Intelligent power generation operation safety management method and system

Abstract

The invention relates to the technical field of artificial intelligence, Internet of Things and power generation, and discloses an intelligent power generation operation safety management method and system, and the method comprises the steps: collecting the multi-dimensional data of an operation site in real time; based on the voice instruction, the operation intention is analyzed through a multi-view semantic understanding model based on a hybrid expert model architecture, real-time position information, environment state data and equipment state data are fused, risk assessment is carried out through a dynamic risk judgment model, and a risk level is generated; according to the operation intention and the risk level, performing retrieval and reasoning in a pre-constructed multi-source knowledge graph through a retrieval enhancement generation framework based on an attention mechanism, and generating operation guidance information; and according to the risk level and the operation guidance information, triggering early warning of a corresponding level, and carrying out feedback and alarm in a voice mode through the intelligent wearable terminal. According to the invention, from passive response to active early warning, the method does not depend on experience, but forms a core intelligent treatment mode by data and knowledge.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence, the Internet of Things and power generation technology, such as an intelligent power generation operation safety management method and system. Background Technology

[0002] The power industry is a core pillar and key energy source for the national economy. Power generation site operations, including unit inspections, equipment maintenance, and high-voltage equipment operation, are core aspects of power production. The power generation site environment is both complex and high-risk, encompassing high-temperature, high-pressure, and high-voltage core equipment, as well as diverse scenarios such as indoor confined spaces, outdoor open areas, and high-altitude operations. Strict adherence to procedures is essential to prevent safety accidents; therefore, operational safety management is of paramount importance for power generation companies' operations and management.

[0003] To address the safety management needs of power generation operations, there are currently two main management solutions: The first is the traditional, manually-led management model, where on-site safety officers monitor operations in real time and trigger alarms according to fixed rules, such as preset equipment temperature thresholds and delineating restricted areas. This model heavily relies on personnel experience and is prone to missed risk assessments due to human error. The second is a preliminary intelligent auxiliary system, such as operation terminals integrating basic voice interaction functions and equipment status monitoring tools based on single models. While these systems incorporate technology, their voice interaction modules lack sufficient semantic understanding of power generation terminology, leading to misinterpretation of some terms, inflexible scheduling of some models, and weak multi-source data fusion capabilities, failing to effectively integrate key information relevant to operators. Therefore, a new power generation operation safety management approach is urgently needed to address these issues.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0006] This disclosure provides an intelligent power generation operation safety management method and system to solve the technical problems of insufficient intelligent interaction accuracy, model adaptability and multi-source data integration capability in existing power generation operation safety management methods.

[0007] In some embodiments, the intelligent power generation operation safety management method includes: Real-time collection of multi-dimensional data from the work site, including voice commands from workers, real-time location information, and environmental status data collected through smart wearable terminals, as well as equipment status data collected through equipment status sensors; Based on voice commands, the system analyzes the work intent using a multi-perspective semantic understanding model based on a hybrid expert model architecture, and integrates real-time location information, environmental status data, and equipment status data. It then uses a dynamic risk assessment model to perform risk evaluation and generate a risk level. Based on the task intent and risk level, the task guidance information is generated by retrieving and reasoning in a pre-built multi-source knowledge graph through an attention-based retrieval enhancement generation framework. Based on the risk level and work instructions, the system triggers the corresponding level of warning and provides feedback and alerts to the workers via voice through a smart wearable terminal.

[0008] In some embodiments, the smart wearable terminal integrates an adaptive beamforming microphone array, a bone conduction speaker, an ultra-wideband-based indoor positioning module, a global navigation satellite system-based outdoor positioning module, and multi-source environmental sensors. Furthermore, the smart wearable terminal employs time-sensitive networking technology to ensure the timeliness of critical data transmission.

[0009] In some embodiments, the smart wearable terminal is configured to execute a hybrid positioning algorithm based on Kalman filtering and particle filtering to achieve the fusion and switching of indoor and outdoor positioning data, wherein the hybrid positioning algorithm includes: The raw observations collected by the outdoor positioning module are preprocessed using Kalman filtering to output an initial position estimate; When the operation scenario is switched to indoor or in an area where global navigation satellite signals are blocked, the algorithm switches to a particle filter algorithm that uses the ranging value of the indoor positioning module as the observation quantity, and uses the initial position estimate as the initial distribution of the particle filter to achieve the switching of positioning mode and the fusion of positioning data.

[0010] In some embodiments, the smart wearable terminal is configured to execute an adaptive noise suppression algorithm based on a deep neural network to perform speech enhancement processing on the acquired speech commands in both the frequency and time domains.

[0011] In some embodiments, the construction of a multi-source knowledge graph includes: A joint learning model of BERT-BiLSTM-CRF is used to perform entity recognition and relation extraction on equipment parameters, operating procedure texts and historical case knowledge in the power generation field. Knowledge representation learning is carried out through the translation model in graph neural networks. The appropriate model is selected based on the complexity of the relational pattern to represent the identified entities and the extracted relations. An attention-based graph convolutional network is used to achieve knowledge graph completion and semantic reasoning. Establish a graph database storage system based on Neo4j, supporting multi-hop semantic retrieval of knowledge graphs through query language.

[0012] In some embodiments, the dynamic risk assessment model is built on a hybrid expert model architecture, including: a base scheduling model, a domain semantic expert model, an equipment state expert model, and an uncertainty estimation module, wherein: The base scheduling model adopts the Transformer-XL architecture, which processes long sequence dependencies through relative position encoding, and dynamically schedules domain semantic expert models and equipment state expert models according to risk assessment task types through a sparse gating mechanism. The domain semantic expert model is based on a pre-defined architecture and uses low-rank adaptive fine-tuning technology. It operates on a specified projection matrix of the Transformer architecture to parse risk association information in the task intent and provide semantic dimension input for risk assessment. The equipment status expert model adopts a hybrid architecture of temporal convolutional network and long short-term memory network to complete equipment fault prediction and real-time status assessment based on equipment status data; The uncertainty estimation module is built on a Gaussian process and is used to receive the output results of the domain semantic expert model and the device state expert model, and calculate the confidence level of each output result.

[0013] In some embodiments, the process of conducting risk assessment and generating risk levels using a dynamic risk assessment model includes: The base scheduling model is triggered, and the domain semantic expert model and the equipment status expert model are synchronously invoked according to the requirements of the operation scenario. The domain semantic expert model receives voice commands, parses the operation intention and extracts risk-related semantics from them, and the equipment status expert model receives equipment status data and completes equipment status assessment. At the same time, it obtains the confidence level output by the uncertainty estimation module, which together form the basic data for risk assessment. A multi-input multi-output fuzzy inference system based on the Takagi-Sugeno model is adopted to integrate basic data, real-time location information and environmental status data into a unified risk dimension, and then complete the preliminary risk level determination based on a preset rule base. The initial risk level is verified by combining the confidence level output by the uncertainty estimation module. After the verification is successful, a Q-learning-based early warning priority scheduling algorithm is used to sort the alarm push order according to the risk level and the urgency of the operation scenario so that critical alarms can be delivered in a timely manner.

[0014] In some embodiments, the method further includes: Collect feedback data from operators regarding work instructions and early warnings, and simultaneously collect data related to early warning accuracy and response delay generated during the model decision-making process; Based on the collected data, the model selection strategy is optimized through a near-end strategy optimization algorithm, and a multi-objective weighted sum reward function including early warning accuracy, response latency, and user satisfaction feedback is constructed. A contrastive learning framework is used to build user behavior profiles to achieve personalized decision-making. A federated learning-based model update mechanism is adopted. When updating model parameters, differential privacy processing is first performed on the client's model update data, and then the processed client-side local model parameters are aggregated to achieve privacy protection. Establish a model monitoring system based on concept drift detection to monitor the distribution changes of multi-dimensional data at the work site and the fluctuations in the accuracy of model output in real time. When concept drift is detected, the model retraining process is automatically triggered.

[0015] In some embodiments, an attention-based retrieval enhancement generation framework is used to perform retrieval and reasoning within a pre-built multi-source knowledge graph to generate assignment guidance information, including: By using a retrieval-enhanced generation framework, a dual DPR encoder is employed to encode the task intent query and knowledge graph fragments. Relevant paragraphs are retrieved from the Faiss index and concatenated with the original query to generate contextualized assignment instructions.

[0016] In some embodiments, the intelligent power generation operation safety management system is used to implement any of the intelligent power generation operation safety management methods described above, including: The intelligent sensing module includes a multimodal sensor array for collecting multi-dimensional data, and a signal processing unit for performing voice enhancement, positioning data fusion, and anomaly detection on the raw data. The knowledge hub module, built on a distributed graph database, is used to store pre-built multi-source knowledge graphs and realize multi-hop semantic retrieval and reasoning. The intelligent decision-making module includes a model repository that stores multi-perspective semantic understanding models, dynamic risk assessment models, and retrieval enhancement generation frameworks; an inference engine that calls models for parsing, evaluation, and generation; and a scheduler for multi-model collaborative scheduling. The interactive execution module includes a speech synthesis unit for generating speech information, an early warning triggering unit for triggering early warnings, and a feedback collection unit for collecting feedback. The optimization learning module is used to optimize the model's policy, update parameters, and monitor retraining based on feedback and performance data through proximal policy optimization, federated learning, and concept drift detection.

[0017] The intelligent power generation operation safety management method and system provided in this disclosure can achieve the following technical effects: This application integrates a smart wearable terminal to achieve synchronous real-time collection of multi-dimensional data, including operator voice commands, real-time location, environmental and equipment status, laying a data foundation for comprehensive perception. By employing a multi-perspective semantic understanding model based on a hybrid expert model architecture, it effectively improves the accuracy of intent parsing and contextual understanding in voice interactions under complex industrial noise environments. By fusing multi-source heterogeneous data and utilizing a dynamic risk assessment model for comprehensive risk evaluation, it overcomes the limitations of single-source data source analysis, significantly enhancing the accuracy and comprehensiveness of risk perception. Furthermore, through an attention-based retrieval enhancement generation framework, it performs retrieval and reasoning within a pre-constructed multi-source knowledge graph, achieving precise matching and intelligent push of massive domain knowledge with real-time operational scenarios, generating highly contextualized operational guidance information. Finally, through a smart wearable terminal, it provides tiered early warnings and feedback via voice, forming an efficient and natural human-computer interaction closed loop. All of these measures collectively address the shortcomings of existing power generation operation safety management methods in terms of intelligent interaction accuracy, model adaptability, and multi-source data integration capabilities, realizing a shift from passive response to proactive early warning. Moreover, it no longer relies solely on experience but rather on a data- and knowledge-based intelligent management model.

[0018] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0019] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a flowchart illustrating an intelligent power generation operation safety management method provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the composition of a smart wearable terminal provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram illustrating the construction process of a multi-source knowledge graph provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of a feedback optimization process provided in an embodiment of this disclosure; Figure 5 This is a framework diagram of an intelligent power generation operation safety management system provided in an embodiment of this disclosure. Detailed Implementation

[0020] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0021] The terms "first," "second," etc., used in the embodiments of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0022] Unless otherwise stated, the term "multiple" means two or more.

[0023] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0024] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0025] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0026] The intelligent power generation operation safety management method provided in the embodiments of this disclosure is described below with reference to the accompanying drawings.

[0027] Figure 1 This is a flowchart illustrating an intelligent power generation operation safety management method provided in an embodiment of this disclosure. Figure 1 As shown, the method includes the following steps: S101: Real-time collection of multi-dimensional data from the work site, including voice commands from workers, real-time location information, and environmental status data collected through smart wearable terminals, as well as equipment status data collected through equipment status sensors.

[0028] In some embodiments, the smart wearable terminal is an integrated interactive device designed specifically for power generation operation scenarios, and its hardware integrates multiple functional modules to adapt to the complex on-site environment. Figure 2 This is a schematic diagram of the composition of a smart wearable terminal provided in an embodiment of this disclosure, combined with... Figure 2The intelligent wearable terminal includes an adaptive beamforming microphone array and a bone conduction speaker. The former is used to directionally collect voice commands from workers, while the latter enables clear voice output in noisy industrial environments (such as machine operation noise and mechanical noise) without affecting workers' reception of other environmental alarm sounds. Positioning utilizes an indoor positioning module based on ultra-wideband (UWB) and an outdoor positioning module based on the Global Navigation Satellite System (GNSS), covering areas with GNSS signal obstruction within the factory and meeting positioning needs in outdoor work areas. It also integrates multi-source environmental sensors to collect environmental parameters such as temperature, humidity, harmful gas concentration, and dust content in real time. Furthermore, to ensure the timeliness of critical data transmission (such as location data and emergency voice commands), the terminal employs Time-Sensitive Networking (TSN) technology, strictly controlling end-to-end data transmission latency to ≤100ms to avoid safety risks due to data lag.

[0029] In some embodiments, during real-time location information acquisition, the smart wearable terminal executes a hybrid positioning algorithm based on Kalman filtering and particle filtering to achieve seamless fusion and smooth switching of indoor and outdoor positioning data. Specifically, when the worker is in an unobstructed outdoor area, the smart wearable terminal prioritizes the outdoor positioning module. It preprocesses the acquired raw observations (including pseudorange, carrier phase, etc.) using a Kalman filtering algorithm to filter out noise such as satellite signal jitter and electromagnetic interference, outputting a high-precision initial position estimate. When the worker enters indoor areas such as factory buildings or switchgear rooms, or areas where global navigation satellite signals are obstructed, the smart wearable terminal automatically switches to the indoor positioning module. It uses the ranging values ​​acquired by the indoor module as the observations for the particle filtering algorithm, and uses the initial position estimate output by the Kalman filter as the initial distribution for the particle filter. Iterative updates of the particle swarm correct the positioning accuracy, ultimately ensuring that the indoor positioning error is ≤0.3m and the outdoor positioning error is ≤1.5m.

[0030] During voice command acquisition, to address the issues of easily interfered and low-recognition voice signals in the high-noise environment of power generation sites, the microphone array of the smart wearable terminal, in conjunction with an adaptive noise suppression algorithm based on deep neural networks (DNN), performs voice enhancement processing. The algorithm first performs joint frequency and time domain analysis on the acquired raw voice signal. Through the DNN model, it learns the characteristic patterns of typical noises at the power generation site (such as wind turbine noise and motor operating noise), thereby suppressing noise frequency bands and enhancing voice frequency bands in the frequency domain, and eliminating signal redundancy interference in the time domain, ultimately achieving an overall signal-to-noise ratio improvement of ≥15dB. Even in a 100dB steady-state noise environment, this technology can improve the signal-to-noise ratio of the target voice to over 20dB, ensuring that the operator's voice commands are clearly acquired and providing high-quality voice data for subsequent work intent analysis.

[0031] In some embodiments, for environmental status data acquisition, the multi-source environmental sensors of the smart wearable terminal collect environmental parameters in real time at a preset sampling frequency. For example, this frequency can be adjusted according to the work scenario: once every 30 seconds in a routine inspection scenario, and once every 5 seconds in a high-risk work scenario. The collected data is preprocessed locally on the terminal and then uploaded. The preprocessing stage adopts an anomaly detection algorithm based on variational autoencoder (VAE). By learning the distribution characteristics of historical normal environmental data through the VAE model, when the real-time collected environmental parameters (such as a sudden increase in the concentration of harmful gases in a certain area or a temperature exceeding the safe operating threshold of the equipment) deviate from the normal distribution, the algorithm can identify and mark the abnormal environmental data in real time, and upload the finally marked abnormal environmental data to the system, providing data input for the environmental risk dimension for subsequent dynamic risk assessment.

[0032] In some embodiments, equipment status data is collected by dedicated equipment status sensors deployed on key parts of power generation equipment, such as voltage sensors, current sensors, and insulation status sensors deployed on electrical equipment (switch cabinets, busbars), to collect parameters such as voltage, current, and insulation resistance in real time. These data, together with voice commands, real-time location information, and environmental status data collected by smart wearable terminals, constitute a multi-dimensional data set at the work site.

[0033] S102: Based on the voice command, the operation intention is analyzed by a multi-perspective semantic understanding model based on a hybrid expert model architecture, and the real-time location information, environmental status data and equipment status data are integrated. A risk assessment is performed by a dynamic risk judgment model to generate a risk level.

[0034] In some embodiments, the dynamic risk assessment model for risk evaluation is built on a hybrid expert model (MoE) architecture, including a base scheduling model, a domain semantic expert model, an equipment status expert model, and an uncertainty estimation module, which together provide accurate and reliable model support for risk assessment at the power generation operation site.

[0035] The base scheduling model adopts the Transformer-XL architecture, which effectively handles two typical long-sequence dependency data in power generation scenarios through relative position encoding technology: first, voice commands from operators involving multiple steps (such as "first check the turbine bearing temperature, then check the circulating water pump pressure"), and second, continuous time-series data collected by equipment status sensors (such as the insulation resistance change sequence of the switchgear within one hour). Simultaneously, the base scheduling model incorporates a sparse gating mechanism, which dynamically schedules downstream expert models based on the type of actual risk assessment task. For example, when determining the "risk correlation between hot work and ambient temperature," the calling weight of the domain semantic expert model is prioritized, thereby achieving precise allocation of computing resources and meeting the real-time requirements of power generation operations.

[0036] This domain semantic expert model is designed to meet the specialized semantic understanding needs of the power generation field. Based on a pre-defined large language model architecture (such as ChatGLM3), it optimizes the specified projection matrices (Query and Value projection matrices) of the Transformer architecture using low-rank adaptation (LoRA) fine-tuning technology. This allows the model to accurately adapt to the terminology system of the power generation field with only a few parameter updates, rather than relying on full-data fine-tuning. Its core function is to receive voice commands from operators, first parsing out structured operational intentions such as "operation object, operation type, and operation stage," and then further mining the risk association information implicit in the intentions. For example, "operation object is a 10kV live busbar, associated with 'personnel proximity' risk; operation type is hot work → associated with 'ambient temperature, flammable and explosive gases' risk." Ultimately, this provides the core semantic input for risk assessment, solving the problem of ambiguous semantic understanding of industrial scenarios by traditional models.

[0037] The equipment status expert model employs a hybrid architecture of Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM). The TCN part uses an exponentially increasing expansion rate sequence (e.g., 1, 2, 4, 8, ...) with a fixed-size convolutional kernel to simultaneously capture the changing patterns of equipment status data on both short and long time scales. The LSTM part further models the temporal dependencies of the feature sequences output by the TCN, effectively avoiding the gradient vanishing problem in long-sequence training and ensuring the model's sensitivity to changes in equipment status. Its specific functions include two aspects: first, real-time status assessment based on equipment status data (e.g., vibration frequency, temperature, pressure, insulation resistance, etc.) to determine the current state of the equipment (normal, abnormal, or faulty); second, fault prediction through trend analysis of historical and real-time data, identifying potential equipment risks in advance (e.g., "the turbine bearing temperature will exceed the safety threshold within the next 30 minutes"), providing crucial equipment-level evidence for risk assessment.

[0038] Finally, the uncertainty estimation module, built on Gaussian Process (GP), receives the output results of the domain semantic expert model and the equipment status expert model. Through the probabilistic modeling capability of Gaussian Process, it calculates the confidence level for each output result (e.g., "confidence level of work intention analysis result is 95%)". When the confidence level is higher than the preset threshold, the result is directly included in the risk assessment. If the confidence level is lower than the threshold (e.g., the confidence level of semantic analysis is less than 80% due to unclear voice commands), data re-acquisition is triggered (e.g., prompting the operator to repeat the command), thereby avoiding risk misjudgment caused by the output bias of a single model.

[0039] In some embodiments, the components of the above-described dynamic risk assessment model work together to complete risk assessment and generate risk levels according to the following process: First, the base scheduling model is triggered, synchronously invoking the domain semantic expert model and the equipment status expert model according to the requirements of the operation scenario. The domain semantic expert model receives voice commands, parses the operation intent, and extracts risk-related semantics. The equipment status expert model receives equipment status data and completes equipment status assessment, while simultaneously obtaining the confidence level output by the uncertainty estimation module. These three elements together form the basic data for risk assessment. Next, a multi-input multi-output fuzzy inference system based on the Takagi-Sugeno model is used to integrate the above basic data with the real-time location information and environmental status data of the operation site into a unified risk dimension. This is then combined with a pre-set safety rule base in the power generation field to complete the preliminary risk level determination. Finally, the preliminary risk level is verified using the confidence level output by the uncertainty estimation module. After successful verification, a Q-learning-based early warning priority scheduling algorithm is used to sort the alarm push order according to the risk level and the urgency of the operation scenario, ensuring that critical alarms can reach operators and monitoring personnel in a timely manner, providing a basis for subsequent operation guidance generation and safety intervention.

[0040] S103: Based on the stated task intent and risk level, the task guidance information is generated by performing retrieval and reasoning in a pre-built multi-source knowledge graph using an attention-based retrieval enhancement generation framework.

[0041] Figure 3 This is a schematic diagram illustrating the construction process of a multi-source knowledge graph provided in an embodiment of this disclosure. Combined with... Figure 3The construction of the multi-source knowledge graph includes: First, a BERT-BiLSTM-CRF joint learning model is used to process multi-source data such as equipment parameter documents, operational safety procedures, and historical accident cases in the power generation field. Through the semantic understanding and sequence labeling capabilities of this model, the identification and extraction of equipment entities (such as 10kV switchgear), procedure entities (such as voltage testing operations), and risk correlations (such as live-line work - insulation protection required) are accurately completed. The model is trained using no less than 50,000 high-quality labeled corpora containing equipment, procedures, and risk points in the power generation field. The core training parameters include: batch size of 32, initial learning rate of 2e-5, cross-entropy loss function with early stopping strategy to prevent overfitting, and F1 value ≥ 0.85. In the knowledge representation learning stage, an appropriate graph neural network translation model is selected based on the complexity of the relational patterns in the data. For simple one-to-one relationships, the TransE model is prioritized for efficient representation, while for complex relationships such as symmetry and inversion, the RotatE model is used to enhance semantic expressiveness. Simultaneously, a negative sampling strategy (generating a corresponding negative sample for each positive sample) optimizes the model training effect. To improve the graph structure and reasoning ability, an attention-based graph convolutional network (Attention-based GCN) is employed to fill in any entity associations or semantic gaps that may have been missed during extraction, ensuring the integrity of the knowledge logic. Finally, a graph database storage system based on Neo4j is established. This database supports the Cypher query language and enables multi-hop semantic retrieval of the knowledge graph (e.g., retrieving from "hot work" to "ambient temperature requirements" and then to "fire extinguishing equipment configuration"), providing a structured and highly relevant domain knowledge foundation for subsequent generation of work guidance information through a retrieval-enhanced generative framework.

[0042] In some embodiments, guided by the parsed work intent (e.g., "10kV switchgear maintenance"), after initiating the Retrieval Enhanced Generation (RAG) framework, the work intent query and knowledge graph fragments are first encoded using a DPR dual encoder, and relevant knowledge is focused using an attention mechanism (e.g., prioritizing the association of safety procedures with high-risk scenarios). Then, relevant paragraphs are retrieved from the Faiss index, and finally, the query and retrieval results are fused by a Fusion-in-Decoder to generate contextualized work guidance information that aligns with the work objectives and mitigates risks.

[0043] S104: Based on the risk level and work instruction information, trigger the corresponding level of warning and provide feedback and alarm to the workers via voice through the smart wearable terminal.

[0044] In some embodiments, based on the determined risk level (e.g., high / medium / low risk) and work instructions, a corresponding early warning strategy is first matched according to the risk level: high risk triggers an "emergency intervention" warning, medium risk triggers an "enhanced monitoring" warning, and low risk triggers a "routine reminder" warning. Then, relying on a smart wearable terminal, speech synthesis technology based on style transfer is used to generate speech content adapted to the scenario: in the case of high risk, a rapid, highly recognizable alarm voice is output (e.g., "Emergency alarm! You have entered a live danger zone, please evacuate immediately!"), and the alarm is simultaneously synchronized to the background monitoring center to notify the monitoring personnel; in the case of medium and low risk, work instructions are first broadcast (e.g., "Please check for voltage first before carrying out switchgear maintenance"), followed by a gentle risk reminder. Finally, synchronous feedback of work instructions and risk warnings is achieved through voice interaction.

[0045] Figure 4 This is a schematic diagram of a feedback optimization process provided in an embodiment of this disclosure. (In conjunction with...) Figure 4 To ensure the reliability of intelligent power generation operation safety management, this method also includes a closed-loop optimization process, as follows: First, it obtains feedback from operators on the effectiveness of operation guidance information and the timeliness of early warnings. Simultaneously, it collects key data generated during the dynamic risk assessment model's decision-making process, such as early warning accuracy (e.g., the correct triggering rate of high-risk alarms) and response delay (e.g., the time from alarm generation to arrival). Based on this data, the Proximal Policy Optimization (PPO) algorithm is used to optimize the selection strategy of the hybrid expert model. A multi-objective weighted sum reward function is designed, incorporating early warning accuracy, the reciprocal of response delay, and user satisfaction feedback. By dynamically adjusting the weights of each indicator, it ensures that model scheduling better aligns with on-site safety priorities. The multi-objective weighted sum reward function is: R = α × early warning accuracy + β × (1 / response delay) + γ × user satisfaction feedback, where α, β, and γ are dynamic weights. Initial values ​​can be set to 0.5, 0.3, and 0.2, respectively, and can be periodically adjusted based on historical performance. To achieve personalized decision-making, a contrastive learning framework (such as SimCSE) is adopted. User behavior profiles are constructed based on operational habits and feedback preferences of workers. The InfoNCE loss function is used to approximate similar behavioral features, making subsequent work guidance and warnings more aligned with individual needs. In the model update stage, a distributed mechanism based on federated learning is employed. Before uploading model update data, Gaussian-distributed noise is added to the client, and gradient pruning is performed. The noise standard deviation δ=0.01, and the gradient pruning threshold C=1.0. The local model parameters are then aggregated and processed by the cloud to prevent leakage of raw data. Simultaneously, a model monitoring system based on concept drift detection is established to monitor the distribution changes of multi-dimensional data at the work site and fluctuations in model output accuracy in real time. When drift is detected (such as a decrease in model accuracy due to new equipment status data), the model retraining process is automatically triggered to ensure the system maintains high adaptability and decision-making accuracy over the long term.

[0046] Based on the same inventive concept as the above-described intelligent power generation operation safety management method, this application also discloses an intelligent power generation operation safety management system in some embodiments, which is used to implement the intelligent power generation operation safety management method disclosed in the above embodiments. Figure 5 This is a framework diagram of an intelligent power generation operation safety management system provided in an embodiment of this disclosure. Combined with... Figure 5 The system includes: an intelligent perception module, comprising a multimodal sensor array for collecting multi-dimensional data, and a signal processing unit for performing speech enhancement, localization data fusion, and anomaly detection on the raw data; a knowledge hub module, built on a distributed graph database, for storing pre-built multi-source knowledge graphs and enabling multi-hop semantic retrieval and reasoning; an intelligent decision-making module, comprising a model repository storing multi-view semantic understanding models, dynamic risk assessment models, and retrieval enhancement generation frameworks, an inference engine for calling models for parsing, evaluation, and generation, and a scheduler for multi-model collaborative scheduling; an interactive execution module, comprising a speech synthesis unit for generating speech information, an early warning triggering unit for triggering warnings, and a feedback collection unit for collecting feedback; and an optimization learning module, for optimizing the model's policy, updating parameters, and monitoring retraining based on feedback and performance data through proximal policy optimization, federated learning, and concept drift detection.

[0047] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions executed by a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code; it can also be a transient storage medium.

[0048] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for descriptive purposes only and is not intended to limit the scope of protection. As used in the description herein, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used herein means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0049] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed between each other may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

Claims

1. A method for intelligent safety management of power generation operation, characterized in that, The method comprises: Real-time acquisition of multi-dimensional data of a work site, including acquisition of voice instructions, real-time position information and environmental state data of workers through intelligent wearable terminals, and acquisition of device state data through device state sensors; Based on the voice instructions, the work intention is analyzed through a multi-view semantic understanding model based on a hybrid expert model architecture, and the real-time position information, environmental state data and device state data are fused, risk assessment is performed through a dynamic risk judgment model, and a risk level is generated; According to the work intention and the risk level, a retrieval and enhancement generation framework based on an attention mechanism is used to search and reason in a pre-built multi-source knowledge graph to generate work guidance information; According to the risk level and the work guidance information, a corresponding level of early warning is triggered, and feedback and alarm are given to the workers through the intelligent wearable terminals in the form of voice. 2.The intelligent power generation operation safety management method according to claim 1, characterized in that, The intelligent wearable terminal integrates an adaptive beamforming microphone array, a bone conduction loudspeaker, an indoor positioning module based on ultra-wideband, an outdoor positioning module based on a global navigation satellite system, and a multi-source environment sensor, and the intelligent wearable terminal uses time-sensitive network technology to ensure the timeliness of key data transmission. 3.The intelligent power generation operation safety management method of claim 2, wherein, The intelligent wearable terminal is configured to execute a hybrid positioning algorithm based on Kalman filtering and particle filtering to realize fusion and switching of indoor and outdoor positioning data, wherein the hybrid positioning algorithm comprises: The original observation value collected by the outdoor positioning module is preprocessed through Kalman filtering to output an initial position estimate; When the work scene switches to an indoor or global navigation satellite signal shielding area, switch to a particle filtering algorithm using indoor positioning module ranging values as observation values, and use the initial position estimate as the initial distribution of particle filtering to realize switching of positioning mode and fusion of positioning data. 4.The intelligent power generation operation safety management method according to claim 2, characterized in that, The intelligent wearable terminal is configured to execute an adaptive noise suppression algorithm based on a deep neural network to perform speech enhancement processing on the collected voice instructions in the frequency domain and the time domain. 5.The intelligent power generation operation safety management method of claim 1, wherein, The construction of the multi-source knowledge graph comprises: A BERT-BiLSTM-CRF joint learning model is used to perform entity recognition and relationship extraction on device parameters, work procedure texts and historical case knowledge in the power generation field; Knowledge representation learning is carried out through a translation model in a graph neural network, and the identified entities and extracted relationships are represented according to the complexity of the relationship mode; A graph convolution network based on an attention mechanism is used to realize completion and semantic reasoning of the knowledge graph; A graph database storage system based on Neo4j is established to support multi-hop semantic retrieval of the knowledge graph through a query language. 6.The intelligent power generation operation safety management method according to claim 1, characterized in that, The dynamic risk judgment model is constructed based on a hybrid expert model architecture, comprising a base scheduling model, a domain semantic expert model, a device state expert model and an uncertainty estimation module, wherein: The base scheduling model uses a Transformer-XL architecture to process long sequence dependencies through relative position encoding, and dynamically schedules the domain semantic expert model and the device state expert model according to the risk assessment task type through a sparse gating mechanism; The field semantic expert model is based on a preset architecture, adopts a low-rank adaptive fine-tuning technology, and acts on a specified projection matrix of a Transformer architecture to analyze risk-related information in a job intent and provide semantic dimension input for risk judgment. The device state expert model adopts a time sequence convolution network-long short-term memory network hybrid architecture to complete device fault prediction and real-time state evaluation based on device state data. The uncertainty estimation module is based on a Gaussian process and is configured to receive output results of the field semantic expert model and the device state expert model and calculate the confidence of each output result. 7.The intelligent power generation operation safety management method of claim 6, wherein, The process of risk assessment by the dynamic risk judgment model and generation of a risk level includes: The base scheduling model is triggered to synchronously call the field semantic expert model and the device state expert model according to job scene requirements, wherein the field semantic expert model receives a voice instruction, analyzes a job intent and extracts risk-related semantics therefrom, the device state expert model receives device state data and completes device state evaluation, and the confidence output by the uncertainty estimation module is obtained, to form basic data for risk assessment; The basic data, real-time location information and environmental state data are fused into a unified risk dimension by a multi-input multi-output fuzzy reasoning system based on a Takagi-Sugeno model, and a preliminary risk level is determined based on a preset rule base; The preliminary risk level is verified in combination with the confidence output by the uncertainty estimation module, and after verification, a pre-warning priority scheduling algorithm based on Q-learning is adopted to sort the risk level and the urgency of the job scene to determine the alarm push order, so that key alarms can be timely reached. 8.The intelligent power generation operation safety management method according to claim 6, characterized in that, The method further includes: Feedback data of the job personnel on the job guidance information and the pre-warning is collected, and pre-warning accuracy and response delay related data generated in the model decision process are collected; Based on the collected data, a model selection strategy is optimized by a proximal policy optimization algorithm, a multi-objective weighted reward function including pre-warning accuracy, response delay and user satisfaction feedback is constructed, and a user behavior portrait is constructed by a contrast learning framework to realize personalized adaptation of decision-making; A model updating mechanism based on federated learning is adopted, and when updating the model parameters, the model updating data of the client is processed for differential privacy, and the processed local model parameters of the client are aggregated to realize privacy protection; A model monitoring system based on concept drift detection is established to monitor the distribution changes of multi-dimensional data in the job site and the fluctuation of model output accuracy in real time, and when concept drift is detected, a model retraining process is automatically triggered. The retrieval and enhancement generation framework based on the attention mechanism retrieves and reasons in a pre-constructed multi-source knowledge graph to generate job guidance information, including: 9.The intelligent power generation operation safety management method of claim 1, wherein, The retrieval and enhancement generation framework encodes the job intent query and the knowledge graph fragment by using a DPR double encoder; The relevant paragraphs are retrieved by the Faiss index library, and are spliced with the original query to generate situational job guidance information. ​ 10. An intelligent power generation operation safety management system, characterized in that, The system for implementing the intelligent power generation operation safety management method in any one of claims 1 to 9 comprises: An intelligent perception module comprising a multi-modal sensor array for collecting multi-dimensional data, and a signal processing unit for performing voice enhancement, positioning data fusion, and abnormality detection processing on the raw data; A knowledge hub module constructed based on a distributed graph database, for storing pre-constructed multi-source knowledge graphs and implementing multi-hop semantic retrieval and reasoning; An intelligent decision-making module comprising a model repository storing multi-perspective semantic understanding models, dynamic risk judgment models, and retrieval enhancement generation frameworks, an inference engine for calling models for analysis, evaluation, and generation, and a scheduler for multi-model collaborative scheduling; An interactive execution module comprising a speech synthesis unit for generating voice information, a pre-warning triggering unit for triggering pre-warning, and a feedback collection unit for collecting feedback; An optimization learning module for implementing strategy optimization, parameter updating, and monitoring retraining of the model based on feedback and performance data through proximal policy optimization, federated learning, and concept drift detection.

Images