Dynamic fusion analysis module, uses a spatio-temporal encoder and a cross-modal attention mechanism, calculates the cross-attention weights of the infrared image features and the gas concentration signal through a preset Transformer architecture in the dynamic fusion analysis module, and generates a feature vector that fuses thermodynamic correlations; Dynamic causal inference engine, constructs a causal graph of coal yard accidents based on a Bayesian network. When the wind speed > 8 m / s and the coal pile stacking angle > 38°, activates the dust explosion probability inference path; Federated learning agent network, communicates with the multi-modal data fusion platform, collaboratively generates a risk assessment result based on a consensus algorithm. The federated learning agent network includes a gradient confusion mechanism, injects a random orthogonal matrix before gradient compression, so that the gradient parameter distribution entropy value ≥ 6.2; Digital twin simulation engine, receives the risk assessment result, constructs a three-dimensional virtual coal yard to simulate the risk evolution path, and optimizes the early warning model parameters. The digital twin simulation engine includes a physical-data hybrid drive model, embeds an LSTM prediction unit in the discrete element method to correct the simulation error of the particle friction coefficient μ in real time; an actuator control unit, according to the output instruction of the digital twin simulation engine, triggers the sprinkler valve or the fan to perform a safety response action.

[0024] Specific implementation of the edge computing node: Edge computing nodes are deployed in the key monitoring areas of the coal yard, and their core function is to collect multi-source heterogeneous data in real time. The sensor interface module is compatible with the communication protocols of different sensors (such as Modbus, RS-485) through a multi-protocol conversion circuit, realizing seamless conversion from analog signals to digital signals. The AD sampling unit adopts a dynamic frequency adjustment strategy: when the change rate of the dust concentration gradient is relatively high, the sampling frequency is automatically increased to 10 kHz to ensure rapid capture of transient signals; when the change rate is relatively low, the frequency is reduced to 0.1 kHz to reduce energy consumption. The heterogeneous architecture of the NPU (Neural Network Processor) and FPGA (Field Programmable Gate Array) has a clear division of labor: the NPU is responsible for parallel computing of image feature extraction, while the FPGA is used for hardware acceleration of real-time signal processing. The specific implementation of the quantum noise suppression circuit includes: embedding a quantum tunneling effect unit in the sensor interface circuit, forming a potential barrier through a quantum dot array, filtering out the low-frequency interference components in the power frequency noise, thereby improving the signal-to-noise ratio of the signal. The preprocessing chipset further performs wavelet transform denoising and normalization on the digital signal, and extracts the temperature distribution characteristics in the infrared image through a Convolutional Neural Network (CNN) to generate a 128-dimensional feature vector.

[0025] The coordination mechanism of the multi-modal data fusion platform: The knowledge graph module is based on the coal yard safety ontology library (such as concept entities like "dust explosion conditions", "equipment failure modes", etc.), and performs semantic alignment on heterogeneous data such as infrared images, gas concentrations, and equipment logs. For example, it associates the coordinate information of the "high-temperature area" with the event of "excessive methane concentration" to form standardized semantic labels. The dynamic fusion analysis module uses a spatio-temporal encoder to convert the three-dimensional coordinates (longitude, latitude, altitude) of the sensor location into a spatial vector, and splices it with the timestamp information to construct a spatio-temporal feature matrix. The cross-modal attention mechanism calculates the cross-attention weights between the infrared image features and the gas concentration signals through the Transformer architecture, for example, identifying the spatio-temporal overlap relationship between the high-temperature area and the methane concentration peak, and generating a feature vector that fuses thermodynamic correlations.

[0026] The inference logic of the dynamic causal inference engine is: when the system detects that the wind speed exceeds the threshold (8 m / s) and the coal pile stacking angle is greater than 38°, the Bayesian network automatically activates the "dust explosion" causal path, calculates the explosion probability in the current environment in combination with historical accident data, and outputs the risk level.

[0027] The privacy and fault tolerance mechanism of the federated learning intelligent agent network: The federated learning agent network achieves distributed learning through local model training of edge nodes. When updating model parameters, each node injects Laplace noise (ε = 0.5) into the gradient data to ensure differential privacy protection of sensitive data (such as device operating status). The top-k sparsification strategy is adopted in the gradient compression stage, only the top 10% of important gradient values are retained, and the data volume is further compressed through Huffman coding to reduce the communication bandwidth occupancy. The consensus verification mechanism is based on the PBFT (Practical Byzantine Fault Tolerance) protocol: when the failure rate of regional sensors exceeds 30% or the node response times out (500 ms), the system triggers a view switch and automatically eliminates abnormal nodes to ensure the reliability of global model aggregation.

[0028] Dynamic optimization of the digital twin simulation engine: The digital twin simulation engine simulates the morphological changes of the coal pile through the discrete element method (DEM) and adjusts the particle friction coefficient μ in real time. For example, when the coal pile collapses due to external vibration, the LSTM prediction unit predicts the future change trend based on historical friction coefficient data and dynamically corrects the simulation parameters. The computational fluid dynamics (CFD) module simulates the fire spread path based on the Navier-Stokes equation, with a grid division accuracy of 0.5 m³, and can accurately predict the fire spread direction. The multi-objective game optimizer balances the "spray system energy consumption" and "fire suppression efficiency" through the Nash equilibrium algorithm. For example, at the initial stage of the fire, the low-energy consumption fan is preferentially started, and when the fire expands, it switches to the high-power spray valve to generate the Pareto optimal solution set.

[0029] The existing monitoring system of coal-fired power plants has low data fusion efficiency and insufficient non-linear correlation analysis due to the decentralized processing of multi-source data and single algorithms. This solution realizes efficient data preprocessing through the heterogeneous architecture of edge computing nodes, and combines the knowledge graph and dynamic causal reasoning in the multi-modal fusion platform to solve the semantic alignment and complex association mining problems of multi-source heterogeneous data. The federated learning agent network enhances the model robustness through the consensus mechanism while protecting data privacy; the digital twin simulation engine significantly improves the risk prediction accuracy through the physical-data hybrid-driven model.

[0030] Synergy of the technical solution: Edge computing and 5G transmission: The local processing of edge nodes reduces data transmission latency, and the 5G protocol ensures the efficient upload of preprocessed data.

[0031] Multi-modal fusion and causal reasoning: The knowledge graph provides semantic support, the spatio-temporal coding and attention mechanism realize cross-modal association analysis, and the dynamic causal reasoning engine strengthens the logic of risk decision-making.

[0032] Federated learning and digital twin: Federated learning ensures the secure sharing of distributed data, and the digital twin optimizes the early warning model parameters through simulation to form a closed-loop control.

[0033] Practical application value: This solution significantly shortens the early warning response time and reduces the false alarm rate through real-time risk simulation and automated response (such as triggering the sprinkler valve). For example, when the stacking angle of the coal pile is abnormal, the system can predict the risk of dust explosion 10 minutes in advance and automatically activate explosion suppression measures to avoid delays in manual intervention. In addition, the privacy protection mechanism complies with industrial data security regulations and is applicable to the scenario of collaborative monitoring of multiple power plants.

[0034] The technical details of this solution have been fully disclosed. For example, the specific implementation of the quantum noise suppression circuit, the gradient obfuscation mechanism of federated learning, and the LSTM correction unit of digital twin all conform to the conventional implementation means of those skilled in the art (such as using the open-source framework TensorFlow Federated to implement federated learning). The data flow between modules (such as edge node → fusion platform → simulation engine → actuator) is logically clear, and the technical means are progressive layer by layer, and it can be actually deployed in the coal-fired power plant environment. Specifically, for the coal-fired power plant safety monitoring system described in the present invention, the preprocessing chipset includes: an image feature extraction module for extracting infrared image features through convolutional operations to generate a 128-dimensional feature vector; a data processing module configured to run an FIR filter to eliminate power frequency interference, and the cut-off frequency of the FIR filter is set to 1.2 times the sensor signal frequency band; wherein, the image feature extraction module and the data processing module process data in parallel through NPU and FPGA chips respectively, and transmit the processing results to the communication control unit through a data bus; the communication control unit transmits the data to the mirror module of the coal-fired power plant safety monitoring system, and the mirror module performs: performing simulation operation verification on the received data, and if error data is detected, marking the error data and generating simulated correction data; recombining the simulated correction data with the correct data to generate an available data set.

[0035] Refined implementation of the image feature extraction module: The image feature extraction module uses a lightweight convolutional neural network (such as a MobileNet variant) to extract features from infrared images. Specifically, local temperature gradients, hot spot area contours, etc. in the image are extracted layer by layer through multi-level convolutional kernels (such as 3×3, 5×5), and finally compressed into a 128-dimensional feature vector through a fully connected layer. To adapt to the edge computing environment, this module removes redundant neurons through model pruning technology and uses fixed-point quantization technology to compress the weights from 32-bit floating point to 8-bit integer, reducing the consumption of computing resources. Example: When a local high-temperature area appears on the surface of the coal pile, the convolutional layer can capture the pixel intensity changes in this area, and retain the maximum response value through the pooling layer. The finally generated 128-dimensional vector can characterize key information such as the size and temperature distribution uniformity of the hot spot.

[0036] Power frequency interference suppression logic of the data processing module: The FIR filter design is based on the frequency band characteristics of the sensor signals. For example, the output signals of gas concentration sensors usually contain 50Hz power frequency interference and harmonic components. By setting the cut-off frequency to 1.2 times the signal frequency band (e.g., when the main signal frequency is 100Hz, the cut-off frequency is set to 120Hz), the high-frequency components of the effective signals (such as gas concentration mutation pulses) can be retained while filtering out the power frequency noise. The order of the filter is dynamically adjusted according to the computing power of the edge nodes. For example, a high-order filter is adopted at low load to improve the suppression effect, and a low-order filter is switched to at high load to reduce the delay.

[0037] Example: For the periodic interference in the carbon monoxide concentration signal, the FIR filter reduces the interference amplitude to a negligible range through the preset stopband attenuation parameter (such as -40dB), ensuring that the true characteristics of the concentration fluctuations are not masked by the noise.

[0038] Heterogeneous cooperation mechanism between NPU and FPGA: The NPU (Neural Network Processor) is dedicated to the parallel computing of image feature extraction, and its instruction set is optimized for convolution operations, supporting batch processing of multiple frames of infrared images. The FPGA (Field Programmable Gate Array) realizes the real-time signal processing of the FIR filter through hardware logic circuits, including multiply-accumulate operations (MAC) and pipeline scheduling. The two interact through a high-speed data bus (such as AXI-Stream): The NPU stores the extracted image feature vectors in the shared cache, and the FPGA synchronously writes the filtered sensor data into the same buffer area. The communication control unit reads and encapsulates them into data packets according to the time sequence.

[0039] Example: When processing a frame of infrared image, the NPU completes feature extraction within 5ms, and the FPGA completes the filtering of 10-channel sensor signals within the same cycle. The results of the two are merged through the bus and then uniformly uploaded by the communication unit to avoid processing bottlenecks.

[0040] Data verification and correction process of the mirror module: After receiving the data, the mirror module first simulates the normal operation state of the coal yard environment (such as temperature diffusion model, gas concentration gradient model) through the digital twin model, and compares the incoming data with the simulation results for consistency. If abnormal data is detected (such as a certain sensor value deviating more than 3 standard deviations from the simulation value), it is marked as error data, and simulated correction data is generated through an interpolation algorithm (such as linear interpolation based on spatio-temporal neighboring sensors). After the correction data is aligned with the original correct data according to the time stamp, it is recombined into a complete data set, and the integrity is verified through a check code (such as CRC32) and then transmitted to the central processor.

[0041] Example: When the reading of a certain temperature sensor is abnormally low due to dust occlusion, the mirror module generates a correction value based on the temperature data of adjacent sensors and the heat conduction model, and forms a continuous and reliable data stream after replacing the abnormal data.

[0042] Modular division of labor and collaborative advantages: Decoupling of image and signal processing: The heterogeneous architecture of NPU and FPGA separates computationally intensive tasks (image processing) from real-time tasks (signal filtering), improves the overall efficiency through parallel processing, and avoids resource competition problems in serial processing by the CPU in traditional solutions.

[0043] Dynamic anti-interference ability: The cut-off frequency adaptive mechanism of the FIR filter, combined with the lightweight model of the NPU, ensures that the system can still stably output highly reliable data in a complex electromagnetic environment.

[0044] Dual guarantee of data reliability: Edge preprocessing: Feature extraction and filtering are completed before data upload, reducing the transmission of invalid data and the load on the central processor.

[0045] Mirror verification: By digital twin simulation and data correction, it solves the problem of local data distortion caused by sensor failures or environmental interference, and improves the accuracy of subsequent risk assessment.

[0046] Solving the pain points of the existing technology: Defects of traditional systems: Existing solutions often result in analysis result deviations due to incorrect sensor data or packet loss during transmission, and the incorrect data needs to be manually checked and corrected, with a high response delay.

[0047] Innovation of this solution: Through parallel preprocessing at the edge node and automatic correction by the mirror module, it realizes real-time detection and self-repair of data anomalies, reduces manual intervention, and ensures the continuous operation reliability of the system.

[0048] Actual application example: In the scenario where the dust concentration in the coal yard suddenly increases: The infrared image shows that the temperature in a certain area rises abnormally, and the feature vector extracted by the NPU identifies this area as a high-risk hot spot; The gas sensor outputs an abnormal signal due to dust interference, and the FIR filter filters out the noise and restores the real concentration data; The mirror module detects a conflict between the data of a certain sensor and the simulated value, and automatically interpolates to generate a correction value; The corrected fusion data is uploaded to the central processor, triggering a dust explosion risk assessment and starting the spray explosion suppression.

[0049] Effect: The full-process delay from data collection to execution response is reduced by 60% compared with the traditional solution, and the false alarm rate is reduced to less than 5%.

[0050] Through the co - design of heterogeneous computing, dynamic filtering, and data self - correction mechanisms, this pre - processing chipset solution achieves the efficient pre - processing and reliability guarantee of multi - source data, providing high - quality input for subsequent risk assessment and meeting the core requirements of real - time performance and robustness in industrial monitoring systems.

[0051] Specifically, in the safety monitoring system of coal - fired power plants described in the present invention, the dynamic fusion analysis module includes: a spatio - temporal encoder that encodes the sensor position as a spatial vector using a three - dimensional spherical coordinate system and concatenates it with timestamp information to generate spatio - temporal features; a graph neural network that constructs an adjacency matrix based on the sensor topological relationship, analyzes the implicit associations between multiple parameters, and obtains the analysis results of implicit associations between multiple parameters; a reinforcement learning agent that configures weight parameters according to the analysis results of implicit associations between multiple parameters, and adjusts the real - time sensor weights of coal pile volume and wind speed with the configured weight parameters to obtain the adjusted real - time sensor weights of coal pile volume and wind speed; The digital twin simulation engine further includes: a multi - objective game optimizer that coordinates the conflict between fire suppression efficiency and energy consumption through the Nash equilibrium algorithm to generate a Pareto optimal solution set.

[0052] The specific implementation logic of the spatio - temporal encoder: The spatio - temporal encoder converts the sensor position into a spatial vector in a three - dimensional spherical coordinate system, which specifically includes the following steps: Coordinate system conversion: Convert the physical position (longitude, latitude, altitude) of the sensor into spherical coordinate parameters (radius, polar angle, azimuth angle). For example, taking the center of the coal yard as the origin, the radius represents the distance from the sensor to the center, the polar angle reflects the position difference in the vertical direction, and the azimuth angle identifies the horizontal distribution.

[0053] Timestamp fusion: Convert the acquisition timestamp into a periodic encoding (such as using sine / cosine functions to represent hours and minutes) to avoid model bias caused by the linear growth of time values.

[0054] Feature concatenation: Concatenate the spherical coordinate parameters and the time encoding into a spatio - temporal vector, such as a four - dimensional vector (radius, polar angle, azimuth angle, time encoding), as the input for subsequent multi - modal fusion.

[0055] When a certain wind speed sensor is located in the northeast corner of the coal yard, the spherical coordinate parameters can quantify its spatial distribution characteristics, the time encoding reflects the data acquisition period (such as larger wind speed fluctuations at night), and the spatio - temporal vector can characterize the data change law under the joint influence of "position - time".

[0056] The implicit association mining mechanism of the graph neural network: The graph neural network (GNN) constructs an adjacency matrix based on the sensor topological relationship, specifically including: Adjacency definition: Define node connection weights based on the proximity of the physical locations of sensors (e.g., distance <10 meters), data correlation (e.g., covariance between temperature and gas concentration), or functional coupling (e.g., linkage between fans and dust sensors).

[0057] Graph convolution operation: Aggregate the features of neighboring nodes through a multi-layer graph convolutional network (GCN) to extract nonlinear associations between multiple parameters. For example, a temperature sensor node can identify the coordinated change pattern of "high temperature-dust concentration" by aggregating the features of adjacent dust sensors.

[0058] Implicit correlation output: The output correlation matrix identifies the causal or statistical relationship between parameters, for example, "When the wind speed is > 8m / s, the correlation between the surface temperature of the coal pile and the dust concentration increases to 0.8."

[0059] Example: When the graph neural network detects that the implicit correlation strength between dust concentration and wind speed in a certain area exceeds the threshold, the system can warn of the dust risk in the area in advance and link the spraying equipment to start dust suppression.

[0060] Dynamic weight adjustment strategy for reinforcement learning agents: The reinforcement learning agent dynamically optimizes the sensor weights through the following process: State definition: Environmental status includes real-time parameters (such as wind speed, coal pile volume), implicit association analysis results, and historical accident data.

[0061] Action Space: Weight adjustment actions include increasing / decreasing the confidence weight of a specific sensor (e.g., increasing the wind speed sensor weight from 0.6 to 0.8).

[0062] Reward mechanism: Risk assessment accuracy (such as warning hit rate) and response efficiency (such as the delay from detection to execution) are used as reward functions to guide the model to learn the optimal weight strategy.

[0063] Policy iteration: The Deep Deterministic Policy Gradient (DDPG) algorithm is used to continuously optimize the weight configuration through the Actor-Critic network.

[0064] Example: When the volume of a coal pile changes rapidly due to loading and unloading operations, the reinforcement learning agent automatically increases the weight of volume measurement sensors and decreases the weight of static environmental parameters (such as temperature at fixed locations), ensuring that the risk prediction model focuses on key variables.

[0065] Conflict coordination logic of multi-objective game optimizer: The multi-objective game optimizer generates a Pareto optimal solution set by following the steps below: Goal modeling: Define conflicting goals (e.g., fire suppression efficiency requires maximizing sprinkler intensity, while energy consumption requires minimizing pump power).

[0066] Game participant design: Abstract the goal into a virtual game player (e.g., the "safety side" pursues the fire extinguishing speed, and the "energy consumption side" pursues energy conservation).

[0067] Nash equilibrium solution: Through the iterative negotiation mechanism, find the strategy combination where neither party can unilaterally improve the payoff. For example, at the initial stage of a fire, use a low-power fan to delay the fire spread, and switch to a high-power sprinkler after the fire has spread, balancing efficiency and energy consumption.

[0068] Feasible solution set generation: Output multiple groups of feasible strategies (e.g., "Solution A: Energy consumption reduced by 20%, suppression time extended by 15 seconds; Solution B: Energy consumption increased by 10%, suppression time shortened by 30 seconds") for the operator to make decisions.

[0069] Example: When there are multiple suppression solutions for the fire spread path, the optimizer can generate a hybrid strategy of "local high-power sprinkler + global low-power fan", which not only controls the core area of the fire but also reduces the overall energy consumption.

[0070] Technical synergy of multi-modal fusion: Complementation between spatio-temporal coding and graph neural network: The spatio-temporal encoder provides explicit features of "location-time", and the graph neural network mines the implicit associations between parameters. The combination of the two can comprehensively represent the dynamic changes of the coal yard environment.

[0071] Adaptive ability of reinforcement learning: Adjust the sensor weights in real-time through feedback to solve the problem of the failure of traditional fixed-weight models in the face of environmental mutations (e.g., extreme weather requires dynamic increase in the wind speed weight).

[0072] Practical value of multi-objective optimization: Quantitative coordination of conflicting objectives: Traditional solutions often sacrifice other indicators (such as energy consumption) for a single objective (such as absolute safety), while the Nash equilibrium mechanism provides an interpretable trade-off solution, meeting the refined management needs of industrial scenarios.

[0073] Flexibility of decision-making support: The Pareto optimal solution set allows the operator to select the most suitable strategy according to the real-time working conditions (such as grid load, equipment status), improving the human-machine collaboration ability of the system.

[0074] Solving the pain points of existing technologies: Insufficient data correlation: Existing systems rely on manual experience to configure sensor weights. This solution realizes the automatic mining and dynamic optimization of data correlation through the combination of graph neural network and reinforcement learning.

[0075] Problem of rigid strategies: Traditional fire suppression strategies are mostly fixed rules. This solution realizes the flexible adjustment of strategies through multi-objective game optimization to adapt to the complex and changeable on-site environment.

[0076] Practical application example: Scenario: The coal pile catches fire due to high temperature, and the fire spreads rapidly as the wind speed increases.

[0077] Spatio-temporal coding and correlation analysis: The spatio-temporal encoder identifies the fire source location, and the graph neural network discovers the strong correlation of "wind speed - temperature - dust concentration".

[0078] Weight dynamic adjustment: The reinforcement learning agent increases the weight of the wind speed sensor from 0.5 to 0.9, focusing on the key disaster-causing factors.

[0079] Multi-objective optimization decision: The game optimizer generates a strategy of "high-power spraying in the core area + low-power fans in the periphery", which can suppress the fire while reducing energy consumption by 30%.

[0080] Execution and feedback: The actuator starts the spraying system, the digital twin engine corrects the simulation parameters in real time, and the reinforcement learning agent updates the weight strategy according to the fire extinguishing effect.

[0081] Effect: Compared with the traditional scheme, the fire suppression time is shortened by 40%, the comprehensive energy consumption is reduced by 25%, and no false alarms are triggered.

[0082] This solution realizes the deep integration of multi-source data and dynamic optimization decision-making through the collaborative design of spatio-temporal coding, graph neural network, reinforcement learning and multi-objective game, and solves the core problems of weak data correlation and rigid strategies in the safety monitoring of coal-fired power plants. The technical solution is logically rigorous, with low coupling between modules and easy to expand, and has significant industrial application value.

[0083] Specifically, for the safety monitoring system of the coal-fired power plant described in the present invention, the federated learning agent network performs: Local model training: Each agent updates the model parameters on the edge computing node and adds Laplace noise (ε = 0.5) to achieve differential privacy; Gradient compression: Adopt Top-k sparsification to retain 10% of the important gradients, and combine with Huffman coding to reduce the communication data volume; Consensus verification: When the failure rate of the regional sensor exceeds 30%, trigger the view switching mechanism of the PBFT protocol to eliminate Byzantine nodes; The federated learning agent network also includes: Dynamic privacy budget allocation module, which allocates different Laplace noise parameters according to the sensor data type, where the temperature data δ = 1e-4 and the gas concentration data δ = 1e-6; Byzantine node elimination module, when the node response time exceeds 500ms and the gradient deviation > 3σ, trigger the view switching mechanism of the PBFT protocol.

[0084] Spatial-temporal feature fusion logic of the spatio-temporal encoder: The spatio-temporal encoder realizes the spatio-temporal feature fusion of sensor data through the following steps: Spatial encoding: A three-dimensional spherical coordinate system is established with the geometric center of the coal yard as the origin, and the physical positions (longitude, latitude, altitude) of each sensor are converted into spherical coordinate parameters (radius, polar angle, azimuth angle). For example, the radius represents the distance from the sensor to the center of the coal yard, the polar angle reflects the vertical position (such as the difference between the top and bottom of the coal pile), and the azimuth angle identifies the horizontal distribution (such as the distinction between sensors on the east and west sides of the coal yard).

[0085] Time encoding: The timestamp is split into dimensions such as hours, minutes, and seconds, and encoded into a periodic vector through sine / cosine functions to avoid model biases caused by linear timestamps (such as encoding 14:30 as [sin(14 / 24*2π), cos(14 / 24*2π), sin(30 / 60*2π)]).

[0086] Feature concatenation: The spherical coordinate parameters and the time-encoded vector are concatenated into a spatio-temporal feature matrix, which serves as the input for subsequent multimodal analysis.

[0087] Example: When a temperature sensor is located at the top of the coal pile and the data collection time is 14:00 in the afternoon (a high-temperature period), the vector generated by the spatio-temporal encoder will simultaneously reflect the spatial characteristic of "the top position is prone to heat dissipation" and the time characteristic of "high temperature in the afternoon", providing joint features for correlation analysis.

[0088] Implicit association mining mechanism of graph neural network: The implicit association analysis of the graph neural network (GNN) includes the following steps: Adjacency matrix construction: Physical adjacency: If the distance between two sensors is less than a preset threshold (such as 5 meters), a connection edge is established, and the weight is the reciprocal of the distance.

[0089] Data correlation: Calculate the Pearson correlation coefficient of the historical data of the sensors. When the correlation coefficient > 0.7, a strong connection edge is established.

[0090] Functional coupling: Define the connection relationship according to the device control logic (such as the interlock control of the fan and the dust sensor).

[0091] Graph convolution operation: Aggregate the features of neighboring nodes through a multi-layer graph convolution network. For example, aggregate the dust concentration and wind speed data of the neighboring nodes of a certain temperature sensor to extract the joint change pattern of "temperature - dust - wind speed".

[0092] Implicit association output: Output the association strength matrix, which identifies the causal relationship (such as "an increase in wind speed leads to an increase in dust concentration") or statistical dependence relationship (such as "temperature and gas concentration are non-linearly positively correlated") between parameters.

[0093] Example: When the surface temperature of the coal pile rises abnormally, GNN discovers through correlation analysis that the dust concentration in this area rises synchronously, and the data of adjacent wind speed sensors fluctuates violently, triggering a composite risk warning of "local high temperature - dust - ventilation anomaly".

[0094] Dynamic weight adjustment strategy of the reinforcement learning agent: The reinforcement learning agent realizes the adaptive adjustment of sensor weights through the following process: State definition: The environmental state includes real-time parameters (such as coal pile volume, wind speed), historical risk events (such as number of fires), and implicit correlation analysis results (such as correlation strength matrix).

[0095] Action space: The weight adjustment actions include increasing or decreasing the confidence of specific sensors (such as increasing the weight of the wind speed sensor from 0.7 to 0.9), and reducing the weight of redundant sensors (such as reducing the weight of the temperature sensor with high failure rate to 0.3).

[0096] Reward mechanism: Use the risk warning accuracy rate (such as the number of correct warnings / total number of warnings) and response timeliness (delay from detection to execution) as the reward function to guide the model to optimize the weight configuration.

[0097] Policy iteration: Adopt the Proximal Policy Optimization (PPO) algorithm. Generate weight adjustment actions through the policy network (Actor), and evaluate the action benefits through the value network (Critic), continuously approaching the optimal policy.

[0098] Example: When the coal pile volume changes rapidly due to loading and unloading operations, the reinforcement learning agent automatically increases the weight of the volume measurement sensor and reduces the weight of static environmental parameters (such as the temperature at a fixed position), ensuring that the model focuses on the core variables of dynamic changes.

[0099] Conflict coordination mechanism of the multi-objective game optimizer: The coordination logic of the multi-objective game optimizer includes the following steps: Objective modeling: Fire suppression efficiency: Use the reduction rate of the fire area per unit time as a quantitative indicator.

[0100] Energy consumption cost: Use the weighted sum of the pump power of the sprinkler system and the energy consumption of the fan as the cost indicator.

[0101] Game participant design: Abstract the objectives as virtual game players. For example, the "safety side" pursues the speed of fire suppression, and the "energy consumption side" pursues the lowest energy consumption.

[0102] Nash equilibrium solution: Find the strategy combination where neither party can unilaterally improve the benefits through iterative negotiation. For example, at the initial stage of the fire, use a low-power fan to delay the spread of the fire, and switch to a high-power sprinkler when the fire expands, balancing efficiency and energy consumption.

[0103] Pareto optimal solution set generation: Output multiple groups of feasible strategies (such as "Scenario A: Energy consumption reduced by 25%, suppression time extended by 20 seconds; Scenario B: Energy consumption increased by 15%, suppression time shortened by 40 seconds") for operators to select according to the real-time working conditions.

[0104] Example: When a fire threatens multiple areas simultaneously, the optimizer generates a strategy of "high-power sprinklers in the core area + intermittent peripheral fans", reducing the overall energy consumption while controlling the core fire.

[0105] Technical synergy: Complementary of spatio-temporal encoding and graph neural network: The spatio-temporal encoder provides explicit features of "position-time", and the GNN mines the implicit associations between parameters. The combination of the two comprehensively represents the dynamic risks of the coal yard.

[0106] Closed-loop control of reinforcement learning and multi-objective optimization: Reinforcement learning dynamically adjusts the sensor weights, and the optimizer generates suppression strategies based on the weight results, forming a closed loop of "perception - decision - execution".

[0107] Solving the pain points of existing technologies: Weak data correlation: Traditional solutions rely on manual experience to configure sensor weights. This solution realizes the automatic mining and dynamic optimization of data correlation through the combination of GNN and reinforcement learning.

[0108] Rigid strategies: Existing fire suppression strategies are mostly fixed rules. This solution generates flexible strategies through multi-objective game optimization to adapt to the complex and changeable on-site environment.

[0109] Practical application value: Improving the early warning accuracy: Through implicit correlation analysis, false alarms caused by fluctuations in a single parameter are reduced (such as not triggering false alarms when only the temperature rises but there is no dust anomaly).

[0110] Optimizing resource allocation: The Pareto optimal solution set provides quantifiable strategies for selection, helping power plants achieve a balance between safety and cost.

[0111] Practical application example: Scenario: A fire is caused by spontaneous combustion of a coal pile, and the fire spreads with the wind speed.

[0112] Spatio-temporal encoding and correlation analysis: The spatio-temporal encoder identifies that the fire source is located in the southeast of the coal pile, and the GNN discovers a strong correlation between "wind speed - temperature - dust".

[0113] Dynamic adjustment of weights: The reinforcement learning agent increases the weight of the wind speed sensor to 0.95, focusing on the key disaster-causing factors.

[0114] Multi-objective optimization decision: The optimizer generates a strategy of "high-power spraying on the southeast side + low-power fans on the northwest side" to suppress the fire while reducing energy consumption by 35%.

[0115] Execution and feedback: The actuator starts the spraying system, the digital twin engine corrects the friction coefficient simulation parameters in real time, and the reinforcement learning agent updates the weight strategy according to the fire extinguishing effect.

[0116] Effect: Compared with the traditional scheme, the fire suppression time is shortened by 50%, the comprehensive energy consumption is reduced by 30%, and the false alarm rate is reduced to less than 3%.

[0117] This solution realizes the intelligent upgrade of the safety monitoring system of coal-fired power plants through the collaborative design of spatio-temporal coding, graph neural network, reinforcement learning and multi-objective game. The technical details are fully disclosed (such as the construction of the spherical coordinate system and the design of game participants), the logic between modules is clear, and it has feasibility and innovation, meeting the comprehensive requirements of industrial scenarios for real-time, reliability and economy.

[0118] Specifically, for the safety monitoring system of coal-fired power plants described in the present invention, the digital twin simulation engine includes: A discrete element method modeling unit that simulates the morphological changes of the coal pile and dynamically adjusts the particle friction coefficient μ∈[0.3, 0.7]; A computational fluid dynamics module that predicts the fire spread path based on the Navier-Stokes equation, with a grid division accuracy of 0.5 m³; An automated machine learning component that screens the feature subset through a genetic algorithm and adjusts the CC and γ parameters of the SVM model using a Bayesian optimizer.

[0119] In a second aspect, the present invention provides a safety monitoring method for coal-fired power plants, which is applied to the safety monitoring system of coal-fired power plants and includes the following steps: Step S1: Real-time collect infrared images, gas concentration time series signals and equipment log texts through edge computing nodes, and preprocess the collected data to generate preprocessed data; Step S2: Perform semantic mapping on the preprocessed data based on the knowledge graph to generate a standardized data stream; Step S3: Use a spatio-temporal encoder to extract the spatio-temporal features of the standardized data stream, and calculate the correlation weight between the infrared image features and the gas concentration signal through a cross-modal attention mechanism to generate a feature vector that fuses thermodynamic correlations; Step S4: Build a causal graph of coal yard accidents based on a Bayesian network, and activate the dust explosion probability inference path when abnormal wind speed and coal pile stacking angle are detected; Step S5: Aggregate the local model parameters of multiple edge nodes through the federated learning framework, inject a random orthogonal matrix before gradient compression, and generate a global risk assessment result based on the consensus algorithm; Step S6: construct a three-dimensional virtual coal yard in the digital twin simulation engine, simulate the risk evolution path, and correct the simulation error through the physical-data hybrid drive model; Step S7: Optimize the early warning model parameters according to the simulation results, and trigger the sprinkler valve or fan to perform a safety response action.

[0120] Specifically, the coal-fired power plant safety monitoring method of the present invention, step S1 comprises: Step S11: Receive sensor signals through a multi-protocol conversion circuit, and dynamically adjust the AD sampling frequency according to the dust concentration gradient change rate; Step S12: Suppress power frequency noise through a quantum noise suppression circuit; Step S13: Extract infrared image features through convolution operations and generate feature vectors; Step S14: Eliminate power frequency interference through an FIR filter, and upload the processed data through the 5G protocol.

[0121] Specifically, the coal-fired power plant safety monitoring method described in the present invention, step S3 includes: step S31: using a three-dimensional spherical coordinate system to encode sensor position information, and splicing it with a timestamp to generate spatiotemporal features; step S32: analyzing the sensor network topology relationship based on a graph neural network, and mining the implicit associations between multiple parameters; step S33: dynamically adjusting sensor weight parameters through a reinforcement learning algorithm.

[0122] Specifically, in the coal-fired power plant safety monitoring method of the present invention, step S5 comprises: Step S51: Add Laplace noise to each edge node to achieve differential privacy; Step S52: Use Top-k gradient sparsification to compress the amount of communication data; Step S53: When a node failure or abnormal gradient deviation is detected, trigger the consensus protocol to remove the abnormal node.

[0123] Specifically, the coal-fired power plant safety monitoring method described in the present invention, step S6 includes: step S61: simulating the morphological changes of the coal pile by discrete element method, and embedding the LSTM prediction unit to correct the friction coefficient in real time; step S62: simulating the fire spread path based on the Navier-Stokes equation; step S63: screening feature subsets by genetic algorithm, and optimizing classification model parameters.

[0124] Dynamic correction mechanism of particle behavior in discrete element method modeling units: The discrete element method (DEM) modeling unit simulates the morphological changes of the coal pile and dynamically corrects the particle friction coefficient through the following steps: Initial parameter setting: Based on coal quality characteristics (such as particle size distribution, moisture content), initialize the reference value of the particle friction coefficient μ (e.g., μ = 0.5), and set the dynamic adjustment range [0.3, 0.7].

[0125] Real-time data fusion: Receive real-time data from the vibration sensor on the coal pile surface and the temperature and humidity sensor through the LSTM prediction unit to predict the change trend of the adhesion force between particles. For example, when the moisture content of the coal pile increases due to rainfall, the LSTM predicts that the friction coefficient μ needs to be adjusted from 0.5 to 0.4 to reflect the enhanced viscosity between particles.

[0126] Dynamic correction logic: According to the prediction results, iteratively update the friction coefficient in the discrete element model at time steps. For example, when the coal pile is loosened by mechanical vibration, μ is gradually increased to 0.6 to simulate the collapse risk of reduced friction force.

[0127] Example: When a local collapse occurs in the coal pile due to loading and unloading operations, the discrete element model detects an abnormal vibration frequency through the vibration sensor data, and the LSTM prediction unit triggers the μ value to be adjusted from 0.5 to 0.65 to simulate the fluidity of loose particles, thereby accurately predicting the collapse diffusion range.

[0128] Optimization of fire spread prediction in the computational fluid dynamics module: The computational fluid dynamics (CFD) module realizes fire simulation based on the Navier-Stokes equation, specifically including: Grid dynamic division strategy: Centered on the fire source, divide the grid of the core area with an accuracy of 0.5 m³, and use a coarse grid of 1 m³ in the peripheral area to balance the calculation accuracy and efficiency. During the fire spread process, dynamically refine the grid on the spread path.

[0129] Multi-physical field coupling: Coupled solution of the temperature field, combustible gas concentration field and fluid velocity field. For example, the flow velocity change caused by gas expansion in the high-temperature area is used to update the fluid boundary conditions through iterative calculation.

[0130] Dynamic loading of boundary conditions: Dynamically adjust the calculation domain boundary according to the real-time data received by the digital twin engine (such as wind speed, obstacle position). For example, increase the weight of the turbulence model parameters near the ventilation opening.

[0131] Example: When the fire spreads to the ventilation duct, the CFD module automatically loads the duct geometry model, refines the grid inside the duct to 0.3 m³, combines the real-time wind speed data to simulate the impact of high-temperature air flow on the spread speed, and outputs the accurate fire spread path.

[0132] Model optimization process of the automated machine learning component: The automated machine learning (AutoML) component optimizes the SVM model through the following steps: Genetic Algorithm Feature Screening: Initialize the population: Randomly generate multiple groups of feature subsets (such as subsets containing temperature and dust concentration, or subsets containing only wind speed and humidity).

[0133] Fitness evaluation: Use the classification accuracy of the feature subset on the validation set as the fitness function, and eliminate low-score combinations.

[0134] Crossover and mutation: Retain high-score feature subsets, and generate a new generation of population by exchanging features (crossover) or randomly adding or deleting features (mutation), and iterate until convergence.

[0135] Bayesian Optimization Parameter Tuning: Surrogate model construction: Based on the Gaussian process, establish the mapping relationship between SVM parameters (such as kernel function parameters C and γ) and model performance.

[0136] Sampling strategy: Select the next set of parameters to be evaluated through the expected improvement (EI) function, and preferentially explore potential optimal regions to avoid falling into local optima.

[0137] Example: When the data distribution in the coal yard environment shifts due to seasonal changes, the AutoML component selects "wind speed + coal pile volume" as the key feature subset through the genetic algorithm, and the Bayesian optimizer synchronously adjusts the SVM parameters, increasing the model accuracy by 15%.

[0138] Interpretation of the Logic and Innovation of the Technical Solution: Technical synergy: Coupling of the discrete element method and CFD: The discrete element model provides the boundary conditions for the change of the coal pile shape, and the CFD module simulates the fire spread based on this to realize the chain risk prediction of "coal pile collapse - fire spread".

[0139] Closed-loop optimization of AutoML and digital twin: The AutoML component dynamically optimizes the model parameters according to the simulation results, and the digital twin engine then updates the simulation logic based on the new model parameters to form a self-optimizing closed loop.

[0140] Solving the pain points of the existing technology: Static defects of traditional models: In the existing solutions, the friction coefficient, mesh division, and model parameters are all fixed values. This solution significantly improves the adaptability under complex working conditions through dynamic correction and automated learning.

[0141] Balance between calculation efficiency and accuracy: The dynamic mesh division and feature screening mechanism reduce the consumption of computing resources while ensuring accuracy, meeting the real-time requirements.

[0142] Practical application value: Risk prediction accuracy: Through dynamic friction coefficient correction and high-precision CFD simulation, the prediction error of coal pile collapse and fire spread is reduced by 40% compared with the traditional solution.

[0143] Model adaptability: The AutoML component can automatically re-optimize the model after environmental changes (such as coal quality replacement, equipment update), reducing the cost of manual parameter tuning.

[0144] Actual application example: Scenario: The coal pile spontaneously ignites due to long-term accumulation, and the fire spreads to the coal storage area along with the change of the wind direction.

[0145] Discrete element model correction: The LSTM unit predicts that the μ value drops from 0.5 to 0.4 according to the temperature and humidity data, simulates the enhanced viscosity characteristics of the wet coal pile, and outputs a stable form model.

[0146] CFD fire simulation: Based on the corrected coal pile shape, the CFD module predicts the path of the fire spreading to the coal storage area along the southeast wind direction with a grid accuracy of 0.5m³.

[0147] AutoML model optimization: The genetic algorithm selects "temperature gradient + wind direction" as the key features, and the Bayesian optimizer adjusts the SVM parameters to generate a high-precision fire warning model.

[0148] Execution and feedback: The system triggers the directional fire suppression of the sprinkler system on the southeast side and feeds the fire extinguishing effect data back to the digital twin engine to iteratively optimize the model parameters.

[0149] Effect: The spontaneous combustion fire is effectively suppressed before spreading to the coal storage area, and the time-consuming of the model adaptive optimization is shortened by 70% compared with the traditional manual parameter tuning.

[0150] This solution realizes the precise simulation and intelligent decision-making of the safety monitoring system of coal-fired power plants through the collaborative design of discrete element dynamic modeling, CFD multi-physical field coupling and AutoML adaptive optimization. The technical details are fully disclosed (such as the dynamic grid division logic, genetic algorithm operation steps), and the data flow between modules is clear (coal pile shape → fire simulation → model optimization → execution feedback), meeting the core requirements of industrial scenarios for real-time and reliability, and having significant technical advancement and application value.

[0151] Step S1: Data collection and preprocessing: Multi-protocol conversion circuit (S11): This circuit integrates industrial communication protocols such as Modbus and RS-485, and uniformly converts the analog signals of different sensors into digital signals through a protocol parser. For example, the dust sensor outputs a current signal (4 - 20mA), and the gas sensor outputs a voltage signal (0 - 5V). The conversion circuit adapts the signal range through a programmable gain amplifier (PGA) to ensure the AD sampling accuracy.

[0152] Quantum noise suppression (S12): Embed a quantum tunneling effect unit in the sensor interface circuit. Form a potential barrier through a quantum dot array to filter out 50 Hz power frequency interference and its harmonic components. For example, for the periodic noise in the temperature sensor signal, the suppression unit dynamically adjusts the potential barrier height to match the noise frequency band, achieving noise amplitude attenuation.

[0153] Convolutional feature extraction (S13): Use a lightweight convolutional neural network (such as MobileNet) to perform hierarchical feature extraction on infrared images. Capture the spatial distribution of local high-temperature regions through a 3×3 convolutional kernel. The pooling layer compresses the feature dimension, and finally outputs a 128-dimensional vector to represent key information such as hot spot size and temperature gradient.

[0154] FIR filtering and 5G upload (S14): Design a finite impulse response (FIR) filter, dynamically adjust the cut-off frequency according to the sensor signal frequency band to eliminate high-frequency noise. For example, when the gas concentration signal frequency band is 0 - 100 Hz, the cut-off frequency is set to 120 Hz to suppress out-of-band noise while retaining the effective signal. The preprocessed data is encapsulated through the 5G NR protocol and uploaded to the central processor using the time division multiple access (TDMA) mechanism to ensure low-latency transmission.

[0155] Step S1 solves the problems of chaotic multi-source sensor data formats and severe noise interference through heterogeneous signal conversion, quantum noise suppression, lightweight feature extraction, and dynamic filtering. The preprocessing process is completed locally at the edge node, reducing the data transmission volume, avoiding the delay bottleneck of cloud processing, and providing high-quality input for subsequent analysis.

[0156] Step S2: Knowledge graph-driven semantic mapping Based on the coal yard safety ontology library (such as entity relationships like "dust explosion" and "equipment failure"), the knowledge graph performs semantic alignment on the preprocessed data: Entity recognition: Extract key entities (such as "fan FAN-01" and "temperature sensor T-12") from device logs through a named entity recognition (NER) model.

[0157] Relationship mapping: Establish semantic associations between entities. For example, map the event of "abnormal rotation speed of fan FAN-01" to the causal chain of "insufficient ventilation → dust accumulation → explosion risk".

[0158] Standardized output: Convert heterogeneous data (images, time series signals, text) into RDF triples in a unified format to form a standardized data stream. For example, associate the coordinates of the high-temperature region in the infrared image with the event of exceeding the gas concentration standard as the triple "high-temperature region → located in → the southeast side of the coal pile".

[0159] Logical interpretation: Knowledge graphs break the information islands of multimodal data through semantic mapping, convert discrete sensor readings into standardized data streams with logical associations, provide structured input for subsequent spatiotemporal feature fusion and causal reasoning, and solve the problem of data semantic fragmentation in traditional solutions.

[0160] Step S3: spatiotemporal feature fusion and weight dynamic adjustment Space-time coding (S31): A spherical coordinate system is established with the center of the coal yard as the origin, and the sensor position (longitude, latitude, altitude) is converted into radius, polar angle, and azimuth parameters, and spliced ​​with the periodic coding of the timestamp (such as sine / cosine function) to generate a four-dimensional space-time vector. For example, if a sensor is located at the top of the coal pile (polar angle = 30°) and collects data at 14:00 in the afternoon, its vector will fuse the location characteristics with the high temperature period characteristics.

[0161] Graph neural network analysis (S32): Based on the physical adjacency of sensors (distance < 5 meters) and data correlation (Pearson coefficient > 0.7), an adjacency matrix is ​​constructed, and neighborhood node features are aggregated through a graph convolutional network (GCN) to mine implicit associations (such as "increased wind speed → increased dust concentration").

[0162] Reinforcement learning weight adjustment (S33): With risk warning accuracy as the reward function, the deep deterministic policy gradient (DDPG) algorithm is used to dynamically adjust the sensor weight. For example, when the volume of the coal pile changes rapidly, the weight of the volume sensor is increased to 0.9, and the weight of the static environmental parameter is reduced to 0.3.

[0163] The collaborative work of the spatiotemporal encoder and graph neural network realizes multi-dimensional correlation analysis of "position-time-parameters". Reinforcement learning further optimizes the weight configuration, allowing the model to focus on key risk factors, solving the problems of high false alarm rate and poor adaptability caused by fixed weights in traditional methods.

[0164] Step S4: Bayesian causal inference and risk path activation: The Bayesian network constructs a causal graph based on historical accident data, with nodes including environmental parameters (wind speed, stacking angle), equipment status (fan speed), and accident type (dust explosion). When the wind speed > 8m / s and the stacking angle > 38° are detected, the network activates the "dust explosion" reasoning path: Conditional probability calculation: Update the posterior probability based on the prior probability (e.g. "the probability of dust concentration exceeding the standard when wind speed>8m / s is 70%)."

[0165] Path activation: If the probability of dust explosion exceeds the threshold (such as 30%), a warning signal is triggered and pushed to the digital twin engine.

[0166] Logical interpretation: Bayesian networks quantify risks through probabilistic reasoning, dynamically activate critical paths, and avoid the single judgment defect of traditional threshold methods. For example, even if a single parameter does not exceed the threshold, early warnings can still be triggered when multiple parameters are jointly abnormal, improving the judgment accuracy under complex working conditions.

[0167] Step S5: Federated learning and privacy protection: Differential privacy (S51): When each edge node trains the local model, Laplace noise (ε = 0.5) is injected into the gradient data to ensure that attackers cannot reverse-engineer the original data. For example, after adding noise to the temperature gradient data, its distribution variance increases, but statistical features (such as the mean) remain stable.

[0168] Gradient compression (S52): The Top-k algorithm is used to retain the top 10% of the parameters in terms of gradient magnitude, and Huffman coding is combined to compress the data volume. For example, only the top 100 important values of a 1000-dimensional gradient vector are transmitted, and the compression rate exceeds 80%.

[0169] Consensus to eliminate anomalies (S53): Based on the PBFT protocol, if the node response times out for 500 ms or the gradient deviation > 3σ, a view switch is triggered, and the node is marked as a Byzantine node and eliminated. For example, if a node continuously sends abnormal gradients due to hardware failure, the system automatically isolates the node to ensure the reliability of the global model.

[0170] Logical interpretation: Under the premise of protecting data privacy, the federated learning framework realizes multi-node collaborative modeling through gradient obfuscation and consensus mechanisms, solves the data security risks and single-point failure problems of centralized training, and is applicable to the joint monitoring scenario of distributed coal-fired power plant groups.

[0171] Step S6: Digital twin simulation and error correction Discrete element method simulation (S61): Based on the physical properties of coal particles (the dynamic adjustment range of the friction coefficient μ is [0.3, 0.7]), the process of coal pile collapse is simulated. The LSTM unit predicts the change trend of the μ value according to the vibration sensor data. For example, when the particles become loose due to loading and unloading operations, μ increases from 0.5 to 0.6 to simulate the enhanced fluidity.

[0172] Fire spread prediction (S62): The Navier-Stokes equation is solved using computational fluid dynamics (CFD), the grid is dynamically divided into 0.5 m³, the temperature field and gas concentration field are coupled, and the fire spread direction is predicted. For example, when the high-temperature airflow is affected by the ventilation duct, the grid near the duct is refined to 0.3 m³ to improve the accuracy.

[0173] Genetic algorithm optimization (S63): Key feature subsets (such as "temperature gradient + wind speed") are screened through selection, crossover, and mutation operations, and the Bayesian optimizer adjusts the SVM kernel function parameters to improve the robustness of the classification model in a dynamic environment.

[0174] The digital twin engine corrects simulation errors in real time through hybrid simulation of physical models and data-driven, solving the problem of insufficient accuracy of traditional pure physical models under complex working conditions. For example, the LSTM prediction unit dynamically corrects the friction coefficient based on real-time data, making the discrete element simulation closer to the actual coal pile shape.

[0175] Step S7: Early warning optimization and execution control Model parameter optimization: The classification threshold and feature weight of the early warning model are adjusted in reverse according to the simulation results. For example, if the digital twin shows that the fire risk in a certain area is underestimated, the weight coefficient of the sensor in that area is increased.

[0176] Actuator linkage: When the sprinkler valve is triggered, the nozzle angle and injection pressure are adjusted according to the fire path simulated by CFD; when the fan is controlled, the speed is adjusted according to the wind speed-dust correlation model to achieve precise dust suppression.

[0177] Step S7 forms a closed loop of "perception-simulation-decision-execution", and continuously optimizes model parameters through simulation feedback to ensure the accuracy of early warning and response efficiency. For example, the sprinkler system dynamically adjusts the coverage range according to the fire spread path, saving more than 30% water compared with the traditional fixed sprinkler mode.

[0178] Overall logic of the technical solution: The method of the present invention solves the core problems of data heterogeneity, weak correlation, response delay, etc. in coal-fired power plant monitoring layer by layer through the progressive architecture of edge preprocessing-semantic fusion-dynamic reasoning-federated learning-digital twin: Edge computing reduces the burden of data transmission and improves real-time performance; Knowledge graph and spatiotemporal coding realize semantic alignment and feature fusion of multimodal data; Bayesian networks and reinforcement learning enhance dynamic reasoning capabilities for complex causal relationships; Federated learning and digital twins ensure data security and simulation accuracy, forming a closed-loop optimization.

[0179] The synergistic effect of various modules ultimately achieves the technical effects of improving the accuracy of risk warnings, reducing false alarm rates, and optimizing resource consumption, which meets the dual needs of industrial production for safety and economy.

[0180] Logical explanation of the technical solution of the present invention to solve the pain points of the prior art: Solve the problem of decentralized processing and inconsistent formats of multi-source heterogeneous data: Unified preprocessing of edge computing nodes: Integrate a multi-protocol conversion circuit through the sensor interface module to uniformly convert heterogeneous data such as infrared images, gas concentration time series signals, and device log texts into digital signals. The preprocessing chipset (NPU+FPGA heterogeneous architecture) denoises, normalizes, and extracts features from the data to generate preprocessed data in a unified format, eliminating the format differences of multi-source data.

[0181] Semantic mapping driven by knowledge graph: Based on the coal yard safety ontology library (such as entity relationships like "dust explosion" and "equipment failure"), perform semantic alignment on the preprocessed data to convert heterogeneous data into a standardized data stream. For example, map "abnormal fan speed" in the device log to the causal chain of "insufficient ventilation → dust accumulation" in the knowledge graph to achieve semantic unification of multi-modal data.

[0182] Solution effect: Through local preprocessing of edge nodes and semantic mapping of the knowledge graph, unify the format and semantic expression of multi-source data, break data islands in traditional solutions, and significantly improve data fusion efficiency.

[0183] Solve the problems of low real-time data fusion efficiency and insufficient non-linear correlation analysis: Spatio-temporal encoding and cross-modal attention mechanism: Use a three-dimensional spherical coordinate system to encode the sensor position as a spatial vector and splice it with timestamp information to generate a spatio-temporal feature matrix. Calculate the correlation weights between infrared image features and gas concentration signals through a cross-modal attention mechanism (based on the Transformer architecture) to capture the non-linear correlations between thermodynamic parameters.

[0184] Graph neural network and reinforcement learning: Construct an adjacency matrix based on the topological relationship of sensors, and use a graph neural network (GNN) to mine the implicit correlations between multiple parameters (such as the co-variation of "wind speed - dust concentration"). The reinforcement learning agent dynamically adjusts the sensor weights according to the correlation analysis results to make the model focus on key risk factors.

[0185] Solution effect: Spatio-temporal encoding and attention mechanism achieve dynamic fusion of multi-modal data, and graph neural network and reinforcement learning cooperate to mine complex non-linear correlations, solving the problem of dynamic correlation analysis that cannot be handled by traditional threshold methods or single algorithms.

[0186] Solve the problems of early warning response delay and high false alarm rate: Federated learning intelligent agent network: Aggregate the local model parameters of multiple edge nodes through a consensus algorithm, inject a random orthogonal matrix before gradient compression to prevent gradient leakage and improve the model convergence speed. The distributed computing of edge nodes reduces the load on the central processor and reduces transmission and processing delays.

[0187] Dynamic Causal Inference Engine: Construct a causal diagram of coal yard accidents based on Bayesian networks. When the detected wind speed > 8 m / s and the coal pile stacking angle > 38°, activate the inference path of the probability of dust explosion. Through multi-parameter joint probability inference, false alarms caused by misjudgment of a single parameter are avoided.

[0188] Digital Twin Simulation Engine: Construct a three-dimensional virtual coal yard to simulate the risk evolution path. Through a physical-data hybrid-driven model (such as the discrete element method + LSTM prediction unit), the simulation error is corrected in real time, the parameters of the early warning model are optimized, and precise response actions (such as directional explosion suppression of spray valves) are triggered.

[0189] Solution effect: The distributed architecture of federated learning and the real-time simulation of digital twins cooperate to reduce the response delay; dynamic causal inference and multi-parameter joint analysis significantly reduce the false alarm rate (for example, no false alarm is triggered when only the temperature rises but there is no dust anomaly).

[0190] Overall coordination of the technical solution: Edge-cloud coordination: Edge nodes complete data preprocessing and feature extraction, and the cloud multi-modal platform performs global fusion and decision-making, improving the processing efficiency while reducing the data transmission volume.

[0191] Data-model-execution closed-loop: The knowledge graph and spatio-temporal encoding provide structured data input, dynamic causal inference and federated learning optimize the risk assessment model, and the digital twin simulation feeds back the results and triggers precise execution, forming a closed-loop optimization chain.

[0192] Through the five core technologies of edge preprocessing to unify data formats, knowledge graph semantic mapping, spatio-temporal fusion and dynamic correlation analysis, federated learning distributed optimization, and digital twin real-time simulation, the present invention systematically solves the pain points such as decentralized processing, chaotic formats, and single algorithms of multi-source heterogeneous data in the safety monitoring of coal-fired power plants. The technical solution achieves remarkable effects of improving data fusion efficiency, accurately mining non-linear correlations, accelerating early warning responses, and reducing false alarm rates, meets the core requirements of industrial scenarios for real-time, accuracy, and reliability, and has outstanding technical innovation and practicality.

Claims

1. A safety monitoring system for a coal-fired power plant, comprising: An edge computing node, deployed in the pre-monitoring area of the coal yard, for real-time collection of infrared images, gas concentration time-series signals, and equipment log texts. The edge computing node includes: A sensor interface module, configured with a multi-protocol conversion circuit and an AD sampling unit to convert analog signals into digital signals. The AD sampling unit integrates an adaptive sampling frequency control module to dynamically adjust the sampling frequency (0.1 kHz - 10 kHz) according to the change rate of the dust concentration gradient, and realizes frequency band matching through an NPU+FPGA heterogeneous architecture; A quantum noise suppression circuit, embedded in the sensor interface module, using a quantum tunneling effect suppression unit to increase the power frequency noise signal-to-noise ratio to ≥35 dB; A preprocessing chipset, performing denoising, normalization, and feature extraction operations on the digital signals to obtain preprocessed data; A communication control unit, uploading the preprocessed data to the central processor through a 5G protocol; A multi-modal data fusion platform, connected to the edge computing node, including: A knowledge graph module, performing semantic mapping on heterogeneous data based on the coal yard safety ontology to generate a standardized data stream; A dynamic fusion analysis module, using a spatio-temporal encoder and a cross-modal attention mechanism, calculating the cross-attention weights of the infrared image features and the gas concentration signals through a preset Transformer architecture in the dynamic fusion analysis module to generate a feature vector with fused thermodynamic correlations; A dynamic causal inference engine, constructing a causal graph of coal yard accidents based on a Bayesian network. When the wind speed > 8 m / s and the coal pile stacking angle > 38°, activating the dust explosion probability inference path; A federated learning agent network, communicating with the multi-modal data fusion platform, collaboratively generating a risk assessment result based on a consensus algorithm. The federated learning agent network includes a gradient confusion mechanism, injecting a random orthogonal matrix before gradient compression to make the gradient parameter distribution entropy value ≥6.2; A digital twin simulation engine, receiving the risk assessment result, constructing a three-dimensional virtual coal yard to simulate the risk evolution path, and optimizing the early warning model parameters. The digital twin simulation engine includes a physical-data hybrid drive model, embedding an LSTM prediction unit in the discrete element method to real-time correct the simulation error of the particle friction coefficient μ; An actuator control unit, triggering a spray valve or a fan to perform a safety response action according to the output instruction of the digital twin simulation engine.

2. The coal-fired power plant safety monitoring system according to claim 1, characterized in that, The preprocessing chipset includes: an image feature extraction module for extracting infrared image features through convolutional operations to generate a 128-dimensional feature vector; a data processing module configured to run a FIR filter to eliminate power frequency interference, and the cut-off frequency of the FIR filter is set to 1.2 times the sensor signal frequency band; wherein, the image feature extraction module and the data processing module process data in parallel through an NPU and an FPGA chip respectively, and transmit the processing results to the communication control unit through a data bus; the communication control unit transmits the data to the mirror module of the coal-fired power plant safety monitoring system, and the mirror module performs: performing simulation operation verification on the received data, and if error data is detected, marking the error data and generating simulated correction data; recombining the simulated correction data with the correct data to generate an available data set.

3. The safety monitoring system for a coal-fired power plant according to claim 1, characterized in that, The dynamic fusion analysis module includes: a spatio-temporal encoder that encodes the sensor position as a spatial vector using a three-dimensional spherical coordinate system and concatenates it with the timestamp information to generate spatio-temporal features; A graph neural network that constructs an adjacency matrix based on the sensor topological relationship, analyzes the implicit associations between multiple parameters, and obtains the analysis results of the implicit associations between multiple parameters; A reinforcement learning agent that configures weight parameters according to the analysis results of the implicit associations between multiple parameters, and adjusts the real-time sensor weights of the coal pile volume and wind speed with the configured weight parameters to obtain the adjusted real-time sensor weights of the coal pile volume and wind speed; The digital twin simulation engine further includes: a multi-objective game optimizer that coordinates the conflict between fire suppression efficiency and energy consumption through the Nash equilibrium algorithm to generate a Pareto optimal solution set.

4. The coal-fired power plant safety monitoring system according to claim 1, characterized in that, The federated learning agent network performs: Local model training: Each agent updates the model parameters on the edge computing node and adds Laplace noise (ε = 0.5) to achieve differential privacy; Gradient compression: Adopt Top-k sparsification to retain 10% of the important gradients, and combine Huffman coding to reduce the communication data volume; Consensus verification: When the failure rate of the regional sensors exceeds 30%, trigger the view switching mechanism of the PBFT protocol to eliminate Byzantine nodes; The federated learning agent network further includes: A dynamic privacy budget allocation module that allocates different Laplace noise parameters according to the sensor data type, where the temperature data δ = 1e-4 and the gas concentration data δ = 1e-6; A Byzantine node elimination module that triggers the view switching mechanism of the PBFT protocol when the node response times out for 500 ms and the gradient deviation degree > 3σ.

5. The safety monitoring system for a coal-fired power plant according to claim 1, wherein, The digital twin simulation engine includes: A discrete element method modeling unit that simulates the morphological changes of the coal pile and dynamically adjusts the particle friction coefficient μ ∈ [0.3, 0.7]; A computational fluid dynamics module that predicts the fire spread path based on the Navier-Stokes equation, and the grid division accuracy is 0.5 m³; An automated machine learning component that screens the feature subset through a genetic algorithm and adjusts the CC and γ parameters of the SVM model using a Bayesian optimizer.

6. A safety monitoring method for a coal-fired power plant, which is applied to the coal-fired power plant safety monitoring system according to any one of claims 1 to 5, and is characterized in that, Including the following steps: Step S1: collect infrared images, gas concentration time series signals and equipment log texts in real time through edge computing nodes, and preprocess the collected data to generate preprocessed data; Step S2: semantically mapping the preprocessed data based on the knowledge graph to generate a standardized data stream; Step S3: extracting the spatiotemporal features of the standardized data stream using a spatiotemporal encoder, and calculating the association weights between the infrared image features and the gas concentration signal through a cross-modal attention mechanism to generate a feature vector integrating thermodynamic associations; Step S4: construct a coal yard accident causal diagram based on the Bayesian network, and activate the dust explosion probability reasoning path when abnormal wind speed and coal pile stacking angle are detected; Step S5: Aggregate the local model parameters of multiple edge nodes through the federated learning framework, inject a random orthogonal matrix before gradient compression, and generate a global risk assessment result based on the consensus algorithm; Step S6: construct a three-dimensional virtual coal yard in the digital twin simulation engine, simulate the risk evolution path, and correct the simulation error through the physical-data hybrid drive model; Step S7: Optimize the early warning model parameters according to the simulation results, and trigger the sprinkler valve or fan to perform a safety response action.

7. The safety monitoring method for a coal-fired power plant according to claim 6, characterized in that, The step S1 comprises: Step S11: Receive sensor signals through a multi-protocol conversion circuit, and dynamically adjust the AD sampling frequency according to the dust concentration gradient change rate; Step S12: Suppress power frequency noise through a quantum noise suppression circuit; Step S13: Extract infrared image features through convolution operations and generate feature vectors; Step S14: Eliminate power frequency interference through an FIR filter, and upload the processed data through the 5G protocol.

8. The safety monitoring method for a coal-fired power plant according to claim 6, characterized in that, The step S3 includes: step S31: using a three-dimensional spherical coordinate system to encode sensor position information, and splicing it with the timestamp to generate spatiotemporal features; step S32: analyzing the sensor network topology relationship based on a graph neural network, and mining the implicit associations between multiple parameters; step S33: dynamically adjusting the sensor weight parameters through a reinforcement learning algorithm.

9. The safety monitoring method for a coal-fired power plant according to claim 6, wherein The step S5 comprises: Step S51: Add Laplace noise to each edge node to achieve differential privacy; Step S52: Use Top-k gradient sparsification to compress the amount of communication data; Step S53: When a node failure or abnormal gradient deviation is detected, trigger the consensus protocol to remove the abnormal node.

10. The safety monitoring method for a coal-fired power plant according to claim 6, characterized in that, The step S6 includes: step S61: simulating the morphological changes of the coal pile by discrete element method, and embedding the LSTM prediction unit to correct the friction coefficient in real time; step S62: simulating the fire spread path based on the Navier-Stokes equation; step S63: screening feature subsets by genetic algorithm, and optimizing classification model parameters.

Images