Nuclear power plant equipment inspection data transmission management method and system

Abstract

The invention relates to the technical field of nuclear power plant equipment inspection and data transmission management, and provides a nuclear power plant equipment inspection data transmission management method and system. According to an existing nuclear power plant equipment inspection system, reliable transmission and intelligent processing of data are difficult to achieve in complex high-radiation, high-humidity and strong-electromagnetic-interference environments, the data are prone to interference in the inspection process, and the fault detection precision is insufficient. The device state information is dynamically collected in real time through the multi-modal data collection technology, the sampling frequency is adjusted according to the environment change through the fuzzy logic control algorithm, and the timeliness and accuracy of data are improved. After data preprocessing, high-dimensional fusion and feature extraction of multi-modal data are realized through a Bayesian network and a deep neural network (DNN). The system realizes intelligent inspection and data transmission management of nuclear power plant equipment in a high-interference environment, has high reliability and real-time adaptability, and can be widely applied to an intelligent safety inspection system of a nuclear power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear power plant equipment inspection and data transmission management, and in particular to a nuclear power plant equipment inspection data transmission management method and system. Background Art

[0002] With the development of nuclear power technology, the operational safety and stability of nuclear power plant equipment have received extensive attention. There are many types of equipment in nuclear power plants, and the operating environment is complex. In particular, factors such as high temperature, high humidity, and strong electromagnetic interference have put forward higher requirements for equipment inspection. Traditional equipment inspection methods mainly rely on manual inspection and regular manual data collection, which results in a long inspection process, making it difficult to obtain equipment status information in real time, and is easily interfered by human factors, thus affecting the accuracy and reliability of the data.

[0003] In the existing automated inspection system, although the automatic collection and monitoring of equipment data has been achieved to a certain extent, it is mostly based on a single data source or a simple data fusion method, which is difficult to cope with the multimodal characteristics of equipment status data in the nuclear power plant environment. In addition, traditional anomaly detection and fault prediction methods mainly rely on static rules or a single model, which is difficult to adapt to the dynamic changes and diversity of equipment status, and there are problems such as untimely response and high false alarm rate.

[0004] In this context, there is an urgent need for a nuclear power plant equipment inspection data transmission management method and system based on multimodal data fusion and intelligent decision-making, which can realize efficient fusion of multiple sensor data under complex environmental conditions, combine dynamic decision-making and fault prediction technology, realize real-time monitoring, intelligent analysis and timely warning of equipment, thereby ensuring the safe operation of nuclear power plants and improving inspection efficiency. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for managing the transmission of nuclear power plant equipment inspection data, which aims to achieve real-time monitoring and fault prediction of equipment by comprehensively analyzing the multimodal data of the status of nuclear power plant equipment. The method utilizes multimodal data fusion technology, combined with intelligent adaptive decision-making algorithms and dynamic feedback mechanisms, to accurately evaluate the operating status of equipment and optimize inspection task priorities, thereby generating personalized inspection and early warning strategies. By solving the problems of data transmission delays, inaccurate equipment status assessment, and insufficient environmental adaptability existing in the prior art, the present invention effectively improves the automation and intelligence level of nuclear power plant equipment inspections, ensuring efficient and reliable equipment management and risk control under complex environmental conditions. To achieve the above-mentioned purpose, the present invention includes:

[0006] A nuclear power plant equipment inspection data transmission management system, the system comprising:

[0007] Multimodal data acquisition module: The module includes multiple sensor nodes, which collect the status data of nuclear power plant equipment in real time. The sensor nodes include vibration sensors, temperature and humidity sensors, and radiation detectors. The module is equipped with a dynamic sampling algorithm, which automatically adjusts the sampling frequency when an abnormality occurs in the environment of the equipment area.

[0008] Data preprocessing module: The module uses wavelet denoising and Kalman filtering to process multimodal data, where wavelet denoising includes multi-layer decomposition, soft threshold processing and reconstruction steps. The threshold setting formula is: Among them, λ is the threshold, which is used for soft threshold processing in wavelet denoising; σ represents the standard deviation of the noise, which is estimated by the known noise characteristics; N is the signal length, which represents the total number of signal samples;

[0009] The Kalman filter predicts the update step based on the state transfer matrix A and the noise covariance matrix R to achieve smoothing of the device state data;

[0010] Multimodal data fusion module: The module includes a Bayesian network and a deep neural network. The Bayesian network uses the conditional probability P(x1, x2, ..., x n ) to generate preliminary fusion features; the deep neural network includes a convolutional neural network (CNN) and a recurrent neural network (RNN), which respectively extract the spatial features of the vibration data and the time series features of the temperature, humidity and radiation data to generate a high-dimensional fusion feature vector;

[0011] Adaptive decision module: The module includes an adaptive decision tree and a reinforcement learning module. The adaptive decision tree dynamically adjusts the path structure based on the device status. The pruning algorithm removes redundant nodes when the status returns to normal to reduce the computational complexity. The reinforcement learning module updates the Q value in real time based on the device status feedback through the Q-learning algorithm. The learning rate α controls the update rate of the Q value. The discount factor γ balances the current reward and future reward to optimize the priority of the inspection task.

[0012] Anomaly detection and fault prediction module: The module includes an isolation forest algorithm and a long short-term memory network (LSTM). The isolation forest algorithm uses multiple random decision trees to perform hierarchical anomaly detection on multimodal data, calculates the path length of data points to identify anomalies, and points with short path lengths are considered high outliers; the long short-term memory network (LSTM) model analyzes the trend of time series data, generates prediction results based on the combination of forget gates, input gates, and output gates, and issues a warning signal when the predicted value exceeds the fault threshold;

[0013] Feedback and continuous learning module: Adopts incremental learning mechanism to dynamically adjust the weight parameters of Bayesian network and deep neural network.

[0014] The dynamic sampling algorithm of the multimodal data acquisition module combines priority scheduling and fuzzy logic control, and uses a fuzzy inference system to determine the sampling frequency f(x). Its fuzzy rule base includes the combination of environmental parameters E and equipment status S. The calculation formula of the sampling frequency f(x) is:

[0015]

[0016] Among them, w i is the fuzzy rule weight, μ i (E, S) represents the fuzzy membership function under the environment and equipment status, and this combination is used to dynamically adjust the data collection frequency.

[0017] The wavelet denoising in the data preprocessing module is combined with the Kalman filter to decompose the original signal x(t) into multiple layers of wavelet coefficients (d1, d2, ..., d n ), perform soft threshold processing on each high-frequency component d, and the threshold is set as:

[0018]

[0019] Among them, λ is the setting parameter of the soft threshold, which is used to suppress high-frequency noise in the wavelet denoising process; σ represents the standard deviation of the noise, which is obtained by estimating the noise level in the signal; N is the signal length, which represents the number of signal samples. The formula dynamically adjusts the threshold size based on the statistical threshold theory in wavelet denoising;

[0020] At the same time, the prediction and update steps of Kalman filtering are:

[0021]

[0022] Among them, the prediction formula is: is the predicted state, which is the estimated value of the state at time k-1 and the control variable uk; A is the state transfer matrix, which describes the process of the system state transferring from time k-1 to k; B is the control input matrix, which is used to adjust the control variable u k Impact on system status; k is the control quantity at time k;

[0023] Among them, the Kalman gain formula is: K k is the Kalman gain matrix, P is the error covariance matrix, and this method is used to achieve joint smoothing of time domain and frequency domain; H is the observation matrix, which is used to map the state space to the observation space; H T is the transpose of the observation matrix; R is the observation noise covariance matrix, which indicates the degree of observation noise.

[0024] The Bayesian network in the multimodal data fusion module combines the Markov chain Monte Carlo (MCMC) method to approximate the conditional probability P(xi|xi-1) and generate the fusion feature vector F:

[0025]

[0026] Among them, the fused feature vector F is generated by approximate sampling of conditional probability; P(xi|xi-1): represents the i-th data x i Given its previous data x i-1 The probability under the condition reflects the dependency between the data; n is the total number of data, indicating the number of features contained in the multimodal data;

[0027] The Markov chain Monte Carlo (MCMC) generates conditional distribution approximation through sampling and acceptance-rejection algorithm. The deep neural network extracts spatial features H by a convolutional neural network (CNN) with a convolution kernel size of 3x3. The hidden state update formula of the recurrent neural network (RNN) is:

[0028] h t =σ(W xh x t +W hh h t-1 +b h )

[0029] Among them, σ is the activation function, which is used to introduce nonlinear relationships; W xh and W hh is the weight matrix, b h is a bias term, and a high-dimensional fusion feature vector is generated through this structure; h t and h t-1 is the hidden state at time t; x t The input data at the current moment.

[0030] The decision tree algorithm of the adaptive decision module combines Q-learning and the multi-armed bandit algorithm (MAB) to update the priority through the Q value under the device state s. The Q value update formula is:

[0031]

[0032] Among them, Q(s,a): represents the Q value corresponding to the action a selected in the current state s, that is, the value evaluation of the action; α is the learning rate, which is used to control the speed of Q value update, and its value range is 0≤α≤10. The larger the value, the faster the update speed; R is the immediate reward, which represents the immediate return brought by the current action a; γ is the discount factor, which ranges from 0≤γ≤10 and is used to balance the importance of current rewards and future rewards. The larger the value, the more the system attaches importance to future rewards; max a′Q(s′,a′) represents the maximum Q value among all possible actions a′ in the next state s′, reflecting the system’s estimate of the optimal future benefit;

[0033] The multi-armed bandit algorithm uses the ε-greedy strategy to select the optimal path and dynamically adjusts the exploration probability ε to ensure that the optimal action is selected under multiple inspection task states.

[0034] The anomaly detection and fault prediction module uses the isolation forest algorithm to detect outliers on the path length L(x) of the data point. The path length formula is:

[0035]

[0036] Where L(x) is the average path length of data point x, which evaluates the degree of abnormality of the data point in the isolation forest algorithm; T represents the total number of trees in the isolation forest, which is defined by the user to control the size of the forest and the stability of anomaly detection; path_length(x,T i ): is the data point x in the i-th random tree T i The path length in is the number of nodes required to go from the root node to the leaf node where the data point is located;

[0037] The LSTM model further predicts the trend of abnormal points. The unit state update formula of the LSTM is:

[0038]

[0039] h t =o t ·tanh(c t )

[0040] Among them, the state update formula is: ct represents the cell state at the current time t, which contains the long-term memory of the time series; ft is the output of the forget gate, which is used to control the cell state c at the previous moment t-1 How much information needs to be retained, the value range is [0,1]; i t The output of the input gate is used to control how much of the new input information needs to be added to the current cell state; is the candidate state, which is calculated by the input data through a nonlinear activation function;

[0041] Among them, the output state formula is: h t is the output state at the current time t; o t : The output of the output gate; tanh(ct) is a nonlinear transformation of the current cell state ct;

[0042] The feedback and continuous learning module updates the weights of the Bayesian network and the deep neural network through an incremental learning mechanism. The incremental learning formula is:

[0043] W new =W old +η·ΔW

[0044] Among them, η is the learning rate, ΔW is the weight adjustment amount, and W is the weight; the federated learning framework is used to realize the distributed aggregation of each node model, and the updated model weights are synchronized to each subsystem to improve the generalization ability of the model in different environments.

[0045] A method for managing inspection data transmission of nuclear power plant equipment, comprising:

[0046] S01, multimodal data acquisition: The device status data is collected in real time through the multimodal data acquisition module. The sensor nodes include vibration sensors, temperature and humidity sensors, and radiation detectors. The collected data is prioritized through the fuzzy logic control algorithm to dynamically adjust the sampling frequency f(x). The environmental parameters E and the device status S determine the high-priority data stream through the membership function μ(E, S) in the fuzzy rule base to ensure the real-time and accuracy of high-risk data. The collected priority data stream is passed to the data preprocessing module.

[0047] S02, data preprocessing: denoising and smoothing the collected data stream; the wavelet denoising algorithm decomposes the original signal into different frequency components through multi-scale wavelet transform, and uses soft threshold Suppress high frequency noise;

[0048] The denoised signal is further processed by Kalman filtering. The prediction and update process of Kalman filtering is based on the Bayesian estimation principle, where the state prediction formula is:

[0049]

[0050] In this step, the denoised and smoothed data provides higher quality input for the multimodal data fusion module, ensuring the accuracy of feature extraction;

[0051] S03, multimodal data fusion: the preprocessed data is input into the Bayesian network and deep neural network (DNN) for multimodal data fusion; the Bayesian network uses the Markov chain Monte Carlo (MCMC) method to approximate the conditional probability distribution P(x i |x i-1), generate a preliminary fusion feature vector F, which is input into a deep neural network, where a convolutional neural network (CNN) is used to extract spatial features and a recurrent neural network (RNN) is used to analyze time series features. The generated high-dimensional fusion feature vector is passed to an adaptive decision-making module for intelligent optimization of inspection tasks;

[0052] S04, Adaptive decision-making and task optimization: Based on the fusion features, the adaptive decision tree and reinforcement learning modules optimize the priority of the inspection tasks; the adaptive decision tree dynamically removes redundant nodes based on the pruning algorithm when the state returns to normal to reduce the computational complexity; the reinforcement learning module uses the Q-learning algorithm to update the inspection task priority, where the update of the Q value follows the Bellman equation:

[0053] Q(s,α)=Q(s,α)+α[R+γmax a′ Q(s′, α′)-Q(s, α)]

[0054] Among them, α is the learning rate, γ is the discount factor, and the priority information of the task path is dynamically adjusted and passed to the anomaly detection module. a′ Q(s′,a′) represents the maximum Q value of all possible actions a′ in the next state s′, reflecting the optimal choice for the next step; R: is the immediate reward; Q(s,a) represents the Q value of the current state s and action a, reflecting the total reward that can be obtained after executing the action;

[0055] S05, anomaly detection and fault prediction: anomaly detection and fault prediction of inspection data are performed through the isolation forest algorithm and the long short-term memory network (LSTM) model; the isolation forest algorithm calculates the path length L(x) of the data point through a randomly generated binary tree structure, and the point with a shorter path length is determined to be an abnormal data point; the detected abnormal data point is then passed to the long short-term memory network (LSTM) model, and the fault is predicted through time series trend analysis; the long short-term memory network (LSTM) model adjusts the state update based on the gating mechanism (forget gate, input gate, output gate), and triggers the warning signal when the abnormal trend exceeds the fault threshold and feeds back to the control center;

[0056] S06, dynamic feedback and continuous learning: The system continuously updates the weights in the Bayesian network and deep neural network based on the inspection feedback data; the incremental learning mechanism uses the learning rate η to control the amplitude of the weight update, and the federated learning framework synchronizes the model weights between distributed nodes to improve the adaptability of the model; the updated model uses a closed-loop feedback mechanism to be reused for data collection, preprocessing and fusion BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0058] Figure 1 It is a schematic diagram of the system structure;

[0059] Figure 2 A method flow chart for managing the data transmission management system for inspection of nuclear power plant equipment;

[0060] Figure 3 This is a sequence diagram of the data transmission management system for nuclear power plant equipment inspection;

[0061] Figure 4 This is the class diagram for the nuclear power plant equipment inspection data transmission management system;

[0062] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0063] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0064] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0065] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0066] Figure 1 The following is a schematic diagram of the system structure. In the nuclear power plant equipment inspection data transmission management system, the flow of data is the core of the entire system architecture, ensuring the coordinated operation of each module. The flow of data starts with multi-modal acquisition data, which is pre-processed and fused to form comprehensive features, and then prioritized by the adaptive decision module, and finally fed back to the anomaly detection module to achieve closed-loop intelligent inspection. The following is a detailed analysis of the specific role of data in each module and how technology supports system functions.

[0067] The data flow starts from the multimodal data acquisition module, which contains multiple nodes such as vibration sensors, temperature and humidity sensors, and radiation detectors to collect equipment status information in real time. Due to the complex environment of nuclear power plants and the large amount of data, the system uses a dynamic sampling algorithm to adjust the sampling frequency in real time at the data acquisition layer. This adjustment is based on the fuzzy logic control principle, and the priority of the sampling frequency is determined by the membership function in the fuzzy rule base to ensure that the sampling frequency is encrypted in high-risk areas or when the equipment status is abnormal. The introduction of fuzzy logic realizes adaptive sampling, that is, maintaining the flexibility of sampling in environmental changes, so that the data collected by the system is both real-time and accurate, providing high-quality input data for subsequent modules.

[0068] The collected data is first processed in the data preprocessing module, which uses a combination of wavelet denoising and Kalman filtering to remove noise and fluctuations. Wavelet denoising decomposes the signal into multi-level frequency components to capture noise and valid signals in different frequency bands and filter out high-frequency interference; Kalman filtering further smoothes the data and ensures the accuracy of equipment status data through state prediction and updating. In this process, wavelet transform provides stability for feature extraction in the time domain and frequency domain, while the recursive estimation of Kalman filtering plays a key role in data smoothing and sudden anomaly detection. The processed data has high reliability, laying the foundation for feature fusion and extraction.

[0069] In the multimodal data fusion module, the system further processes the preprocessed data and constructs a conditional probability model of multimodal data through a Bayesian network. The Bayesian network evaluates the preliminary fusion features of multi-source data based on the conditional probabilities of each sensor. The Markov chain Monte Carlo method performs conditional probability sampling here to generate fusion features to make up for the missing information. Next, the system inputs the fused features into a deep neural network, extracts spatial features through a convolutional neural network (CNN) and extracts time series features through a recurrent neural network (RNN). The vibration data features extracted by the convolutional neural network (CNN) can identify the tiny vibration patterns of the equipment, while the time features of the temperature, humidity and radiation data extracted by the recurrent neural network (RNN) capture the regularity of environmental changes. Combining these high-dimensional feature vectors provides comprehensive data support for adaptive decision-making. The fused high-dimensional features can not only comprehensively reflect the equipment status, but also enhance the decision-making of the model through feature screening.

[0070] The adaptive decision module receives the fused feature data and performs optimization decisions on task priorities. This module combines adaptive decision trees and reinforcement learning (Q-learning) to achieve dynamic task adjustment. The adaptive decision tree prunes redundant paths based on real-time feedback and optimizes the decision-making process; the reinforcement learning module updates the Q value of the inspection task under different task states through the Bellman equation, and adjusts the inspection priority in time according to changes in equipment status. This process dynamically adjusts the task path through equipment status feedback to ensure flexible task allocation and improve inspection efficiency and response speed. The introduction of Q-learning enables the system to continuously optimize the current strategy based on historical experience, and then intelligently manage the allocation and execution of inspection tasks.

[0071] After receiving the task instructions from the decision-making module, the anomaly detection and fault prediction module further analyzes the inspection data to ensure the safety of the equipment. The module uses the isolation forest algorithm to detect outliers in the inspection data, generates random trees to evaluate the path length of the data points, and determines whether the data is abnormal by calculating the length of each path. Short paths are considered high outliers. The long short-term memory network (LSTM) model predicts equipment failures based on the trend of time series data, and identifies potential equipment failures in multi-dimensional time data through the synergy of input gates, forget gates, and output gates. When the data stream shows a fault trend that exceeds the threshold in the long short-term memory network (LSTM) model, the system immediately triggers an early warning signal. The isolation forest algorithm is used to identify sudden anomalies, while the long short-term memory network (LSTM) ensures accurate prediction of future states through long-term time series analysis, thus forming an accurate anomaly detection and fault prediction closed loop.

[0072] The feedback and continuous learning module receives inspection results and feedback data, and adjusts the weight parameters of the aforementioned modules through incremental learning and federated learning mechanisms. Incremental learning gradually optimizes the model based on real-time feedback from data streams, so that the weights of the Bayesian network and deep neural network are continuously updated to improve the adaptability and prediction accuracy of each module. Federated learning synchronizes model weights in multiple distributed nodes to enhance the system's generalization ability in different inspection areas. Through this closed-loop feedback mechanism, the system maintains linkage between various links such as data collection, processing, fusion, decision-making and detection, and data flows throughout, ensuring the adaptability and intelligence of inspection data transmission management.

[0073] exist Figure 2 In the method flow chart, the nuclear power plant equipment inspection data transmission management method realizes efficient monitoring of equipment status through multi-level data processing and intelligent algorithms. The algorithm logic in each step is closely related and strengthens the accuracy and real-time performance of the data, including:

[0074] S01: Multimodal Data Acquisition

[0075] In this step, the system uses fuzzy logic control algorithms to dynamically adjust the acquisition frequency to achieve priority acquisition of device status in complex environments. The fuzzy logic control system calculates the priority of the sampling frequency based on environmental parameters (such as temperature and humidity) and device status (such as vibration amplitude). The fuzzy rule base combines environmental parameters with device status to generate membership, and generates the sampling frequency through the calculation formula of the membership function. The membership function uses nonlinear membership rules to adapt to the volatility of complex environments, and adaptively adjusts the sampling frequency under different conditions through the fuzzy reasoning system. This acquisition frequency adjustment mechanism ensures the timeliness of data acquisition under high-risk device conditions and avoids the loss of important data.

[0076] S02: Data Preprocessing

[0077] In the preprocessing step, the system uses multi-scale wavelet transform and Kalman filtering to achieve data denoising and smoothing. Wavelet transform decomposes the signal into frequency components of different scales. The detail component is used to extract high-frequency noise, and the basic component retains the key characteristic signal. This method suppresses high-frequency noise through the formula of soft threshold processing to ensure the fidelity of low-frequency signals. Kalman filtering further optimizes data smoothing, updates the prediction results through the state transfer matrix and observation equation, and uses the covariance matrix R to control the prediction error. The logic of Kalman filtering recursively processes data is based on the Bayesian estimation principle, which balances the weight of current data and historical state, ensures the smoothness and time consistency of data, and provides high-quality input for subsequent fusion.

[0078] S03: Multimodal Data Fusion

[0079] The system integrates high-dimensional features of data of different modalities in the data fusion module, and extracts key information of each data through Bayesian network and deep neural network (DNN). The Bayesian network generates preliminary fusion features through conditional probability. The Markov chain Monte Carlo (MCMC) method approximates the conditional distribution of multi-source data and obtains conditional probability through random sampling. In the deep neural network, the convolutional neural network (CNN) processes vibration data, the convolution kernel extracts spatial features, and the key feature vectors are retained after dimensionality reduction through the pooling layer. The recurrent neural network (RNN) is used to extract the time series features of temperature, humidity and radiation data. The hidden state update formula of the recurrent neural network (RNN) generates the final features through time step iteration. This time-series-based feature extraction method can effectively capture the changing trend of the equipment status, and the generated high-dimensional fusion features are suitable for multi-dimensional decision optimization.

[0080] S04: Adaptive Decision Making and Task Optimization

[0081] The adaptive decision module optimizes the priority of inspection tasks through decision trees and reinforcement learning Q-learning. The adaptive decision tree adjusts its structure according to the device status and reduces the computational burden by removing redundant paths through the pruning algorithm. The Q-learning algorithm dynamically updates the Q value based on the Bellman equation, controls the balance between current rewards and future rewards through the discount factor γ, and adjusts the Q value update speed through the learning rate α. The algorithm adaptively selects the optimal inspection path between different states based on real-time feedback, improving the system's task allocation efficiency in a dynamic environment. The multi-round learning mechanism of Q-learning enables the system to continuously optimize strategies between historical data and real-time feedback to ensure flexible allocation of inspection tasks.

[0082] S05: Anomaly Detection and Fault Prediction

[0083] This step uses the isolation forest algorithm and the long short-term memory network (LSTM) to predict equipment anomalies and failures. The isolation forest evaluates the path length of data points by generating a random binary tree structure. The point with a shorter path length is an abnormal point. The long short-term memory network (LSTM) model controls the state update of time series data through the input gate, forget gate, and output gate. The input gate controls the proportion of input data, the forget gate adjusts the influence of past states, and the output gate generates the next step prediction. The gating mechanism of the long short-term memory network (LSTM) effectively retains the key state changes in long time series and is suitable for predicting equipment failure trends. When an abnormal path length is detected or the predicted value exceeds the threshold, the system triggers a warning signal to ensure that nuclear power plant equipment has an opportunity to intervene before potential failures.

[0084] S06: Dynamic Feedback and Continuous Learning

[0085] This step continuously optimizes the data collection and processing process through the feedback mechanism. The incremental learning mechanism dynamically updates the weights of the Bayesian network and the deep neural network, and realizes continuous learning based on the current data feedback. Federated learning synchronously updates the model weights between distributed nodes to ensure the adaptability of the model between different inspection areas. The weight update in incremental learning follows the gradient descent method to control the learning rate η to optimize the amplitude of weight adjustment. The closed-loop feedback mechanism forms a cycle of data, processing, decision-making, and feedback between nodes, continuously improving the generalization ability of the model and ensuring the intelligent adaptability of the system in complex environments.

[0086] Driven by data flow, the entire method process completes the closed-loop management of the entire process from data collection, preprocessing, multimodal fusion, decision optimization, anomaly detection to continuous learning. Through rigorous algorithm design and logical construction, the effectiveness and intelligence of the system in the high-risk environment of nuclear power plants are guaranteed.

[0087] in accordance with Figure 3Sequence diagram,In the system initialization phase, the data acquisition module is activated and started, and multimodal sensor nodes, including vibration sensors, temperature and humidity sensors, and radiation detectors, are deployed in the inspection area of ​​the nuclear power plant. The system uses a dynamic sampling algorithm based on fuzzy logic to adjust the sampling frequency to ensure the timeliness and accuracy of the data. The core of the fuzzy logic system lies in the membership function and fuzzy reasoning process. The membership function is a way to map environmental parameters (such as temperature, humidity, and radiation level) to a fuzzy set. Through the membership function, the environmental conditions are converted into fuzzy values ​​to characterize the degree of influence of the environment on the operation of the system. For example, when the temperature and radiation level are high, the fuzzy value generated by the membership function is also high, which means that the sampling frequency needs to be increased. Fuzzy reasoning combines these fuzzy conditions by integrating different membership values ​​to calculate a final sampling frequency. The system sets priorities according to the fuzzy rule base and dynamically adjusts the sampling frequency to ensure that the data can be updated in real time when the environment changes suddenly.

[0088] In the data preprocessing stage, the system applies wavelet denoising and Kalman filtering techniques to process the signal. Wavelet denoising decomposes the signal into components of different frequency bands, and suppresses high-frequency noise through soft threshold processing to retain the main signal components. Specifically, wavelet decomposition generates multiple wavelet coefficients of different frequencies. For each high-frequency component, the system applies soft thresholding to reduce the impact of noise and retain the core characteristics of the signal. The denoised signal is further input into the Kalman filter module, which performs signal smoothing based on the prediction and update process. Kalman filtering uses a state transition model to estimate the current state of the signal in the prediction stage; in the update stage, the prediction result is adjusted by introducing real-time measurement data. The Kalman gain is the key, which is used to determine the weight balance between the predicted value and the real-time measurement value, and dynamically adjust the output result to generate a smooth signal.

[0089] In this system, the core role of the federated learning framework is to achieve the synchronous update of model weights in a distributed environment. Its working principle is: each distributed node has a localized model, which is trained in its own environment to generate specific model parameters. Federated learning aggregates the local model weights of these distributed nodes on the central server to generate a global model. Specifically, during the training process of each node, the system will not directly transmit sensitive raw data, but will ensure privacy and security by uploading the trained weight parameters to the central server. After receiving the weights of each node, the central server uses the weighted averaging method (Weighted Averaging) to calculate the update of the global model, so that the model can adapt to different inspection environments. The formula for weight aggregation is as follows: For the weight parameter of the i-th node is W i , the training data size of the node is n i, the update formula of global weight is W = Σ(n i / N)*W i , where N is the sum of data of all nodes. This structure enables the system to achieve global optimization of the model without centralizing data, effectively coping with the challenge of uneven data distribution.

[0090] In the data fusion module, Bayesian networks and deep neural networks (DNNs) form the data processing skeleton of the system. The Bayesian network adopts the structure of a directed acyclic graph (DAG), where nodes represent different data sources (such as vibration sensors and temperature and humidity sensors), and edges represent the causal relationship between these data sources. Specifically, the Bayesian network represents the correlation between each sensor data through the conditional probability P(X|Y). For example, if there is a significant correlation between vibration data X and temperature and humidity data Y, the value of P(X|Y) will be higher. During the reasoning process of the Bayesian network, the system uses the Markov Chain Monte Carlo (MCMC) method to generate samples, and repeatedly samples to approximate the conditional probability distribution, and the generated conditional probability results are used as preliminary fusion features.

[0091] The deep neural network (DNN) structure consists of a convolutional neural network (CNN) and a recurrent neural network (RNN), which are used to extract features of different modal data. The convolutional neural network (CNN) part uses a 3x3 convolution kernel to extract the spatial features of vibration data through layer-by-layer convolution and pooling operations. Each layer of convolution extracts local features of the input data and passes low-level features to high-level layers to generate feature maps. The pooling layer is used to reduce the data dimension, retain important features, and reduce the amount of calculation. The recurrent neural network (RNN) part is mainly used to process time series data, such as temperature, humidity, and radiation data. Its structure consists of a series of hidden units, and the output of each unit depends on the state of the previous moment. Through the loop mechanism, the recurrent neural network (RNN) can capture the long-term dependencies of time series and effectively identify the trend of temperature, humidity, or radiation changes over time.

[0092] These two structures work together to achieve comprehensive analysis of multimodal data through the conditional probability integration of the Bayesian network and the multi-layer feature extraction of the DNN. The Bayesian network provides statistical associations during the integration process, and the DNN further refines and strengthens the feature expression, providing support for subsequent decision optimization.

[0093] exist Figure 4 In the class diagram, each technical module is defined as an independent class and is interconnected through attributes, methods and data flow to achieve the complete functions of the nuclear power plant equipment inspection system. The following is a detailed description of how the various classes interact through data flow and method calls:

[0094] As the core module of data acquisition, the DataAcquisition class is responsible for collecting data from sensors, including multimodal data such as vibration, temperature and humidity, and radiation. The properties in the DataAcquisition class, such as sampling frequency and data type, define the sampling mode of each sensor. The setSamplingFrequency() method of this class dynamically adjusts the sampling frequency according to the environmental conditions of the sensor, ensuring that the sampling frequency is increased when the environment changes or the device is abnormal, thereby improving the timeliness and accuracy of data acquisition. The data collected by the DataAcquisition class will be passed to the DataPreprocessing class.

[0095] After receiving the data from the DataAcquisition class, the DataPreprocessing class begins to preprocess the data, mainly including denoising and smoothing. This class contains properties such as threshold and noise standard deviation to ensure the accuracy of preprocessing operations. The preprocess() method is called when the data flows through the DataPreprocessing class to perform wavelet denoising and Kalman filtering. Wavelet denoising suppresses noise through multi-layer decomposition and soft threshold processing, while Kalman filtering smoothes data based on state prediction and update operations. The processed data will be sent to the DataFusion class for multimodal fusion.

[0096] The DataFusion class plays a key role in the data fusion stage. This class integrates the characteristics of Bayesian networks and deep neural networks (DNNs), and calls the feature extraction capabilities of Bayesian networks and DNNs through the BayesianNetwork() and DeepNeuralNetwork() methods respectively to generate high-dimensional fused feature vectors. The BayesianNetwork() method is called after the data enters the DataFusion class to calculate the conditional probability of each modal data and generate preliminary fused features. Next, the DeepNeuralNetwork() method extracts the spatial and temporal series features of the data, and combines the preliminary fused features to generate the final fused feature vector. This fused feature vector will be passed to the DecisionModule class for task decision making.

[0097] The DecisionModule class is the core of adaptive decision making. Its optimizeDecision() method uses an algorithm that combines decision trees and Q-learning to optimize task priorities based on real-time feedback from inspection tasks. The optimizeDecision() method calls the DecisionTree and Q-learning modules to update the Q value based on the current system status and device status feedback, thereby dynamically adjusting the path and priority of the inspection task. The update of task priority is passed to other modules of the system through feedback to achieve adaptive adjustment. The decision information output by the DecisionModule class is sent to the AnomalyDetection class for further anomaly detection.

[0098] After receiving the output of the decision module, the AnomalyDetection class detects anomalies and predicts faults through the detectAnomaly() and predictFailure() methods. The detectAnomaly() method uses the isolation forest algorithm to perform hierarchical anomaly detection on the data and identifies the outliers of the data by calculating the path length. If an anomaly is detected, the predictFailure() method will further call the long short-term memory network (LSTM) model to analyze the trend of the abnormal data points to predict the probability of future faults. If the fault prediction result exceeds the set threshold, a warning signal will be triggered and fed back to the FeedbackLearning class for continuous learning.

[0099] The FeedbackLearning class receives feedback from the anomaly detection module and continuously optimizes the system using the incrementalUpdate() and federatedLearning() methods. The incrementalUpdate() method dynamically updates the model weights based on the incremental learning mechanism to improve the adaptability of the system. The federatedLearning() method synchronously updates the model weights between multiple distributed nodes to ensure that the data and models between subsystems remain consistent. Through the dual mechanisms of incremental update and federated learning, the FeedbackLearning class ultimately achieves real-time adjustment of various algorithm parameters and model weights, and reapplies the optimized model to the data collection, preprocessing and fusion processes to form a closed-loop feedback loop to ensure the efficiency and accuracy of the system in a dynamic environment.

[0100] In this system architecture, data flows from the DataAcquisition class to the DataPreprocessing class, and then passes through the DataFusion class to generate multimodal fusion features. It is then optimized by the DecisionModule class and anomaly analyzed in the AnomalyDetection class, and finally feedback and continuous learning are performed through the FeedbackLearning class.

Claims

1. A nuclear power plant equipment inspection data transmission management system, characterized in that: The system includes: Multimodal data acquisition module: The module includes multiple sensor nodes, which collect the status data of nuclear power plant equipment in real time. The sensor nodes include vibration sensors, temperature and humidity sensors, and radiation detectors. The module is equipped with a dynamic sampling algorithm, which automatically adjusts the sampling frequency when an abnormality occurs in the environment of the equipment area. Data preprocessing module: The module uses wavelet denoising and Kalman filtering to process multimodal data, where wavelet denoising includes multi-layer decomposition, soft threshold processing and reconstruction steps. The threshold setting formula is: Among them, λ is the threshold, which is used for soft threshold processing in wavelet denoising; σ represents the standard deviation of the noise, which is estimated by the known noise characteristics; N is the signal length, which represents the total number of signal samples; The Kalman filter uses the state transfer matrix A and the noise covariance matrix R to predict the update step to achieve smoothing of the device state data; Multimodal data fusion module: The module includes a Bayesian network and a deep neural network. The Bayesian network uses the conditional probability P(x1, x2, ..., x n ) to generate preliminary fusion features; the deep neural network includes a convolutional neural network (CNN) and a recurrent neural network (RNN), which respectively extract the spatial features of the vibration data and the time series features of the temperature, humidity and radiation data to generate a high-dimensional fusion feature vector; Adaptive decision module: The module includes an adaptive decision tree and a reinforcement learning module. The adaptive decision tree dynamically adjusts the path structure based on the device status. The pruning algorithm removes redundant nodes when the status returns to normal to reduce the computational complexity. The reinforcement learning module updates the Q value in real time based on the device status feedback through the Q-learning algorithm. The learning rate α controls the update rate of the Q value. The discount factor γ balances the current reward and future reward to optimize the priority of the inspection task. Anomaly detection and fault prediction module: The module includes an isolation forest algorithm and a long short-term memory network (LSTM). The isolation forest algorithm uses multiple random decision trees to perform hierarchical anomaly detection on multimodal data, calculates the path length of data points to identify anomalies, and points with short path lengths are considered high outliers; the long short-term memory network (LSTM) model analyzes the trend of time series data, generates prediction results based on the combination of forget gates, input gates, and output gates, and issues a warning signal when the predicted value exceeds the fault threshold; Feedback and continuous learning module: Adopts incremental learning mechanism to dynamically adjust the weight parameters of Bayesian network and deep neural network.

2. The nuclear power plant equipment inspection data transmission management system according to claim 1, characterized in that: The dynamic sampling algorithm of the multimodal data acquisition module combines priority scheduling and fuzzy logic control, and uses a fuzzy inference system to determine the sampling frequency f(x). Its fuzzy rule base includes a combination of environmental parameters E and device status S. The calculation formula of the sampling frequency f(x) is: Among them, w i is the fuzzy rule weight, μ i (E, S) represents the fuzzy membership function under the environment and equipment status, and this combination is used to dynamically adjust the data collection frequency.

3. The nuclear power plant equipment inspection data transmission management system according to claim 1, characterized in that: The wavelet denoising in the data preprocessing module is combined with the Kalman filter to decompose the original signal x(t) into multiple layers of wavelet coefficients (d1, d2, ..., d n ), perform soft threshold processing on each high-frequency component d, and the threshold is set as: Among them, λ is the setting parameter of the soft threshold, which is used to suppress high-frequency noise in the wavelet denoising process; σ represents the standard deviation of the noise, which is obtained by estimating the noise level in the signal; N is the signal length, which represents the number of signal samples; the formula dynamically adjusts the threshold size based on the statistical threshold theory in wavelet denoising; At the same time, the prediction and update steps of Kalman filtering are: Among them, the prediction formula is: is the predicted state, which is the estimated value of the state at time k-1 and the control variable uk; A is the state transfer matrix, which describes the process of the system state transferring from time k-1 to k; B is the control input matrix, which is used to adjust the control variable u k Impact on system status; k is the control quantity at time k; Among them, the Kalman gain formula is: K k is the Kalman gain matrix, P is the error covariance matrix, and this method is used to achieve joint smoothing of time domain and frequency domain; H is the observation matrix, which is used to map the state space to the observation space; H T is the transpose of the observation matrix; R is the observation noise covariance matrix, which indicates the degree of observation noise.

4. The nuclear power plant equipment inspection data transmission management system according to claim 1, characterized in that: The Bayesian network in the multimodal data fusion module combines the Markov chain Monte Carlo (MCMC) method to perform approximate sampling on the conditional probability P(xi|xi-1) to generate a fusion feature vector F: Among them, the fused feature vector F is generated by approximate sampling of conditional probability; P(xi|xi-1) represents the i-th data x i Given its previous data x i-1 The probability under the condition reflects the dependency relationship between data; n is the total number of data, indicating the number of features contained in multimodal data; The Markov chain Monte Carlo (MCMC) generates conditional distribution approximation through sampling and acceptance-rejection algorithm. The deep neural network extracts spatial features H by a convolutional neural network (CNN) with a convolution kernel size of 3x3. The hidden state update formula of the recurrent neural network (RNN) is: h t =σ(W xh x t +W hh h t-1 +b h ) Among them, σ is the activation function, which is used to introduce nonlinear relationships; W xh and W hh is the weight matrix, b h is a bias term, and a high-dimensional fusion feature vector is generated through this structure; h t and h t-1 is the hidden state at time t; x t The input data at the current moment.

5. The nuclear power plant equipment inspection data transmission management system according to claim 1, characterized in that: The decision tree algorithm of the adaptive decision module combines Q-learning and the multi-armed bandit algorithm (MAB) to update the priority by the Q value under the device state s. The Q value update formula is: Among them, Q(s,a): represents the Q value corresponding to the action a selected in the current state s, that is, the value evaluation of the action; α is the learning rate, which is used to control the speed of Q value update, and its value range is 0≤α≤10. The larger the value, the faster the update speed; R is the immediate reward, which represents the immediate return brought by the current action a; γ is the discount factor, which ranges from 0≤γ≤10 and is used to balance the importance of current rewards and future rewards. The larger the value, the more the system attaches importance to future rewards; max a′ Q(s′,a′) represents the maximum Q value among all possible actions a′ in the next state s′, reflecting the system’s estimate of the optimal future benefit; The multi-armed bandit algorithm uses the ε-greedy strategy to select the optimal path and dynamically adjusts the exploration probability ε to ensure that the optimal action is selected under multiple inspection task states.

6. The nuclear power plant equipment inspection data transmission management system according to claim 1, characterized in that: The anomaly detection and fault prediction module performs outlier detection on the path length L(x) of the data point through the isolation forest algorithm. The path length formula is: Where L(x) is the average path length of data point x, which evaluates the degree of abnormality of the data point in the isolation forest algorithm; T represents the total number of trees in the isolation forest, which is defined by the user to control the size of the forest and the stability of anomaly detection; path_length(x,T i ): is the data point x in the i-th random tree T i The path length in is the number of nodes required to go from the root node to the leaf node where the data point is located; The LSTM model further predicts the trend of abnormal points. The unit state update formula of the LSTM is: h t =o t ·tanh(c t ) Among them, the state update formula is: ct represents the cell state at the current time t, which contains the long-term memory of the time series; ft is the output of the forget gate, which is used to control the cell state c at the previous moment t-1 How much information needs to be retained, the value range is [0, 1]; i t The output of the input gate is used to control how much of the new input information needs to be added to the current cell state; is the candidate state, which is calculated by the input data through a nonlinear activation function; Among them, the output state formula is: h t is the output state at the current time t; o t is the output of the output gate; tanh(c t ) is the current cell state c t Nonlinear transformation.

7. The nuclear power plant equipment inspection data transmission management system according to claim 1, characterized in that: The feedback and continuous learning module updates the weights of the Bayesian network and the deep neural network through an incremental learning mechanism. The incremental learning formula is: IN new =In old +η ΔW Among them, η is the learning rate, ΔW is the weight adjustment amount, and W is the weight; the federated learning framework is used to realize the distributed aggregation of each node model, and the updated model weights are synchronized to each subsystem to improve the generalization ability of the model in different environments.

8. A method for managing the transmission of nuclear power plant equipment inspection data, applied to the nuclear power plant equipment inspection data transmission management system according to any one of claims 1 to 7, characterized in that: include: S01, multimodal data acquisition: The device status data is collected in real time through the multimodal data acquisition module. The sensor nodes include vibration sensors, temperature and humidity sensors, and radiation detectors. The collected data is prioritized through the fuzzy logic control algorithm to dynamically adjust the sampling frequency f(x). The environmental parameters E and the device status S determine the high-priority data stream through the membership function μ(E, S) in the fuzzy rule base to ensure the real-time and accuracy of high-risk data. The collected priority data stream is passed to the data preprocessing module. S02, data preprocessing: denoising and smoothing the collected data stream; The wavelet denoising algorithm decomposes the original signal into different frequency components through multi-scale wavelet transform and uses soft threshold Suppress high frequency noise; The denoised signal is further processed by Kalman filtering. The prediction and update process of Kalman filtering is based on the Bayesian estimation principle, where the state prediction formula is: In this step, the denoised and smoothed data provides higher quality input for the multimodal data fusion module, ensuring the accuracy of feature extraction; S03, multimodal data fusion: the preprocessed data is input into the Bayesian network and deep neural network (DNN) for multimodal data fusion; the Bayesian network uses the Markov chain Monte Carlo (MCMC) method to approximate the conditional probability distribution P(x i |x i-1 ), generate a preliminary fusion feature vector F, which is input into a deep neural network, where a convolutional neural network (CNN) is used to extract spatial features and a recurrent neural network (RNN) is used to analyze time series features. The generated high-dimensional fusion feature vector is passed to an adaptive decision-making module for intelligent optimization of inspection tasks; S04, Adaptive decision-making and task optimization: Based on the fusion features, the adaptive decision tree and reinforcement learning modules optimize the priority of the inspection tasks; Adaptive decision trees dynamically remove redundant nodes based on pruning algorithms when the state returns to normal to reduce computational complexity; The reinforcement learning module uses the Q-learning algorithm to update the inspection task priority, where the update of the Q value follows the Bellman equation: Q(s,a)=Q(s,a)+α[R+γmax a′ Q(s′,a′)-Q(s,a)] Among them, α is the learning rate, γ is the discount factor, and the priority information of the task path is dynamically adjusted and passed to the anomaly detection module. a′ Q(s′,a′) represents the maximum Q value of all possible actions a′ in the next state s′, reflecting the optimal choice for the next step; R: is the immediate reward; Q(s,a) represents the Q value of the current state s and action a, reflecting the total reward that can be obtained after executing the action; S05, anomaly detection and fault prediction: anomaly detection and fault prediction of inspection data are performed through the isolation forest algorithm and the long short-term memory network (LSTM) model; the isolation forest algorithm calculates the path length L(x) of the data point through a randomly generated binary tree structure, and the point with a shorter path length is determined to be an abnormal data point; the detected abnormal data point is then passed to the long short-term memory network (LSTM) model, and the fault is predicted through time series trend analysis; The long short-term memory network (LSTM) model adjusts the state update based on the gating mechanism (forget gate, input gate, output gate), predicts that when the abnormal trend exceeds the fault threshold, it triggers an early warning signal and feeds back to the control center; S06, dynamic feedback and continuous learning: The system continuously updates the weights in the Bayesian network and deep neural network based on the inspection feedback data; the incremental learning mechanism uses the learning rate η to control the amplitude of the weight update, and the federated learning framework synchronizes the model weights between distributed nodes to improve the adaptability of the model; the updated model uses a closed-loop feedback mechanism to reuse for data collection, preprocessing and fusion.

Images