A coal-fired power plant coal conveying sampling packaging operation and maintenance method and system combined with deep learning

Abstract

The application provides a kind of coal-fired power plant coal conveying sampling packaging operation and maintenance method and system combined with deep learning in the cross technical field of coal-fired power plant automation and artificial intelligence, the method comprises the following steps: step S10, a large number of historical multi-source heterogeneous data of coal-fired power plant are acquired to construct first, second, third and fourth data sets;Step S20, create a collaborative operation intelligent agent including a sampling strategy generation model, a packaging state detection model, a fault detection model and an intrusion detection model;Step S30, train the collaborative operation intelligent agent by the first, second, third and fourth data sets;Step S40, input the collaborative operation intelligent agent after preprocessing the collected real-time multi-source heterogeneous data, obtain the sampling strategy adjustment suggestion, the packaging state detection result, the fault detection result and the intrusion detection result, and then generate the comprehensive operation instruction.The advantage of the present application lies in: greatly improving the operation efficiency and safety and reliability of the coal conveying system.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of thermal power plant automation and artificial intelligence, specifically to a method and system for coal conveying sampling and encapsulation operation and maintenance in thermal power plants that combines deep learning. Background Technology

[0002] As a critical infrastructure for global energy supply, the safe and stable operation of thermal power plants is directly related to national energy security and socio-economic stability. In an energy structure where coal is the primary fuel, thermal power plants bear an irreplaceable responsibility for ensuring power supply, and their operational efficiency and safety are highly dependent on the stability and reliability of the coal supply chain. The coal conveying system, as the core link throughout the entire process of coal receiving, transportation, sampling, packaging, and maintenance, directly impacts fuel cost control, quality consistency, environmental compliance, and the overall stability of the unit operation.

[0003] Currently, most thermal power plants still rely primarily on manual operation or automated systems based on fixed thresholds for the sampling, packaging, and maintenance of coal conveying systems. These traditional operating models struggle to meet the demands of modern power industry for efficient, precise, and sustainable operation, and pose significant risks to data security and system reliability. Specifically, existing technologies have the following main shortcomings: 1. Poor representativeness and lack of real-time performance in the sampling process make it difficult to guarantee data authenticity: Manual sampling is easily affected by operational procedures, and there is a risk that the data may be deliberately tampered with or unintentionally misoperated; Existing mechanical sampling equipment mostly adopts an independent and closed control architecture, and the sampling data often lacks an effective encryption and integrity verification mechanism when uploaded to the monitoring system, making it vulnerable to man-in-the-middle attacks or data theft, which leads to distortion of key coal quality parameters (such as calorific value and sulfur content), thereby affecting the accuracy of combustion optimization and cost accounting, and even providing the possibility for emission data tampering.

[0004] 2. Low reliability of the packaging process and weak traceability and anti-tampering capabilities of the data chain: Existing packaging operations lack reliable means of monitoring and recording the sealing status; relevant quality data (such as images and sensor readings) are usually stored in the form of scattered files, lacking a unified and secure data management and access control mechanism, making the data susceptible to unauthorized modification or deletion, and making it difficult to provide legally valid traceability evidence in the event of coal shortages or pollution disputes.

[0005] 3. Operation and maintenance management relies on post-event alarms, and the system itself faces network security threats: Current systems mostly use alarm mechanisms based on fixed rules, which cannot achieve predictive maintenance; at the same time, sensor networks and industrial control networks used for status monitoring generally lack effective boundary protection and intrusion detection capabilities. Attackers may penetrate the system and inject false equipment status information, creating false alarms or covering up real faults, thereby causing unplanned downtime or even equipment damage; in addition, the large amount of operational data generated during operation and maintenance is a core asset of the enterprise, and it also faces the risk of leakage in the transmission and storage links.

[0006] In-depth analysis reveals that existing coal conveying systems in thermal power plants not only lack a unified data management architecture and intelligent analysis capabilities, but more importantly, they lack a systematic data security protection framework. Data at each stage—sampling, packaging, and operation and maintenance—is isolated and poorly protected, making it difficult to guarantee the confidentiality, integrity, and availability of the data. This results in the entire process collaborative optimization facing both technical bottlenecks and security risks.

[0007] With the continuous development of Industry 4.0 technologies, artificial intelligence methods such as deep learning have shown significant potential in areas such as industrial data analysis and predictive maintenance. However, when introduced into coal conveying systems of thermal power plants, in addition to addressing traditional challenges such as complex operating environments and strong data heterogeneity, it is also crucial to pay attention to the new security risks arising from this. For example, deep learning models themselves may be vulnerable to adversarial attacks, leading to distorted decision-making; once a centralized intelligent system is compromised, it can cause a complete shutdown of the entire process. Existing research mostly focuses on optimizing the efficiency of local processes and lacks an integrated solution that deeply integrates intelligent analysis capabilities with inherent security mechanisms, covering the entire data lifecycle.

[0008] Therefore, how to provide a sampling, packaging, operation and maintenance method and system for coal conveying in thermal power plants that combines deep learning, while improving the level of intelligence throughout the entire process, and building an endogenous safety protection system to comprehensively solve the three major problems of poor authenticity of sampling data, weak traceability of packaging, and passive defense in operation and maintenance, and ultimately synergistically improve the operating efficiency and safety reliability of the coal conveying system, has become an urgent technical problem to be solved. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a method and system for sampling, packaging and operation and maintenance of coal conveying in thermal power plants that combines deep learning. While improving the level of intelligence of the whole process, it builds an endogenous safety protection system to comprehensively solve the three major problems of poor authenticity of sampling data, weak traceability of packaging and passive defense of operation and maintenance, and ultimately improve the operating efficiency and safety reliability of the coal conveying system.

[0010] In a first aspect, the present invention provides a method for sampling, encapsulating, and maintaining coal conveying systems in thermal power plants that incorporates deep learning, comprising the following steps: Step S10: Obtain a large amount of historical multi-source heterogeneous data from the thermal power plant, preprocess the historical multi-source heterogeneous data, and construct the first dataset, the second dataset, the third dataset, and the fourth dataset. Step S20: Create a collaborative operation and maintenance intelligent agent that includes a sampling strategy generation model, an encapsulation state detection model, a fault detection model, and an intrusion detection model; The sampling strategy generation model is built based on a multimodal feature extraction module, a multimodal feature fusion module, and a strategy decision module. It is used to output sampling strategy adjustment suggestions based on input visual image data, physical attribute data, and infrared spectral data. The encapsulation state detection model is built based on an image feature extraction module, a weight feature extraction module, a multimodal feature aggregation module, and a multi-task output module. It is used to output encapsulation state detection results based on input visual image data and physical attribute data. The fault detection model is built based on a vibration feature extraction module, a temperature feature extraction module, a humidity feature extraction module, a visual feature extraction module, a contour feature extraction module, a heterogeneous feature fusion module, and a fault prediction module. It is used to output fault detection results based on input visual image data, physical attribute data, and equipment status data. The intrusion detection model is built upon a network traffic analysis module, an identity authentication analysis module, a log analysis module, a multimodal representation learning module, a comprehensive risk assessment module, and a protection suggestion generation module. It is used to output intrusion detection results based on input network communication data. Step S30: Train the collaborative operation and maintenance agent using the first dataset, the second dataset, the third dataset, and the fourth dataset; Step S40: After preprocessing the collected real-time multi-source heterogeneous data, input it into the collaborative operation and maintenance intelligent agent to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, generate comprehensive operation and maintenance instructions.

[0011] Secondly, this invention provides a coal conveying sampling and packaging operation and maintenance system for thermal power plants that incorporates deep learning, comprising the following modules: The dataset construction module is used to acquire a large amount of historical multi-source heterogeneous data from thermal power plants, preprocess the aforementioned historical multi-source heterogeneous data, and construct the first dataset, the second dataset, the third dataset, and the fourth dataset. The agent creation module is used to create a collaborative operation and maintenance agent that includes a sampling strategy generation model, an encapsulated state detection model, a fault detection model, and an intrusion detection model. The sampling strategy generation model is built based on a multimodal feature extraction module, a multimodal feature fusion module, and a strategy decision module. It is used to output sampling strategy adjustment suggestions based on input visual image data, physical attribute data, and infrared spectral data. The encapsulation state detection model is built based on an image feature extraction module, a weight feature extraction module, a multimodal feature aggregation module, and a multi-task output module. It is used to output encapsulation state detection results based on input visual image data and physical attribute data. The fault detection model is built based on a vibration feature extraction module, a temperature feature extraction module, a humidity feature extraction module, a visual feature extraction module, a contour feature extraction module, a heterogeneous feature fusion module, and a fault prediction module. It is used to output fault detection results based on input visual image data, physical attribute data, and equipment status data. The intrusion detection model is built upon a network traffic analysis module, an identity authentication analysis module, a log analysis module, a multimodal representation learning module, a comprehensive risk assessment module, and a protection suggestion generation module. It is used to output intrusion detection results based on input network communication data. The agent training and deployment module is used to train the collaborative operation and maintenance agent using the first dataset, the second dataset, the third dataset, and the fourth dataset. The model inference module is used to preprocess the collected real-time multi-source heterogeneous data and input it into the collaborative operation and maintenance intelligent agent to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, a comprehensive operation and maintenance instruction is generated.

[0012] The advantages of this invention are: 1. A large amount of historical, multi-source, heterogeneous data from thermal power plants is acquired to construct four datasets: the first, second, third, and fourth datasets. Then, using these datasets, the created sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model are trained and deployed to the industrial control computer (ICC) of the power plant. Based on input sampling control commands, the ICC samples and encapsulates the coal conveyed by the conveyor belt, simultaneously collecting real-time multi-source heterogeneous data through an IoT sensing device array. This preprocessed real-time multi-source heterogeneous data is then input into the deployed sampling strategy generation model, encapsulation state detection model, and fault detection model. The model and intrusion detection model are used to obtain sampling strategy adjustment suggestions, encapsulation state detection results, fault detection results, and intrusion detection results. Then, based on these results, a comprehensive operation and maintenance command is generated, executed, and execution feedback is recorded. Next, real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation state detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance commands, and execution feedback are encrypted into an encrypted runtime data packet. The data fingerprint of the encrypted runtime data packet is calculated, uploaded to the blockchain, and then uploaded to the server. The server verifies and decrypts the encrypted runtime data packet to obtain... Real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback are stored in the power plant operation and maintenance knowledge base. After authentication between the mobile terminal and the server, the power plant operation and maintenance knowledge base is managed online. Specifically, by deploying four deep learning models—a sampling strategy generation model, an encapsulation status detection model, a fault detection model, and an intrusion detection model—as the core of intelligence, real-time data-driven decision-making and predictive maintenance are achieved for the sampling, encapsulation, and operation and maintenance processes, significantly improving the overall level of intelligence. Simultaneously, security mechanisms such as blockchain immutability, encrypted data transmission and storage, and authentication access control are embedded within the system. The entire lifecycle of data generation, transmission, storage, and access is used to build an intrinsic security protection system. On the one hand, this system fundamentally ensures the authenticity of sampled data and the traceability of the encapsulation process through blockchain notarization and encryption technology. On the other hand, it achieves proactive defense against network threats by leveraging intrusion detection models and secure communication mechanisms. Ultimately, through the deep integration of intelligent analysis and security protection, it collaboratively overcomes the three major challenges faced by traditional systems: improving the level of intelligence throughout the entire process while building an intrinsic security protection system to comprehensively solve the three major problems of poor authenticity of sampled data, weak traceability of encapsulation, and passive defense in operation and maintenance. Ultimately, it collaboratively improves the operating efficiency and safety reliability of the coal conveying system.

[0013] 2. By deploying multiple dedicated deep learning models, dynamic optimization of sampling strategies, fine-grained detection of encapsulation status, early warning of equipment failures, and real-time analysis of network intrusions were achieved. This transformed the operation and maintenance mode from passive response to proactive prediction and intelligent decision-making, significantly enhancing the overall intelligence level. Simultaneously, an intrinsic security protection system was constructed. Starting from the IoT sensing end, hardware security modules, streaming encryption, and blockchain evidence storage technologies were used to ensure the confidentiality, integrity, and immutability of data during collection, transmission, and storage. Furthermore, through mobile authentication and incremental model update mechanisms, closed-loop security management of data access and model iteration was achieved. This not only improved operational efficiency but also comprehensively addressed the three core challenges of poor sampling data authenticity, weak encapsulation traceability, and passive defense in operation and maintenance, collaboratively ensuring the efficient and reliable operation of the coal conveying system.

[0014] 3. By systematically acquiring and preprocessing historical multi-source heterogeneous data from thermal power plants and constructing multiple dedicated datasets, and simultaneously creating four deep learning models for sampling strategy generation, encapsulation status detection, fault detection, and intrusion detection, this integrated design avoids the problems of data silos and model dispersion in traditional methods. It can comprehensively process data from different sources, thereby improving data utilization efficiency, reducing redundant operations, and making operation and maintenance decisions more comprehensive and accurate.

[0015] 4. The industrial control computer collects real-time multi-source heterogeneous data based on the IoT sensing device array, and uses a pre-trained deep learning model to quickly generate outputs such as sampling strategy adjustment suggestions and packaging status detection results. This real-time capability ensures that the coal conveying process of thermal power plants can dynamically adapt to changes, such as timely adjustment of sampling frequency or packaging parameters, thereby reducing human delays, improving the level of automation in operation and maintenance, and effectively preventing operational errors or equipment failures.

[0016] 5. The server stores the decrypted data in the power plant operation and maintenance knowledge base, and also supports online management via mobile terminals, forming a closed-loop feedback system. This allows historical data and execution feedback to continuously enrich the power plant operation and maintenance knowledge base, providing data support for subsequent model optimization and decision-making. It also enables the accumulation and sharing of operation and maintenance experience, and improves the system's adaptability and long-term reliability.

[0017] 6. By acquiring multi-source heterogeneous data such as visual image data, physical attribute data, infrared spectral data, equipment status data, and network communication data, it covers multiple dimensions of coal transportation sampling, packaging, operation, and maintenance in thermal power plants, such as coal flow status, equipment operation, and network security. This comprehensiveness ensures that the model training does not rely on a single data source, reducing bias caused by data limitations, thereby improving the generalization ability and robustness of the deep learning model. For example, combining visual data and physical attribute data can more accurately identify sampling timing, avoid misjudgments caused by environmental changes, and help solve the problem of data isolation in existing technologies.

[0018] 7. By implementing specialized preprocessing steps for different types of data, such as cleaning, denoising, and enhancing visual image data, handling outliers and aligning time series for physical attribute data, and performing baseline correction and normalization for infrared spectral data, this targeted processing eliminates noise and inconsistencies, improves the reliability and usability of the data, provides clean and standardized input for model training, reduces the risk of model overfitting, and improves the accuracy of operation and maintenance decisions.

[0019] 8. Four independent datasets were built based on preprocessed data, targeting sampling strategy generation, encapsulation status detection, fault detection, and intrusion detection models respectively. This modular design allows each model to focus on a specific task, while achieving collaborative optimization through data sharing. For example, the fault detection model combines mechanical vibration and visual data, which can provide earlier warnings of equipment anomalies, avoiding the shortcomings of traditional single models in handling complex scenarios and improving the overall automation level of operation and maintenance.

[0020] 9. By performing fine-grained annotations on each dataset, such as including optimal sampling timing and location labels in the sampling strategy dataset, and fault detection dataset including fault type and handling suggestion labels, these annotations not only enrich the information content of supervised learning, but also enable the model to output more specific operational guidance, such as automatically generating protection suggestions. This goes beyond simple data collection, realizes knowledge injection, and helps the model converge quickly and be deployed in practice.

[0021] 10. By integrating multi-source data preprocessing, dataset construction, and annotation into a coherent process, it demonstrates the deep integration of deep learning and thermal power plant operation and maintenance. Through this integration, end-to-end intelligent management can be achieved, such as full automation from data collection to model training, which significantly improves operation and maintenance efficiency and safety.

[0022] 11. By acquiring multi-source heterogeneous data from thermal power plants (such as visual images, physical attributes, infrared spectra, equipment status, and network communication data) and performing targeted preprocessing (such as data cleaning, outlier handling, and normalization), high-quality and consistent data were ensured. Furthermore, modular datasets (such as those required for sampling strategies, encapsulation status detection, fault detection, and intrusion detection models) were constructed and supplemented with refined annotations (such as optimal sampling timing, fault type, and threat level labels). This improved the robustness, learning accuracy, and practicality of the deep learning model, enabling multi-task collaborative optimization. Ultimately, deep learning was deeply integrated with industrial operation and maintenance, significantly improving the automation efficiency, security, and scalability of coal conveying sampling and encapsulation operation and maintenance in thermal power plants.

[0023] 12. By integrating multiple data sources such as images, numerical values, spectra, vibration, and temperature, and employing advanced methods such as multi-head attention mechanisms and cross-attention fusion networks for multimodal feature extraction and fusion, the richness and robustness of feature representation are significantly improved. For example, in the sampling strategy generation model, combining coal flow images, coal volume data, coal contour data, and infrared spectral data can more comprehensively capture coal characteristics, thereby generating more accurate sampling strategies. This multimodal fusion approach solves the limitations of traditional methods that rely on a single data source, improving the accuracy and adaptability of coal transportation operation and maintenance in thermal power plants.

[0024] 13. By using lightweight neural networks (such as MobileNetV2) and modular designs (such as multilayer perceptrons and one-dimensional CNNs), the computational resources and storage requirements are reduced while ensuring model performance. For example, the image feature extraction unit uses MobileNetV2, which is suitable for real-time processing scenarios in industrial sites and avoids the latency problems caused by complex models. This design is easy to deploy in resource-constrained thermal power plant environments, improving practicality and scalability, and meeting the needs of industrial IoT for efficient operation and maintenance.

[0025] 14. By sharing a fully connected layer and multi-task output modules (such as encapsulating a state detection model to simultaneously handle integrity, clarity, and morphology detection), parallel learning and inference of multiple related tasks are achieved. This not only reduces model redundancy and training costs, but also improves the overall generalization ability through the correlation between tasks. For example, the fault detection model can simultaneously predict time, type, and level, providing a more complete fault analysis and helping maintenance personnel to respond quickly. This multi-task integration optimizes the workflow and improves maintenance efficiency.

[0026] 15. By integrating multiple cutting-edge neural network architectures such as Transformer, LSTM, GRU, and PointNet, and combining attention mechanisms (such as self-attention and cross-modal attention), the model's ability to capture temporal, spatial, and heterogeneous features is enhanced. For example, in the intrusion detection model, using bidirectional LSTM and graph attention networks to analyze network traffic can more effectively identify complex threat patterns. This integration of technologies improves the model's prediction accuracy and robustness.

[0027] 16. By covering the entire process from sampling strategy generation, encapsulation state detection to fault prediction and intrusion detection, and optimizing model training through custom loss functions (such as weighted multi-task loss), a complete closed-loop system is formed. This integrated design avoids the drawbacks of traditional segmented processing, realizes the automation and intelligence of coal conveying operation and maintenance in thermal power plants, reduces the cost of manual intervention, and improves system reliability and security.

[0028] 17. By innovatively integrating multimodal features (such as image, numerical, and spectral data) and employing lightweight network architectures (such as MobileNetV2) and advanced deep learning technologies (such as attention mechanisms and Transformers), efficient and accurate feature extraction and fusion are achieved, thereby improving the accuracy and real-time performance of coal conveying operation and maintenance in thermal power plants. Simultaneously, its multi-task learning mechanism and end-to-end intelligent design cover the entire process of sampling strategy generation, encapsulation state detection, fault prediction, and intrusion detection. By optimizing the loss function and integrating multiple neural networks, the robustness and generalization ability of the model are significantly enhanced, ultimately forming a comprehensive and automated operation and maintenance solution that effectively reduces labor costs and improves system reliability and security.

[0029] 18. By partitioning the dataset according to time order, the temporal relevance of the data is ensured, and the data leakage problem caused by random partitioning is avoided, thereby improving the prediction accuracy of the model in real industrial scenarios. Through the systematic partitioning of training, validation and test sets, and combined with gradient descent to optimize the loss function, the model can learn data patterns more effectively, reduce the risk of overfitting, and make the sampling strategy generation model and the encapsulated state detection model more robust, which is suitable for the dynamic operation and maintenance environment of thermal power plants.

[0030] 19. By applying dynamic pruning techniques after each training batch, the importance of neurons or connections is evaluated and redundant structures are pruned, significantly reducing the complexity of the model and the computational resource requirements. This model compression operation not only speeds up inference but also makes the model lighter, making it suitable for deployment on resource-constrained industrial control computers in thermal power plants, thereby reducing hardware costs and improving the system's energy efficiency ratio and real-time response capability.

[0031] 20. The deployment of tested models using containerization technology achieves environmental isolation and consistency, and simplifies the installation and update process of models on industrial control computers. This deployment method improves the scalability and maintainability of the system, facilitates remote monitoring and model iteration in thermal power plants, and reduces operation and maintenance downtime.

[0032] 21. By collecting actual multi-source heterogeneous data for model drift compensation training, the model parameters can be dynamically adjusted to compensate for performance degradation caused by factors such as equipment aging and environmental changes. This ensures that the model maintains high accuracy during long-term operation, improves the reliability and safety of coal transportation in thermal power plants, and reduces the risk of unexpected shutdowns.

[0033] 22. By collecting real-time multi-source heterogeneous data through an array of IoT sensing devices (including sampling cameras, packaging cameras, weight sensors, vibration sensors, and other devices) and storing it in a hardware security module, comprehensive and high-precision monitoring of the coal conveying process in thermal power plants is achieved. This enables the capture of subtle changes at each stage from sampling to packaging, thereby improving the integrity and reliability of the data. The introduction of the hardware security module further enhances data security, preventing unauthorized access or leakage, providing a high-quality data foundation for subsequent analysis, and meeting the high standards of data integrity and confidentiality required in industrial environments.

[0034] 23. The industrial control computer uses the eKuiper streaming computing engine to preprocess real-time multi-source heterogeneous data. This avoids the latency problem caused by traditional batch processing, realizes instant data cleaning and transformation, significantly improves data processing efficiency, ensures that data can be quickly input into deep learning models, and reduces system response time. This streaming processing method is particularly suitable for scenarios such as thermal power plants that require real-time monitoring, and can respond to emergencies in a timely manner, improving the agility and accuracy of operation and maintenance.

[0035] 24. By deploying multiple deep learning models (such as sampling strategy generation models, encapsulated state detection models, etc.) and using hardware acceleration technology to perform high-concurrency real-time inference on GPUs, the computational efficiency is greatly improved. It can handle multiple tasks simultaneously without sacrificing performance, thereby supporting complex decision-making processes. This design enables the system to quickly output adjustment suggestions and detection results, such as timely identification of faults or intrusions, reducing the operational risks caused by latency and improving the overall level of automation.

[0036] 25. The industrial control computer monitors performance indicators (such as computing load, memory usage, etc.) in real time and dynamically adjusts the model execution frequency based on preset inference priorities when resources are scarce. This ensures the real-time performance of high-priority models (such as fault detection and intrusion detection) and prioritizes critical tasks even under high load, thereby avoiding system crashes or performance degradation. This intelligent scheduling mechanism improves the robustness and adaptability of the system, enabling it to operate stably in variable industrial environments and reducing the need for manual intervention.

[0037] 26. By using an array of IoT sensing devices, real-time multi-source heterogeneous data can be comprehensively collected and securely stored. Combined with the eKuiper streaming computing engine for efficient preprocessing and GPU-accelerated high-concurrency real-time inference, multiple deep learning models can work together to automatically optimize sampling strategies and detect anomalies. At the same time, its dynamic resource management mechanism ensures the real-time performance of critical tasks (such as fault and intrusion detection) based on preset priorities. This improves the efficiency, accuracy, and safety of coal conveying operations in thermal power plants, while enhancing the reliability and automation level of the system and significantly reducing the need for human intervention and operational risks.

[0038] 27. The industrial control computer generates comprehensive operation and maintenance instructions in real time based on sampling strategy adjustment suggestions, encapsulation status detection results, fault handling suggestions, and intrusion prevention suggestions, and controls the operation of the execution unit while recording execution feedback. This integrated processing avoids the delay of manual intervention, realizes closed-loop automated operation and maintenance of the coal conveying process in thermal power plants, significantly improves system response speed and operational accuracy, reduces the risk of downtime due to human error, and thus improves overall operational efficiency.

[0039] 28. By employing multiple encryption measures (such as SHA256 hashing and AES-128-GCM symmetric encryption) and a blockchain notarization mechanism, data is serialized, compressed, scrambled, and encrypted. Before transmission, data fingerprints are calculated and stored on the blockchain. This not only ensures the confidentiality and integrity of data during transmission and storage but also provides verifiable anti-tampering protection through timestamps and authentication tags, reducing the risk of data leakage or malicious attacks and meeting the high security standards of industrial control systems.

[0040] 29. By adopting MessagePack serialization and LZ4 compression algorithms, multi-source heterogeneous data is converted into binary streams and compressed, which significantly reduces data volume and lowers network bandwidth requirements. Combined with TLS protocol uploading, efficient and low-latency data transmission is achieved, which is particularly important for real-time monitoring scenarios in thermal power plants. It can support the rapid processing of massive amounts of data and improve system resource utilization.

[0041] 30. The server uses incremental data from the power plant operation and maintenance knowledge base to periodically train the baseline model (e.g., generate a baseline model using a sampling strategy) and sends encrypted model change parameters to the industrial control computer to optimize the local model. This mechanism enables the system to learn from historical data and adjust itself to adapt to changes in operating conditions, gradually improve the accuracy of detection and decision-making, extend the service life of the system, and reduce later maintenance costs.

[0042] 31. By generating comprehensive operation and maintenance instructions in real time through industrial control computers and controlling execution units, a high degree of automation in coal conveying operation and maintenance of thermal power plants has been achieved, significantly improving response efficiency and operational accuracy. At the same time, the use of MessagePack serialization, LZ4 compression, multiple encryption (such as AES-128-GCM), and blockchain notarization mechanisms has optimized data processing and transmission speed and ensured the confidentiality, integrity, and tamper-proof capabilities of data. In addition, the server regularly trains the baseline model based on incremental data and encrypts and distributes updates, supporting the system's adaptive optimization and continuous learning, thereby enhancing reliability, traceability, and end-to-end collaborative operation and maintenance capabilities, and overall promoting the intelligentization, safety, and efficiency of thermal power plant operation and maintenance.

[0043] 32. By adopting a composite identity credential consisting of a digital certificate and a dynamic password, the authentication strength of mobile terminals accessing the server is significantly improved. The digital certificate provides non-repudiable identity proof based on public key infrastructure, while the dynamic password introduces a time-sensitive one-time factor, effectively preventing credential leakage or replay attacks. This two-factor authentication mechanism reduces the risk of unauthorized access, complies with current network security standards, and provides solid identity protection for critical operation and maintenance of thermal power plants.

[0044] 33. By negotiating temporary session keys based on the TLS protocol, the confidentiality and integrity of data transmission between the mobile terminal and the server are ensured. TLS is a widely recognized encryption protocol in the industry that can prevent eavesdropping and man-in-the-middle attacks. The use of temporary session keys further reduces the risk of long-term key leakage and achieves session-level dynamic security protection. This not only improves the reliability of operation and maintenance communication, but also meets the high security requirements of the mobile environment.

[0045] 34. The mobile terminal calculates the Message Authentication Code (MAC) through the temporary session key and sends the operation content, terminal identifier and operation time together. The server executes the operation after verifying the MAC. This method ensures that the data is not tampered with during transmission. Combined with timeliness verification (operation time) and legality verification (terminal identifier), it effectively defends against replay attacks and malicious modifications, and improves the credibility and execution accuracy of operation and maintenance instructions.

[0046] 35. By combining composite identity credentials with the TLS protocol, high-strength identity authentication and session encryption are achieved, fundamentally improving system access security. A multi-layered anti-tampering mechanism is constructed using message authentication codes, timeliness verification, and terminal identification verification to ensure the integrity and reliability of operation and maintenance instructions. At the same time, online operation and operation log recording significantly improve the real-time performance, traceability, and overall efficiency of operation and maintenance management, forming a highly efficient operation and maintenance solution that integrates security protection, reliable execution, and audit supervision.

[0047] 36. By deeply integrating multimodal data (such as visual images, physical attributes, and network communication data) and utilizing customized deep learning models (such as sampling strategy generation models and fault detection models), intelligent decision-making and automated processing of coal conveying sampling and packaging operation and maintenance in thermal power plants have been achieved, significantly improving decision-making accuracy and adaptive optimization capabilities. Combined with real-time inference, dynamic resource management, and multi-layered security measures (such as data encryption and blockchain notarization), the system ensures operation and maintenance efficiency, data integrity, and system stability. At the same time, through integrated management and scalable design, manual intervention has been reduced, enhancing the reliability, security, and long-term maintainability of thermal power plant operation and maintenance. Attached Figure Description

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0049] Figure 1 This is a flowchart of a coal sampling, encapsulation, and maintenance method for thermal power plants that combines deep learning, according to the present invention.

[0050] Figure 2 This is a schematic diagram of the structure of a coal conveying sampling and packaging operation and maintenance system for thermal power plants that incorporates deep learning, according to the present invention.

[0051] Figure 3 This is a schematic diagram of the sampling strategy generation model of the present invention.

[0052] Figure 4 This is a schematic diagram of the packaging state detection model of the present invention.

[0053] Figure 5 This is a schematic diagram of the fault detection model of the present invention.

[0054] Figure 6 This is a schematic diagram of the intrusion detection model of the present invention. Detailed Implementation

[0055] The overall approach of the technical solution in this application is as follows: By deploying four deep learning models—a sampling strategy generation model, a packaging state detection model, a fault detection model, and an intrusion detection model—as the core of intelligence, real-time data-driven decision-making and predictive maintenance are achieved for the sampling, packaging, and operation and maintenance processes, thereby significantly improving the overall intelligence level. Simultaneously, security mechanisms such as blockchain immutability, encrypted data transmission and storage, and authentication access control are embedded into the entire lifecycle of data generation, transmission, storage, and access, constructing an intrinsic security protection system. This system, on the one hand, fundamentally guarantees the authenticity of the sampled data and the traceability of the packaging process through blockchain notarization and encryption technology; on the other hand, it achieves proactive defense against network threats through intrusion detection models and secure communication mechanisms. Ultimately, through the deep integration of intelligent analysis and security protection, it collaboratively overcomes the three major challenges faced by traditional systems, and collaboratively improves the operating efficiency and reliability of the coal conveying system.

[0056] Please refer to Figures 1 to 6 As shown, a preferred embodiment of the present invention, a coal conveying sampling and encapsulation operation and maintenance method for thermal power plants combining deep learning, includes the following steps: Step S10: Obtain a large amount of historical multi-source heterogeneous data from the thermal power plant, preprocess the historical multi-source heterogeneous data, and construct the first dataset, the second dataset, the third dataset, and the fourth dataset. Step S20: Create a collaborative operation and maintenance intelligent agent that includes a sampling strategy generation model, an encapsulation state detection model, a fault detection model, and an intrusion detection model; The sampling strategy generation model is built based on a multimodal feature extraction module, a multimodal feature fusion module, and a strategy decision module. It is used to output sampling strategy adjustment suggestions based on input visual image data, physical attribute data, and infrared spectral data. The encapsulation state detection model is built based on an image feature extraction module, a weight feature extraction module, a multimodal feature aggregation module, and a multi-task output module. It is used to output encapsulation state detection results based on input visual image data and physical attribute data. The fault detection model is built based on a vibration feature extraction module, a temperature feature extraction module, a humidity feature extraction module, a visual feature extraction module, a contour feature extraction module, a heterogeneous feature fusion module, and a fault prediction module. It is used to output fault detection results based on input visual image data, physical attribute data, and equipment status data. The intrusion detection model is built upon a network traffic analysis module, an identity authentication analysis module, a log analysis module, a multimodal representation learning module, a comprehensive risk assessment module, and a protection suggestion generation module. It is used to output intrusion detection results based on input network communication data. Simultaneously deploying and running four complex deep learning models in real time on resource-constrained industrial control computers presents a significant practical challenge. This invention does not directly apply computationally intensive large-scale models, but unexpectedly tailors a lightweight network structure for each task (e.g., extensively employing MobileNetV2 and Multilayer Perceptron (MLP), and introducing dynamic pruning techniques). This "constrained design" approach, tailored to the computing resources of industrial environments, ensures the feasibility of high-concurrency real-time inference. Furthermore, a dynamic priority scheduling mechanism based on performance monitoring further guarantees the absolute priority of core tasks (such as fault and intrusion detection) when resource bottlenecks occur. This resource adaptability is key to the invention's applicability to harsh industrial environments.

[0057] Step S30: Train the collaborative operation and maintenance agent using the first dataset, the second dataset, the third dataset, and the fourth dataset; Step S40: After preprocessing the collected real-time multi-source heterogeneous data, input it into the collaborative operation and maintenance intelligent agent to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, generate comprehensive operation and maintenance instructions.

[0058] Step S10 specifically involves: Acquire a large amount of historical multi-source heterogeneous data from thermal power plants, including visual image data, physical attribute data, infrared spectral data, equipment status data, and network communication data; The visual image data includes coal flow images and packaging status images; the physical property data includes sample weight data, coal volume data, and coal outline data; the equipment status data includes mechanical vibration data, temperature data, and humidity data; and the network communication data includes network traffic data, equipment authentication information, and abnormal access logs. Coal flow images are acquired at sampling points using sampling cameras to analyze coal particle size distribution, color, and the presence of large foreign objects (such as wood or stones). Packaging status images are acquired using packaging cameras, including images of the sampling containers before, during, and after packaging. Sample weight data is acquired using weight sensors, measuring the weight of each batch of coal in real time; this is the basis for calculating coal quality indicators (such as calorific value) and is crucial for determining whether the sampling volume meets standards. Coal volume and contour data are acquired using laser scanners, performing a 3D scan of the coal flow on the conveyor belt to generate point cloud data, which is then used to calculate coal volume and contour. Mechanical vibration data is acquired using vibration sensors installed on rotating or reciprocating components such as samplers, crushers, and dividers, to determine early faults such as wear, imbalance, or loosening in bearings and gears. Temperature data is acquired using temperature sensors, including equipment temperature monitoring (for motors, gearboxes, etc.). The data collected includes the surface or internal temperature of key equipment (to prevent overheating damage) and ambient temperature (which affects coal moisture evaporation and equipment operation); humidity data is collected through humidity sensors, which monitor the air humidity in the sampling and packaging areas. Excessive humidity may cause coal to adhere and clog the equipment, and also affect the accuracy of infrared spectroscopy analysis; network communication data is collected through the gateway, and the network traffic data records the communication logs between the equipment and the industrial control computer, and between the industrial control computer and the server, including source / destination IP addresses, port numbers, communication protocols, packet size and frequency, etc.; equipment authentication information includes authentication data when each device accesses the gateway; abnormal access logs include records of unauthorized IP attempts to access the equipment, abnormal high-frequency commands, etc.; infrared spectral data is the spectral curve obtained by quickly scanning the sampled coal with an infrared spectrometer. The key quality parameters of the coal can be inferred from the spectral curve, such as: calorific value, moisture content, ash content, sulfur content, and volatile matter content.

[0059] The visual image data undergoes the following processes: data cleaning and denoising (invalid image removal: automatically identifying and deleting invalid images caused by camera malfunction, lens occlusion, excessive blurring, or severe under / overexposure; using algorithms such as median filtering and Gaussian filtering to reduce random noise introduced during image acquisition and transmission), data augmentation (to increase the diversity and scale of the dataset and improve the generalization ability of the model, image enhancement operations are performed, such as random rotation, flipping, scaling, brightness / contrast adjustment, adding slight noise, etc.), and size standardization and normalization (scaling all images to a uniform fixed size to adapt to the input requirements of the deep learning model; normalizing pixel values ​​from [0, 255] to [0, 1] or [-1, 255]). Preprocessing of the physical attribute data and equipment status data includes outlier handling (using statistical methods or machine learning-based methods to identify sensor readings that significantly exceed the normal range, and removing, smoothing, or filling out outliers using interpolation of preceding and following data), missing value handling (for brief interruptions or loss in the data stream, depending on the situation, using forward / backward filling, linear interpolation, or fitting with data from adjacent sensors), time series alignment (there may be slight differences in the data acquisition frequency and clock of different sensors, so it is necessary to use a unified time axis as a reference to resample (e.g., unify to once per second) and align all time series data to ensure that data points are meaningful at the same time), smoothing and filtering (for high-frequency data such as mechanical vibration data, use low-pass filters (such as moving average filters, Butterworth filters)). The preprocessing includes: removing high-frequency noise and retaining trend information reflecting the health status of the equipment; feature engineering (extracting meaningful features from the original time-series data for model use, such as: mechanical vibration data: extracting time-domain features (mean, root mean square, peak value, kurtosis), frequency-domain features (obtaining the spectrum through Fourier transform and analyzing the dominant frequency component); temperature data: extracting trends (slope), fluctuations (variance), etc.); and preprocessing the infrared spectral data including baseline correction (using adaptive iterative reweighted penalized least squares method to eliminate baseline drift of the spectrum), standard normal transformation (eliminating spectral intensity changes caused by sample particle size and surface scattering), smoothing and denoising (using methods such as Savitzky-Golay convolution smoothing to reduce noise while maintaining the spectral shape), and normalization (normalizing the entire spectral vector to a specific range to eliminate the influence of dimensions). Based on the preprocessed coal flow images, coal volume data, coal contour data, and infrared spectral data, a first dataset is constructed for training a sampling strategy generation model; based on the preprocessed encapsulation state images and sample weight data, a second dataset is constructed for training an encapsulation state detection model; based on the preprocessed mechanical vibration data, temperature data, humidity data, coal flow images, and coal contour data, a third dataset is constructed for training a fault detection model; and based on the preprocessed network traffic data, device authentication information, and abnormal access logs, a fourth dataset is constructed for training an intrusion detection model. The first dataset is labeled with at least the following: optimal sampling time (marking the specific time point that is the best sampling time under the current coal flow conditions), optimal sampling location (marking the lateral position on the conveyor belt that best represents the quality of the entire batch of coal (e.g., center, left, right)), and sampling action identifier (for some data segments, it may be labeled "no sampling required" because the coal quality is stable; while for other data segments, it may be labeled "sampling required"). The second dataset is labeled with at least the following: packaging integrity status label (good sealing: tight seal, no damage, no signs of leakage, appearing on the image as a flat seal line, no opening, no tearing of the packaging; stable weight data, consistent with the standard package weight and without a continuous decreasing trend; incomplete / damaged sealing: indicating physical damage to the packaging, which may appear on the image as holes, tears, cracked seals, or severe damage). Damaged or obscured markings; weight data may show abnormal weight loss (suggesting leakage of contents); incomplete sealing: labels indicate that the sealing operation was not completed or failed, which is shown in the image as an open seal, not pressed together, or labels not pasted or improperly pasted; weight data may be significantly lower than the standard value); clear labeling of packaging markings (clear and readable markings: indicating that key information such as sample number, sampling time, batch number, etc. on the packaging is clear, complete, and unobstructed in the image); abnormal packaging shape labels (normal shape: indicating that the packaging shape is regular and meets expectations; abnormal shape: indicating that the packaging has bulges, dents, deformations, etc., suggesting changes in the internal sample state or damage from external compression); the third dataset should be labeled with at least the fault time, fault type, fault severity level, and fault handling recommendations; the fourth dataset should be labeled with at least the risk behavior, threat level, and protection recommendations.

[0060] Existing technologies either employ a single data source for analysis or, while using multiple data sources, fail to establish a deep, specific correlation between the data and the actual operation and maintenance tasks. This invention creatively binds specific types of data (such as coal flow images and coal contour data) to specific operation and maintenance decisions (such as sampling strategy generation) and designs dedicated deep learning models (such as sampling strategy generation models) for them. This precise mapping relationship between "data-task-model" ensures that each model can learn the most effective features from the most relevant data, greatly improving the accuracy of decision-making. More importantly, the real-time decision-making of the industrial control computer, the knowledge base updates of the server, and the incremental learning of the model form a complete closed loop from data acquisition to model optimization, enabling the system to possess a continuous evolutionary capability that traditional methods lack.

[0061] In step S20, the multimodal feature extraction module contains three parallel units that process image, numerical, and spectral data respectively, and extract high-level feature vectors (coal image features, numerical features, and spectral features); the multimodal feature fusion module uses an attention mechanism to weightedly fuse the high-level feature vectors of the multimodal data, highlighting key information and obtaining fused features; the strategy decision module outputs a sampling strategy adjustment suggestion based on the fused features. The multimodal feature extraction module is constructed based on an image feature extraction unit, a numerical feature extraction unit, and a spectral feature extraction unit. The image feature extraction unit is used to extract coal image features from coal flow images using a lightweight first MobileNetV2 network (through convolutional layers, activation layers (ReLU), and pooling layers, outputting a 256-dimensional feature vector). The numerical feature extraction unit is used to extract numerical features from coal volume data and coal contour data using a first multilayer perceptron (containing two fully connected layers (128-dimensional and 64-dimensional), with a Dropout layer (ratio 0.2) in between to prevent overfitting). The spectral feature extraction unit is used to extract spectral features from infrared spectral data using a first one-dimensional convolutional neural network (containing two convolutional layers (kernel size 3) and a max-pooling layer, followed by an LSTM unit to handle sequence dependencies). MobileNetV2 networks have the advantages of having few parameters and being suitable for real-time processing, and their linear bottleneck structure avoids information loss; multilayer perceptrons have the advantages of simple structure and fast computation, and the combination of Dropout improves generalization ability. The multi-modal feature fusion module is used to fuse coal image features, numerical features, and spectral features through a multi-head attention mechanism (lightweight version, number of heads = 4) to obtain fused features; The strategy decision-making module is constructed based on a first shared fully connected layer, a timing output unit, a position output unit, a frequency output unit, and a suggestion output unit; The first shared fully connected layer is used to reduce the dimensionality of the fused features to obtain the first shared features; the timing output unit is used to infer the first shared features through the first fully connected layer to obtain the sampling timing (normalized time offset); the position output unit is used to infer the first shared features through the second fully connected layer to obtain the sampling position (normalized coordinates x, y); the frequency output unit is used to infer the first shared features through the third fully connected layer to obtain the sampling frequency (normalized interval); the suggestion output unit is used to output a sampling strategy adjustment suggestion carrying the sampling timing, sampling position, and sampling frequency. The sampling loss function of the model generated by the sampling strategy is: ; in, This represents the loss value of the sampling loss function; Indicates the actual sampling timing; Indicates the actual sampling location; Indicates the actual sampling frequency; Indicates the predicted sampling timing; Indicates the predicted sampling location; Indicates the predicted sampling frequency; All represent weighting coefficients; The image feature extraction module is used to extract encapsulated image features from the encapsulated state image through a lightweight second MobileNetV2 network; The weight feature extraction module is used to extract weight features from sample weight data using a second multilayer sensor. The multimodal feature aggregation module is used to aggregate encapsulated image features and weight features through a lightweight cross-attention fusion network to obtain aggregated features. The lightweight cross-attention fusion network introduces a cross-attention mechanism, which allows features of one modality to "query" relevant information of another modality, achieving more refined fusion without significantly increasing the number of parameters. The multi-task output module is constructed based on a second shared fully connected layer, an integrity output unit, a clarity output unit, a morphology output unit, and a state detection result output unit. The second shared fully connected layer is used to reduce the dimensionality of the aggregated features to obtain the second shared features; the integrity output unit is used to infer the second shared features through the fourth fully connected layer to obtain the encapsulation integrity status label; the clarity output unit is used to infer the second shared features through the fifth fully connected layer to obtain the encapsulation identification clarity label; the morphology output unit is used to infer the second shared features through the sixth fully connected layer to obtain the packaging morphology anomaly label; the state detection result output unit is used to output the packaging state detection result carrying the encapsulation integrity status label, the encapsulation identification clarity label, and the packaging morphology anomaly label; The packaging loss function of the packaging state detection model is: ; ; ; ; in, This represents the loss value of the encapsulated loss function; Represents the loss of the encapsulation integrity state sub-value; This indicates a loss of clarity in the packaging designation; This indicates a loss due to abnormal packaging shape; All represent weighting coefficients; A label indicating the actual encapsulation integrity status; This indicates the clarity of the packaging label. Labels indicating abnormal packaging conditions; A label indicating the predicted encapsulation integrity status; Indicates the predicted clarity of the encapsulation label; This indicates a predicted abnormal packaging shape label.

[0062] In step S20, the vibration feature extraction module is used to extract long-term dependent features from mechanical vibration data through a long short-term memory network. The self-attention mechanism in the first Transformer encoder enhances the focus on key time points in the long-term dependent features, and the vibration feature vector is output after feature compression through the seventh fully connected layer. The long short-term memory network processes sequence data, can remember long-term patterns, and avoids gradient vanishing. The self-attention mechanism dynamically weights important time points, improving feature representativeness and adapting to variable-length sequences. Dimensionality reduction is performed through the seventh fully connected layer, reducing the subsequent computational burden. The temperature feature extraction module is used to extract temperature feature vectors from temperature data through a gated recurrent unit and an eighth fully connected layer; the gated recurrent unit is used to output sequence summaries and is a simplified version of LSTM, which has the advantages of fewer parameters and faster training. The humidity feature extraction module is used to extract local features from humidity data through a second one-dimensional convolutional neural network, and then perform max pooling dimensionality reduction before outputting a humidity feature vector through the ninth fully connected layer. The one-dimensional convolutional neural network extracts features through a sliding window, which has the advantages of high computational efficiency and suitability for periodic signals. Max pooling can prevent overfitting. The visual feature extraction module is used to extract multi-level features from coal flow images using a convolutional neural network (CNN). A convolutional block attention module focuses on key regions within these multi-level features, and after global average pooling, a visual feature vector is output through a tenth fully connected layer. The hierarchical feature extraction using a convolutional neural network has the advantages of translation invariance and suitability for images. The convolutional block attention module adaptively focuses on abnormal regions, improving small target detection capabilities. Global average pooling reduces parameters and avoids overfitting. The contour feature extraction module is used to extract local geometric features from coal contour data using PointNet. After global max pooling of the local geometric features, the contour feature vector is output through the eleventh fully connected layer. The advantage of global max pooling is that it is insensitive to the order of points and has strong robustness. The heterogeneous feature fusion module is used to concatenate vibration feature vectors, temperature feature vectors, humidity feature vectors, visual feature vectors, and contour feature vectors into a long vector. After feature interaction and weighting by the second Transformer encoder, it is then subjected to dimensionality reduction and nonlinear fusion by the third multilayer perceptron to output the fused embedding. The second Transformer encoder processes different modes in parallel, which has the advantages of enhanced interaction and adaptability to dynamic weights. The deep fusion through the third multilayer perceptron has the advantages of high flexibility and improved generalization ability. The fault prediction module is constructed based on a time prediction unit, a type classification unit, a level classification unit, a suggestion generation unit, and a fault prediction output unit. The time prediction unit is used to perform regression output on the fusion embedding through the twelfth fully connected layer to obtain the fault prediction time (e.g., number of hours). The type classification unit is used to infer the fusion embedding through the thirteenth fully connected layer and the first Softmax layer to obtain the fault prediction type (such as wear, blockage, overheating, etc.). The classification unit is used to infer the fusion embedding through the fourteenth fully connected layer and the second Softmax layer to obtain the fault prediction severity level (such as low, medium, high). The suggestion generation unit is used to generate fault handling suggestions based on the fusion embedding through a rule-based suggestion generator. The fault prediction output unit is used to output fault detection results carrying fault prediction time, fault prediction type, fault prediction severity level, and fault handling suggestions. The fault loss function of the fault detection model is: ; in, This represents the loss value of the fault loss function; The regression loss for failure time prediction is represented by the mean squared error function; The classification loss, representing the fault type classification, is expressed using the cross-entropy loss function; The classification loss, representing the severity level of the fault, is expressed using the cross-entropy loss function. All represent weighting coefficients; The network traffic analysis module extracts temporal features (such as traffic bursts or periodic patterns) from network traffic data using a Bi-Short Short-Term Memory (Bi-LSTM) network (capturing bi-directional time dependencies and effectively detecting slow attacks such as low-speed DDoS). It then enhances the spatial relationship awareness of these temporal features through a Graph Attention Network (GAT) (mapping traffic data to a network topology graph (nodes are IP devices, edges are connections), learning attention weights between nodes to enhance spatial relationship awareness), thus obtaining network traffic features. GAT processes graph-structured data, improving the detection capability for complex network attacks (such as lateral movement); it creatively integrates graph networks to adapt to dynamic topologies. The identity authentication analysis module is used to extract identity authentication features from device identity authentication information through a third Transformer encoder (with self-attention mechanism); that is, to learn the long-term dependencies between authentication events and to weight important events (such as consecutive failed logins) through self-attention; the Transformer processes long sequences to avoid gradient vanishing in RNNs; and the attention mechanism improves the sensitivity to abnormal authentication. The log analysis module is used to extract contextual features (such as event sequence patterns) from abnormal access logs through a bidirectional gated recurrent unit (Bi-GRU) and to highlight abnormal log events (such as unconventional access times) through a self-attention mechanism to obtain log features; Bi-GRU balances computational efficiency and sequence modeling, making it suitable for real-time log streams; The multimodal representation learning module is used to unify the dimensions of network traffic features, authentication features, and log features through the fifteenth fully connected layer (mapping features from different modalities to the same dimension for easy fusion), and then perform weighted fusion through a cross-modal attention network to obtain a first aggregated representation. The first aggregated representation is then compressed and encoded (preserving key information) through the sixteenth fully connected layer and the ReLU activation function to obtain a second aggregated representation. This module creatively utilizes cross-modal attention to enhance multi-source data collaboration, improves the detection of complex attacks (such as multi-stage intrusions), introduces non-linearity through ReLU to enhance representational capabilities, and optimizes the output dimension to reduce subsequent computational burden. The comprehensive risk assessment module is used to extract high-level features from the second aggregated representation by combining a fourth multilayer perceptron with Dropout, and predict risk behavior and threat level based on the high-level features by combining a dual-branch network of Softmax and Sigmoid; it learns shared features through multi-task learning to improve efficiency; and it creatively combines classification and regression to adapt to dynamic threat environments. The protection suggestion generation module is used to map the risk behaviors and threat levels to a predefined rule base through a rule engine simulation layer (based on lookup tables or lightweight MLP), generate preliminary suggestion keywords, and generate natural language protection suggestions (text descriptions, such as "suggest blocking IP address X") based on the preliminary suggestion keywords through a sequence-to-sequence model combined with an attention mechanism, and output intrusion detection results carrying risk behaviors, threat levels, and protection suggestions. The intrusion loss function of the intrusion detection model is: ; in, This represents the loss value of the intrusion loss function; The loss is represented by the cross-entropy loss function, which indicates the loss from predicting risky behavior. The loss for threat level prediction is represented by a binary cross-entropy loss function. The loss for generating protection recommendations is represented by the sequence cross-entropy loss function; All of these represent weighting coefficients.

[0063] Step S30 specifically involves: The first dataset is divided into a first training set, a first validation set, and a first test set according to a preset first ratio based on time sequence, in order to train, validate, and test the sampling strategy generation model in the collaborative operation and maintenance agent; during the training process of the sampling strategy generation model, the sampling loss value of the sampling loss function is calculated, and the model parameters of the sampling strategy generation model are updated in reverse based on the sampling loss value using the gradient descent method to minimize the sampling loss function; after each training batch, the importance of neurons or connections in the sampling strategy generation model is evaluated by combining dynamic pruning technology, and redundant structures are pruned to perform model compression operation; The second dataset is divided into a second training set, a second validation set, and a second test set according to a preset second ratio based on time sequence, in order to train, validate, and test the encapsulation state detection model in the collaborative operation and maintenance agent. During the training process of the encapsulation state detection model, the encapsulation loss value of the encapsulation loss function is calculated, and the model parameters of the encapsulation state detection model are updated in reverse based on the encapsulation loss value using gradient descent to minimize the encapsulation loss function. After each training batch, the importance of neurons or connections in the encapsulation state detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The third dataset is divided into a third training set, a third validation set, and a third test set according to a preset third ratio based on time sequence, in order to train, validate, and test the fault detection model in the collaborative operation and maintenance agent. During the training process of the fault detection model, the fault loss value of the fault loss function is calculated, and the model parameters of the fault detection model are updated in reverse based on the fault loss value using the gradient descent method to minimize the fault loss function. After each training batch, the importance of neurons or connections in the fault detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The fourth dataset is divided into a fourth training set, a fourth validation set, and a fourth test set according to a preset fourth ratio based on time sequence, in order to train, validate, and test the intrusion detection model in the collaborative operation and maintenance agent. During the training process of the intrusion detection model, the intrusion loss value of the intrusion loss function is calculated, and the model parameters of the intrusion detection model are updated in reverse based on the intrusion loss value using the gradient descent method to minimize the intrusion loss function. After each training batch, the importance of neurons or connections in the intrusion detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The tested sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model were deployed to the industrial control computer of the thermal power plant using containerization technology. The actual multi-source heterogeneous data from thermal power plants are collected to train the sampling strategy generation model, encapsulation status detection model, fault detection model, and intrusion detection model deployed on the industrial control computer for model drift compensation training.

[0064] Step S40 specifically includes: Step S41: Based on the input sampling control command, the industrial control computer samples and packages the coal conveyed by the conveyor belt, and simultaneously collects real-time multi-source heterogeneous data through the Internet of Things sensing device array. Step S42: The industrial control computer preprocesses the real-time multi-source heterogeneous data and inputs it into the deployed collaborative operation and maintenance intelligent agent. The collaborative operation and maintenance intelligent agent performs inference through the sampling strategy generation model, encapsulation status detection model, fault detection model and intrusion detection model to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Step S43: The industrial control computer generates a comprehensive operation and maintenance instruction based on the sampling strategy adjustment suggestion, the packaging status detection result, the fault detection result, and the intrusion detection result, executes the comprehensive operation and maintenance instruction, and records the execution feedback; Step S44: The industrial control computer encrypts the real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback into an encrypted running data packet, calculates the data fingerprint of the encrypted running data packet, uploads it to the blockchain, and uploads the encrypted running data packet to the server. Step S45: The server verifies and decrypts the received encrypted operation data packet to obtain real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback, and stores them in a pre-created power plant operation and maintenance knowledge base; Step S46: After the mobile terminal authenticates with the server, it performs online management of the power plant operation and maintenance knowledge base.

[0065] One of the most outstanding innovations of this invention lies in transforming security protection from an "external" component to an "internal" one. Security is no longer an additional module independent of business processes, but is deeply embedded in every link from data acquisition (hardware security modules), transmission (encryption, blockchain notarization) to access control (strong authentication on mobile devices). For example, the intrusion detection model itself, as one of the four core intelligent models, directly participates in the generation of comprehensive operation and maintenance instructions, which means that the detection of security threats and operation and maintenance responses are seamlessly linked. In addition, by using blockchain notarization to sample and encapsulate key data fingerprints throughout the entire process, underlying support is provided for the authenticity of data and the immutability of operations. This solution, which combines the reliability of intelligent analysis results with the trust mechanism of blockchain, provides a brand-new technical path to solve the long-standing "trust" problem in the thermal power plant industry.

[0066] Step S41 specifically involves: The industrial control computer receives the input sampling control command and, based on the initial sampling timing, initial sampling point, and initial sampling frequency carried in the sampling control command, controls the coal sampler to sample the coal conveyed by the conveyor belt, loads the sampled coal into a sampling bucket, and controls the packaging machine to package the sampled bucket. Simultaneously, it collects real-time multi-source heterogeneous data through an array of IoT sensing devices, including a sampling camera, a packaging camera, a weight sensor, a vibration sensor, a temperature sensor, a humidity sensor, a laser scanner, an infrared spectrometer, and a gateway, and stores the collected real-time multi-source heterogeneous data in a hardware security module. Step S42 specifically involves: The industrial control computer uses the eKuiper streaming computing engine to preprocess the real-time multi-source heterogeneous data within the hardware security module, and then inputs it into the deployed sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. The sampling strategy generation model, packaging state detection model, fault detection model, and intrusion detection model are based on hardware acceleration technology and perform high-concurrency real-time inference on the GPU, respectively outputting sampling strategy adjustment suggestions, packaging state detection results, fault detection results, and intrusion detection results. During the inference process, the industrial control computer monitors in real time the performance indicators of the local machine, including at least computing load, memory usage, GPU memory usage, and inference latency. When one of the performance indicators exceeds the corresponding threshold, it is determined that the sampling strategy generation model, packaging state detection model, fault detection model, and intrusion detection model cannot be supported for synchronous inference. Based on the preset inference priority, the execution frequency of low-priority models is dynamically reduced to ensure the real-time performance of high-priority models. The inference priority is as follows: fault detection model > intrusion detection model > encapsulation state detection model > sampling strategy generation model.

[0067] Step S43 specifically involves: The industrial control computer generates comprehensive operation and maintenance instructions in real time based on the sampling strategy adjustment suggestions, the packaging status detection results, the fault handling suggestions carried by the fault detection results, and the protection suggestions carried by the intrusion detection results. It controls the corresponding execution unit to execute the comprehensive operation and maintenance instructions, records the execution feedback of the execution unit, encrypts the comprehensive operation and maintenance instructions into encrypted operation and maintenance instructions, and backs up the encrypted operation and maintenance instructions to the instruction library of the server. The specific steps for encrypting the comprehensive operation and maintenance instructions into encrypted operation and maintenance instructions are as follows: The integrated operation and maintenance instructions are converted into a byte sequence using UTF-8 encoding; a temporary encryption key is derived using HMAC-SHA256 based on a preset shared key and the length of the byte sequence; the byte sequence is encrypted using the ChaCha20 stream encryption algorithm with the temporary encryption key to obtain intermediate ciphertext data of the same length as the byte sequence; the MD5 value of the last 64 bits of the shared key is calculated using the MD5 algorithm, the last two characters of the MD5 value are converted into hexadecimal data, and the first character of the hexadecimal data is used as the displacement; based on the displacement, each character of the intermediate ciphertext data is cyclically shifted to the right to obtain obfuscated data, and the obfuscated data is Base64 encoded to obtain the encrypted operation and maintenance instructions; Both the server and the industrial control computer pre-store the shared key.

[0068] In practice, the integrated operation and maintenance instructions have interactive and weight allocation logic. For example, when the intrusion detection result shows a high-risk network attack, the protection suggestions carried by the intrusion detection result are executed first, and even if the fault detection result shows a low-level fault, a shutdown inspection instruction is generated. When the packaging status detection result shows a serious anomaly, the fault sensitivity of the packaging mechanical components in the fault detection model is increased.

[0069] Step S44 specifically involves: The industrial control computer serializes the real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback through MessagePack to obtain a binary data stream D_serial. The binary data stream D_serial is then compressed using the LZ4 compression algorithm to obtain compressed data D_compressed. Obtain the current timestamp, calculate the hash value of the timestamp using the SHA256 algorithm, and perform an XOR operation on the first 128 bits of the hash value with the compressed data D_compressed to obtain the scrambled data D_confused. A random initialization vector is generated. The random initialization vector and the preset shared key are called through the symmetric encryption algorithm AES-128-GCM to encrypt the scrambling data D_confused to obtain the ciphertext C_cipher and the authentication tag Tag_auth. The ciphertext C_cipher, the authentication tag Tag_auth, the random initialization vector, and the timestamp are packaged into an encrypted runtime data packet; The encrypted running data packet is calculated using the SHA-3 algorithm, the data fingerprint is stored on the blockchain, and the encrypted running data packet is uploaded to the server in real time via the TLS protocol.

[0070] Step S45 specifically involves: The server receives the encrypted running data packet in real time. After verifying the integrity of the encrypted running data packet by the data fingerprint stored on the blockchain, it parses the encrypted running data packet to obtain the ciphertext C_cipher, the authentication tag Tag_auth, the random initialization vector, and the timestamp. The random initialization vector and the preset shared key are invoked using the symmetric encryption algorithm AES-128-GCM to decrypt the ciphertext C_cipher and verify the authentication tag Tag_auth to obtain the scrambling data D_confused. The hash value of the timestamp is calculated using the SHA256 algorithm. The first 128 bits of the hash value are XORed with the scrambled data D_confused to obtain compressed data D_compressed. The compressed data D_compressed is decompressed using the LZ4 compression algorithm to obtain a binary data stream D_serial. The binary data stream D_serial is deserialized using MessagePack to obtain real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback. The real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback are stored in a pre-created power plant operation and maintenance knowledge base. The server deploys baseline models for sampling strategy generation, encapsulation state detection, fault detection, and intrusion detection, corresponding to the sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. Based on a preset iteration cycle, incremental data is read from the power plant operation and maintenance knowledge base. An incremental dataset is constructed based on each incremental data set. The sampling strategy generation baseline model, encapsulation state detection baseline model, fault detection baseline model, and intrusion detection baseline model are trained based on the incremental dataset. The trained model change parameters are obtained, and the model change parameters are encrypted and sent to the industrial control computer to optimize the sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. When the model change parameters are issued, they are encrypted using the industrial control computer's asymmetric encryption public key, and the transaction hash of the parameter distribution is recorded through the blockchain to ensure the integrity and traceability of the model update process and prevent malicious model implantation.

[0071] Step S46 specifically involves: The mobile terminal sends an access request to the server carrying a composite identity credential, which consists of a digital certificate and a dynamic password. After receiving the access request, the server verifies the composite identity credential carried in the access request and negotiates a temporary session key with the mobile terminal for this session based on the TLS protocol. The mobile terminal obtains the input operation content for the power plant operation and maintenance knowledge base, the terminal identifier of the mobile terminal, and the current operation time. It calculates the message authentication code of the operation content, terminal identifier, and operation time through the temporary session key, generates an operation request based on the operation content, terminal identifier, operation time, and message authentication code, and sends the operation request to the server. After verifying the message authentication code carried in the operation request using the temporary session key, the server performs timeliness verification using the operation time and legality verification using the terminal identifier before executing the operation content to manage the power plant operation and maintenance knowledge base online and record operation logs.

[0072] A preferred embodiment of the coal conveying sampling and packaging operation and maintenance system for thermal power plants that incorporates deep learning according to the present invention includes the following modules: The dataset construction module is used to acquire a large amount of historical multi-source heterogeneous data from thermal power plants, preprocess the aforementioned historical multi-source heterogeneous data, and construct the first dataset, the second dataset, the third dataset, and the fourth dataset. The agent creation module is used to create a collaborative operation and maintenance agent that includes a sampling strategy generation model, an encapsulated state detection model, a fault detection model, and an intrusion detection model. The sampling strategy generation model is built based on a multimodal feature extraction module, a multimodal feature fusion module, and a strategy decision module. It is used to output sampling strategy adjustment suggestions based on input visual image data, physical attribute data, and infrared spectral data. The encapsulation state detection model is built based on an image feature extraction module, a weight feature extraction module, a multimodal feature aggregation module, and a multi-task output module. It is used to output encapsulation state detection results based on input visual image data and physical attribute data. The fault detection model is built based on a vibration feature extraction module, a temperature feature extraction module, a humidity feature extraction module, a visual feature extraction module, a contour feature extraction module, a heterogeneous feature fusion module, and a fault prediction module. It is used to output fault detection results based on input visual image data, physical attribute data, and equipment status data. The intrusion detection model is built upon a network traffic analysis module, an identity authentication analysis module, a log analysis module, a multimodal representation learning module, a comprehensive risk assessment module, and a protection suggestion generation module. It is used to output intrusion detection results based on input network communication data. Simultaneously deploying and running four complex deep learning models in real time on resource-constrained industrial control computers presents a significant practical challenge. This invention does not directly apply computationally intensive large-scale models, but unexpectedly tailors a lightweight network structure for each task (e.g., extensively employing MobileNetV2 and Multilayer Perceptron (MLP), and introducing dynamic pruning techniques). This "constrained design" approach, tailored to the computing resources of industrial environments, ensures the feasibility of high-concurrency real-time inference. Furthermore, a dynamic priority scheduling mechanism based on performance monitoring further guarantees the absolute priority of core tasks (such as fault and intrusion detection) when resource bottlenecks occur. This resource adaptability is key to the invention's applicability to harsh industrial environments.

[0073] The agent training and deployment module is used to train the collaborative operation and maintenance agent using the first dataset, the second dataset, the third dataset, and the fourth dataset. The model inference module is used to preprocess the collected real-time multi-source heterogeneous data and input it into the collaborative operation and maintenance intelligent agent to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, a comprehensive operation and maintenance instruction is generated.

[0074] The dataset construction module is specifically used for: Acquire a large amount of historical multi-source heterogeneous data from thermal power plants, including visual image data, physical attribute data, infrared spectral data, equipment status data, and network communication data; The visual image data includes coal flow images and packaging status images; the physical property data includes sample weight data, coal volume data, and coal outline data; the equipment status data includes mechanical vibration data, temperature data, and humidity data; and the network communication data includes network traffic data, equipment authentication information, and abnormal access logs. Coal flow images are acquired at sampling points using sampling cameras to analyze coal particle size distribution, color, and the presence of large foreign objects (such as wood or stones). Packaging status images are acquired using packaging cameras, including images of the sampling containers before, during, and after packaging. Sample weight data is acquired using weight sensors, measuring the weight of each batch of coal in real time; this is the basis for calculating coal quality indicators (such as calorific value) and is crucial for determining whether the sampling volume meets standards. Coal volume and contour data are acquired using laser scanners, performing a 3D scan of the coal flow on the conveyor belt to generate point cloud data, which is then used to calculate coal volume and contour. Mechanical vibration data is acquired using vibration sensors installed on rotating or reciprocating components such as samplers, crushers, and dividers, to determine early faults such as wear, imbalance, or loosening in bearings and gears. Temperature data is acquired using temperature sensors, including equipment temperature monitoring (for motors, gearboxes, etc.). The data collected includes the surface or internal temperature of key equipment (to prevent overheating damage) and ambient temperature (which affects coal moisture evaporation and equipment operation); humidity data is collected through humidity sensors, which monitor the air humidity in the sampling and packaging areas. Excessive humidity may cause coal to adhere and clog the equipment, and also affect the accuracy of infrared spectroscopy analysis; network communication data is collected through the gateway, and the network traffic data records the communication logs between the equipment and the industrial control computer, and between the industrial control computer and the server, including source / destination IP addresses, port numbers, communication protocols, packet size and frequency, etc.; equipment authentication information includes authentication data when each device accesses the gateway; abnormal access logs include records of unauthorized IP attempts to access the equipment, abnormal high-frequency commands, etc.; infrared spectral data is the spectral curve obtained by quickly scanning the sampled coal with an infrared spectrometer. The key quality parameters of the coal can be inferred from the spectral curve, such as: calorific value, moisture content, ash content, sulfur content, and volatile matter content.

[0075] The visual image data undergoes the following processes: data cleaning and denoising (invalid image removal: automatically identifying and deleting invalid images caused by camera malfunction, lens occlusion, excessive blurring, or severe under / overexposure; using algorithms such as median filtering and Gaussian filtering to reduce random noise introduced during image acquisition and transmission), data augmentation (to increase the diversity and scale of the dataset and improve the generalization ability of the model, image enhancement operations are performed, such as random rotation, flipping, scaling, brightness / contrast adjustment, adding slight noise, etc.), and size standardization and normalization (scaling all images to a uniform fixed size to adapt to the input requirements of the deep learning model; normalizing pixel values ​​from [0, 255] to [0, 1] or [-1, 255]). Preprocessing of the physical attribute data and equipment status data includes outlier handling (using statistical methods or machine learning-based methods to identify sensor readings that significantly exceed the normal range, and removing, smoothing, or filling out outliers using interpolation of preceding and following data), missing value handling (for brief interruptions or loss in the data stream, depending on the situation, using forward / backward filling, linear interpolation, or fitting with data from adjacent sensors), time series alignment (there may be slight differences in the data acquisition frequency and clock of different sensors, so it is necessary to use a unified time axis as a reference to resample (e.g., unify to once per second) and align all time series data to ensure that data points are meaningful at the same time), smoothing and filtering (for high-frequency data such as mechanical vibration data, use low-pass filters (such as moving average filters, Butterworth filters)). The preprocessing includes: removing high-frequency noise and retaining trend information reflecting the health status of the equipment; feature engineering (extracting meaningful features from the original time-series data for model use, such as: mechanical vibration data: extracting time-domain features (mean, root mean square, peak value, kurtosis), frequency-domain features (obtaining the spectrum through Fourier transform and analyzing the dominant frequency component); temperature data: extracting trends (slope), fluctuations (variance), etc.); and preprocessing the infrared spectral data including baseline correction (using adaptive iterative reweighted penalized least squares method to eliminate baseline drift of the spectrum), standard normal transformation (eliminating spectral intensity changes caused by sample particle size and surface scattering), smoothing and denoising (using methods such as Savitzky-Golay convolution smoothing to reduce noise while maintaining the spectral shape), and normalization (normalizing the entire spectral vector to a specific range to eliminate the influence of dimensions). Based on the preprocessed coal flow images, coal volume data, coal contour data, and infrared spectral data, a first dataset is constructed for training a sampling strategy generation model; based on the preprocessed encapsulation state images and sample weight data, a second dataset is constructed for training an encapsulation state detection model; based on the preprocessed mechanical vibration data, temperature data, humidity data, coal flow images, and coal contour data, a third dataset is constructed for training a fault detection model; and based on the preprocessed network traffic data, device authentication information, and abnormal access logs, a fourth dataset is constructed for training an intrusion detection model. The first dataset is labeled with at least the following: optimal sampling time (marking the specific time point that is the best sampling time under the current coal flow conditions), optimal sampling location (marking the lateral position on the conveyor belt that best represents the quality of the entire batch of coal (e.g., center, left, right)), and sampling action identifier (for some data segments, it may be labeled "no sampling required" because the coal quality is stable; while for other data segments, it may be labeled "sampling required"). The second dataset is labeled with at least the following: packaging integrity status label (good sealing: tight seal, no damage, no signs of leakage, appearing on the image as a flat seal line, no opening, no tearing of the packaging; stable weight data, consistent with the standard package weight and without a continuous decreasing trend; incomplete / damaged sealing: indicating physical damage to the packaging, which may appear on the image as holes, tears, cracked seals, or severe damage). Damaged or obscured markings; weight data may show abnormal weight loss (suggesting leakage of contents); incomplete sealing: labels indicate that the sealing operation was not completed or failed, which is shown in the image as an open seal, not pressed together, or labels not pasted or improperly pasted; weight data may be significantly lower than the standard value); clear labeling of packaging markings (clear and readable markings: indicating that key information such as sample number, sampling time, batch number, etc. on the packaging is clear, complete, and unobstructed in the image); abnormal packaging shape labels (normal shape: indicating that the packaging shape is regular and meets expectations; abnormal shape: indicating that the packaging has bulges, dents, deformations, etc., suggesting changes in the internal sample state or damage from external compression); the third dataset should be labeled with at least the fault time, fault type, fault severity level, and fault handling recommendations; the fourth dataset should be labeled with at least the risk behavior, threat level, and protection recommendations.

[0076] Existing technologies either employ a single data source for analysis or, while using multiple data sources, fail to establish a deep, specific correlation between the data and the actual operation and maintenance tasks. This invention creatively binds specific types of data (such as coal flow images and coal contour data) to specific operation and maintenance decisions (such as sampling strategy generation) and designs dedicated deep learning models (such as sampling strategy generation models) for them. This precise mapping relationship between "data-task-model" ensures that each model can learn the most effective features from the most relevant data, greatly improving the accuracy of decision-making. More importantly, the real-time decision-making of the industrial control computer, the knowledge base updates of the server, and the incremental learning of the model form a complete closed loop from data acquisition to model optimization, enabling the system to possess a continuous evolutionary capability that traditional methods lack.

[0077] In the intelligent agent creation module, the multimodal feature extraction module contains three parallel units that process image, numerical, and spectral data respectively to extract high-level feature vectors (coal image features, numerical features, and spectral features); the multimodal feature fusion module uses an attention mechanism to weightedly fuse the high-level feature vectors of the multimodal data, highlighting key information to obtain fused features; the policy decision module, based on the fused features, regresses and outputs suggestions for adjusting the sampling policy. The multimodal feature extraction module is constructed based on an image feature extraction unit, a numerical feature extraction unit, and a spectral feature extraction unit. The image feature extraction unit is used to extract coal image features from coal flow images using a lightweight first MobileNetV2 network (through convolutional layers, activation layers (ReLU), and pooling layers, outputting a 256-dimensional feature vector). The numerical feature extraction unit is used to extract numerical features from coal volume data and coal contour data using a first multilayer perceptron (containing two fully connected layers (128-dimensional and 64-dimensional), with a Dropout layer (ratio 0.2) in between to prevent overfitting). The spectral feature extraction unit is used to extract spectral features from infrared spectral data using a first one-dimensional convolutional neural network (containing two convolutional layers (kernel size 3) and a max-pooling layer, followed by an LSTM unit to handle sequence dependencies). MobileNetV2 networks have the advantages of having few parameters and being suitable for real-time processing, and their linear bottleneck structure avoids information loss; multilayer perceptrons have the advantages of simple structure and fast computation, and the combination of Dropout improves generalization ability. The multi-modal feature fusion module is used to fuse coal image features, numerical features, and spectral features through a multi-head attention mechanism (lightweight version, number of heads = 4) to obtain fused features; The strategy decision-making module is constructed based on a first shared fully connected layer, a timing output unit, a position output unit, a frequency output unit, and a suggestion output unit; The first shared fully connected layer is used to reduce the dimensionality of the fused features to obtain the first shared features; the timing output unit is used to infer the first shared features through the first fully connected layer to obtain the sampling timing (normalized time offset); the position output unit is used to infer the first shared features through the second fully connected layer to obtain the sampling position (normalized coordinates x, y); the frequency output unit is used to infer the first shared features through the third fully connected layer to obtain the sampling frequency (normalized interval); the suggestion output unit is used to output a sampling strategy adjustment suggestion carrying the sampling timing, sampling position, and sampling frequency. The sampling loss function of the model generated by the sampling strategy is: ; in, This represents the loss value of the sampling loss function; Indicates the actual sampling timing; Indicates the actual sampling location; Indicates the actual sampling frequency; Indicates the predicted sampling timing; Indicates the predicted sampling location; Indicates the predicted sampling frequency; All represent weighting coefficients; The image feature extraction module is used to extract encapsulated image features from the encapsulated state image through a lightweight second MobileNetV2 network; The weight feature extraction module is used to extract weight features from sample weight data using a second multilayer sensor. The multimodal feature aggregation module is used to aggregate encapsulated image features and weight features through a lightweight cross-attention fusion network to obtain aggregated features. The lightweight cross-attention fusion network introduces a cross-attention mechanism, which allows features of one modality to "query" relevant information of another modality, achieving more refined fusion without significantly increasing the number of parameters. The multi-task output module is constructed based on a second shared fully connected layer, an integrity output unit, a clarity output unit, a morphology output unit, and a state detection result output unit. The second shared fully connected layer is used to reduce the dimensionality of the aggregated features to obtain the second shared features; the integrity output unit is used to infer the second shared features through the fourth fully connected layer to obtain the encapsulation integrity status label; the clarity output unit is used to infer the second shared features through the fifth fully connected layer to obtain the encapsulation identification clarity label; the morphology output unit is used to infer the second shared features through the sixth fully connected layer to obtain the packaging morphology anomaly label; the state detection result output unit is used to output the packaging state detection result carrying the encapsulation integrity status label, the encapsulation identification clarity label, and the packaging morphology anomaly label; The packaging loss function of the packaging state detection model is: ; ; ; ; in, This represents the loss value of the encapsulated loss function; Represents the loss of the encapsulation integrity state sub-value; This indicates a loss of clarity in the packaging designation; This indicates a loss due to abnormal packaging shape; All represent weighting coefficients; A label indicating the actual encapsulation integrity status; Indicates the clarity of the actual packaging label; Labels indicating abnormal packaging conditions; A label indicating the predicted encapsulation integrity status; Indicates the predicted clarity of the encapsulation label; This indicates a predicted abnormal packaging shape label.

[0078] In the intelligent agent creation module, the vibration feature extraction module is used to extract long-term dependent features from mechanical vibration data through a long short-term memory network. The self-attention mechanism in the first Transformer encoder enhances the focus on key time points in the long-term dependent features, and the vibration feature vector is output after feature compression through the seventh fully connected layer. The long short-term memory network processes sequential data, can remember long-term patterns, and avoids gradient vanishing. The self-attention mechanism dynamically weights important time points, improving feature representativeness and adapting to variable-length sequences. Dimensionality reduction is performed through the seventh fully connected layer to reduce subsequent computational burden. The temperature feature extraction module is used to extract temperature feature vectors from temperature data through a gated recurrent unit and an eighth fully connected layer; the gated recurrent unit is used to output sequence summaries and is a simplified version of LSTM, which has the advantages of fewer parameters and faster training. The humidity feature extraction module is used to extract local features from humidity data through a second one-dimensional convolutional neural network, and then perform max pooling dimensionality reduction before outputting a humidity feature vector through the ninth fully connected layer. The one-dimensional convolutional neural network extracts features through a sliding window, which has the advantages of high computational efficiency and suitability for periodic signals. Max pooling can prevent overfitting. The visual feature extraction module is used to extract multi-level features from coal flow images using a convolutional neural network (CNN). A convolutional block attention module focuses on key regions within these multi-level features, and after global average pooling, a visual feature vector is output through a tenth fully connected layer. The hierarchical feature extraction using a convolutional neural network has the advantages of translation invariance and suitability for images. The convolutional block attention module adaptively focuses on abnormal regions, improving small target detection capabilities. Global average pooling reduces parameters and avoids overfitting. The contour feature extraction module is used to extract local geometric features from coal contour data using PointNet. After global max pooling of the local geometric features, the contour feature vector is output through the eleventh fully connected layer. The advantage of global max pooling is that it is insensitive to the order of points and has strong robustness. The heterogeneous feature fusion module is used to concatenate vibration feature vectors, temperature feature vectors, humidity feature vectors, visual feature vectors, and contour feature vectors into a long vector. After feature interaction and weighting by the second Transformer encoder, it is then subjected to dimensionality reduction and nonlinear fusion by the third multilayer perceptron to output the fused embedding. The second Transformer encoder processes different modes in parallel, which has the advantages of enhanced interaction and adaptability to dynamic weights. The deep fusion through the third multilayer perceptron has the advantages of high flexibility and improved generalization ability. The fault prediction module is constructed based on a time prediction unit, a type classification unit, a level classification unit, a suggestion generation unit, and a fault prediction output unit. The time prediction unit is used to perform regression output on the fusion embedding through the twelfth fully connected layer to obtain the fault prediction time (e.g., number of hours). The type classification unit is used to infer the fusion embedding through the thirteenth fully connected layer and the first Softmax layer to obtain the fault prediction type (such as wear, blockage, overheating, etc.). The classification unit is used to infer the fusion embedding through the fourteenth fully connected layer and the second Softmax layer to obtain the fault prediction severity level (such as low, medium, high). The suggestion generation unit is used to generate fault handling suggestions based on the fusion embedding through a rule-based suggestion generator. The fault prediction output unit is used to output fault detection results carrying fault prediction time, fault prediction type, fault prediction severity level, and fault handling suggestions. The fault loss function of the fault detection model is: ; in, This represents the loss value of the fault loss function; The regression loss for failure time prediction is represented by the mean squared error function; The classification loss, representing the fault type classification, is expressed using the cross-entropy loss function; The classification loss, representing the severity level of the fault, is expressed using the cross-entropy loss function. All represent weighting coefficients; The network traffic analysis module extracts temporal features (such as traffic bursts or periodic patterns) from network traffic data using a Bi-Short Short-Term Memory (Bi-LSTM) network (capturing bi-directional time dependencies and effectively detecting slow attacks such as low-speed DDoS). It then enhances the spatial relationship awareness of these temporal features through a Graph Attention Network (GAT) (mapping traffic data to a network topology graph (nodes are IP devices, edges are connections), learning attention weights between nodes to enhance spatial relationship awareness), thus obtaining network traffic features. GAT processes graph-structured data, improving the detection capability for complex network attacks (such as lateral movement); it creatively integrates graph networks to adapt to dynamic topologies. The identity authentication analysis module is used to extract identity authentication features from device identity authentication information through a third Transformer encoder (with self-attention mechanism); that is, to learn the long-term dependencies between authentication events and to weight important events (such as consecutive failed logins) through self-attention; the Transformer processes long sequences to avoid gradient vanishing in RNNs; and the attention mechanism improves the sensitivity to abnormal authentication. The log analysis module is used to extract contextual features (such as event sequence patterns) from abnormal access logs through a bidirectional gated recurrent unit (Bi-GRU) and to highlight abnormal log events (such as unconventional access times) through a self-attention mechanism to obtain log features; Bi-GRU balances computational efficiency and sequence modeling, making it suitable for real-time log streams; The multimodal representation learning module is used to unify the dimensions of network traffic features, authentication features, and log features through the fifteenth fully connected layer (mapping features from different modalities to the same dimension for easy fusion), and then perform weighted fusion through a cross-modal attention network to obtain a first aggregated representation. The first aggregated representation is then compressed and encoded (preserving key information) through the sixteenth fully connected layer and the ReLU activation function to obtain a second aggregated representation. This module creatively utilizes cross-modal attention to enhance multi-source data collaboration, improves the detection of complex attacks (such as multi-stage intrusions), introduces non-linearity through ReLU to enhance representational capabilities, and optimizes the output dimension to reduce subsequent computational burden. The comprehensive risk assessment module is used to extract high-level features from the second aggregated representation by combining a fourth multilayer perceptron with Dropout, and predict risk behavior and threat level based on the high-level features by combining a dual-branch network of Softmax and Sigmoid; it learns shared features through multi-task learning to improve efficiency; and it creatively combines classification and regression to adapt to dynamic threat environments. The protection suggestion generation module is used to map the risk behaviors and threat levels to a predefined rule base through a rule engine simulation layer (based on lookup tables or lightweight MLP), generate preliminary suggestion keywords, and generate natural language protection suggestions (text descriptions, such as "suggest blocking IP address X") based on the preliminary suggestion keywords through a sequence-to-sequence model combined with an attention mechanism, and output intrusion detection results carrying risk behaviors, threat levels, and protection suggestions. The intrusion loss function of the intrusion detection model is: ; in, This represents the loss value of the intrusion loss function; The loss is represented by the cross-entropy loss function, which indicates the loss from predicting risky behavior. The loss for threat level prediction is represented by a binary cross-entropy loss function. The loss for generating protection recommendations is represented by the sequence cross-entropy loss function; All of these represent weighting coefficients.

[0079] The agent training and deployment module is specifically used for: The first dataset is divided into a first training set, a first validation set, and a first test set according to a preset first ratio based on time sequence, in order to train, validate, and test the sampling strategy generation model in the collaborative operation and maintenance agent; during the training process of the sampling strategy generation model, the sampling loss value of the sampling loss function is calculated, and the model parameters of the sampling strategy generation model are updated in reverse based on the sampling loss value using the gradient descent method to minimize the sampling loss function; after each training batch, the importance of neurons or connections in the sampling strategy generation model is evaluated by combining dynamic pruning technology, and redundant structures are pruned to perform model compression operation; The second dataset is divided into a second training set, a second validation set, and a second test set according to a preset second ratio based on time sequence, in order to train, validate, and test the encapsulation state detection model in the collaborative operation and maintenance agent. During the training process of the encapsulation state detection model, the encapsulation loss value of the encapsulation loss function is calculated, and the model parameters of the encapsulation state detection model are updated in reverse based on the encapsulation loss value using gradient descent to minimize the encapsulation loss function. After each training batch, the importance of neurons or connections in the encapsulation state detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The third dataset is divided into a third training set, a third validation set, and a third test set according to a preset third ratio based on time sequence, in order to train, validate, and test the fault detection model in the collaborative operation and maintenance agent. During the training process of the fault detection model, the fault loss value of the fault loss function is calculated, and the model parameters of the fault detection model are updated in reverse based on the fault loss value using the gradient descent method to minimize the fault loss function. After each training batch, the importance of neurons or connections in the fault detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The fourth dataset is divided into a fourth training set, a fourth validation set, and a fourth test set according to a preset fourth ratio based on time sequence, in order to train, validate, and test the intrusion detection model in the collaborative operation and maintenance agent. During the training process of the intrusion detection model, the intrusion loss value of the intrusion loss function is calculated, and the model parameters of the intrusion detection model are updated in reverse based on the intrusion loss value using the gradient descent method to minimize the intrusion loss function. After each training batch, the importance of neurons or connections in the intrusion detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The tested sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model were deployed to the industrial control computer of the thermal power plant using containerization technology. The actual multi-source heterogeneous data from thermal power plants are collected to train the sampling strategy generation model, encapsulation status detection model, fault detection model, and intrusion detection model deployed on the industrial control computer for model drift compensation training.

[0080] The model inference module specifically includes: The real-time multi-source heterogeneous data acquisition unit is used by the industrial control computer to sample and package the coal conveyed by the conveyor belt based on the input sampling control commands, and simultaneously collect real-time multi-source heterogeneous data through the Internet of Things sensing device array. The reasoning result output unit is used by the industrial control computer to input the preprocessed real-time multi-source heterogeneous data into the deployed collaborative operation and maintenance intelligent agent. The collaborative operation and maintenance intelligent agent performs reasoning through the sampling strategy generation model, encapsulation status detection model, fault detection model and intrusion detection model to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. The integrated operation and maintenance instruction execution unit is used by the industrial control computer to generate integrated operation and maintenance instructions based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, execute the integrated operation and maintenance instructions and record execution feedback. The data encryption upload unit is used by the industrial control computer to encrypt the real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions and execution feedback into an encrypted running data packet, calculate the data fingerprint of the encrypted running data packet and upload it to the blockchain, and upload the encrypted running data packet to the server; The power plant operation and maintenance knowledge base update unit is used by the server to verify and decrypt the received encrypted operation data packets to obtain real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions and execution feedback, and store them in the pre-created power plant operation and maintenance knowledge base. The power plant operation and maintenance knowledge base management unit is used to manage the power plant operation and maintenance knowledge base online after the mobile terminal authenticates with the server.

[0081] One of the most outstanding innovations of this invention lies in transforming security protection from an "external" component to an "internal" one. Security is no longer an additional module independent of business processes, but is deeply embedded in every link from data acquisition (hardware security modules), transmission (encryption, blockchain notarization) to access control (strong authentication on mobile devices). For example, the intrusion detection model itself, as one of the four core intelligent models, directly participates in the generation of comprehensive operation and maintenance instructions, which means that the detection of security threats and operation and maintenance responses are seamlessly linked. In addition, by using blockchain notarization to sample and encapsulate key data fingerprints throughout the entire process, underlying support is provided for the authenticity of data and the immutability of operations. This solution, which combines the reliability of intelligent analysis results with the trust mechanism of blockchain, provides a brand-new technical path to solve the long-standing "trust" problem in the thermal power plant industry.

[0082] The real-time multi-source heterogeneous data acquisition unit is specifically used for: The industrial control computer receives the input sampling control command and, based on the initial sampling timing, initial sampling point, and initial sampling frequency carried in the sampling control command, controls the coal sampler to sample the coal conveyed by the conveyor belt, loads the sampled coal into a sampling bucket, and controls the packaging machine to package the sampled bucket. Simultaneously, it collects real-time multi-source heterogeneous data through an array of IoT sensing devices, including a sampling camera, a packaging camera, a weight sensor, a vibration sensor, a temperature sensor, a humidity sensor, a laser scanner, an infrared spectrometer, and a gateway, and stores the collected real-time multi-source heterogeneous data in a hardware security module. The reasoning result output unit is specifically used for: The industrial control computer uses the eKuiper streaming computing engine to preprocess the real-time multi-source heterogeneous data within the hardware security module, and then inputs it into the deployed sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. The sampling strategy generation model, packaging state detection model, fault detection model, and intrusion detection model are based on hardware acceleration technology and perform high-concurrency real-time inference on the GPU, respectively outputting sampling strategy adjustment suggestions, packaging state detection results, fault detection results, and intrusion detection results. During the inference process, the industrial control computer monitors in real time the performance indicators of the local machine, including at least computing load, memory usage, GPU memory usage, and inference latency. When one of the performance indicators exceeds the corresponding threshold, it is determined that the sampling strategy generation model, packaging state detection model, fault detection model, and intrusion detection model cannot be supported for synchronous inference. Based on the preset inference priority, the execution frequency of low-priority models is dynamically reduced to ensure the real-time performance of high-priority models. The inference priority is as follows: fault detection model > intrusion detection model > encapsulation state detection model > sampling strategy generation model.

[0083] The integrated operation and maintenance instruction execution unit is specifically used for: The industrial control computer generates comprehensive operation and maintenance instructions in real time based on the sampling strategy adjustment suggestions, the packaging status detection results, the fault handling suggestions carried by the fault detection results, and the protection suggestions carried by the intrusion detection results. It controls the corresponding execution unit to execute the comprehensive operation and maintenance instructions, records the execution feedback of the execution unit, encrypts the comprehensive operation and maintenance instructions into encrypted operation and maintenance instructions, and backs up the encrypted operation and maintenance instructions to the instruction library of the server. The specific steps for encrypting the comprehensive operation and maintenance instructions into encrypted operation and maintenance instructions are as follows: The integrated operation and maintenance instructions are converted into a byte sequence using UTF-8 encoding; a temporary encryption key is derived using HMAC-SHA256 based on a preset shared key and the length of the byte sequence; the byte sequence is encrypted using the ChaCha20 stream encryption algorithm with the temporary encryption key to obtain intermediate ciphertext data of the same length as the byte sequence; the MD5 value of the last 64 bits of the shared key is calculated using the MD5 algorithm, the last two characters of the MD5 value are converted into hexadecimal data, and the first character of the hexadecimal data is used as the displacement; based on the displacement, each character of the intermediate ciphertext data is cyclically shifted to the right to obtain obfuscated data, and the obfuscated data is Base64 encoded to obtain the encrypted operation and maintenance instructions; Both the server and the industrial control computer pre-store the shared key.

[0084] In practice, the integrated operation and maintenance instructions have interactive and weight allocation logic. For example, when the intrusion detection result shows a high-risk network attack, the protection suggestions carried by the intrusion detection result are executed first, and even if the fault detection result shows a low-level fault, a shutdown inspection instruction is generated. When the packaging status detection result shows a serious anomaly, the fault sensitivity of the packaging mechanical components in the fault detection model is increased.

[0085] The data encryption and uploading unit is specifically used for: The industrial control computer serializes the real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback through MessagePack to obtain a binary data stream D_serial. The binary data stream D_serial is then compressed using the LZ4 compression algorithm to obtain compressed data D_compressed. Obtain the current timestamp, calculate the hash value of the timestamp using the SHA256 algorithm, and perform an XOR operation on the first 128 bits of the hash value with the compressed data D_compressed to obtain the scrambled data D_confused. A random initialization vector is generated. The random initialization vector and the preset shared key are called through the symmetric encryption algorithm AES-128-GCM to encrypt the scrambling data D_confused to obtain the ciphertext C_cipher and the authentication tag Tag_auth. The ciphertext C_cipher, the authentication tag Tag_auth, the random initialization vector, and the timestamp are packaged into an encrypted runtime data packet; The encrypted running data packet is calculated using the SHA-3 algorithm, the data fingerprint is stored on the blockchain, and the encrypted running data packet is uploaded to the server in real time via the TLS protocol.

[0086] The power plant operation and maintenance knowledge base update unit is specifically used for: The server receives the encrypted running data packet in real time. After verifying the integrity of the encrypted running data packet by the data fingerprint stored on the blockchain, it parses the encrypted running data packet to obtain the ciphertext C_cipher, the authentication tag Tag_auth, the random initialization vector, and the timestamp. The random initialization vector and the preset shared key are invoked using the symmetric encryption algorithm AES-128-GCM to decrypt the ciphertext C_cipher and verify the authentication tag Tag_auth to obtain the scrambling data D_confused. The hash value of the timestamp is calculated using the SHA256 algorithm. The first 128 bits of the hash value are XORed with the scrambled data D_confused to obtain compressed data D_compressed. The compressed data D_compressed is decompressed using the LZ4 compression algorithm to obtain a binary data stream D_serial. The binary data stream D_serial is deserialized using MessagePack to obtain real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback. The real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback are stored in a pre-created power plant operation and maintenance knowledge base. The server deploys baseline models for sampling strategy generation, encapsulation state detection, fault detection, and intrusion detection, corresponding to the sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. Based on a preset iteration cycle, incremental data is read from the power plant operation and maintenance knowledge base. An incremental dataset is constructed based on each incremental data set. The sampling strategy generation baseline model, encapsulation state detection baseline model, fault detection baseline model, and intrusion detection baseline model are trained based on the incremental dataset. The trained model change parameters are obtained, and the model change parameters are encrypted and sent to the industrial control computer to optimize the sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. When the model change parameters are issued, they are encrypted using the industrial control computer's asymmetric encryption public key, and the transaction hash of the parameter distribution is recorded through the blockchain to ensure the integrity and traceability of the model update process and prevent malicious model implantation.

[0087] The power plant operation and maintenance knowledge base management unit is specifically used for: The mobile terminal sends an access request to the server carrying a composite identity credential, which consists of a digital certificate and a dynamic password. After receiving the access request, the server verifies the composite identity credential carried in the access request and negotiates a temporary session key with the mobile terminal for this session based on the TLS protocol. The mobile terminal obtains the input operation content for the power plant operation and maintenance knowledge base, the terminal identifier of the mobile terminal, and the current operation time. It calculates the message authentication code of the operation content, terminal identifier, and operation time through the temporary session key, generates an operation request based on the operation content, terminal identifier, operation time, and message authentication code, and sends the operation request to the server. After verifying the message authentication code carried in the operation request using the temporary session key, the server performs timeliness verification using the operation time and legality verification using the terminal identifier before executing the operation content to manage the power plant operation and maintenance knowledge base online and record operation logs.

[0088] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A sampling and encapsulation operation and maintenance method for coal conveying in thermal power plants that combines deep learning, characterized in that: Includes the following steps: Step S10: Obtain a large amount of historical multi-source heterogeneous data from the thermal power plant, preprocess the historical multi-source heterogeneous data, and construct the first dataset, the second dataset, the third dataset, and the fourth dataset. Step S20: Create a collaborative operation and maintenance intelligent agent that includes a sampling strategy generation model, an encapsulation state detection model, a fault detection model, and an intrusion detection model; The sampling strategy generation model is built based on a multimodal feature extraction module, a multimodal feature fusion module, and a strategy decision module. It is used to output sampling strategy adjustment suggestions based on input visual image data, physical attribute data, and infrared spectral data. The encapsulation state detection model is built based on an image feature extraction module, a weight feature extraction module, a multimodal feature aggregation module, and a multi-task output module. It is used to output encapsulation state detection results based on input visual image data and physical attribute data. The fault detection model is built based on a vibration feature extraction module, a temperature feature extraction module, a humidity feature extraction module, a visual feature extraction module, a contour feature extraction module, a heterogeneous feature fusion module, and a fault prediction module. It is used to output fault detection results based on input visual image data, physical attribute data, and equipment status data. The intrusion detection model is built upon a network traffic analysis module, an identity authentication analysis module, a log analysis module, a multimodal representation learning module, a comprehensive risk assessment module, and a protection suggestion generation module. It is used to output intrusion detection results based on input network communication data. Step S30: Train the collaborative operation and maintenance agent using the first dataset, the second dataset, the third dataset, and the fourth dataset; Step S40: After preprocessing the collected real-time multi-source heterogeneous data, input it into the collaborative operation and maintenance intelligent agent to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, generate comprehensive operation and maintenance instructions.

2. The method for sampling, encapsulating, and maintaining coal conveying systems in thermal power plants using deep learning as described in claim 1, characterized in that: Step S10 specifically involves: Acquire a large amount of historical multi-source heterogeneous data from thermal power plants, including visual image data, physical attribute data, infrared spectral data, equipment status data, and network communication data; The visual image data includes coal flow images and packaging status images; the physical property data includes sample weight data, coal volume data, and coal outline data; the equipment status data includes mechanical vibration data, temperature data, and humidity data; and the network communication data includes network traffic data, equipment authentication information, and abnormal access logs. The visual image data is preprocessed including data cleaning and denoising, data augmentation, size standardization and normalization; the physical attribute data and device status data are preprocessed including outlier handling, missing value handling, time series alignment, smoothing and filtering, and feature engineering; the infrared spectral data are preprocessed including baseline correction, standard normal transformation, smoothing and denoising, and normalization. Based on the preprocessed coal flow images, coal volume data, coal contour data, and infrared spectral data, a first dataset is constructed for training the sampling strategy generation model. Based on the preprocessed images of each of the encapsulation states and the sample weight data, a second dataset is constructed for training the encapsulation state detection model. Based on the preprocessed mechanical vibration data, temperature data, humidity data, coal flow images, and coal contour data, a third dataset is constructed for training the fault detection model. Based on the preprocessed network traffic data, device authentication information, and abnormal access logs, a fourth dataset is constructed for training the intrusion detection model. The first dataset is annotated with at least the optimal sampling time, optimal sampling location, and sampling action identifier; The second dataset is labeled with at least the following tags: packaging integrity status, packaging identification clarity, and packaging form abnormality. The third dataset is labeled with at least the following tags: failure time, failure type, failure severity level, and failure handling suggestions. The fourth dataset is labeled with at least the following tags: risk behavior, threat level, and protection suggestions.

3. The method for sampling, encapsulating, and maintaining coal conveying systems in thermal power plants using deep learning as described in claim 1, characterized in that: In step S20, the multimodal feature extraction module is constructed based on an image feature extraction unit, a numerical feature extraction unit, and a spectral feature extraction unit. The image feature extraction unit is used to extract coal image features from coal flow images using a lightweight first MobileNetV2 network. The numerical feature extraction unit is used to extract numerical features from coal volume data and coal contour data using a first multilayer perceptron. The spectral feature extraction unit is used to extract spectral features from infrared spectral data using a first one-dimensional convolutional neural network. The multi-modal feature fusion module is used to fuse coal image features, numerical features, and spectral features through a multi-head attention mechanism to obtain fused features; The strategy decision-making module is constructed based on a first shared fully connected layer, a timing output unit, a position output unit, a frequency output unit, and a suggestion output unit; The first shared fully connected layer is used to reduce the dimensionality of the fused features to obtain the first shared features; the timing output unit is used to infer the first shared features through the first fully connected layer to obtain the sampling timing; the location output unit is used to infer the first shared features through the second fully connected layer to obtain the sampling location; the frequency output unit is used to infer the first shared features through the third fully connected layer to obtain the sampling frequency; and the suggestion output unit is used to output a sampling strategy adjustment suggestion carrying the sampling timing, sampling location, and sampling frequency. The sampling loss function of the model generated by the sampling strategy is: ; in, This represents the loss value of the sampling loss function; Indicates the actual sampling timing; Indicates the actual sampling location; Indicates the actual sampling frequency; Indicates the predicted sampling timing; Indicates the predicted sampling location; Indicates the predicted sampling frequency; , , All represent weighting coefficients; The image feature extraction module is used to extract encapsulated image features from the encapsulated state image through a lightweight second MobileNetV2 network; The weight feature extraction module is used to extract weight features from sample weight data using a second multilayer sensor. The multimodal feature aggregation module is used to aggregate encapsulated image features and weight features through a lightweight cross-attention fusion network to obtain aggregated features; The multi-task output module is constructed based on a second shared fully connected layer, an integrity output unit, a clarity output unit, a morphology output unit, and a state detection result output unit. The second shared fully connected layer is used to reduce the dimensionality of the aggregated features to obtain the second shared features; the integrity output unit is used to infer the second shared features through the fourth fully connected layer to obtain the encapsulation integrity status label; the clarity output unit is used to infer the second shared features through the fifth fully connected layer to obtain the encapsulation identification clarity label; the morphology output unit is used to infer the second shared features through the sixth fully connected layer to obtain the packaging morphology anomaly label; the state detection result output unit is used to output the packaging state detection result carrying the encapsulation integrity status label, the encapsulation identification clarity label, and the packaging morphology anomaly label; The packaging loss function of the packaging state detection model is: ； ； ； ； in, This represents the loss value of the encapsulated loss function; Represents the loss of the encapsulation integrity state sub-value; This indicates a loss of clarity in the packaging designation; This indicates a loss due to abnormal packaging shape; , , All represent weighting coefficients; A label indicating the actual encapsulation integrity status; Indicates the clarity of the actual packaging label; Labels indicating abnormal packaging conditions; A label indicating the predicted encapsulation integrity status; Indicates the predicted clarity of the encapsulation label; This indicates a predicted abnormal packaging shape label.

4. The method for coal conveying sampling and encapsulation operation and maintenance of thermal power plants combining deep learning as described in claim 1, characterized in that: In step S20, the vibration feature extraction module is used to extract long-term dependent features from mechanical vibration data through a long short-term memory network, enhance the attention to key time points in the long-term dependent features through the self-attention mechanism in the first Transformer encoder, and output the vibration feature vector after feature compression through the seventh fully connected layer. The temperature feature extraction module is used to extract temperature feature vectors from temperature data through a gated loop unit and an eighth fully connected layer. The humidity feature extraction module is used to extract local features from humidity data through a second one-dimensional convolutional neural network, and then perform max pooling dimensionality reduction before outputting a humidity feature vector through the ninth fully connected layer. The visual feature extraction module is used to extract multi-level features from the coal flow image through a convolutional neural network, focus the key regions in the multi-level features through the convolutional block attention module, and after global average pooling, output the visual feature vector through the tenth fully connected layer. The contour feature extraction module is used to extract local geometric features from coal contour data using PointNet. After global max pooling of the local geometric features, the contour feature vector is output through the eleventh fully connected layer. The heterogeneous feature fusion module is used to concatenate vibration feature vector, temperature feature vector, humidity feature vector, visual feature vector and contour feature vector into a long vector. After feature interaction and weighting by the second Transformer encoder, it is then dimensionality reduction and nonlinear fusion by the third multilayer perceptron to output the fused embedding. The fault prediction module is constructed based on a time prediction unit, a type classification unit, a level classification unit, a suggestion generation unit, and a fault prediction output unit. The time prediction unit is used to perform regression output on the fused embedding through the twelfth fully connected layer to obtain the fault prediction time; The type classification unit is used to infer the fusion embedding through the thirteenth fully connected layer and the first Softmax layer to obtain the fault prediction type; The classification unit is used to infer the fusion embedding through the fourteenth fully connected layer and the second Softmax layer to obtain the fault prediction severity level. The suggestion generation unit is used to generate fault handling suggestions based on the fusion embedding through a rule-based suggestion generator. The fault prediction output unit is used to output fault detection results carrying fault prediction time, fault prediction type, fault prediction severity level, and fault handling suggestions. The fault loss function of the fault detection model is: ; in, This represents the loss value of the fault loss function; The regression loss for failure time prediction is represented by the mean squared error function; The classification loss, representing the fault type classification, is expressed using the cross-entropy loss function; The classification loss, representing the severity level of the fault, is expressed using the cross-entropy loss function. , , All represent weighting coefficients; The network traffic analysis module is used to extract temporal features from network traffic data through a bidirectional long short-term memory network, and to enhance the spatial relationship perception of the temporal features through a graph attention network to obtain network traffic features. The identity authentication analysis module is used to extract identity authentication features from the device identity authentication information through the third Transformer encoder; The log analysis module is used to extract context features from abnormal access logs through a bidirectional gated loop unit and to highlight abnormal log events through a self-attention mechanism to obtain log features. The multimodal representation learning module is used to unify the dimensions of network traffic features, identity authentication features and log features through the fifteenth fully connected layer, and then perform weighted fusion through a cross-modal attention network to obtain the first aggregated representation. The first aggregated representation is then compressed and encoded through the sixteenth fully connected layer and the ReLU activation function to obtain the second aggregated representation. The comprehensive risk assessment module is used to extract high-level features from the second aggregated representation by combining a fourth multilayer perceptron with Dropout, and predict risk behavior and threat level based on the high-level features by combining a dual-branch network of Softmax and Sigmoid. The protection suggestion generation module is used to map the risk behaviors and threat levels to a predefined rule base through a rule engine simulation layer, generate preliminary suggestion keywords, generate natural language protection suggestions based on the preliminary suggestion keywords through a sequence-to-sequence model combined with an attention mechanism, and output intrusion detection results carrying risk behaviors, threat levels, and protection suggestions. The intrusion loss function of the intrusion detection model is: ; in, This represents the loss value of the intrusion loss function; The loss is represented by the cross-entropy loss function, which indicates the risk behavior prediction loss. The loss for threat level prediction is represented by a binary cross-entropy loss function. The loss for generating protection recommendations is represented by the sequence cross-entropy loss function; , , All of these represent weighting coefficients.

5. The method for sampling, encapsulating, and maintaining coal conveying systems in thermal power plants using deep learning as described in claim 1, characterized in that: Step S30 specifically involves: The first dataset is divided into a first training set, a first validation set, and a first test set according to a preset first ratio based on time sequence, in order to train, validate, and test the sampling strategy generation model in the collaborative operation and maintenance agent; during the training process of the sampling strategy generation model, the sampling loss value of the sampling loss function is calculated, and the model parameters of the sampling strategy generation model are updated in reverse based on the sampling loss value using the gradient descent method to minimize the sampling loss function; after each training batch, the importance of neurons or connections in the sampling strategy generation model is evaluated by combining dynamic pruning technology, and redundant structures are pruned to perform model compression operation; The second dataset is divided into a second training set, a second validation set, and a second test set according to a preset second ratio based on time order, so as to train, validate, and test the encapsulated state detection model in the collaborative operation and maintenance intelligent agent. During the training of the encapsulation state detection model, the encapsulation loss value of the encapsulation loss function is calculated, and the model parameters of the encapsulation state detection model are updated in reverse based on the encapsulation loss value using the gradient descent method to minimize the encapsulation loss function. After each training batch, the importance of neurons or connections in the encapsulation state detection model is evaluated by combining dynamic pruning technology, and redundant structures are pruned to perform model compression operations. The third dataset is divided into a third training set, a third validation set, and a third test set according to a preset third ratio based on time order, so as to train, validate, and test the fault detection model in the collaborative operation and maintenance intelligent agent. During the training process of the fault detection model, the fault loss value of the fault loss function is calculated, and the model parameters of the fault detection model are updated in reverse based on the fault loss value using the gradient descent method to minimize the fault loss function. After each training batch, the importance of neurons or connections in the fault detection model is evaluated by combining dynamic pruning technology, and redundant structures are pruned to perform model compression operation. The fourth dataset is divided into a fourth training set, a fourth validation set, and a fourth test set according to a preset fourth ratio based on time sequence, in order to train, validate, and test the intrusion detection model in the collaborative operation and maintenance agent. During the training process of the intrusion detection model, the intrusion loss value of the intrusion loss function is calculated, and the model parameters of the intrusion detection model are updated in reverse based on the intrusion loss value using the gradient descent method to minimize the intrusion loss function. After each training batch, the importance of neurons or connections in the intrusion detection model is evaluated using dynamic pruning techniques, and redundant structures are pruned to perform model compression operations. The tested sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model were deployed to the industrial control computer of the thermal power plant using containerization technology. The actual multi-source heterogeneous data from thermal power plants are collected to train the sampling strategy generation model, encapsulation status detection model, fault detection model, and intrusion detection model deployed on the industrial control computer for model drift compensation training.

6. The method for coal conveying sampling and encapsulation operation and maintenance of thermal power plants combining deep learning as described in claim 1, characterized in that: Step S40 specifically includes: Step S41: Based on the input sampling control command, the industrial control computer samples and packages the coal conveyed by the conveyor belt, and simultaneously collects real-time multi-source heterogeneous data through the Internet of Things sensing device array. Step S42: The industrial control computer preprocesses the real-time multi-source heterogeneous data and inputs it into the deployed collaborative operation and maintenance intelligent agent. The collaborative operation and maintenance intelligent agent performs inference through the sampling strategy generation model, encapsulation status detection model, fault detection model and intrusion detection model to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Step S43: The industrial control computer generates a comprehensive operation and maintenance instruction based on the sampling strategy adjustment suggestion, the packaging status detection result, the fault detection result, and the intrusion detection result, executes the comprehensive operation and maintenance instruction, and records the execution feedback; Step S44: The industrial control computer encrypts the real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback into an encrypted running data packet, calculates the data fingerprint of the encrypted running data packet, uploads it to the blockchain, and uploads the encrypted running data packet to the server. Step S45: The server verifies and decrypts the received encrypted operation data packet to obtain real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback, and stores them in a pre-created power plant operation and maintenance knowledge base; Step S46: After the mobile terminal authenticates with the server, it performs online management of the power plant operation and maintenance knowledge base.

7. The method for coal conveying sampling and encapsulation operation and maintenance in thermal power plants combining deep learning as described in claim 6, characterized in that: Step S41 specifically involves: The industrial control computer receives the input sampling control command and, based on the initial sampling timing, initial sampling point, and initial sampling frequency carried in the sampling control command, controls the coal sampler to sample the coal conveyed by the conveyor belt, loads the sampled coal into a sampling bucket, and controls the packaging machine to package the sampled bucket. Simultaneously, it collects real-time multi-source heterogeneous data through an array of IoT sensing devices, including a sampling camera, a packaging camera, a weight sensor, a vibration sensor, a temperature sensor, a humidity sensor, a laser scanner, an infrared spectrometer, and a gateway, and stores the collected real-time multi-source heterogeneous data in a hardware security module. Step S42 specifically involves: The industrial control computer uses the eKuiper streaming computing engine to preprocess the real-time multi-source heterogeneous data within the hardware security module, and then inputs it into the deployed sampling strategy generation model, encapsulation state detection model, fault detection model, and intrusion detection model. The sampling strategy generation model, packaging state detection model, fault detection model, and intrusion detection model are based on hardware acceleration technology and perform high-concurrency real-time inference on the GPU, respectively outputting sampling strategy adjustment suggestions, packaging state detection results, fault detection results, and intrusion detection results. During the inference process, the industrial control computer monitors in real time the performance indicators of the local machine, including at least computing load, memory usage, GPU memory usage, and inference latency. When one of the performance indicators exceeds the corresponding threshold, it is determined that the sampling strategy generation model, packaging state detection model, fault detection model, and intrusion detection model cannot be supported for synchronous inference. Based on the preset inference priority, the execution frequency of low-priority models is dynamically reduced to ensure the real-time performance of high-priority models. The inference priority is as follows: fault detection model > intrusion detection model > encapsulation state detection model > sampling strategy generation model.

8. The method for coal conveying sampling and encapsulation operation and maintenance in thermal power plants combining deep learning as described in claim 6, characterized in that: Step S43 specifically involves: The industrial control computer generates comprehensive operation and maintenance instructions in real time based on the sampling strategy adjustment suggestions, the packaging status detection results, the fault handling suggestions carried by the fault detection results, and the protection suggestions carried by the intrusion detection results. It controls the corresponding execution unit to execute the comprehensive operation and maintenance instructions, records the execution feedback of the execution unit, encrypts the comprehensive operation and maintenance instructions into encrypted operation and maintenance instructions, and backs up the encrypted operation and maintenance instructions to the instruction library of the server. Step S44 specifically involves: The industrial control computer serializes the real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback through MessagePack to obtain a binary data stream D_serial. The binary data stream D_serial is then compressed using the LZ4 compression algorithm to obtain compressed data D_compressed. Obtain the current timestamp, calculate the hash value of the timestamp using the SHA256 algorithm, and perform an XOR operation on the first 128 bits of the hash value with the compressed data D_compressed to obtain the scrambled data D_confused. A random initialization vector is generated. The random initialization vector and the preset shared key are called through the symmetric encryption algorithm AES-128-GCM to encrypt the scrambling data D_confused to obtain the ciphertext C_cipher and the authentication tag Tag_auth. The ciphertext C_cipher, the authentication tag Tag_auth, the random initialization vector, and the timestamp are packaged into an encrypted runtime data packet; The encrypted running data packet is calculated using the SHA-3 algorithm, the data fingerprint is stored on the blockchain, and the encrypted running data packet is uploaded to the server in real time via the TLS protocol.

9. The method for sampling, encapsulating, and maintaining coal conveying systems in thermal power plants using deep learning as described in claim 6, characterized in that: Step S45 specifically involves: The server receives the encrypted running data packet in real time. After verifying the integrity of the encrypted running data packet by the data fingerprint stored on the blockchain, it parses the encrypted running data packet to obtain the ciphertext C_cipher, the authentication tag Tag_auth, the random initialization vector, and the timestamp. The random initialization vector and the preset shared key are invoked using the symmetric encryption algorithm AES-128-GCM to decrypt the ciphertext C_cipher and verify the authentication tag Tag_auth to obtain the scrambling data D_confused. The hash value of the timestamp is calculated using the SHA256 algorithm. The first 128 bits of the hash value are XORed with the scrambled data D_confused to obtain compressed data D_compressed. The compressed data D_compressed is decompressed using the LZ4 compression algorithm to obtain a binary data stream D_serial. The binary data stream D_serial is deserialized using MessagePack to obtain real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback. The real-time multi-source heterogeneous data, sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results, intrusion detection results, comprehensive operation and maintenance instructions, and execution feedback are stored in a pre-created power plant operation and maintenance knowledge base. The server deploys the sampling strategy generation baseline model, the packaging state detection baseline model, the fault detection baseline model, and the intrusion detection baseline model corresponding to the sampling strategy generation model, the packaging state detection baseline model, the fault detection baseline model, and the intrusion detection baseline model. Based on a preset iteration cycle, incremental data is read from the power plant operation and maintenance knowledge base. An incremental dataset is constructed based on each incremental data. The sampling strategy generation baseline model, the encapsulation status detection baseline model, the fault detection baseline model, and the intrusion detection baseline model are trained based on the incremental dataset. The model change parameters after training are obtained. The model change parameters are encrypted and sent to the industrial control computer to optimize the sampling strategy generation model, the encapsulation status detection model, the fault detection model, and the intrusion detection model. Step S46 specifically involves: The mobile terminal sends an access request to the server carrying a composite identity credential, which consists of a digital certificate and a dynamic password. After receiving the access request, the server verifies the composite identity credentials carried in the access request and negotiates a temporary session key with the mobile terminal for this session based on the TLS protocol. The mobile terminal obtains the input operation content for the power plant operation and maintenance knowledge base, the terminal identifier of the mobile terminal, and the current operation time. It calculates the message authentication code of the operation content, terminal identifier, and operation time through the temporary session key, generates an operation request based on the operation content, terminal identifier, operation time, and message authentication code, and sends the operation request to the server. After verifying the message authentication code carried in the operation request using the temporary session key, the server performs timeliness verification using the operation time and legality verification using the terminal identifier before executing the operation content to manage the power plant operation and maintenance knowledge base online and record operation logs.

10. A coal conveying sampling and packaging operation and maintenance system for thermal power plants that combines deep learning, characterized in that: Includes the following modules: The dataset construction module is used to acquire a large amount of historical multi-source heterogeneous data from thermal power plants, preprocess the aforementioned historical multi-source heterogeneous data, and construct the first dataset, the second dataset, the third dataset, and the fourth dataset. The agent creation module is used to create a collaborative operation and maintenance agent that includes a sampling strategy generation model, an encapsulated state detection model, a fault detection model, and an intrusion detection model. The sampling strategy generation model is built based on a multimodal feature extraction module, a multimodal feature fusion module, and a strategy decision module. It is used to output sampling strategy adjustment suggestions based on input visual image data, physical attribute data, and infrared spectral data. The encapsulation state detection model is built based on an image feature extraction module, a weight feature extraction module, a multimodal feature aggregation module, and a multi-task output module. It is used to output encapsulation state detection results based on input visual image data and physical attribute data. The fault detection model is built based on a vibration feature extraction module, a temperature feature extraction module, a humidity feature extraction module, a visual feature extraction module, a contour feature extraction module, a heterogeneous feature fusion module, and a fault prediction module. It is used to output fault detection results based on input visual image data, physical attribute data, and equipment status data. The intrusion detection model is built upon a network traffic analysis module, an identity authentication analysis module, a log analysis module, a multimodal representation learning module, a comprehensive risk assessment module, and a protection suggestion generation module. It is used to output intrusion detection results based on input network communication data. The agent training and deployment module is used to train the collaborative operation and maintenance agent using the first dataset, the second dataset, the third dataset, and the fourth dataset. The model inference module is used to preprocess the collected real-time multi-source heterogeneous data and input it into the collaborative operation and maintenance intelligent agent to obtain sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results. Based on the sampling strategy adjustment suggestions, encapsulation status detection results, fault detection results and intrusion detection results, a comprehensive operation and maintenance instruction is generated.

Images