Real-time alarm and statistical analysis system for power plant safety inspection

Abstract

This invention discloses a real-time alarm and statistical analysis system for power plant safety inspections, relating to the fields of smart power plants and industrial artificial intelligence. It solves the technical problems of traditional inspection systems, such as passive response, difficulty in knowledge accumulation, and inability to adapt to complex risk scenarios. The system includes: a multimodal data acquisition module for real-time acquisition of time-series, video, and text data from power plant equipment; a large-scale power plant model module that uses a Transformer-based multimodal fusion model to comprehensively analyze data, identify cross-equipment risks, and predict evolution trends; an intelligent agent task execution module that dynamically synthesizes composite inspection tasks based on risk assessments and optimizes the allocation of inspection resources; and a knowledge base and model iteration module that stores contextualized cases and continuously optimizes the large-scale power plant model through comparative learning. This invention significantly improves the accuracy of risk warnings, the timeliness of response, and the overall adaptive capability of the system.

Description

Technical Field

[0001] This invention belongs to the field of smart power plant and industrial artificial intelligence technology, specifically involving a real-time alarm and statistical analysis system for power plant safety inspection. Background Technology

[0002] With the advancement of the construction of new power systems, thermal power generating units are facing increasingly frequent demands for deep peak shaving and rapid changes in operating conditions, making equipment operation increasingly complex and increasing the pressure on safety management.

[0003] Traditional power plant safety inspections mainly rely on regular manual inspections and fixed video surveillance, which have significant shortcomings: manual inspections have blind spots and intervals; fixed monitoring systems rely heavily on preset rules, making it difficult to perceive the dynamic risks of equipment operation in real time and comprehensively, especially the risks related to cross-equipment and cross-systems; existing intelligent analysis technologies based on small models have limited ability to identify complex and nonlinear risk patterns, resulting in high false alarm and false negative rates; and inspection tasks are usually based on fixed plans or post-event responses, making it impossible to dynamically adjust according to the real-time risk situation. When faced with complex risk scenarios, the system lacks autonomous analysis and planning capabilities, making it difficult to quickly generate comprehensive inspection plans covering multiple devices and steps, resulting in low response efficiency. The experience and knowledge of inspection personnel and experts are mostly in unstructured form, unable to be effectively digitized and accumulated. The system lacks the ability to learn autonomously and extract effective patterns from massive amounts of inspection data and handling cases, making it difficult to transform the "data mine" into "knowledge assets" and hindering the continuous improvement of the system's intelligence level. Furthermore, when facing sudden or complex inspection tasks, the scheduling of different types of inspection resources (such as robots, drones, and fixed cameras) relies heavily on human experience, lacking scientific optimization models, making it difficult to achieve optimal matching of resources and tasks, thus affecting overall inspection efficiency and cost.

[0004] Therefore, there is an urgent need for a real-time alarm and statistical analysis system for power plant safety inspections. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a real-time alarm and statistical analysis system for power plant safety inspections, used to solve the following technical problems: Existing technologies struggle to accurately and proactively perceive the complex and dynamic operational risks of power plants, and they also struggle to transform abstract risk assessments into efficient and actionable inspection actions. The system lacks the ability to continuously learn and self-optimize, and it also lacks the technical capacity to integrate dispersed multi-source information.

[0006] To address the above problems, this invention provides a real-time alarm and statistical analysis system for power plant safety inspections, comprising: The multimodal data acquisition module is used to collect power plant operation data in real time, including equipment time-series operation data, monitoring video data, and text data; The power plant large model module is used to construct a power plant large model, perform fusion analysis on the data collected by the multimodal data acquisition module, identify potential risk points and their associated equipment, predict the risk evolution trend, and output safety risk assessment results. The intelligent agent task execution module is communicatively connected to the power plant large model module and includes a task synthesis unit and a resource dynamic adaptation unit. The task synthesis unit is used to dynamically combine basic inspection items to generate a new composite inspection task based on the cross-equipment related risks or complex risk scenarios identified in the safety risk assessment results. The resource dynamic adaptation unit is used to calculate and allocate the optimal combination of inspection resources for the composite inspection task in real time. The knowledge base and model iteration module is used to store and manage data samples and algorithm models generated during the inspection process, record the composite inspection tasks and their execution results, store them in the knowledge base as contextualized cases, and train the power plant large model based on the cases.

[0007] Optionally, the multimodal data acquisition module includes: The timing operation data of the equipment is obtained in real time from the power plant's distributed control system or plant-level monitoring information system through a standard industrial interface or a specific interface protocol. The text data includes structured work information obtained from the power plant work order and operation ticket system, and unstructured text reports obtained from equipment maintenance records.

[0008] Optionally, the construction and training process of the large-scale power plant model includes: The large-scale power plant model adopts a Transformer architecture multimodal encoder-decoder model. The encoder part includes: a time-series data encoding sub-network: using a recurrent neural network with an attention mechanism to process the equipment's time-series operating data and extract time-series features. Visual data encoding sub-network: This network combines convolutional neural networks and visual Transformers to process surveillance video data and extract spatial and spatiotemporal features. Text data encoding sub-network: Employs the encoding layer of a pre-trained language model to process text data and extract semantic features. ; A cross-membrane attention fusion layer is configured to receive the features. , and The association weights between features of different modalities are calculated through a multi-head attention mechanism, and the fused unified features are output. ; Labeled multimodal data extracted from the power plant's historical operation and maintenance database was used as the training set. The labels included risk type, associated equipment identification, and risk evolution stage. Supervised training was conducted by minimizing the weighted sum of the cross-entropy loss function for risk prediction and the loss function for equipment association prediction.

[0009] Optionally, the task synthesis unit includes: The process by which the task synthesis unit generates composite inspection tasks is as follows: There are N basic inspection items, each inspection item Each basic inspection action corresponds to a required resource type. When a complex risk scenario involving M related devices is identified, the task synthesis unit retrieves a set of all necessary basic inspection items covering the M related devices based on a predefined task map. ; Define task composition function ,in, The context feature vector of the current risk scenario is output by the power plant large model. The derived function calculates the execution priority score of each basic inspection item in the current context using a neural network. The composite inspection task, which includes a unique task ID, a sequence of task steps, and the expected target, is generated according to priority and logical relationship.

[0010] Optionally, the resource dynamic adaptation unit calculates and allocates the optimal combination of inspection resources, including: Let the available inspection resources be set. Each resource It has type, current location, set of functional capabilities, and status attributes; For the first in the composite inspection task Task steps Its required capability vector is ; The unit allocates resources by solving an optimization problem. The objective function is to minimize the total expected completion time and the total resource scheduling cost, specifically:

[0011] in, Let be the objective function. For the first in the composite inspection task Each task step The total number of steps in the task. To assign to step Inspection resources, Let be the allocation function, representing the first... The resource index allocated to each task step. The expected completion time function, The set of inspection resources that are scheduled and activated in the task. To enable resources Cost weighting coefficient and These are the normalized balance coefficients; The constraints of the optimization problem include: each step must be allocated a resource that can meet its capacity requirements, and the same resource does not conflict in time.

[0012] Optionally, training the large-scale power plant model based on the case study includes the following steps: Contextualized cases are extracted from the knowledge base, each case including: risk scenario description features. The composite inspection tasks performed Verification data collected during task execution And the final risk management outcome label Y; based on the outcome label Y, positive examples of successfully managed risks will be identified. Negative cases that were unsuccessful or irrelevant Composition of training sample pairs ; Through trainable encoding networks Extract verification data The key state features are denoted as verification state features. ; Calculate the risk scenario characteristics of positive samples and their own verification state characteristics. Cosine similarity as the first similarity ; Calculate the risk scenario characteristics of positive samples and the first Validation state feature vector of each negative sample Cosine similarity as the second similarity A temperature coefficient hyperparameter is introduced to construct a contrast loss function.

[0013] Optionally, constructing the contrastive loss function includes: The construction of the contrastive loss function is specifically as follows:

[0014] in, To compare loss functions, The first similarity score, For the second similarity, This is the temperature coefficient hyperparameter. This represents the total number of negative samples paired with positive samples in the current training sample group. Indexed by quantity.

[0015] Optionally, the following modules may also be included: Real-time alarm and situation display module: used to receive the safety risk assessment results output by the power plant large model module, and generate an alarm event when the risk level exceeds the preset risk threshold; The module integrates a visualization engine to overlay and display the potential risk points, related equipment, predicted risk evolution paths, and the real-time location and status of the inspection resources dispatched by the intelligent agent task execution module on the power plant's digital twin map.

[0016] The beneficial effects of this invention are: This invention constructs a large power plant model and performs in-depth fusion and correlation analysis of multimodal data. The system can identify hidden risks and cross-equipment chain risks that are difficult to detect by traditional methods, and predict risk trends, thus transforming the safety management mode from "post-event remediation" to "pre-event warning and in-event intervention". This invention synthesizes composite inspection tasks in a "tailor-made" manner based on real-time risks and uses mathematical optimization models to achieve efficient and economical allocation of inspection resources, which greatly improves the emergency response speed and overall inspection efficiency in complex scenarios and reduces operating costs. This invention, through the accumulation of contextualized cases in a knowledge base and a model iteration mechanism based on comparative learning, enables the system to continuously learn from practical experience. As runtime increases, its risk identification accuracy, task synthesis rationality, and resource scheduling efficiency will continuously optimize itself, achieving a sustained increase in the system's intelligence level. This invention transforms abstract data and analysis results into intuitive visualizations, helping managers to "control the safety status and resource dynamics of the entire plant with one screen," significantly improving the accuracy of decision-making and the efficiency of command and dispatch. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the module process of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a real-time alarm and statistical analysis system for power plant safety inspections, comprising the following modules: Multimodal data acquisition module: used to collect power plant operation data in real time, including equipment time-series operation data, monitoring video data and text data; Large-scale power plant model module: Used to construct a large-scale power plant model, perform fusion analysis on the data collected by the multimodal data acquisition module, identify potential risk points and their associated equipment, predict risk evolution trends, and output safety risk assessment results; Intelligent agent task execution module: Communicatively connected to the power plant large model module, including a task synthesis unit and a resource dynamic adaptation unit; the task synthesis unit is used to dynamically combine basic inspection items to generate a new composite inspection task based on the cross-equipment related risks or complex risk scenarios identified in the safety risk assessment results; the resource dynamic adaptation unit is used to calculate and allocate the optimal combination of inspection resources for the composite inspection task in real time; Knowledge base and model iteration module: used to store and manage data samples and algorithm models generated during the inspection process, record the composite inspection tasks and their execution results, store them in the knowledge base as contextualized cases, and train the power plant large model based on the cases.

[0020] Specifically, the multimodal data acquisition module, acting as the system's perception layer, simultaneously accesses and collects the following three types of core data in real time through standard industrial interfaces and network protocols: Equipment time-series operation data: Real-time acquisition of operating parameters of key equipment (e.g., turbines, boilers, feedwater pumps) from the power plant's distributed control system (DCS) and plant-level monitoring information system (SIS), including high-frequency time-series data with timestamps such as pressure, temperature, flow rate, vibration, and valve opening; Monitoring video data: Real-time acquisition of video stream data from high-definition network cameras deployed in key areas of the power plant (boiler room, turbine room, booster station, coal conveying corridor); Text data: Through integration with the power plant's existing production management system. The system automatically acquires structured and unstructured text data, primarily including work tasks, personnel, and location information extracted from the work order / operation ticket ("two-ticket") system, as well as fault descriptions and handling reports extracted from equipment maintenance records and operation logs. The collected multimodal data is synchronously transmitted to the power plant large-scale model module for processing. The core of this module is a dedicated large-scale model pre-trained with large-scale power industry data and fine-tuned for a specific power plant. The implementation process is as follows: independent sub-networks within the model extract features from time-series data, video frames, and text, obtaining a high-dimensional time-series feature vector, visual feature vector, and semantic feature vector. Subsequently, a cross-modal injection... The inference fusion layer calculates the correlation weights between different modal features and merges them into a unified joint feature representation rich in contextual information. Based on the fused features, the power plant large model performs two core inference tasks: first, risk identification, determining and classifying the existence of current safety hazards (equipment overheating, valve leakage, personnel violations, etc.); second, correlation analysis and trend prediction, identifying related equipment affected by the risk and predicting the possible evolution path (deterioration, stabilization, or mitigation) of the risk in the future based on historical data patterns. Finally, the module outputs a structured safety risk assessment report, which includes at least: risk point location (equipment / area), risk type and level, and a list of related equipment. The system also includes a predicted risk evolution trend curve. The intelligent agent task execution module receives a risk assessment report from the power plant large model and initiates an autonomous decision-making process. When the assessment report indicates the existence of "cross-equipment related risks" or "complex risk scenarios," the task synthesis unit is activated. Based on a predefined power plant safety task map, it intelligently combines and sorts several basic inspection items required to cover all related risk points (e.g., "checking the bearing temperature of pump A," "taking a picture of the status of valve B," and "reading the gas concentration in area C"), and adds logical dependencies to dynamically generate a new, clearly defined composite inspection task. This task has a unique task ID and contains an ordered sequence of steps.For the generated composite inspection task, the resource dynamic adaptation unit initiates optimization calculations. It first inventories all available resources, including inspection robots, drones, and fixed cameras controllable by a gimbal, each with its location, capabilities, and status attributes. Then, the unit solves an optimization problem, aiming to minimize the overall task completion time and resource scheduling cost while meeting the capability requirements of each task step. The solution allocates the most suitable resources to each task step, forming an optimal inspection resource combination scheduling scheme, and immediately issues execution instructions. The knowledge base and model iteration module are responsible for the system's continuous learning and optimization, forming a closed loop of "practice-learning-evolution": for each composite inspection task planned and executed by the agent, its complete context information is captured and structured by the system; this includes: the triggering factors for this task... The risk scenario characteristics, the generated composite tasks, all verification data collected during task execution, and the final handling results and expert confirmation labels (success / failure / risk resolution, etc.) constitute a complete contextualized case, which is stored in the system knowledge base. The system periodically (weekly) retrains the power plant large-scale model using new cases accumulated in the knowledge base. The training employs a contrastive learning strategy: successfully handled cases are used as positive examples, and failed or irrelevant cases are used as negative examples. The training objective is to enable the model to distinguish between them, maximizing the correlation (similarity) between "risk scenario characteristics" and "post-task verification status characteristics" in positive examples, while minimizing the correlation in negative examples. By minimizing the contrastive loss function, the power plant large-scale model continuously absorbs new practical experience, allowing its risk identification accuracy, correlation analysis capabilities, and task synthesis rationality to continuously iterate and evolve.

[0021] In one embodiment of the present invention, the multimodal data acquisition module includes: The timing operation data of the equipment is obtained in real time from the power plant's distributed control system or plant-level monitoring information system through a standard industrial interface or a specific interface protocol. The text data includes structured work information obtained from the power plant work order and operation ticket system, and unstructured text reports obtained from equipment maintenance records.

[0022] Specifically, this system establishes a data channel with the power plant's core control and monitoring system to acquire real-time time-series data reflecting the dynamic processes of equipment. For modern power plants with a high degree of digitalization, this embodiment prioritizes OPC UA as the standard industrial interface protocol. OPC UA features platform independence, built-in security mechanisms (encryption and authentication), and powerful information modeling capabilities, enabling efficient and secure access to real-time and historical databases in DCS (Distributed Control System) and SIS (Plant Monitoring Information System). For some older systems using proprietary or specific communication protocols, a dedicated protocol conversion gateway is developed or configured. This gateway is deployed at the security I / II zone boundary of the power plant's information network and is responsible for converting data from proprietary protocols into the standard OPC UA information model or the common MQTT / HTTP JSON. The system maintains a list of key measurement points, jointly developed by power plant process experts and system engineers. This list covers key parameters characterizing the health and safety status of core equipment such as boilers, turbines, generators, and main auxiliary equipment, including main steam temperature and pressure, drum water level, bearing vibration, winding temperature, flue gas composition, valve opening commands and feedback, etc. Based on this list, the data acquisition service initiates data subscription to the DCS / SIS server through established OPC UA sessions or API interfaces.The system employs either variable rate transmission or fixed period (1 second) reading strategies to ensure data real-time performance and network load balance. All collected data is accompanied by a high-precision timestamp. Video data provides an intuitive understanding of personnel behavior, equipment appearance, and environmental conditions. The system connects to existing security or industrial monitoring network cameras in the factory area via standard streaming media protocols. For older cameras that do not support standard protocols, protocol conversion is performed by deploying video encoders. During system deployment, each connected camera undergoes spatial registration, i.e., its installation location, orientation, and field of view are precisely marked in a Geographic Information System (GIS) or digital twin model, providing spatial context for subsequent intelligent analysis and situational awareness fusion. The connected video streams first enter the video management service, which is responsible for forwarding, load balancing, and lightweight preprocessing (bitrate adaptation, keyframe extraction). Some video streams are directly sent to the storage system for post-event review, while others serve as a source for real-time analysis. The visual analysis subnetwork of the power plant's large model module is pushed to the system. Text data, carrying rich procedural knowledge and experience records, is key to semantic understanding. The system interfaces with the power plant's "Work Permit and Operation Permit System" via an enterprise service bus or dedicated API interface. When a new permit ("#3 Boiler Superheater Overhaul Work Permit" or "#1 Feedwater Pump Start-up Operation Permit") is created, approved, or executed, the system automatically captures its structured field information. The captured information includes, but is not limited to: permit number, work type, risk level, planned work time, work equipment KKS code, work supervisor, work team members, safety measures list, and permit conditions. This highly structured data provides accurate input for the agent to perform task planning and permission verification. For unstructured or semi-structured text such as equipment maintenance records, operation logs, and event reports, the system obtains it through the following two methods: direct database connection: for data stored in relational databases (Oracle, ... Records in SQL Server are periodically queried through read-only accounts to extract fields such as report content, time, device, and responsible person; document parsing: for offline reports stored in formats such as PDF and Word, the system deploys a document parsing engine.The engine first uses OCR technology to recognize the text in the scanned document, then uses natural language processing technology for named entity recognition to extract key information such as "fault phenomenon," "cause analysis," "handling measures," and "involved equipment," and stores it in the system knowledge base after semi-structured processing. All collected multimodal data must undergo unified entry processing before entering the core analysis module: a high-throughput message middleware is used as the data bus, and each data acquisition sub-service converts the acquired data into messages in a unified format and publishes them to the corresponding topics. The system has a built-in data quality probe to continuously monitor the health status of the data stream, including: breakpoints and jump anomalies in time-series data, frame drops and interruptions in video streams, data latency, etc. Once an anomaly is detected, an alarm is immediately triggered to notify the operation and maintenance personnel to intervene, ensuring that the data input into the analysis model is continuous, complete, and reliable.

[0023] In one embodiment of the present invention, the construction and training process of the large-scale power plant model includes: The large-scale power plant model adopts a Transformer architecture multimodal encoder-decoder model. The encoder part includes: a time-series data encoding sub-network: using a recurrent neural network with an attention mechanism to process the equipment's time-series operating data and extract time-series features. Visual data encoding sub-network: This network combines convolutional neural networks and visual Transformers to process surveillance video data and extract spatial and spatiotemporal features. Text data encoding sub-network: Employs the encoding layer of a pre-trained language model to process text data and extract semantic features. ; A cross-membrane attention fusion layer is configured to receive the features. , and The association weights between features of different modalities are calculated through a multi-head attention mechanism, and the fused unified features are output. .

[0024] Labeled multimodal data extracted from the power plant's historical operation and maintenance database was used as the training set. The labels included risk type, associated equipment identification, and risk evolution stage. Supervised training was conducted by minimizing the weighted sum of the cross-entropy loss function for risk prediction and the loss function for equipment association prediction.

[0025] Specifically, the sub-network construction includes: a time-series data encoding sub-network: Given the strong temporal dependence and nonlinear characteristics of power plant equipment time-series data, this embodiment employs a deep network combining bidirectional gated recurrent units and a self-attention mechanism. The BiGRU layer is used to capture the forward and backward long-term dependencies of the sequence, and its output hidden state sequence is then fed into a lightweight Transformer encoder. The self-attention mechanism focuses on the features at key time points. The output of this sub-network is a fixed-dimensional feature vector rich in temporal dynamic information. The visual data encoding sub-network, to balance local details and global context, adopts a hybrid architecture of "CNN backbone network + Vision Transformer (ViT) patch". First, a pre-trained CNN such as ResNet is used to extract low-level spatial features from video frames or image sequences. Then, these feature maps are divided into sequential "visual lexical units" and fed into a lightweight ViT module. Through self-attention calculation across "visual lexical units", the spatiotemporal relationships within and between frames are modeled, and finally, a visual feature vector that integrates local and global information is output. Text data encoding sub-network: This sub-network directly utilizes the encoder part of a language model (BERT, RoBERTa, or similar model) pre-trained on a large-scale general corpus as its foundation. For power industry terminology, it uses a vast amount of power industry literature, regulations, and reports for domain-adaptive pre-training. For the input text ("The report shows that the vibration of the #2 blower bearing continues to increase, accompanied by abnormal noise"), this sub-network outputs a contextual feature vector representing the complete semantics. ; To align and deeply fuse features from different modalities in the semantic space, this embodiment constructs a cross-modal multi-head attention fusion layer. Specifically, the feature vector of each modality is linearly projected into Query(Q), Key(K), and Value(V) matrices. In the attention calculation, Q of one modality is used to query K and V of another modality. For example, temporal features are used as Q to query K and V of visual features, thereby learning the association of "what visual anomalies may correspond to a specific temperature change pattern". By executing multiple sets of such cross-modal attention calculations (multi-head) in parallel, and concatenating and linearly transforming the outputs of all heads, a unified joint feature representation that reflects the intrinsic association between the cross-modalities is finally generated. From the power plant's historical operation and maintenance database, event logging system, video storage server, and work order system accumulated over many years, time-aligned multimodal data samples are extracted. Each sample contains equipment time-series data for a specific time period (1 hour before the fault occurs), corresponding monitoring video clips, and related operation logs or event report texts. Expert annotation: Senior operation, maintenance, and safety experts from the power plant are organized to perform three levels of annotation on each data sample: Risk type label: such as "bearing overheating," "steam and water leakage," "insulation aging," "personnel violation," etc.; Related equipment identifier: listing the KKS codes of related equipment directly affected by the risk and those that may trigger chain reactions; Risk evolution stage label: marking the risk as "emergence stage," "development stage," or "critical stage," which usually requires post-event analysis for judgment. End-to-end model training is performed using the annotated data. The training objective of the model is to simultaneously complete risk identification and equipment correlation analysis. Therefore, the overall loss function is defined as the weighted sum of the two sub-tasks, specifically: ,in, For risk prediction, cross-entropy loss, For equipment-related predicted losses, and The hyperparameters are used to balance the importance of the two tasks and are typically determined through cross-validation. A distributed training framework is used on a server cluster equipped with multiple high-performance GPUs. The AdamW optimizer is employed, and a cosine annealing learning rate scheduling strategy with hot restart is set to promote more stable convergence of the model to a better local minimum. During training, an independent validation set is used to monitor the model's performance on metrics such as risk identification accuracy and associated device recall, and early stopping is used to prevent overfitting. After training, the model is subjected to knowledge distillation or quantization compression to reduce its computational and storage overhead, enabling it to be deployed on the edge computing servers of the power plant and meet the low latency requirements of real-time inference.

[0026] In one embodiment of the present invention, the task synthesis unit includes: The process by which the task synthesis unit generates composite inspection tasks is as follows: It has Each basic inspection item, and each inspection item A basic inspection action corresponds to a required resource type; when an investigation is identified... In complex risk scenarios involving multiple related devices, the task synthesis unit retrieves, based on a predefined task map, items covering the aforementioned devices. A set of all necessary basic inspection items for each associated device ; Define task composition function ,in, The context feature vector of the current risk scenario is output by the power plant large model. The derived function calculates the execution priority score of each basic inspection item in the current context using a neural network. The composite inspection task, which includes a unique task ID, a sequence of task steps, and the expected target, is generated according to priority and logical relationship.

[0027] Specifically, in the initial stage of system deployment, in conjunction with power plant process experts, safety engineers, and equipment management personnel, a comprehensive review and deconstruction of all routine and non-routine inspection work throughout the plant was conducted. Each finest-grained, atomic-level inspection or operation was defined as a basic inspection item. Each inspection item is defined by a set of structured attributes, including at least: Action description: for example, "Measure the bearing temperature [equipment identifier]", "Visually inspect [equipment identifier] for leaks", "Record the pressure reading [instrument identifier]"; Required resource type: Indicates the equipment capabilities necessary to complete the action, such as "thermal imager", "high-definition zoom camera", "acoustic sensor", "mobile robot platform"; Standard time consumption: The average estimated time to complete this inspection under typical operating conditions; Preconditions / safety constraints: "Must be performed while the equipment is stopped" "Must hold a specific work permit"; All basic inspection items constitute a queryable and scalable basic inspection item library, which is the "Lego brick" of task synthesis; This embodiment constructs a knowledge base based on a graph structure—the task graph. The nodes of this graph represent entities such as equipment, systems, risk types, and safety procedures, while edges represent relationships between entities, such as "belonging to the same system," "historical fault association," "physical spatial proximity," and "operational logic dependency." In the graph, for each equipment node or risk node, a set of the most relevant basic inspection items are pre-associated. For example, for the "feed pump bearing" equipment node, its associated inspection items might include I1 (measure temperature), I2 (measure vibration), and I3 (listen for abnormal noises). The graph also stores the execution logic relationships (sequential, parallel, mutually exclusive) between different inspection items. When the power plant large model module outputs a related... When the assessment results of a complex risk scenario involving M related devices (M≥2) are obtained ("suspected leakage in the boiler water-cooled wall, which may lead to overheating of the adjacent superheater tube wall"), the task synthesis unit first extracts a list of unique identifiers for these M devices from the assessment results. Using these M device identifiers as input, the task synthesis unit performs a breadth / depth-first traversal on the task graph, and along the edges of "device-related inspection items", quickly retrieves all basic inspection items that cover all M devices and are highly related to the current risk type ("leakage" "overheating"), forming a preliminary, non-redundant set of candidate basic inspection items. In order to dynamically adjust the task, the unit derives a context feature vector that is more focused on the current decision from the fused features output by the power plant big model. The context feature vector encodes information such as the degree of risk urgency, the current operating load of the equipment, environmental conditions (light, noise) and the real-time status of available resources. This embodiment designs and trains a lightweight priority scoring neural network. The network input is a concatenation of the feature representation (vectorized encoding from the inspection item library) of each candidate inspection item and the context feature vector. After training, the network outputs a scalar score. The training objective is to make this score reflect the overall utility of performing the inspection item under the current context feature vector. Factors considered include: information gain: the contribution of performing this inspection to confirming or eliminating risk; timeliness: the urgency of the inspection result; feasibility: the probability of success based on the current resource status and environmental conditions; cost / risk: the potential disturbance or cost of performing this inspection; and priority scoring for all candidates. The priority sequence is obtained by sorting the tasks in descending order. Combined with the logical relationships read from the task graph (e.g., gas concentration must be measured before entering a confined space), the priority sequence is logically rearranged to form a linear sequence of task steps that conforms to safety specifications and operating procedures. Finally, the system assigns a globally unique task ID to this newly generated composite task, clarifies its expected objectives (e.g., "confirm the leak point of the #3 boiler water-cooled wall and assess its impact on the superheater"), and encapsulates the task step sequence, estimated required resource types, and total expected time into a structured task object.

[0028] In one embodiment of the present invention, the resource dynamic adaptation unit calculates and allocates the optimal inspection resource combination, including: Let the available inspection resources be set. Each resource It has type, current location, set of functional capabilities, and status attributes; For the first in the composite inspection task Task steps Its required capability vector is ; The unit allocates resources by solving an optimization problem. The objective function is to minimize the total expected completion time and the total resource scheduling cost, specifically:

[0029] in, Let be the objective function. For the first in the composite inspection task Each task step The total number of steps in the task. To assign to step Inspection resources, Let be the allocation function, representing the first... The resource index allocated to each task step. The expected completion time function, The set of inspection resources that are scheduled and activated in the task. To enable resources Cost weighting coefficient and These are the normalized balance coefficients; The constraints of the optimization problem include: each step must be allocated a resource that can meet its capacity requirements, and the same resource does not conflict in time.

[0030] Specifically, the system performs unified modeling of all schedulable inspection resources within the plant area, assuming the set of available resources is... Each resource Maintain a dynamic attribute model, which should include at least: Type: such as "wheeled inspection robot", "quadcopter drone", "fixed gimbal camera", "handheld thermal imager (carried by personnel)"; Unique identifier and current location: its precise coordinates in the map coordinate system are obtained in real time through UWB, laser SLAM, GPS / BeiDou or visual positioning system; Functional capability set: a vectorized representation describing the functions of the resource; Status attributes: including battery level / endurance, whether it is currently idle, whether it is under fault or maintenance, and the current task queue; For composite inspection tasks from the task synthesis unit, each step has a quantified capability requirement vector, which has the same dimension as the resource capability set, but each dimension represents the minimum required capability level; for example, the requirement vector for the step "measure the temperature at point A" might be Need = [thermal imaging accuracy > 0.1°C, mobility = reachable, others = 0]; This embodiment formalizes the resource allocation problem as a constrained multi-objective combinatorial optimization problem and solves it precisely using a mathematical model. Specifically, it introduces decision variables. , indicating steps Is it assigned to a resource? (1 for yes, 0 for no); Also, for each step Define its start time and finish time as continuous variables; Total expected completion time: ,in, As a key function, compute resources Execution steps The estimated time is as follows: ,in, For the time of movement, based on the current location of the resource. With the target location of the steps The map path distance is divided by the average movement speed of the resource, taking into account factors such as factory access routes and gate access control. The task execution time is estimated based on the historical average efficiency model of the steps (number of photos, number of measurement points) and resources used to execute this type of task. To enable cost weighting, it can be dynamically set based on resource type (robot depreciation / energy consumption > drone > fixed camera rotation) and current status (e.g., increasing cost when battery is low); The value is 0.6. The value is 0.4; This embodiment models the aforementioned optimization model (containing 0-1 integer variables and continuous variables) as a mixed-integer linear programming or constraint satisfaction problem. The system calls an efficient mathematical programming solver (such as Gurobi, CPLEX) or a dedicated scheduling algorithm library to solve the problem. The solver outputs the optimal allocation scheme and the planned time for each step. Based on this, the system generates a detailed inspection resource scheduling scheme, which includes at least: which resource will execute each step, the task sequence (steps and planned time) for each resource, and the critical path and estimated total completion time of the entire composite task. The scheduling scheme is distributed to the local controllers or central control systems of each inspection resource (robot, drone, etc.) via a secure wireless network. The instructions include target locations, operation actions, safety parameters, and timing requirements. During task execution, the unit continuously monitors the resource status and task progress. If an unexpected event occurs (such as resource failure, channel congestion, or severe timeout of a step), the system can quickly trigger a local rescheduling based on the current state, resolving the allocation of the remaining steps to ensure the achievement of the overall goal.

[0031] In one embodiment of the present invention, training the large-scale power plant model based on the case study includes the following steps: Contextualized cases are extracted from the knowledge base, each case including: risk scenario description features. The composite inspection tasks performed Verification data collected during task execution and the final risk management outcome label Based on the label of the disposal result Positive examples of successfully managing risks Negative cases that were unsuccessful or irrelevant Composition of training sample pairs ; Through trainable encoding networks Extract verification data The key state features are denoted as verification state features. ; Calculate the risk scenario characteristics of positive samples and their own verification state characteristics. Cosine similarity as the first similarity ; Calculate the risk scenario characteristics of positive samples and the first Validation state feature vector of each negative sample Cosine similarity as the second similarity A temperature coefficient hyperparameter is introduced to construct a contrast loss function.

[0032] Specifically, recent contextualized cases are periodically (e.g., daily or weekly) extracted from the system knowledge base. Each case is a complete "perception-decision-verification" closed-loop data package, strictly containing four elements: Risk scenario description features: a high-dimensional vector snapshot of the fusion features directly from the power plant's large-scale model that triggered the task, encoding the multimodal on-site state when the risk occurred; The composite inspection task executed: the specific task object generated and executed by the intelligent agent's task execution module, including its step sequence, resource allocation, and other structured information; Task verification data: all data actually collected and transmitted back by the scheduled resources during task execution, such as high-definition photos and temperature heat maps of fault points taken by the inspection robot, equipment status videos recorded by drones, and precise values ​​read by sensors; Final handling result label: the result confirmed and entered into the system by operators or safety experts afterward, such as "Risk confirmed and eliminated ( =Success) "Risk is false alarm ( =Failure / False Alarm) "Task Not Completed ( =Failure / Abort); based on the tag The cases are strictly divided into two categories: positive cases (Y = success, representing an effective risk identification and handling) and negative cases (Y = failure / false alarm / abortion, representing model misjudgment or task invalidity). In each round of training, a positive case is randomly selected, and K negative cases are randomly sampled from the negative case pool to form a training sample group. The system maintains a trainable verification encoding network, which (usually a convolutional neural network or a small Transformer) is specifically responsible for extracting essential features that can characterize the "on-site state after task execution" from multimodal verification data (mainly images, videos, and sensor readings) to obtain a verification state feature vector; the network participates in training together with the power plant large model; In this process, a hyperparameter temperature coefficient greater than zero is introduced to adjust the scale of the similarity score before the exponential operation. A smaller hyperparameter temperature coefficient will amplify the differences between similarities, making the model pay more attention to difficult negative examples; a larger hyperparameter temperature coefficient will make the distribution smoother and the learning more stable. Among them, iterative training: using a large number of accumulated contextualized cases to construct training sample groups, the model is trained in multiple rounds of iterative training. Each round of training prompts the model to deepen its understanding: what risk features of which patterns should be associated with the verification results of which patterns; when the model's performance on independent verification case sets no longer improves significantly (such as the loss function decreasing slowly, or the mean similarity of positive examples reaching a preset threshold), the training is considered to have converged; after the trained new version of the power station large model and the verification coding network are tested, the old model in the online inference module is replaced by a hot update, thereby achieving a seamless upgrade and evolution of the system's core intelligence.

[0033] In one embodiment of the present invention, constructing the contrastive loss function includes: The construction of the contrastive loss function is specifically as follows:

[0034] in, To compare loss functions, The first similarity score, For the second similarity, This is the temperature coefficient hyperparameter. This represents the total number of negative samples paired with positive samples in the current training sample group. Indexed by quantity.

[0035] In one embodiment of the present invention, the following modules are also included: Real-time alarm and situation display module: used to receive the safety risk assessment results output by the power plant large model module, and generate an alarm event when the risk level exceeds the preset risk threshold; The module integrates a visualization engine to overlay and display the potential risk points, related equipment, predicted risk evolution paths, and the real-time location and status of the inspection resources dispatched by the intelligent agent task execution module on the power plant's digital twin map.

[0036] Specifically, the real-time alarm and situation display module serves as the "command and control view" for system-user interaction. Its core objective is to transform the complex intelligent analysis and decision-making processes in the background into intuitive, real-time, and operable command information, achieving a "one-screen overview and one-map dispatch" of the power plant's safety situation. The implementation steps are as follows: The system administrator configures a multi-dimensional alarm rule base in the background. This rule base not only includes general risk level thresholds (dividing risks into three levels: "early warning," "alarm," and "emergency"), but also supports refined threshold settings based on risk type, the importance of associated equipment, and the speed of risk evolution. For example, the alarm threshold for "boiler tube leakage" is much higher than that for "abnormal auxiliary area lighting." Alarm rules support logical combinations; for example, when "risk level > alarm" and "is in deep peak shaving condition," it automatically escalates to an "emergency" alarm. The module monitors the safety risk assessment result stream output by the power plant large model module in real time. When it detects that the risk level in any risk assessment result exceeds the preset threshold of its corresponding rule... Upon activation, the alarm generation process is immediately triggered. The system automatically generates a structured alarm event object, including: a unique alarm ID, trigger time, geographical coordinates of the risk point (latitude and longitude or in-plant coordinates), risk type, level, list of associated equipment, credibility score, and a snapshot of the original assessment data. Based on the alarm event object and preset push rules, the system performs precise multi-channel, role-based push notifications: Control room screen / audio-visual alarm: All "emergency" level alarms immediately trigger the control room audio-visual alarm and display a strong pop-up notification on the central screen; Mobile application push: Alarm information (including a brief description, location, and images) is pushed in real time to relevant responsible persons, such as the on-duty shift leader, equipment manager, and safety inspector, via WeChat, DingTalk, or a dedicated APP. The push list automatically matches personnel responsibilities based on "associated equipment" and "risk type"; SMS / telephone notification (optional): For particularly serious or nighttime "emergency" alarms, SMS or voice calls can be automatically triggered to ensure that the information is delivered. In this embodiment, the existing building information model, 3D point cloud scan data, or high-precision 2D / 3D plant area map of the power plant are used to construct the system's visual base map—the power plant digital twin map. Through a spatial registration process, each pixel / vertex in the map is precisely mapped to the actual geographic coordinate system (WGS-84) or the plant's internal engineering coordinate system. At the same time, all key equipment models (such as boilers, turbines, and transformers) and inspection resource models (robot and drone 3D models) are accurately placed in their corresponding positions on the map. The visualization engine establishes a data subscription interface with other core modules of the system to receive the following dynamic data streams in real time: a list of risk points (including coordinates, type, and level) output by the power plant's large model, and a list of composite inspection tasks, resource allocation schemes, and real-time GPS / UWB positioning data and status information (power, task progress, online / offline) of each inspection resource output by the intelligent agent task execution module. On the digital twin map, risk points are dynamically plotted using icons of different colors and shapes (e.g., a flashing red triangle for "emergency" and an orange circle for "alarm"). Clicking an icon brings up a details panel displaying the risk description, associated equipment, and predicted trend charts. Equipment associated with the current risk point is automatically highlighted in the 3D scene (with a red border or semi-transparent red) to visually represent the risk's impact range. For the risk evolution trend predicted by the large-scale power plant model, dynamic particle flow, color temperature diffusion maps, or arrow path diagrams are used to visualize the potential direction of risk spread on the map. The system displays the expected impact area and uses personalized 3D icons to dynamically move and move on the map based on real-time location data for mobile resources such as inspection robots and drones. The icons can display simple statuses ("In Task" or "Idle") next to them. When a resource performs a specific inspection step, its operation can be simulated near the corresponding equipment location with dynamic effects (such as scanning aperture and focus frame). The lines connecting resources, tasks, and risk points intuitively show "who is performing which task and which risk is being addressed". Clicking on any resource icon will display its detailed status panel, including current task, battery life, and historical trajectory playback. The default view is a macro-level situation map covering the entire plant area, displaying an overview of the distribution of all active alarms and online resources. Users can quickly focus on a specific workshop or equipment group by clicking or selecting a box to view the detailed situation of that local area, achieving seamless zooming from macro to micro. In the visualization interface, administrators can directly click on an idle inspection resource icon on the map and then drag it to a risk point or equipment to manually issue an emergency inspection command. The system will automatically generate a temporary task and issue it, achieving agile command with "what you see is what you get".

[0037] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A real-time alarm and statistical analysis system for power plant safety inspections, characterized in that, include: The multimodal data acquisition module is used to collect power plant operation data in real time, including equipment time-series operation data, monitoring video data, and text data; The power plant large model module is used to construct a power plant large model, perform fusion analysis on the data collected by the multimodal data acquisition module, identify potential risk points and their associated equipment, predict the risk evolution trend, and output safety risk assessment results. The intelligent agent task execution module is communicatively connected to the power plant large model module and includes a task synthesis unit and a resource dynamic adaptation unit. The task synthesis unit is used to dynamically combine basic inspection items to generate a new composite inspection task based on the cross-equipment related risks or complex risk scenarios identified in the safety risk assessment results. The resource dynamic adaptation unit is used to calculate and allocate the optimal combination of inspection resources in real time for the composite inspection task. The knowledge base and model iteration module is used to store and manage data samples and algorithm models generated during the inspection process, record the composite inspection tasks and their execution results, store them in the knowledge base as contextualized cases, and train the power plant large model based on the cases.

2. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 1, characterized in that, The multimodal data acquisition module includes: The timing operation data of the equipment is obtained in real time from the power plant's distributed control system or plant-level monitoring information system through a standard industrial interface or a specific interface protocol. The text data includes structured work information obtained from the power plant work order and operation ticket system, and unstructured text reports obtained from equipment maintenance records.

3. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 1, characterized in that, The construction and training process of the large-scale power plant model includes: The large-scale power plant model adopts a Transformer architecture multimodal encoder-decoder model. The encoder part includes: a time-series data encoding sub-network: using a recurrent neural network with an attention mechanism to process the equipment's time-series operating data and extract time-series features. Visual data encoding sub-network: This network combines convolutional neural networks and visual Transformers to process surveillance video data and extract spatial and spatiotemporal features. Text data encoding sub-network: Employs the encoding layer of a pre-trained language model to process text data and extract semantic features. ; A cross-membrane attention fusion layer is configured to receive the features. , and The association weights between features of different modalities are calculated through a multi-head attention mechanism, and the fused unified features are output. ; Labeled multimodal data extracted from the power plant's historical operation and maintenance database was used as the training set. The labels included risk type, associated equipment identification, and risk evolution stage. Supervised training was conducted by minimizing the weighted sum of the cross-entropy loss function for risk prediction and the loss function for equipment association prediction.

4. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 1, characterized in that, The task synthesis unit includes: The process by which the task synthesis unit generates composite inspection tasks is as follows: There are N basic inspection items, each inspection item Each basic inspection action corresponds to a required resource type. When a complex risk scenario involving M related devices is identified, the task synthesis unit retrieves a set of all necessary basic inspection items covering the M related devices based on a predefined task map. ; Define task composition function ,in, The context feature vector of the current risk scenario is output by the power plant large model. The derived function calculates the execution priority score of each basic inspection item in the current context using a neural network. The composite inspection task, which includes a unique task ID, a sequence of task steps, and the expected target, is generated according to priority and logical relationship.

5. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 1, characterized in that, The resource dynamic adaptation unit calculates and allocates the optimal combination of inspection resources, including: Let the available inspection resources be set. Each resource It has type, current location, set of functional capabilities, and status attributes; For the first in the composite inspection task Task steps Its required capability vector is ; The unit allocates resources by solving an optimization problem. The objective function is to minimize the total expected completion time and the total resource scheduling cost, specifically: in, Let be the objective function. For the first in the composite inspection task Each task step The total number of steps in the task. To assign to step Inspection resources Let be the allocation function, representing the first... The resource index allocated to each task step. The expected completion time function, The set of inspection resources that are scheduled and activated in the task. To enable resources Cost weighting coefficient and These are the normalized balance coefficients; The constraints of the optimization problem include: each step must be allocated a resource that can meet its capacity requirements, and the same resource does not conflict in time.

6. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 1, characterized in that, The training of the large-scale power plant model based on the aforementioned case includes the following steps: Contextualized cases are extracted from the knowledge base, each case including: risk scenario description features. The composite inspection tasks performed Verification data collected during task execution And the final risk management outcome label Y; based on the outcome label Y, positive examples of successfully managed risks will be identified. Negative cases that were unsuccessful or irrelevant Composition of training sample pairs ; Through trainable encoding networks Extract verification data The key state features are denoted as verification state features. ; Calculate the risk scenario characteristics of positive samples and their own verification state characteristics. Cosine similarity as the first similarity ; Calculate the risk scenario characteristics of positive samples and the first Validation state feature vector of each negative sample Cosine similarity as the second similarity A temperature coefficient hyperparameter is introduced to construct a contrast loss function.

7. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 6, characterized in that, The construction of the contrastive loss function includes: The construction of the contrastive loss function is specifically as follows: in, To compare loss functions, The first similarity score, For the second similarity, This is the temperature coefficient hyperparameter. This represents the total number of negative samples paired with positive samples in the current training sample group. Indexed by quantity.

8. The real-time alarm and statistical analysis system for power plant safety inspections according to claim 1, characterized in that, It also includes the following modules: Real-time alarm and situation display module: used to receive the safety risk assessment results output by the power plant large model module, and generate an alarm event when the risk level exceeds the preset risk threshold; The module integrates a visualization engine to overlay and display the potential risk points, related equipment, predicted risk evolution paths, and the real-time location and status of the inspection resources dispatched by the intelligent agent task execution module on the power plant's digital twin map.

Images