A method and device for intelligently diagnosing the operating state of an excitation system of a hydropower plant

Abstract

The application discloses a kind of water power plant excitation system operating state intelligent diagnosis method and device, it is related to water power plant intelligent monitoring and fault diagnosis technical field, wherein method includes: build integrated infrared temperature measurement, visualization, voice, storage and wireless communication module hardware platform, real-time temperature monitoring and multimodal data acquisition are carried out to water power plant excitation system and upload host computer;Based on the data acquisition constructs DDPG operating state prediction model, builds strategy network, value network and target network, data characteristics are mapped into state and action space.By interactive iteration, reward evaluation and gradient update complete model training, obtain pre-training diagnosis model;Real-time data are input into model, output fault diagnosis, fault probability and state evaluation result, and generate corresponding disposal measures, realize excitation system integrated intelligent diagnosis.The application builds modular hardware platform and DDPG prediction model, solves the technical problems of complex working condition data distortion, early warning lag.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and fault diagnosis technology for hydropower plants, and in particular to an intelligent diagnostic method and device for the operating status of a hydropower plant excitation system. Background Technology

[0002] During the operation of hydropower generating units, the excitation system, as a core device for voltage regulation and excitation control, directly affects the safe and stable operation of the entire unit and even the entire hydropower plant due to changes in its internal structural temperature and operational stability. With the development of power systems towards intelligence and automation, traditional monitoring methods relying on manual inspections, fixed-point temperature measurements, and post-event investigations are no longer adequate for the highly dynamic, complex, and continuous operation scenarios of hydropower plants. These methods suffer from problems such as monitoring lag, incomplete data, and untimely anomaly detection, failing to meet the requirements of modern hydropower plants for refined, real-time, and intelligent operation and maintenance of the excitation system.

[0003] Current temperature monitoring methods for hydropower plant excitation systems primarily focus on acquiring external surface temperatures and visualizing infrared images. These methods only provide a simple presentation of temperature data and intuitive image display, failing to capture in-depth information on temperature changes in critical internal structures of the excitation system. Furthermore, existing monitoring solutions generally lack stable and reliable multimodal data fusion mechanisms. Under complex operating conditions, monitoring data is prone to deviation, compromising accuracy and reliability, and lacking robust real-time online interaction and remote collaborative control capabilities.

[0004] More notably, most existing technologies remain at the level of data monitoring and visualization analysis, without combining advanced intelligent algorithms to deeply mine and learn from massive amounts of monitoring data. They lack the ability to predict and analyze equipment operating trends and to diagnose faults intelligently. They cannot identify potential anomalies in advance, and it is difficult to determine whether there are hidden faults or risks of "operating with defects" in the excitation system. This can easily lead to the expansion of faults and affect the safe operation of the unit.

[0005] In the current field of safe operation and maintenance of hydropower plant excitation systems, there is an urgent need for an integrated online monitoring method that can achieve accurate acquisition of internal structural temperature, adaptive monitoring under complex operating conditions, reliable transmission of multimodal data, and real-time remote human-machine interaction, while also having intelligent diagnosis and status prediction functions. This method is intended to fill the gaps in existing technologies in terms of all-dimensional monitoring, forward-looking early warning, and intelligent diagnosis, and ensure the long-term safe and reliable operation of the excitation system. Summary of the Invention

[0006] The main objective of this invention is to provide an intelligent diagnostic method for the operating status of a hydropower plant excitation system.

[0007] Another objective of this invention is to provide an intelligent diagnostic device for the operating status of a hydropower plant excitation system.

[0008] The third objective of this invention is to provide an electronic device.

[0009] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0010] To achieve the above objectives, a first aspect of the present invention proposes an intelligent diagnostic method for the operating status of a hydropower plant excitation system, comprising:

[0011] A hardware circuit platform integrating infrared temperature measurement module, visualization module, voice module, storage module and wireless communication module was built to perform real-time temperature monitoring and multimodal data acquisition of the excitation system of hydropower plant, and the acquired data was transmitted to the host computer. Based on real-time acquired multimodal data, a running state prediction model with the DDPG algorithm as its core is constructed. The policy network, value network and corresponding target network structures are built, and the multimodal data features are mapped to the model state space and action space. The operating state prediction model interacts and iterates with the operating environment of the excitation system. The value of actions is evaluated and the policy gradient is updated through a reward mechanism. The parameters of the policy network and the value network are iteratively trained and optimized to obtain a pre-trained diagnostic model. The real-time collected multimodal data is input into the pre-trained diagnostic model, which outputs the fault diagnosis results, fault occurrence probability and operation status assessment conclusions of the excitation system, and generates corresponding processing measures based on the conclusions, thereby realizing integrated intelligent diagnosis of the hydropower plant excitation system.

[0012] Optionally, a hardware circuit platform integrating an infrared temperature measurement module, a visualization module, a voice module, a storage module, and a wireless communication module is built to perform real-time temperature monitoring and multi-modal data acquisition of the hydropower plant's excitation system, and to transmit the acquired data to a host computer, including: The infrared temperature measurement module, visualization module, voice module, storage module and wireless communication module are integrated and assembled to form an integrated hardware circuit platform; The infrared temperature measurement module performs continuous infrared scanning and temperature data acquisition on various equipment parts of the hydropower plant's excitation system to obtain real-time raw infrared temperature data. The acquired multimodal temperature data is transmitted to the visualization module, which then drives the visualization module to draw data curves and display the status interface in real time. The voice module provides real-time voice broadcast of the collected data, enabling voice feedback of status information. The collected multimodal data is stored on-site through a storage module to form a local data cache; The hardware circuit platform is connected to the host computer via a wireless communication module, and the real-time collected multimodal data is uploaded to the host computer for centralized display and persistent storage.

[0013] Optionally, based on real-time acquired multimodal data, a runtime state prediction model with the DDPG algorithm at its core is constructed, and a policy network, value network, and corresponding target network structure are built, including: Based on the multimodal data collected and transmitted to the host computer from the hardware circuit platform, a prediction model of the excitation system operation status with the DDPG algorithm as the core is constructed. Within the operational state prediction model, an Actor policy network and a Critic value network are built to form the basic network core. Simultaneously build the target strategy network and the target value network to form a complete network architecture.

[0014] Optionally, the multimodal data features are mapped to the model state space and action space, including: Extract the real-time maximum temperature, minimum temperature, average temperature, temperature standard deviation of the excitation system, and the distance between the acquisition device and the device under test, and combine them to form the model state space; The rotation angle parameters of the infrared acquisition equipment are uniformly summarized to form the model's motion space; The state space and action space are normalized in dimension and standardized in format to make the data structure suitable for the input and output format of the DDPG algorithm network.

[0015] Optionally, the operating state prediction model interacts and iterates with the excitation system operating environment, and the value of actions is evaluated and the policy gradient is updated through a reward mechanism, including: The state space is input into the running state prediction model, the network weights are loaded and propagated forward to drive the model to generate the corresponding action space. The action space is output to the excitation system simulation environment to control the agent to perform corresponding actions, thus completing the real-time interaction process between the agent and the environment. Collect system state change information from environmental feedback after interaction, quantify and score interactive behavior through a preset reward mechanism, and complete the evaluation of action value; The value network loss is calculated based on the action value obtained from the evaluation, and the value network parameters are updated by backpropagation. The policy gradient is calculated based on the updated value network output, and the policy gradient is updated and corrected in real time.

[0016] Optionally, the parameters of the policy network and value network are iteratively trained and optimized to obtain a pre-trained diagnostic model, including: The updated policy gradient is fed into the policy network to perform policy network parameter gradient updates and optimize the action output logic. The interactive data of state, action, reward, and next state are stored in the experience replay pool, and samples are randomly selected for batch training. The parameters of the target policy network and the target value network are updated smoothly with a fixed step size to maintain stable convergence during the training process. Repeatedly execute the processes of state input, action generation, environment interaction, reward evaluation, gradient calculation and parameter update to complete multiple rounds of iterative training; After the model converges and the error meets the preset threshold, the network parameters are solidified to form a pre-trained diagnostic model with state prediction and fault diagnosis capabilities.

[0017] Optionally, the real-time collected multimodal data is input into the pre-trained diagnostic model, which outputs the fault diagnosis results, fault occurrence probability, and operating status assessment conclusions of the excitation system, and generates corresponding processing measures based on the conclusions, including: Multimodal data collected in real time by the hardware circuit platform is imported into the pre-trained diagnostic model to complete data feature extraction and state encoding; The model identifies and classifies the current state of the system, and outputs the diagnostic results of the excitation system fault type and fault location. Based on the temperature time series evolution characteristics, trend inference is performed to calculate the probability of failure in future periods and output the corresponding values; The health level of the excitation system is comprehensively scored by combining the diagnostic results and the failure probability, the operating status level is divided, and a complete evaluation conclusion is formed. Based on the assessment results, pre-set handling rules are matched to generate targeted fault handling measures, thereby realizing integrated intelligent diagnosis of the hydropower plant's excitation system.

[0018] To achieve the above objectives, a second aspect of the present invention provides an intelligent diagnostic device for the operating status of a hydropower plant excitation system, comprising: The hardware building unit is used to build a hardware circuit platform that integrates an infrared temperature measurement module, a visualization module, a voice module, a storage module, and a wireless communication module. It is used to perform real-time temperature monitoring and multi-modal data acquisition of the excitation system of the hydropower plant and transmit the acquired data to the host computer. The model building unit, based on real-time acquired multimodal data, constructs a runtime state prediction model with the DDPG algorithm as its core. This model is used to build the policy network, value network, and corresponding target network structures, and maps the multimodal data features into the model's state space and action space. The iterative training unit is used to interact and iterate with the operating environment of the excitation system using the operating state prediction model. It evaluates the value of actions and updates the policy gradient through a reward mechanism, iteratively trains and optimizes the parameters of the policy network and the value network to obtain a pre-trained diagnostic model. The intelligent diagnostic unit is used to input the real-time collected multimodal data into the pre-trained diagnostic model, output the fault diagnosis results, fault occurrence probability and operation status assessment conclusion of the excitation system, and generate corresponding processing measures based on the conclusions to realize integrated intelligent diagnosis of the hydropower plant excitation system.

[0019] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0020] To achieve the above objectives, a third aspect of this application provides an electronic device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement an intelligent diagnostic method for the operating status of a hydropower plant excitation system as described in the first aspect embodiment.

[0021] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an intelligent diagnostic method for the operating status of a hydropower plant excitation system as described in the first aspect embodiment.

[0022] The embodiments of the present invention have the following beneficial effects: 1. This invention establishes a real-time data acquisition and monitoring hardware circuit platform for hydropower plant excitation systems, integrating multiple sensing and interaction modules. It can stably achieve real-time acquisition, voice broadcasting, local storage, visualization display, and human-computer interaction of infrared temperature data under highly dynamic and complex operating conditions, significantly improving the accuracy and reliability of monitoring data, providing high-quality data support for excitation system status analysis, and effectively improving the problems of data distortion and poor scenario adaptability of traditional monitoring methods.

[0023] 2. This invention uses the DDPG algorithm to construct a predictive model for the operating status of the excitation system. Through continuous interaction and iterative optimization of the control strategy between the agent and the operating environment, the model's adaptability to complex operating conditions and the accuracy of status prediction are greatly improved. Based on the pre-trained model, the fault probability, diagnostic conclusions and operating status assessment results can be quickly output, enabling early identification and proactive warning of potential faults in the excitation system. This effectively improves the stability of equipment operation and maintenance efficiency, reduces the risk of failure and downtime, and extends the service life of the equipment. Attached Figure Description

[0024] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of an intelligent diagnostic method for the operating status of a hydropower plant excitation system provided in an embodiment of the present invention; Figure 2 A schematic diagram of the hardware circuit platform provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an infrared temperature measurement module provided in an embodiment of the present invention; Figure 4 A schematic diagram of a voice module provided in an embodiment of the present invention; Figure 5 This is a visual display and human-computer interaction diagram provided for embodiments of the present invention; Figure 6 This is a schematic diagram of real-time online monitoring and data acquisition provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the strategy update for the excitation system operation state prediction model based on the DDPG algorithm provided in an embodiment of the present invention. Figure 8 This is a structural diagram of an intelligent diagnostic device for the operating status of a hydropower plant excitation system, provided in an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0027] The following describes, with reference to the accompanying drawings, an intelligent diagnostic method and apparatus for the operating status of a hydropower plant excitation system according to an embodiment of the present invention.

[0028] Example 1 This invention provides an intelligent diagnostic method for the operating status of a hydropower plant's excitation system. Figure 1 This is a flowchart illustrating an intelligent diagnostic method for the operating status of a hydropower plant excitation system, provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step S1: Build a hardware circuit platform integrating an infrared temperature measurement module, a visualization module, a voice module, a storage module, and a wireless communication module to perform real-time temperature monitoring and multi-modal data acquisition of the hydropower plant's excitation system, and transmit the acquired data to the host computer.

[0029] In this embodiment, to achieve full-dimensional monitoring of the excitation system under highly dynamic and complex operating conditions, the hardware circuit platform adopts a modular integrated design, using a Raspberry Pi as the core main control module, which undertakes the tasks of coordinated control, data processing, and instruction distribution of various functional sub-modules. Its overall architecture is as follows: Figure 2 As shown, the main control module establishes stable electrical connections with the infrared temperature measurement module, storage module, voice module, Bluetooth module, visualization module, and WiFi module through dedicated interfaces. Each module has a clear division of labor and works collaboratively to form an integrated, highly reliable real-time monitoring hardware system, providing hardware support for subsequent data acquisition, transmission, and intelligent diagnosis.

[0030] Among them, the infrared temperature measurement module, as the core component for temperature data acquisition, has the following structural diagram: Figure 3 As shown, this module achieves bidirectional communication with the Raspberry Pi main control module via a ribbon cable. Deployed inside the excitation system control cabinet of the hydropower plant, it precisely targets key equipment parts such as the rotor, stator, and rectifier bridge of the excitation system, continuously performing full-area infrared scanning. It can not only collect temperature data of each equipment part in real time, but also complete infrared thermal imaging. The collected raw infrared temperature data and thermal imaging pixel information are synchronously transmitted to the main control module, which performs preliminary filtering and noise reduction of the data to remove errors caused by environmental interference, laying the foundation for subsequent data utilization.

[0031] The specific structure of the voice module is as follows: Figure 4 As shown, it also establishes a connection with the Raspberry Pi main control module through a dedicated communication interface. Driven by the main control module, it receives and broadcasts key information such as data acquisition status, real-time device temperature, and anomaly warnings in real time, realizing voice feedback of the on-site operating status. This allows on-site maintenance personnel to grasp the device monitoring status without having to check the terminal, improving maintenance efficiency. For example, when the collected temperature data exceeds the preset threshold, the voice module will automatically broadcast "Warning: Excitation system rectifier bridge temperature exceeds limit, current temperature XX℃, please check in time," and simultaneously send a warning signal to the main control module, triggering subsequent linkage processing.

[0032] The storage module is integrated into the core area of ​​the hardware circuit platform and adopts a high-speed storage chip design. The main control module performs local real-time read and write, classification and archiving, and persistent storage of the collected multimodal data (including infrared temperature time series data, infrared thermal imaging images, equipment operation status identification data, etc.) according to preset instructions, forming a complete field data archive. It can support data backtracking and query, and avoid data loss due to sudden power outages, communication interruptions, etc.

[0033] The visualization module communicates bidirectionally with the main control module, receiving pre-processed real-time monitoring data from the main control module and driving its own display unit to complete functions such as data curve plotting, pseudo-color rendering of the temperature field, and display of equipment operating status icons. Its visualization display effect is as follows: Figure 5 As shown, the field terminal and the host computer can simultaneously display monitoring data, infrared thermal imaging images, and equipment operating status parameters, enabling intuitive presentation of monitoring data and facilitating maintenance personnel to quickly capture equipment temperature change trends.

[0034] The wireless communication module adopts a dual-mode design of Bluetooth and WiFi, with the WiFi module serving as the core remote communication channel. Its communication link diagram is shown below. Figure 5 and Figure 6 As shown, a stable TCP / IP communication link is established between the hardware circuit platform and the host computer via the WiFi module. Real-time collected multimodal data, infrared imaging information, and equipment operating status data are synchronously uploaded to the host computer, enabling remote real-time data display, centralized storage, data query, and human-computer interaction on the host computer. The Bluetooth module serves as an auxiliary short-range communication channel, supporting Bluetooth connection with the mobile terminals (such as mobile phones and tablets) of maintenance personnel. This enables portable monitoring needs such as close-range viewing of monitoring data and parameter setting, improving the flexibility of on-site maintenance.

[0035] In actual operation, the hardware circuit platform's workflow proceeds in an orderly manner: First, the main control module initiates a self-test program to initialize each functional sub-module, ensuring that the infrared temperature measurement module, voice module, and storage module are all in normal working condition; then, the main control module sends a start command to drive the infrared temperature measurement module to perform a full-area infrared scan of the equipment in the excitation system cabinet, collecting temperature data and infrared images frame by frame according to a preset sampling frequency (set to 1 time / second in this embodiment), while simultaneously correcting the collected temperature data in real time using a temperature measurement error correction formula; after collection, the main control module synchronously transmits the data to the visualization module, driving the visualization module to complete the real-time visualization display of the field terminal, including temperature time-series curves, infrared thermal images, current average temperature, and other information; simultaneously, the main control module drives the voice module to synchronously broadcast information such as collected data, average temperature, and abnormal states, such as... Figure 6As shown, the host computer terminal can display real-time announcements such as "Current average temperature 21.8 degrees Celsius, equipment is operating normally" and complete time-series monitoring data, achieving dual status feedback both on-site and remotely. Simultaneously, the storage module locally caches and categorizes all raw data, corrected data, and status information, forming a traceable data archive. Finally, the WiFi module uploads all processed multimodal data to the host computer in real-time, enabling remote monitoring, secondary data storage, and preliminary status analysis. This ensures the accuracy and reliability of measurement data under highly dynamic and complex operating conditions, providing a high-quality and reliable data foundation for subsequent model training and intelligent diagnosis based on the DDPG algorithm.

[0036] Step S2: Based on real-time collected multimodal data, construct a running state prediction model with the DDPG algorithm as the core, build a policy network, a value network and corresponding target network structure, and map the multimodal data features into the model state space and action space.

[0037] In this embodiment, the core premise of model construction is to ensure the validity and standardization of input data. Therefore, firstly, through a preset data reading interface, the system calls and reads the multimodal monitoring data path address stored in the host computer and collected and uploaded by the front-end hardware circuit platform, accurately acquiring full multimodal data such as infrared temperature time-series data, infrared thermal imaging feature data, and equipment operation status identification data. To eliminate the impact of data noise, redundant information, and dimensional differences on model training, this embodiment performs system preprocessing on the read data, sequentially performing data cleaning, outlier removal, missing value completion, and data standardization operations. Outlier removal adopts the 3σ criterion, that is, when the deviation of a data point from the mean exceeds three times the standard deviation, it is judged as an outlier and removed. Missing value completion uses linear interpolation, with the interpolation formula as follows:

[0038] in, For missing data, , The valid data adjacent to the missing data are preprocessed to regularize the data, providing high-quality data support for subsequent model building and training.

[0039] After preprocessing, this embodiment of the application constructs an operation state prediction model adapted to the operating characteristics of the hydropower plant excitation system based on the DDPG (Deep Deterministic Policy Gradient) deep reinforcement learning algorithm. The overall policy update architecture of this model is as follows: Figure 7As shown, the model enables adaptive optimization of the strategy and accurate prediction of the state. Internally, an Actor-Critic dual-network architecture is used as the basic network, with a target policy network and a target value network built simultaneously, forming a complete four-network parallel structure to ensure the stability and convergence of model training. The Actor policy network, as the action generation unit, outputs continuous action commands based on the real-time input system state to regulate the operating posture of the infrared acquisition device, achieving adaptive adjustment of the monitoring angle. The Critic value network, as the action evaluation unit, is responsible for quantifying the value of each action output by the Actor policy network, determining its contribution to system state optimization and accurate fault detection, and providing a basis for policy updates.

[0040] To avoid overfitting during model training, reduce fluctuations in network parameter updates, and improve the model's generalization ability and convergence performance, this embodiment simultaneously constructs a target policy network and a target value network, corresponding to the Actor policy network and the Critic value network. The target policy network and the Actor policy network have completely identical structures, as do the target value network and the Critic value network. The core difference lies in their parameter update mechanisms: the Actor network and the Critic network use a real-time update mode, while the target policy network and the target value network use a delayed soft update mode. The update formula is as follows:

[0041] in, For the current network parameters, For the target network parameters, These are soft update coefficients. This delayed update mechanism smooths the parameter update process, avoids training oscillations, and ensures that the model can converge stably.

[0042] After completing the network structure construction, the key step is to accurately map the preprocessed multimodal data features into a state space and action space that the model can recognize and process, thus constructing the model's input-output system. Regarding state space construction, this embodiment combines the operating characteristics and monitoring requirements of the excitation system to extract the most representative feature parameters from the multimodal data. Specifically, these include: the real-time maximum temperature, minimum temperature, average temperature, and temperature standard deviation of each key component of the excitation system, as well as the real-time distance between the infrared acquisition device and the device under test. These five feature parameters are combined to construct the model's state space S.

[0043] In terms of action space construction, the adjustable characteristics of infrared acquisition equipment are combined, and the rotation angle of the infrared acquisition equipment is uniformly summarized into the action space A of the model, forming a set of single control actions that the intelligent agent can execute. By adjusting the rotation angle of the infrared acquisition equipment, it is ensured that the temperature changes of key parts of the excitation system can be accurately captured, thereby improving the comprehensiveness and accuracy of monitoring.

[0044] To ensure that the data formats of the state space and action space are compatible with the input and output requirements of the DDPG algorithm network and to avoid model training bias due to differences in data dimensionality, this embodiment performs dimensionality normalization and format standardization on the state space and action space. The Min-Max normalization method is used to map all feature parameters to the [0,1] interval. The normalization formula is as follows:

[0045] in, These are the normalized parameter values. These are the original parameter values. , These are the minimum and maximum values ​​of the parameter, respectively. Through standardization, the data structures of the state space and action space are made to fully adapt to the input and output specifications of the DDPG algorithm network, thus completing the final definition of the model's input and output system.

[0046] In this embodiment, to clarify the working logic of the Actor policy network and the Critic value network, corresponding core functions are defined respectively. The action value function of the Actor network is defined as follows:

[0047] in, Current state Next action The value of the action, For expectation operator, The immediate reward for performing this action in the current state. This is a discount factor used to measure the weight of future rewards on current decisions. For the Actor network in the next state The function outputs the optimal action, which quantifies the long-term value of the current action and provides direction for policy optimization.

[0048] The evaluation loss function of the Critic network is defined as:

[0049] in, This represents the loss value of the Critic network. For the parameters of the Critic network, The target Q-value is used to calculate the error between the current action value and the target value. The smaller the error, the higher the evaluation accuracy of the Critic network. This loss function is the core basis for updating the parameters of the Critic network. By minimizing the loss function, the evaluation capability of the Critic network is continuously optimized, ensuring that the value evaluation of the actions output by the Actor network is more accurate.

[0050] Step S3: The operating state prediction model interacts and iterates with the operating environment of the excitation system. The value of actions is evaluated and the policy gradient is updated through a reward mechanism. The parameters of the policy network and the value network are iteratively trained and optimized to obtain a pre-trained diagnostic model.

[0051] In this embodiment, the model training process relies on the interactive iterative mechanism of deep reinforcement learning, driven by the continuous dynamic interaction between the agent and the excitation system simulation environment. The model accuracy is improved through continuous trial and error, evaluation, and optimization. The overall training process and strategy update logic are as follows: Figure 7 As shown.

[0052] First, the normalized state space S is input into the pre-built operational state prediction model. The Actor policy network loads and initializes the network weights, performs forward propagation calculations, and automatically generates a matching continuous action space A based on the current input system state characteristics. This action directly corresponds to the rotation angle control command of the infrared acquisition device. Then, the generated action space A is output to the excitation system simulation environment. The agent performs corresponding attitude adjustment operations according to the action commands, completing real-time interaction with the simulation environment and transitioning the system from the current state to a new operational state. After the interaction, the model collects the new state S′ and corresponding reward value R from the environment in real time, storing a complete set of interaction samples (S, A, R, S′) in the experience replay pool for caching, providing sufficient data support for subsequent batch offline training.

[0053] To achieve precise guidance and quantitative evaluation of agent behavior, this application embodiment establishes a dedicated value reward mechanism. A preset reward function scores the performance of each action taken by the agent. The reward function is defined as follows:

[0054] in, These represent the gradient feature values ​​of the temperature anomaly region in the horizontal and vertical directions in the infrared thermal imaging image, respectively. This function quantifies the uniformity of temperature distribution and imaging clarity, giving high rewards to actions that reasonably adjust the monitoring angle and obtain high-quality temperature data, and giving low rewards or even penalties to actions that deviate from the optimal monitoring position, thereby guiding the agent to quickly converge towards the optimal monitoring angle and optimal state control direction.

[0055] After completing the action value assessment, the network parameter update phase begins. First, the Critic value network performs a value assessment based on the reward function calculation results and the target Q-value, where the formula for calculating the target Q-value y_t is:

[0056] in, The output of the target value network, The output of the target policy network, This is a discount factor used to balance the weights of immediate rewards and long-term benefits. The Critic network constructs a loss function based on the difference between the current action value and the target value, and uses gradient descent to update its network parameters, continuously improving the accuracy of action evaluation.

[0057] After the Critic network is optimized, the Actor policy network is updated using the policy gradient algorithm. The policy gradient update formula is expressed as: The model adjusts the Actor network weights through gradient ascent, strengthening the output probability of high-reward actions and weakening the selection bias of low-reward actions, thus achieving continuous policy optimization. To ensure stable and non-divergent training, this embodiment employs a soft update mechanism to update the parameters of the target network, with the update formula being: In the formula, τ is a fixed update step size, which is much smaller than 1. By smoothly and slowly updating the parameters of the target policy network and the target value network, the parameter oscillation during the training process is effectively suppressed, and the convergence speed and prediction stability of the model are significantly improved.

[0058] Throughout the iterative training phase, the model continuously draws batches of historical interaction samples randomly from the experience replay pool, repeatedly executing the complete process of state input, action generation, environmental interaction, reward evaluation, gradient calculation, and parameter update, achieving multi-round, highly efficient offline training. When the model's loss function value continuously decreases and stably converges to within the preset error threshold, and the loss fluctuation amplitude of consecutive iterations meets the preset stability condition, it indicates that the model has excellent state prediction and fault identification capabilities. At this point, training is stopped and all network parameters are solidified, ultimately forming a pre-trained diagnostic model that can be directly used in engineering sites, providing core support for subsequent real-time online fault diagnosis and state assessment of the excitation system.

[0059] Step S4: Input the real-time collected multimodal data into the pre-trained diagnostic model, output the fault diagnosis results, fault occurrence probability and operation status assessment conclusion of the excitation system, and generate corresponding processing measures based on the conclusions to realize integrated intelligent diagnosis of the hydropower plant excitation system.

[0060] In this embodiment, S4 serves as the core implementation stage of the entire intelligent diagnostic method. It follows the data acquisition in S1, the model building in S2, and the model training in S3. It deeply integrates the pre-trained diagnostic model with the real-time monitoring scenario on site, realizing a closed loop from data input to diagnostic decision-making and disposal execution. This effectively solves the technical pain points of lagging online diagnosis, insufficient accuracy, and untimely disposal in the excitation system, providing intelligent support for the safe and stable operation of the hydropower plant's excitation system.

[0061] After the online diagnostic phase begins, the hardware circuit platform built by S1 continuously collects multimodal real-time data from the hydropower plant's excitation system, including infrared temperature time-series data, infrared thermal imaging images, and equipment operating status identification data. This data undergoes preliminary filtering and noise reduction by the hardware platform and is then transmitted in real-time to the host computer via a WiFi module. The host computer then imports all multimodal data synchronously into the pre-trained diagnostic model trained by S3 through a preset data interface. To ensure a complete match between the input data and the model input specifications, the pre-trained diagnostic model first performs feature extraction and state encoding operations on the imported real-time data, replicating the core logic of data preprocessing and feature mapping in S2. This accurately maps the real-time collected multimodal data into a state space input that the model can recognize, ensuring that the data format and dimensions remain consistent with those during model training. This avoids diagnostic errors due to data mismatch and guarantees the accuracy of subsequent diagnostic results.

[0062] After data input and encoding, the pre-trained diagnostic model, relying on the network parameters fixed in S3 and the excitation system state evolution laws learned during training, accurately identifies and classifies the current operating state of the system. The model deeply mines the temperature characteristics and temporal variation patterns in real-time multimodal data, comparing and matching them with the normal state and various fault state characteristics learned during training. It quickly outputs specific fault diagnosis results for the excitation system, clearly identifying the fault type (such as excitation winding overheating, rectifier bridge poor contact, rotor insulation damage, etc.) and the precise location of the fault. Simultaneously, it labels the corresponding abnormal temperature characteristics, providing clear fault location information for on-site maintenance personnel, completely changing the limitation of traditional monitoring that can only detect anomalies but cannot accurately locate faults.

[0063] While outputting fault diagnosis results, the pre-trained diagnostic model, based on its learned temperature time-series evolution characteristics, extrapolates and predicts the temperature change trend of the excitation system over a preset period, outputting the probability of the corresponding fault occurrence. This probability value is strictly controlled between 0 and 1; the closer the value is to 1, the higher the risk of fault occurrence, and the closer the value is to 0, the more stable the system operation. To further improve the accuracy of fault probability prediction, this embodiment of the application incorporates real-time temperature fluctuations to correct the probability results, comprehensively considering factors such as temperature distribution uniformity and temperature change rate, making the probability prediction results more closely match the actual on-site working conditions and providing a reliable basis for subsequent condition assessment.

[0064] Subsequently, combining the output fault diagnosis results and fault occurrence probabilities, the pre-trained diagnostic model comprehensively scores and classifies the current health level of the excitation system, constructing a scientific and comprehensive state assessment system. In this embodiment, the operating status of the excitation system is divided into four levels: normal, warning, abnormal, and fault. The normal level corresponds to a low probability of fault occurrence, stable system operation, and no potential abnormalities; the warning level corresponds to a moderate probability of fault occurrence, with minor system anomalies and potential fault risks; the abnormal level corresponds to a high probability of fault occurrence, with significant system anomalies and prominent fault risks; and the fault level corresponds to an extremely high probability of fault occurrence, with a clear system fault requiring immediate shutdown. The comprehensive score is calculated by weighting factors such as the severity of the fault type and the probability of fault occurrence. This comprehensive assessment method forms a complete and accurate operational status assessment conclusion, providing a clear basis for the formulation of subsequent handling measures.

[0065] Based on the above fault diagnosis results, fault occurrence probability, and operational status assessment conclusions, the pre-trained diagnostic model automatically matches the preset fault handling rule library to generate targeted and implementable fault handling measures, achieving seamless integration of diagnosis and handling. Specifically, for the normal level, routine maintenance reminders are generated to prompt maintenance personnel to regularly inspect and calibrate equipment to ensure continuous and stable system operation; for the warning level, targeted maintenance suggestions are generated, such as adjusting the infrared monitoring angle, strengthening temperature monitoring of key parts, and shortening the inspection cycle, to identify potential hazards in advance; for the abnormal level, load adjustment suggestions are generated, such as appropriately reducing the excitation system load and adjusting operating parameters, while triggering a warning notification to remind maintenance personnel to promptly investigate anomalies and prevent the escalation of hazards; for the fault level, an emergency shutdown command is immediately generated to simultaneously disconnect relevant operating circuits to prevent the fault from escalating, prevent equipment damage or safety accidents, and minimize the economic losses caused by the fault.

[0066] Throughout the entire online diagnosis and handling process, all modules work together to form a closed-loop intelligent diagnosis system: the host computer terminal receives and displays in real time the fault diagnosis results, fault occurrence probability, comprehensive evaluation conclusions, and handling measures output by the pre-trained diagnostic model, and simultaneously stores all diagnostic data and handling records for easy data backtracking and operation and maintenance analysis, providing data support for optimizing operation and maintenance strategies; the voice module of the hardware circuit platform simultaneously broadcasts abnormal warning information, fault types, and handling suggestions, enabling on-site operation and maintenance personnel to quickly grasp the on-site situation without checking the terminal, improving handling efficiency; the visualization module updates the equipment temperature field status, fault location identification, and operating status level in real time, and intuitively presents the system operating status through pseudo-color rendering, abnormal area marking, etc., providing clear visual support for the handling operations of operation and maintenance personnel.

[0067] In summary, the embodiments of this application, through the online diagnosis and handling process of S4, organically integrate real-time monitoring, intelligent diagnosis, fault prediction, status assessment and decision execution to form an integrated intelligent management and control solution for hydropower plant excitation systems. This effectively improves the real-time performance, accuracy and relevance of excitation system fault diagnosis, significantly reduces the risk of outages due to faults, extends equipment service life, reduces manual maintenance costs, and promotes the upgrading of hydropower plant excitation system operation and maintenance towards intelligence, automation and refinement.

[0068] Example 2 This invention provides an intelligent diagnostic device for the operating status of a hydropower plant's excitation system. Figure 8 This is a flowchart illustrating an intelligent diagnostic device for the operating status of a hydropower plant excitation system, provided in an embodiment of the present invention. Figure 8 As shown, the device includes: Hardware building unit 100 is used to build a hardware circuit platform integrating infrared temperature measurement module, visualization module, voice module, storage module and wireless communication module, to perform real-time temperature monitoring and multi-modal data acquisition of the hydropower plant excitation system, and transmit the acquired data to the host computer. Model building unit 200, based on real-time acquired multimodal data, constructs a running state prediction model with the DDPG algorithm as its core, which is used to build the policy network, value network and corresponding target network structure, and maps the multimodal data features into the model state space and action space; The iterative training unit 300 is used to interact and iterate with the operating environment of the excitation system using the operating state prediction model. It evaluates the value of actions and updates the policy gradient through a reward mechanism, iteratively trains and optimizes the parameters of the policy network and the value network to obtain a pre-trained diagnostic model. The intelligent diagnostic unit 400 is used to input the real-time collected multimodal data into the pre-trained diagnostic model, output the fault diagnosis results, fault occurrence probability and operation status assessment conclusion of the excitation system, and generate corresponding processing measures based on the conclusions, so as to realize the integrated intelligent diagnosis of the hydropower plant excitation system.

[0069] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0070] Example 3 To implement the methods of the above embodiments, the present invention also provides an electronic device, which includes a memory and a processor; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the various steps of the methods described above.

[0071] Example 4 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0073] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0074] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for intelligent diagnosis of the operating status of a hydropower plant excitation system, characterized in that, include: S1 is a hardware circuit platform that integrates an infrared temperature measurement module, a visualization module, a voice module, a storage module, and a wireless communication module. It is used to monitor the temperature of the hydropower plant's excitation system in real time and collect multimodal data, and then transmit the collected data to the host computer. S2, based on real-time collected multimodal data, constructs a running state prediction model with the DDPG algorithm as the core, builds policy network, value network and corresponding target network structure, and maps multimodal data features into model state space and action space; S3 utilizes the interactive iteration between the operating state prediction model and the excitation system operating environment, evaluates the value of actions and updates the policy gradient through a reward mechanism, iteratively trains and optimizes the parameters of the policy network and value network, and obtains a pre-trained diagnostic model. S4. Input the real-time collected multimodal data into the pre-trained diagnostic model, output the fault diagnosis results, fault occurrence probability and operation status assessment conclusion of the excitation system, and generate corresponding processing measures based on the conclusions to realize integrated intelligent diagnosis of the hydropower plant excitation system.

2. The method according to claim 1, characterized in that, A hardware circuit platform integrating infrared temperature measurement, visualization, voice, storage, and wireless communication modules was built to perform real-time temperature monitoring and multimodal data acquisition of the hydropower plant's excitation system, and to transmit the acquired data to a host computer, including: The infrared temperature measurement module, visualization module, voice module, storage module and wireless communication module are integrated and assembled to form an integrated hardware circuit platform; The infrared temperature measurement module performs continuous infrared scanning and temperature data acquisition on various equipment parts of the hydropower plant's excitation system to obtain real-time raw infrared temperature data. The acquired multimodal temperature data is transmitted to the visualization module, which then drives the visualization module to draw data curves and display the status interface in real time. The voice module provides real-time voice broadcast of the collected data, enabling voice feedback of status information. The collected multimodal data is stored on-site through a storage module to form a local data cache; The hardware circuit platform is connected to the host computer via a wireless communication module, and the real-time collected multimodal data is uploaded to the host computer for centralized display and persistent storage.

3. The method according to claim 2, characterized in that, Based on real-time acquired multimodal data, a runtime state prediction model with the DDPG algorithm at its core is constructed, and a policy network, value network, and corresponding target network structure are built, including: Based on the multimodal data collected and transmitted to the host computer from the hardware circuit platform, a prediction model of the excitation system operation status with the DDPG algorithm as the core is constructed. Within the operational state prediction model, an Actor policy network and a Critic value network are built to form the basic network core. Simultaneously build the target strategy network and the target value network to form a complete network architecture.

4. The method according to claim 3, characterized in that, Mapping multimodal data features to model state and action spaces includes: Extract the real-time maximum temperature, minimum temperature, average temperature, temperature standard deviation of the excitation system, and the distance between the acquisition device and the device under test, and combine them to form the model state space; The rotation angle parameters of the infrared acquisition equipment are uniformly summarized to form the model's motion space; The state space and action space are normalized in dimension and standardized in format to make the data structure suitable for the input and output format of the DDPG algorithm network.

5. The method according to claim 4, characterized in that, The system utilizes an operational state prediction model to iterate and interact with the excitation system's operating environment. A reward mechanism is used to evaluate the value of actions and update the policy gradient, including: The state space is input into the running state prediction model, the network weights are loaded and propagated forward to drive the model to generate the corresponding action space. The action space is output to the excitation system simulation environment to control the agent to perform corresponding actions, thus completing the real-time interaction process between the agent and the environment. Collect system state change information from environmental feedback after interaction, quantify and score interactive behavior through a preset reward mechanism, and complete the evaluation of action value; The value network loss is calculated based on the action value obtained from the evaluation, and the value network parameters are updated by backpropagation. The policy gradient is calculated based on the updated value network output, and the policy gradient is updated and corrected in real time.

6. The method according to claim 5, characterized in that, The parameters of the policy network and value network are iteratively trained and optimized to obtain a pre-trained diagnostic model, including: The updated policy gradient is fed into the policy network to perform policy network parameter gradient updates and optimize the action output logic. The interactive data of state, action, reward, and next state are stored in the experience replay pool, and samples are randomly selected for batch training. The parameters of the target policy network and the target value network are updated smoothly with a fixed step size to maintain stable convergence during the training process. Repeatedly execute the processes of state input, action generation, environment interaction, reward evaluation, gradient calculation and parameter update to complete multiple rounds of iterative training; After the model converges and the error meets the preset threshold, the network parameters are solidified to form a pre-trained diagnostic model with state prediction and fault diagnosis capabilities.

7. The method according to claim 6, characterized in that, The real-time acquired multimodal data is input into the pre-trained diagnostic model, which outputs the fault diagnosis results, fault occurrence probability, and operational status assessment conclusions of the excitation system. Based on these conclusions, corresponding processing measures are generated, including: Multimodal data collected in real time by the hardware circuit platform is imported into the pre-trained diagnostic model to complete data feature extraction and state encoding; The model identifies and classifies the current state of the system, and outputs the diagnostic results of the excitation system fault type and fault location. Based on the temperature time series evolution characteristics, trend inference is performed to calculate the probability of failure in future periods and output the corresponding values; The health level of the excitation system is comprehensively scored by combining the diagnostic results and the failure probability, the operating status level is divided, and a complete evaluation conclusion is formed. Based on the assessment results, pre-set handling rules are matched to generate targeted fault handling measures, thereby realizing integrated intelligent diagnosis of the hydropower plant's excitation system.

8. An intelligent diagnostic device for the operating status of a hydropower plant excitation system, characterized in that, include: The hardware building unit is used to build a hardware circuit platform that integrates an infrared temperature measurement module, a visualization module, a voice module, a storage module, and a wireless communication module. It is used to perform real-time temperature monitoring and multi-modal data acquisition of the excitation system of the hydropower plant and transmit the acquired data to the host computer. The model building unit, based on real-time acquired multimodal data, constructs a runtime state prediction model with the DDPG algorithm as its core. This model is used to build the policy network, value network, and corresponding target network structures, and maps the multimodal data features into the model's state space and action space. The iterative training unit is used to interact and iterate with the operating environment of the excitation system using the operating state prediction model. It evaluates the value of actions and updates the policy gradient through a reward mechanism, iteratively trains and optimizes the parameters of the policy network and the value network to obtain a pre-trained diagnostic model. The intelligent diagnostic unit is used to input the real-time collected multimodal data into the pre-trained diagnostic model, output the fault diagnosis results, fault occurrence probability and operation status assessment conclusion of the excitation system, and generate corresponding processing measures based on the conclusions to realize integrated intelligent diagnosis of the hydropower plant excitation system.

9. An electronic device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images