Self-attention-based nuclear power plant equipment predictive evaluation maintenance method and system

Abstract

The application relates to the technical field of equipment life prediction, and provides a nuclear power plant equipment predictive evaluation maintenance method and system based on self-attention, which comprises the following steps: obtaining multi-source sensing data of equipment by adopting a sliding window mechanism, and obtaining an input time window sequence after preprocessing; performing reasoning on the time window sequence data by a model to obtain a state vector for regression analysis, and generating equipment residual service life; performing visual processing on an attention matrix to obtain an equipment health state interpretation atlas of attention weights, and generating interpretation information; and fusing the obtained equipment residual service life and corresponding interpretation information to generate equipment maintenance strategy suggestions. The application innovatively fuses a sliding window and a Transformer model, realizes deep modeling and explainability analysis of multi-source equipment data, and constructs an equipment health state atlas based on an attention mechanism, which is used for accurate prediction and maintenance strategy generation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment life prediction, in particular to a nuclear power plant equipment predictive evaluation maintenance method and system based on self-attention. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] In high-safety-level industrial scenarios such as nuclear power plants, the health status of key equipment is directly related to the safety, reliability and economy of the system. The remaining useful life (RUL) prediction technology of equipment, as an important means to realize intelligent maintenance and predictive maintenance, can significantly improve the efficiency of equipment operation and maintenance, reduce the risk of unplanned downtime, and promote the strategic transformation from traditional post-maintenance to predictive maintenance. There are special challenges in data collection and processing in nuclear power plants: many monitoring points, long data transmission link and high collection frequency, resulting in large data volume, high real-time requirement and strict processing accuracy requirement, which puts higher requirements on the deployment architecture and response efficiency of the RUL prediction technology.

[0004] The current mainstream remaining useful life (RUL) prediction methods mainly include physical modeling, statistical analysis and deep learning based paths. Among them, the physical modeling method relies on the mathematical description of the degradation mechanism of the equipment, which is often difficult to cover the multi-source coupling and strong nonlinear degradation behavior in complex systems, and has limited applicability and scalability; the statistical analysis method shows insufficient modeling capability when facing high-dimensional variables, multi-condition disturbance and nonlinear degradation trend, and the prediction accuracy is unstable; while the deep learning model (such as LSTM) has made progress in sequence modeling, but still has weak modeling ability for long-term time series dependence, is prone to gradient disappearance in the training process, and has poor result interpretability, which is difficult to meet the transparency and trustworthiness requirements of high-safety-level industrial systems. In addition, most of the existing RUL prediction schemes assume continuous data, clear structure and stable processing environment, which is difficult to adapt to the industrial characteristics of long-distance sensor deployment, unstable data link, high-frequency update of massive time series data and other industrial characteristics in nuclear power plants and other practical application scenarios, and lack the deployment ability and processing efficiency of coordination with the distributed control system (Distributed Control System, DCS). SUMMARY

[0005] The application proposes a nuclear power plant equipment predictive evaluation maintenance method and system based on self-attention to solve the above problems, innovatively combines sliding window and Transformer model, realizes deep modeling and explainability analysis of multi-source equipment data, and constructs an equipment health state atlas based on an attention mechanism to realize equipment life prediction and maintenance.

[0006] In order to achieve the above purpose, the application adopts the following technical solutions:

[0007] One or more embodiments provide a nuclear power plant equipment predictive evaluation maintenance method based on self-attention, comprising the following steps:

[0008] A sliding window mechanism is used to obtain multi-source sensing data of the nuclear power plant equipment, and the preprocessed data is constructed into an input time window sequence;

[0009] The time window sequence data is inferred through a Transformer model to obtain a state vector for regression analysis, and the remaining useful life of the equipment is generated;

[0010] The attention matrix obtained by the attention mechanism of the Transformer model in the inference process is visualized to obtain an equipment health state explanation atlas of attention weights; and a dynamic knowledge graph is constructed to generate explanation information for abnormal deviation points in the equipment health state explanation atlas;

[0011] The obtained remaining useful life of the equipment and the corresponding explanation information are fused to generate equipment maintenance strategy suggestions.

[0012] One or more embodiments provide a nuclear power plant equipment predictive evaluation maintenance system based on self-attention, comprising:

[0013] A sliding window unit configured to use a sliding window mechanism to obtain multi-source sensing data of the nuclear power plant equipment, and to construct an input time window sequence after preprocessing;

[0014] An inference unit configured to infer time window sequence data through a Transformer model to obtain a state vector for regression analysis, and to generate the remaining useful life of the equipment;

[0015] An explanation unit configured to visualize the attention matrix obtained by the attention mechanism of the Transformer model in the inference process to obtain an equipment health state explanation atlas of attention weights; and to construct a dynamic knowledge graph to generate explanation information for abnormal deviation points in the equipment health state explanation atlas;

[0016] A fusion unit configured to fuse the obtained remaining useful life of the equipment and the corresponding explanation information to generate equipment maintenance strategy suggestions.

[0017] Compared with the prior art, the present application has the following advantages:

[0018] The present application overcomes the problems of insufficient modeling capability, poor nonlinear feature processing and poor interpretability in traditional RUL prediction technology, and improves the prediction accuracy and reliability. By deeply integrating the dynamic sliding window mechanism and the Transformer model, the structural stability and time consistency management of multi-dimensional time series input data are realized, the perception ability of the model to long sequence and multi-variable equipment degradation trend is significantly improved, and the prediction stability and robustness are enhanced. The attention matrix visualization technology is introduced to realize the transparency of the prediction process, the system can automatically identify the key time period and key feature channel in the input data, realize the self-explaining modeling of the equipment abnormal state and its evolution mode, and enhance the transparency and expert understanding of the model output.

[0019] The advantages of the present application and the advantages of the additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0020] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification. The embodiments of these drawings are set forth to explain the present application and are not limiting of the present application. In the drawings:

[0021] Figure 1 is a flow chart block diagram of the equipment predictive maintenance and life assessment method of embodiment 1 of the present application;

[0022] Figure 2 is a schematic diagram of the edge-center joint modeling mechanism in the equipment predictive maintenance and life assessment method of embodiment 1 of the present application;

[0023] Figure 3 is a schematic diagram of the inference process based on the Transformer model of embodiment 1 of the present application;

[0024] Figure 4 is a flow chart of the equipment health state interpretation atlas generation method of embodiment 1 of the present application. DETAILED DESCRIPTION

[0025] The present application will be further described below in conjunction with the drawings and embodiments.

[0026] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0027] It is to be noted that the terms used herein are merely for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise, and it should be further understood that the terms "comprise" and / or "include" when used in this specification, specify the presence of stated features, steps, operations, devices, components and / or combinations thereof. It should be noted that the various embodiments and features in the present application can be combined with each other without conflict, and the embodiments will be described in detail below with reference to the accompanying drawings.

[0028] Embodiment 1

[0029] In the technical solutions disclosed in one or more embodiments, as shown in Figures 1 to 4 The self-attention-based nuclear power plant equipment predictive evaluation maintenance method includes the following steps:

[0030] Step 1: Obtain multi-source sensor data of the nuclear power plant equipment by using a sliding window mechanism, and construct an input time window sequence after preprocessing;

[0031] Step 2: Inference the time window sequence data through a Transformer model to obtain a state vector for regression analysis, and generate the remaining useful life of the equipment;

[0032] Step 3: Visualize the attention matrix obtained in the inference process of the attention mechanism of the Transformer model to obtain an equipment health state interpretation atlas of attention weights; and construct a dynamic knowledge graph to generate explanation information for abnormal deviation points in the equipment health state interpretation atlas;

[0033] Step 4: Fuse the obtained remaining useful life of the equipment and the corresponding explanation information to generate a device maintenance strategy suggestion;

[0034] In this embodiment, the problems of insufficient modeling capability, poor nonlinear feature processing and poor interpretability in traditional RUL prediction technology are overcome, and the prediction accuracy and reliability are improved. By deeply integrating the dynamic sliding window mechanism with the Transformer model, the structure stability and time consistency management of multi-dimensional time series input data are realized, the perception ability of the model to the degradation trend of long sequence and multi-variable equipment is significantly improved, and the prediction stability and robustness are enhanced. The attention matrix visualization technology is introduced to realize the transparency of the prediction process, the system can automatically identify the key time period and key feature channel in the input data, realize self-explaining modeling of the abnormal state and its evolution mode of the equipment, and enhance the transparency and expert interpretability of the model output. Further, by constructing a device health state explanation atlas integrating trend curve, abnormal frequency and risk level distribution, the current state and historical evolution path of the equipment are presented comprehensively, and the auditability and decision support capability in engineering scenarios are enhanced. Combined with the RUL prediction result and the explanation information, a maintenance strategy is formulated to realize the refinement and intelligentization of predictive maintenance, which is especially suitable for high safety level industrial scenarios such as nuclear power plants with high frequency, numerous equipment and complex state, and improves the operation efficiency and safety of the system.

[0035] In step 1, a real-time sliding input implementation method for equipment remaining useful life (RUL) prediction is provided, which is suitable for the inference input construction of Transformer type model, especially suitable for high frequency, remote and unstable data collection environment of nuclear power plant, and has the characteristics of strong real-time, high anti-interference ability and good prediction continuity.

[0036] Multi-source sensor data is collected by field sensors, including real-time collection of multi-dimensional operating parameters from nuclear power plant equipment, including temperature, vibration, current, flow, pressure and other operating parameter data, which are transmitted to edge computing devices through industrial communication protocols (such as OPC UA, Modbus);

[0037] Nuclear power plant equipment, including but not limited to main pump, cooling system, generator, etc.

[0038] The data processing in step 1 is realized in the edge computing device, which is deployed in the DCS node or industrial control rack of the nuclear power plant, has high performance data processing and low delay response ability, and is responsible for receiving, caching and preprocessing field sensor data to ensure data integrity and real-time performance.

[0039] In the DCS edge side of the nuclear power plant, the traditional fixed-beat FIFO sliding window will mix different working conditions such as start-stop, power-up, steady state and SCRAM in the same window when processing sensor data stream, resulting in working condition aliasing and feature drift, thereby reducing the discrimination and robustness of abnormal detection and remaining useful life (RUL) modeling.

[0040] To solve the above problems, the embodiment proposes an event alignment sliding window construction method, the core of which is to introduce working condition signal recognition events as window division trigger boundaries to realize dynamic alignment and sampling of windows.

[0041] Specifically, the multi-source sensor data of the equipment is obtained by using the sliding window mechanism, which can be configured to be implemented in the edge device, as shown in Figure 2 The method comprises the following steps:

[0042] Step 11, synchronously listening to the received multi-source sensor data and DCS working condition signals and interlocking signals to identify events;

[0043] The events identified by the DCS working condition signals and interlocking signals include device or system start-stop, load increase-decrease, loop or valve position switching, interlocking action and protection action, etc.

[0044] Specifically, after the DCS working condition signals and interlocking signals are timestamped, de-bounced and merged, the identified events are recorded as .

[0045] Among them, is the event timestamp; is the event type, including start-stop, power increase / decrease, loop switching, interlocking and protection action, etc.

[0046] Step 12, according to the event occurrence point identified, the time period set before and after the event occurrence point is set as a transition window according to the event type, and the time period set after the event occurrence is divided into a steady state window, the transition window is smaller than the steady state window, the sizes of the transition window and the steady state window are set to be different according to the event type, and the time period without events is set as a fixed window as a time window, to obtain a dynamically changing time window;

[0047] Specifically, when an event is identified, two types of windows are generated according to the event alignment:

[0048] Transition window: Resample all channel process quantities at a higher sampling rate to capture rapid changes before and after the event;

[0049] Steady state window: Collect at a lower sampling rate , and statistical quantities such as mean, variance and spectral energy can be extracted to depict the stable trend after the event.

[0050] The non-event interval generates a regular window by sliding at a set beat to ensure continuous inference.

[0051] Among them: Pre-event observation duration, Post-event transition observation duration, Stable delay duration, Indicates the length of the steady state window.

[0052] Before an event occurs, such as main pump start-stop, valve switching, power-up instruction, trace back to seconds of data, and include this data in the transition window. This can capture early signs or slow drift before the event triggers; for example, 30 seconds of vibration and temperature trends before the event triggers. seconds of data after the event triggers to form a complete transition window. This part can record the rapid response characteristics of the device after the event occurs; for example, the transient changes in pressure and flow within 10 seconds after valve switching.

[0053] After an event occurs, the steady state is not immediately started, but a delay of seconds is waited to ensure that the system response is complete and the key quantities enter the relatively stable interval. This delay avoids mixing strong transient fluctuations into the steady state analysis. For example, after the main pump speed-up for 60 seconds, the current and pressure fluctuations gradually subside, and at this time it is considered to enter the steady state. After the system enters the stable stage, the length of the data collection time is The window is mainly used to extract statistical features such as mean, variance, and spectral energy to reflect the long-term running trend of the device after the event. For example, 5 minutes of temperature and vibration data after entering the steady state.

[0054] Step 13, based on the dynamic update of the time window, the obtained multi-source sensor data is divided to obtain a plurality of time window data, and each time window data is pre-processed as an input time window sequence.

[0055] Further, the method for pre-processing the input time window sequence includes the following steps:

[0056] Step 101, for the obtained time window data, align the data of different source sensors according to the time window, mask mark and interpolate the missing values in the data of each time window, so that the data length is consistent.

[0057] The data of this embodiment is calculated at the same time as the data is received, and the data is divided according to the time window. After division, the time alignment is performed through this step;

[0058] Step 1011, align the time stamp of each sensor data with the grid of the time window, and unify the data grabbed by different sampling frequencies to one time window. If the time difference between data points is within the allowable error, it is considered to be continuous, and the discontinuous is interpolated and completed by mask, so that the data length is consistent.

[0059] Specifically, the mask data can be estimated by linear interpolation or Kalman update to complete the missing data.

[0060] Step 1012, for each time window data, check whether the various types of sensing data are missing field by field;

[0061] Specifically, it can be judged whether temperature, vibration, current, pressure, etc. are missing in the event window, and if missing, the mask interpolation is used to complete it.

[0062] Step 102, for each time window, the data of each data point is constructed into a triple data including metadata, mask mark and confidence, and a multivariate sensor state vector of each time window is obtained.

[0063] Specifically, all data points are constructed into triple input ; wherein, is a numerical value, may be original data or mask data; is a mask mark, and 0 is set for missing or interpolation; is the confidence;

[0064] The confidence is determined as follows: the original data is set to high confidence; the interpolation data is set to medium confidence; and the estimated numerical value is set to low confidence, and is written into the triple data. The transition window prohibits cross-event backfilling to avoid introducing incorrect context.

[0065] Further, the obtained triple data can be updated and managed by window: each window buffer is updated according to the first-in-first-out (FIFO) principle: for each active window maintain a circular buffer. When a new sample arrives, generate and then press it in to ensure that the feature matrix dimension and time sequence are consistent.

[0066] Step 103, for the triple data of each time window, the mask mark and confidence of each data point in the window are encoded, the action mode, event type and time position corresponding to each time window data are encoded, and after encoding, the metadata of the triple data are spliced to obtain the Token of each time window as the input time window sequence.

[0067] In the above embodiment, after the dynamic update of the sliding window, the time alignment and the data completion are completed, the multivariate sensor state vector of each time is further enhanced and encoded to form a unified model input sequence.

[0068] Step 1031, the metadata in the triple data is normalized.

[0069] Specifically, at time , the sensor state vector contains multiple physical measurement point values such as temperature, vibration, current, pressure, etc.; the sensor state vector is normalized to form a standardized vector , ensuring that data of different dimensions are comparable;

[0070] Optionally, the normalization of the sensor state vector can use Min-Max or Z-score, etc.

[0071] Step 1032, extract the mask label and confidence data in the triple data, calculate the weighted value as the first encoding feature ;

[0072] Specifically, the missing indication and the confidence weight are concatenated into an enhanced token (Token) through a function for dynamically adjusting the attention of each channel in the model inference process. The function is used to fuse the missing indication and the confidence into part of the enhanced Token, represented as:

[0073] ;

[0074] wherein is a learnable mapping matrix.

[0075] Step 1033, identify the type of time window, encode the time window type to obtain the second encoding feature ;

[0076] In this embodiment, the time window type includes a transition window, a steady state window, and a no event window;

[0077] Step 1034, identify the event type corresponding to the time window, and encode to obtain a time type sub-vector ;

[0078] This step encodes the trigger event category corresponding to the time step, such as load rise, valve switching, etc., to enhance the model's ability to recognize event-driven degradation characteristics.

[0079] Step 1035, encode the timestamp position of the data to obtain a time position encoding sub-vector ; so that the model can maintain the perception of the sequence relationship;

[0080] Step 1036, sequentially splice the normalized metadata with the first encoding feature, the second encoding feature, the time type sub-vector, and the time position encoding sub-vector to obtain an enhanced token, and combine each time window to obtain an input time window sequence;

[0081] Finally, the enhanced token formed at the time is defined as the input time window sequence:

[0082] ;

[0083] Wherein, the symbol represents a vector splicing operation. The structure of the input time window sequence ensures that the input sequence not only contains numerical information, but also introduces data quality, operation mode, event category and time position information, which can fully represent the operation state of the nuclear power plant equipment.

[0084] Through the above enhanced token construction steps, the embodiment realizes the unified expression of multi-source and multi-dimensional sensor data without relying on artificial setting rules, and provides a complete and interpretable input basis for the Q / K / V vector generation and attention calculation in the subsequent Transformer encoder.

[0085] Through the above improvements, in view of the reality that the nuclear power plant has multiple different working conditions, the window alignment and classification sampling mechanism under event triggering is realized, which can alleviate the working condition aliasing problem at the source, and strengthen the timing details in the transition state and the trend stability in the steady state, thereby significantly improving the distinguishability and robustness of nuclear power plant edge side anomaly detection and RUL prediction.

[0086] Step 2, input the time window sequence into the Transformer model for inference, perform multi-head self-attention calculation on the input time window sequence data, generate a predicted query vector (Q), an operation state feature vector (K), and a historical information carrier vector (V) through linear transformation, perform dot product matching, Softmax normalization and attention weighted fusion to generate a current device health state vector.

[0087] In step 2, the time window sequence data is inferred through the Transformer model, as shown in Figure 3 , including the following steps:

[0088] Step 21, input the time window sequence data into the attention mechanism module, and generate a query vector Query, a key vector Key and a value vector Value through three groups of linear projection weight matrices , and , respectively:​

[0089] ;

[0090] This step provides a vector basis for the attention mechanism, and completes the initial feature expression of the sequence information. This structure allows the model to automatically learn the implicit dependency between device states without manually setting state rules.

[0091] Step 22, calculate the attention weight of each time point by using the multi-head attention mechanism, and perform softmax function normalization to obtain the attention weight vector of each time point;

[0092] The correlation score is calculated by calculating the inner product of the query vector Query and the key vector KeyQuery:

[0093] ;

[0094] wherein, Queryj represents the query vector at the jth time point; Keyj represents the key vector at the jth time point; D represents the dimension;

[0095] After softmax function normalization, the attention weight vector of each time point is obtained :

[0096] ;

[0097] In the above embodiment, the dot product attention mechanism is used to explicitly model the state information between different time segments. The attention weight is obtained by calculating the similarity between Query and Key, and then the corresponding Value is weighted and fused to form the dynamic understanding representation of the current device state. This mechanism has explicit time weight output capability, which is the basis for ensuring the model interpretability and long-range dependency modeling capability; it adaptively learns the complex nonlinear relationship between device states from multi-dimensional time series data, significantly reducing the dependence on artificial construction of degradation indicators, setting thresholds or configuring rule logic, and improving the modeling automation degree and development efficiency.

[0098] Step 23, weight and fuse the attention weight vector and the value vector to obtain the state vector of the fused context information at the current time point :

[0099] ;

[0100] State representation features Capture the degradation trend features of the device throughout the cycle, i.e., the degradation information of the key device.

[0101] An alternative technical solution replaces the attention mechanism with a convolutional neural network (CNN) to extract features; the attention mechanism of the Transformer implements inter-time slice dependency modeling, and a convolutional neural network can also be used to extract features from time series and input into a regression layer to complete prediction.

[0102] Specifically, a one-dimensional convolution kernel is constructed to slide and extract local features from device historical state data, and a multi-layer convolution and pooling structure is used to model the degradation trend. This solution is particularly suitable for devices with short prediction periods and obvious feature changes, and has fast model training speed and convenient deployment.

[0103] This solution implicitly models time dependence in the convolution kernel weights, although it does not have explicit and interpretable attention output, but performs well in inference efficiency and small data scenarios, and has application feasibility.

[0104] Step 24, multi-head attention splicing and feature integration are performed, and the state vectors obtained by all h attention heads are spliced along the feature dimension to form global comprehensive features

[0105] ;

[0106] This splicing operation enables the model to simultaneously aggregate features of multiple sensor signals at different time scales and physical dimensions, enhancing the ability to perceive the state of complex nuclear power equipment.

[0107] Step 25, the spliced global comprehensive features are linearly mapped to align the dimensions and output a unified format feature vector for subsequent module processing.

[0108] Step 26, residual connection is performed to add the feature vector to the original input and perform normalization to obtain the global state vector :

[0109] ;

[0110] This operation preserves the direct path of the original device features, preventing information loss during deep model training and ensuring the integrity of the nuclear power plant operation data representation.

[0111] Layer normalization is performed on to balance the mean and variance across feature dimensions, improving the robustness and generalization ability of the model for high-dimensional complex nuclear signals, especially in multi-feature coupling scenarios. The formula is:

[0112] ​​ ;

[0113] The final output is the global state vector fused with the multi-head attention and the multi-dimensional operating features of the nuclear power plant, which contains both local degradation features and global health information, serving as the input of the subsequent feedforward network to extract complex degradation patterns and RUL prediction. After multi-head self-attention splicing, residual connection and normalization processing, the global state vector output by the Transformer model has already contained the comprehensive features of the multi-dimensional sensor signals of the nuclear power plant, such as the joint variation features of the main pump vibration, cooling water temperature and motor current.

[0114] In order to further extract the complex degradation patterns that may exist in the equipment, such as long-term vibration recession trend, temperature abnormal fluctuation and electrical abnormality accompanied by load change, the input feedforward neural network module.

[0115] Step 27, the obtained global state vector After the feedforward network feature extraction, full connection operation, activation operation, residual connection and layer normalization processing, the final state vector is obtained.

[0116] Specifically, the feedforward network includes two full connection layers and ReLU activation function layers, which have strong non-linear fitting capability and can identify degradation patterns and coupled features. The output of the feedforward network is subjected to residual connection and layer normalization again to form a stable feature representation.

[0117] Further, the RUL regression prediction, the extracted state vector is input into the regression prediction module, and the remaining useful life (RUL) prediction value of the equipment under the current state is output.

[0118] Optionally, the regression prediction module adopts a single-layer or multi-layer full connection network, and the formula is represented as:

[0119] ;

[0120] wherein, is a regression mapping function that can accurately predict the remaining useful life value according to the historical data and current features of the equipment after training.

[0121] In some embodiments, the uncertainty range of the prediction result can also be estimated by a Bayesian regression layer, an MC Dropout method or a double-head network structure, and the upper and lower confidence limits of the RUL, such as the 95% confidence interval, are output.

[0122] Specifically, a Bayesian regression layer is introduced, which can introduce a Bayesian regression structure in the output layer of the Transformer model. The Transformer model outputs the mean prediction value μ and the variance prediction value σ of the RUL at the same time 2 . In the training process, the loss function based on negative log likelihood is optimized to realize the adaptive modeling of the prediction confidence range. The confidence interval can be expressed as:

[0123] ;

[0124] This deformation scheme is suitable for systems with extremely high safety levels, and can assist operation and maintenance personnel in evaluating models and risk boundaries, improving decision-making robustness, and is especially suitable for high-safety-level scenarios such as nuclear power plants. The operation and maintenance strategy is more inclined to the conservative principle of early maintenance rather than risk omission.

[0125] In the above process, the complete process of multi-dimensional sensor data stream, real-time feature extraction, self-attention calculation, multi-head information fusion, nonlinear feature recognition, and remaining life prediction is realized.

[0126] Further technical solutions, in the training process of the Transformer model, use a data set containing complete degradation life records, with real RUL as the supervision label, and continuously optimize the model parameters by minimizing the loss function. The loss function includes a prediction loss term, a long-term error penalty term, and a confidence modeling term. The loss function formula is as follows:

[0127] ;

[0128] Wherein, represents the true life of the ith sample, i.e. the label; represents the prediction value of the ith sample; represents a dynamic weight coefficient; represents a confidence modeling term, represents an uncertainty loss coefficient, represents a control long-term error influence coefficient; represents the prediction value of the backward step and the label, used to calculate the penalty of the long-term prediction error;

[0129] The above loss function includes:

[0130] Prediction loss term : used to judge the accuracy of the prediction using weighted mean square error;

[0131] Confidence modeling term : used to encourage the model to explicitly estimate the prediction uncertainty and improve the adaptability to unknown states;

[0132] Long-term error penalty term : for suppressing the drift of the model in judging future trends, improving the robustness of prediction;

[0133] The Transformer model of the embodiment can be periodically iteratively trained to support online incremental learning and multi-device migration.

[0134] Compared with the traditional recurrent neural network (such as LSTM) which models the previous state in time sequence, the embodiment introduces the multi-head attention mechanism of the Transformer, so that the model can automatically learn the dependency relationship and attention weight between time points during the training process, thereby breaking the strict time sequence restriction and being able to directly establish a global context dependency within a long sequence, thereby improving the response capability to key state changes. In the actual industrial environment, the sensor signal is easily affected by noise, delay or data jitter. Through the dynamic attention mechanism, the attention weight of the data at different times can be adaptively adjusted during the inference stage, thereby having strong robustness and error suppression capability, and significantly improving the stability and reliability of the prediction output.

[0135] Since the attention mechanism generates an explicit attention matrix during the inference process, the embodiment can visualize the attention matrix to generate an attention weight atlas to generate explanation information, which is used to assist the operation and maintenance personnel to understand the basis of the model prediction, and enhance the trust and actual availability of the system in the nuclear power plant scene.

[0136] In the embodiment, the attention explanation mechanism in the nuclear power plant background specifically: using the attention matrix generated in the foregoing steps, extracting key sensor features and time periods related to RUL prediction, such as the time window corresponding to the change of the main pump vibration frequency, the duration of the abnormal cooling temperature, and outputting the explanation information together with the RUL prediction result. In this way, not only the device life prediction can be given, but also intuitive degradation mode explanation and decision basis can be provided for the operation and maintenance personnel.

[0137] Further, in step 3, the attention matrix obtained by the attention mechanism of the Transformer model during the inference process is visualized to obtain an equipment health state explanation atlas of attention weight, and the method for generating explanation information, as shown in Figure 4 , includes the following steps:

[0138] Step 31, obtaining the attention matrix output by the Transformer model, based on the combination of time steps and feature channels with high weight values in the attention matrix, reversely positioning the input data region with high attention in the Transformer model as a high attention region, that is, identifying the key working condition signals and the corresponding abnormal patterns on which the judgment result of the Transformer model is significantly dependent;

[0139] For example, if the "cooling temperature rise" is given a very high attention weight in a certain period of time, it means that the change of this variable plays a decisive role in the model prediction, which can be explained as a potential "abnormal signal" or "key working condition";

[0140] Further, as shown in Figure 2 The attention matrix can be uploaded, stored and processed through the set central platform;

[0141] Step 32, based on the feature channel of the high attention area and the time window, extract the abnormal features of the key equipment, identify the abnormal behavior patterns with physical meaning, extract the degradation features and their time sequence information to construct the degradation feature set;

[0142] Specifically, the construction process of the degradation feature set includes:

[0143] Step 321, according to the high attention area obtained in step 31, select the time points with weight values significantly higher than the set threshold, and take the time points as the center to intercept the fixed length time window in the original multi-dimensional input sequence, and extract the corresponding original sensor data subsequence;

[0144] Step 322, channel index mapping to physical quantity: mapping the feature channel index in the attention matrix to the actual physical quantity name corresponding to the sensor, such as the transverse vibration of the main pump, the outlet temperature of the cooling system, the current signal, etc., to clarify the physical meaning of each high weight signal and the device part it belongs to;

[0145] Step 323, execute feature extraction operation: execute multi-dimensional feature calculation on the original sensor data subsequence intercepted in step 321 to obtain the extracted feature A;

[0146] Among them, the feature A includes but is not limited to:

[0147] Time domain statistics, including mean, maximum, variance, change rate, etc.;

[0148] Frequency domain indicators, including power spectral density, frequency peak value, etc.;

[0149] Variability indicators, including the number of mutation points, skewness, kurtosis, etc.;

[0150] Abnormal point detection results, such as outlier judgment based on IQR and z-score.

[0151] Step 324, reference baseline difference analysis: compare the extracted current feature A with the historical feature statistical range of the equipment under normal working conditions to evaluate its deviation degree, and take the signal features with over-limit or significant abnormalities as abnormal signal items;

[0152] Step 325, generating a structured abnormal feature set: based on the signal items judged as abnormal and their corresponding data and device record information, a degradation feature set is constructed;

[0153] Optionally, the content of the degradation feature set can include: abnormal feature name and dimension, abnormal starting and duration, belonging device subsystem, abnormal amplitude or confidence score, and labeled feature category;

[0154] The degradation feature set will serve as the input basic data for subsequent health state atlas generation, risk level assessment, alarm rule matching, etc.

[0155] In the above steps, the position of the model's attention is found through the attention matrix obtained in step 31, and the device abnormal features are further extracted using the results of step 31 in step 32;

[0156] Step 33, the abnormal features in the degradation feature set are fused with the historical operation data to construct a device health state interpretation atlas of the Transformer model attention weight;

[0157] Specifically, the device health state interpretation atlas includes device degradation trend curves, abnormal frequency distribution, and risk distribution maps, etc. The atlas intuitively shows the current state and historical evolution path of the device, assisting experts in understanding and operation and maintenance judgment;

[0158] Step 331, aligning abnormal features with historical working conditions: the degradation feature set generated in step 32 is time-dimensionally aligned with the historical operation data, and a mapping index is established according to the device, working condition parameter and time stamp, forming a related data set of the current state and historical evolution path;

[0159] Step 332, degradation trend modeling: for each key device feature of the degradation feature set, an evolution curve of its change over time is fitted combining the corresponding historical data, and the current abnormal position is superimposed, generating a device degradation trend graph for representing the dynamic evolution process of the device health state;

[0160] Wherein, the fitting of the evolution curve can adopt linear regression, polynomial trend line or moving average, etc.

[0161] Step 333, abnormal frequency statistics: based on the historical data, the occurrence frequency of various abnormal events in different time periods and different working conditions is counted, and an abnormal intensity heat map or a fault frequency map is constructed to reveal the high-risk areas and frequent patterns.

[0162] Step 334, risk level distribution calculation: combining the abnormal feature type, severity and duration information, the health score or risk level of the device in each time period is calculated, and a risk level distribution map is constructed to visualize the risk concentration trend of the device operation state;

[0163] Step 335, the equipment degradation trend chart, the failure frequency chart and the risk level distribution chart are fused and integrated based on the unified time axis with the equipment identification to generate the nuclear power plant equipment health state interpretation atlas to provide health state interpretation information.

[0164] Step 336, the global state vector output by the Transformer model is compared with the health baseline to obtain the health state information of the equipment at each time, to judge the health deviation degree, and to be associated with each data point of the nuclear power plant equipment health state interpretation atlas to obtain the final nuclear power plant equipment health state interpretation atlas.

[0165] Specifically, in the embodiment, the Transformer model outputs a global state quantity after inputting each time window.

[0166] ;

[0167] The sequence corresponds to the health state of the equipment at different times respectively.

[0168] The deviation degree of the current state from the health baseline is calculated by the formula:

[0169] ;

[0170] Wherein, represents the health baseline, which can be determined by historical health sample mean, artificial calibration, etc.

[0171] Through the above steps, the obtained equipment health state interpretation atlas at least contains the following information: display the historical evolution trajectory of the key feature vector of the equipment; mark the position and influence range of the key abnormal event set ; and visualize the deviation degree of the current state from the health baseline ;

[0172] Further, if the deviation degree is large, the reason needs to be explored, and the dynamic knowledge graph can be used to further generate interpretation information.

[0173] In the above embodiments, a visual nuclear power plant health state interpretation atlas is given, the before-and-after operation trajectories of the change in the health state of the plant are intuitively represented by a change curve, and the change trend is intuitively displayed; in order to generate dynamic instructions in the form of text, the embodiment further proposes a method of a dynamic knowledge atlas to realize dynamic interpretation of data and generate health state interpretation information fused with text description;

[0174] Further, to realize interpretation alignment at the knowledge level, when the global state vector displayed in the plant health state interpretation atlas deviates by more than a set value, the abnormal deviation point is identified, and a dynamic knowledge atlas is constructed for the abnormal deviation point to generate health state interpretation information in the form of text;

[0175] For the abnormal deviation point in the plant health state interpretation atlas, a dynamic knowledge atlas is constructed based on the abnormal deviation point to generate interpretation information, including the following steps:

[0176] Step 361, a directed graph is constructed with device nodes, component nodes, fault mode nodes and working condition nodes as nodes and time sequence correlation weights between the nodes as edges As a dynamic knowledge atlas:

[0177] ;

[0178] Wherein: the node set , respectively represents the device node, the component node, the fault mode node and the working condition node; the edge set , represents the time sequence correlation between the device, the component, the fault mode and the working condition;

[0179] The time sequence correlation weight of each edge is calculated by the following formula:

[0180] ;

[0181] Wherein, is an attention score, which is calculated by a multi-head attention mechanism of a Transformer model and is used to measure the dependence of an input feature u on an input feature v, and the sum of all is equal to 1, and the calculation formula is as follows: ;

[0182] Wherein, is a Query vector, which is mapped from the input feature ; for example, the input feature may be a certain sensor signal or time feature; is a Key vector, which is mapped from the input feature ; represents the dot product of two input features, which measures the correlation of the input feature u and the input feature v; denotes a normalization factor to prevent dimension overflows when the inner product is too large; denotes the operation of converting all correlation scores into a probability distribution; denotes the event intensity, which is used to represent the actual impact of the abnormal event on the equipment operation state at a certain time;

[0183] In a specific embodiment, the event intensity can be determined based on one or more of the following factors:

[0184] the exceeding amplitude, which is the degree to which the monitoring signal exceeds the set threshold, and the event intensity increases with the increase of the exceeding degree;

[0185] the duration, which is the time length of the abnormal state maintaining the abnormal state, and the event intensity increases with the increase of the time length.

[0186] the occurrence frequency, which is the number of times of the abnormal event in a single time window, and the event intensity increases with the increase of the occurrence frequency when the number of times of the abnormal event in a certain time window is large.

[0187] In practical applications, the event intensity can also be determined by the weighted combination of the above factors to comprehensively reflect the influence of the abnormal event in amplitude, persistence and frequency.

[0188] Step 362, based on the constructed directed graph , an explanation path with time evolution characteristics is generated, i.e. textual explanation information, which is used to represent the corresponding relationship of the equipment from the current state to the potential failure mode, as follows:

[0189] ;

[0190] wherein, denotes the explanation path, which is used to represent the reasoning chain and its intensity of the equipment health state at time t; denotes the equipment node; denotes the component node, which represents the key components in the equipment; denotes the failure mode node, which represents the failure mode that the component may appear, such as wear, corrosion, crack, jam, leakage, etc.; denotes the working condition node, which represents the working condition environment of the equipment when running, such as start-stop, power-up, steady-state operation, emergency shutdown (SCRAM), high temperature and high pressure operation, etc.

[0191] denotes the edge weight, which represents the intensity of the association relationship between the equipment, the component, the failure mode and the working condition at time t.

[0192] ​Step 363, obtaining the operation feedback data and operation annotation results of the operation and maintenance personnel as feedback, merging the feedback information with the information of the knowledge graph in a union set manner to dynamically update the knowledge graph .

[0193] According to the health state interpretation information, combined with the device type such as pump, cooling, pipeline, valve, and electrical equipment, a maintenance alarm rule matched with the degradation trend is established, and a maintenance strategy suggestion and an alarm signal are pushed to the DCS monitoring interface of the nuclear power plant in real time, and a corresponding operation and maintenance strategy suggestion is generated, such as immediate shutdown inspection, advance maintenance arrangement, and regular monitoring, which is checked, confirmed, and responded by the on-site operation and maintenance personnel or the main control room operator.

[0194] Specifically, the operation feedback data and operation annotation results will be fed back to the Transformer model learning pool to realize dynamic updating of the knowledge graph :

[0195] ;

[0196] Among them: the set of operation feedback and operation annotation results can include confirming alarms, modifying alarm levels, and manually annotating fault causes;

[0197] Dynamic updating can continuously optimize the graph structure and weight over time, thereby reflecting the real degradation process of the device and human-computer interaction experience.

[0198] Further, the dynamic updating of the knowledge graph includes updating or supplementing the attributes of the nodes, updating the node set, and updating the edge set, which will be described in detail as follows.

[0199] 1) At time , there is a knowledge graph in the system.

[0200] Among them: the node set includes device nodes, component nodes, fault mode nodes, and working condition nodes; the edge set represents the association relationship between the above nodes, and the edge has attribute information including attention score, event intensity, risk level, etc.

[0201] 2) At time , a new feature set is obtained.

[0202] Among them: For newly identified node information, including potential failure modes inferred by the model, component status labeled by operation and maintenance personnel, and newly added working condition categories; For newly generated or modified edge information, including the correspondence between abnormal events and nodes, the dependency relationship corresponding to the attention score, the causal relationship confirmed by operation and maintenance feedback, etc.

[0203] 3) The update rule is:

[0204] ;

[0205] Specifically, the following three cases are included:

[0206] When the node in already exists in , update or supplement the attributes of the node, for example, add artificial confirmation marks or correct health status labels;

[0207] When contains new devices, components or failure modes, add them to the node set ;

[0208] When contains new association relationships, add them to the edge set ; when overlaps with existing edges, update the attributes of the edge such as weight value, risk label, and timestamp.

[0209] Through the above union method, the graph can not only maintain historical evolution information, but also dynamically absorb new model inference results and operation and maintenance personnel feedback, realizing the rolling update and continuous expansion of the graph over time. Therefore, the graph can reflect the historical trajectory and current state of the device health status at any moment, and support subsequent explanation path generation and risk assessment.

[0210] Further, according to the health state explanation information, combined with the type of nuclear power equipment such as pump, cooling, pipeline, valve, electrical equipment, etc., establish maintenance alarm rules matched with the equipment degradation trend, and generate alarm signals according to the alarm trigger conditions and corresponding risk levels.

[0211] Specifically, combined with historical equipment failure records and maintenance experience library, determine the risk level of the current state according to the equipment operation characteristics;

[0212] In step 4, based on the risk level determined by the health state explanation information, generate device maintenance strategy suggestions and corresponding alarm thresholds;

[0213] Specifically, the device maintenance strategy suggestion includes immediate shutdown inspection, advance maintenance arrangement, regular monitoring, and other operation and maintenance strategies, and the alarm threshold is used to trigger the alarm system of the distributed control system (DCS) of the nuclear power plant.

[0214] The device maintenance strategy suggestion and the alarm signal can be realized to be pushed to the DCS monitoring interface of the nuclear power plant in real time, and viewed, confirmed, and responded by the on-site operation and maintenance personnel or the main control room operator. The operation and maintenance personnel or the main control room operator manually confirms, adjusts, or labels the maintenance suggestion and the alarm signal generated by the system according to the actual situation, to form feedback data and operation records. The feedback data and the operation labeling results of the operation and maintenance personnel are returned to the Transformer model learning pool for continuous optimization of model parameters and decision logic, so as to improve the accuracy and adaptability of subsequent attention interpretation and maintenance alarm strategy.

[0215] Through the above process, the nuclear power plant device realizes closed-loop management from real-time data acquisition, abnormal feature identification, health state interpretation, to risk assessment and maintenance alarm generation, significantly improving the accuracy and efficiency of predictive maintenance of the nuclear power plant.

[0216] Further technical solutions also include a model correction mechanism based on asynchronous feedback, and the specific process includes the following steps:

[0217] Step 51, model inference and output: the Transformer model deployed on the edge node or the center server outputs the RUL prediction result at the current time after receiving the latest sliding window input, while retaining the attention weight information inside the model;

[0218] Step 52, result pushing and manual labeling: the prediction result output by the Transformer model and the health state interpretation information of the attention weight information are manually confirmed, adjusted, and corrected labeled;

[0219] Specifically, the model prediction result and the corresponding attention visualization information are pushed to the terminal interface of the operation and maintenance personnel or the central DCS monitoring platform in the form of an interpretable graphical interface. The operation and maintenance personnel can manually confirm, evaluate, or correct the labeling of the model output result in combination with the actual working condition, device maintenance history, and experience knowledge. The corrective labeling can include false positives, time period corrections, and weight anomaly explanations;

[0220] Step 53, feedback information asynchronous return: the prediction result after manual confirmation, adjustment, and corrective labeling is returned to the Transformer model in the form of asynchronous data stream as a fine-tuning data set to fine-tune and adjust the Transformer model;

[0221] The revised model of the embodiment is used for subsequent inference tasks, realizes continuous absorption of field experience information and update of prediction strategy of the model under the premise of uninterrupted service, avoids the problem of long update cycle and slow response of traditional offline training, and significantly improves the practicality and synchronization of the model in the engineering field.

[0222] Further technical solutions, an edge-center joint modeling mechanism is adopted, specifically: including edge nodes and center nodes.

[0223] The edge node: a lightweight Transformer model is deployed on the edge node of the nuclear power plant industrial distributed control system (DCS). This edge node model can use the data collected by the field sensors in real time, and through the local computing device, it can perform dynamic and small-scale model parameter update and real-time inference in a sliding window manner, ensuring that the model prediction has extremely low computational delay and can quickly respond to changes in the field state.

[0224] The center node, i.e., the center platform, is used to perform lightweight synchronization functions, such as obtaining the attention weight pattern and discriminant logic summary information output by the Transformer model from each edge node to realize cross-site data feature and weight sharing.

[0225] The scheme of the embodiment realizes the structural stability and time consistency management of multi-dimensional time series input data by deeply integrating the dynamic sliding window mechanism with the Transformer model, significantly improves the model's perception ability of long sequence and multi-variable equipment degradation trend, and enhances the prediction stability and robustness. At the same time, with the help of the attention mechanism of the Transformer, the system can automatically identify the key time period and key feature channel in the input data, realize self-explaining modeling of the equipment abnormal state and its evolution mode, and enhance the transparency and expert understandability of the model output. On this basis, the system further constructs a health state atlas integrating trend curve, abnormal frequency and risk level distribution, fully presents the current state and historical evolution path of the equipment, and enhances the auditability and auxiliary decision-making ability in the engineering scene. To realize the continuous optimization of the model and the operation and maintenance experience, the asynchronous feedback correction mechanism is introduced, so that the artificial annotation information can be returned to the edge node asynchronously and drive the model to update slightly, overcoming the hysteresis problem of traditional offline retraining of the model. In addition, the scheme supports the uncertainty quantization output of the RUL prediction result, provides confidence interval expression based on Bayesian regression, MC Dropout or double output structure, and improves the confidence evaluation ability and robust decision support ability of the model in high-risk scenarios.

[0226] Embodiment 2

[0227] Based on embodiment 1, the present embodiment provides 8. A self-attention-based nuclear power plant equipment predictive evaluation and maintenance system, comprising:

[0228] a sliding window unit configured to obtain multi-source sensing data of the nuclear power plant equipment by using a sliding window mechanism, and to construct an input time window sequence after preprocessing;

[0229] an inference unit configured to perform inference on the time window sequence data by using a Transformer model, to obtain a state vector for regression analysis, and to generate equipment remaining useful life;

[0230] an explanation unit configured to visualize an attention matrix obtained by the attention mechanism of the Transformer model in the inference process, to obtain an equipment health state explanation graph of attention weights, and to construct a dynamic knowledge graph to generate explanation information for an abnormal deviation point in the equipment health state explanation graph;

[0231] a fusion unit configured to fuse the obtained equipment remaining useful life and corresponding explanation information to generate equipment maintenance strategy suggestions.

[0232] Further, the multi-source sensing data of the nuclear power plant equipment is obtained by using a sliding window mechanism, and the input time window sequence is constructed after preprocessing, including the following steps:

[0233] synchronously listening to the received multi-source sensing data, DCS working condition signals and interlocking signals, and identifying events;

[0234] According to the identified event occurrence point , the time period set before and after the event occurrence point is set as a transition window, and the time period set after the event occurrence is divided into a steady state window, the transition window is smaller than the steady state window, and the sizes of the transition window and the steady state window are set to be different according to the event type. When there is no event, the fixed window is set as the time window, and a dynamically changing time window is obtained;

[0235] Based on the dynamically updated time window, the obtained multi-source sensing data is divided to obtain a plurality of time window data, and each time window data is preprocessed as an input time window sequence.

[0236] Further, the attention matrix obtained by the attention mechanism of the Transformer model in the inference process is visualized to obtain an equipment health state explanation graph of attention weights, and a method for generating explanation information, including the following steps:

[0237] obtain the attention matrix output by the Transformer model, and based on the combination of time steps and feature channels with high weight values in the attention matrix, reversely locate the input data region with high attention in the Transformer model as a high attention region;

[0238] Based on the feature channel and time window of the high attention area, the abnormal features of the key equipment are extracted, the abnormal behavior patterns with physical meaning are identified, the degradation feature set is constructed by extracting the equipment degradation features and the time sequence information thereof;

[0239] The abnormal features in the degradation feature set are fused with the historical operation data to construct the device health state interpretation atlas of the attention weight of the Transformer model.

[0240] It should be noted that each module in the embodiment corresponds to each step in Embodiment 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0241] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0242] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or changes made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A predictive assessment and maintenance method for nuclear power plant equipment based on self-attention, characterized in that, Includes the following steps: A sliding window mechanism is used to acquire multi-source sensor data from nuclear power plant equipment, which is then preprocessed and constructed into an input time window sequence. A sliding window mechanism is used to acquire multi-source sensor data from nuclear power plant equipment. After preprocessing, the data is constructed into an input time window sequence, including the following steps: Synchronously monitor the received multi-source sensor data, DCS operating status signals, and interlocking signals to identify events; Based on the identified event occurrence point, the time period before and after the event occurrence point is set as a transition window according to the event type. The time period after the event occurs is divided into a steady state window. The transition window is smaller than the steady state window. The size of the transition window and the steady state window are set to different sizes according to the event type. The time window without events is set as a fixed window, resulting in a dynamically changing time window. The acquired multi-source sensor data is divided into multiple time window data based on dynamically updated time windows. Each time window data is preprocessed and then used as the input time window sequence. The time window sequence data is inferred through the Transformer model to obtain state vectors, which are then used for regression analysis to generate the remaining service life of the equipment. The attention matrix obtained during the inference process of the Transformer model's attention mechanism is visualized to obtain a device health status interpretation graph with attention weights; for abnormal deviations in the device health status interpretation graph, a dynamic knowledge graph is constructed to generate explanatory information. For abnormal deviations in the device health status interpretation graph, a dynamic knowledge graph is constructed based on these abnormal deviations to generate explanatory information, including the following steps: Using equipment nodes, component nodes, fault mode nodes, and operating condition nodes as nodes, and the temporal correlation weights between nodes as edges, a directed graph is constructed as a dynamic knowledge graph. Based on the constructed directed graph, an explanation path with time evolution characteristics is generated as a health status explanation information with textual description, which is used to characterize the correspondence between the current state of the device and potential failure modes. The system obtains feedback data and operation annotation results from operation and maintenance personnel as feedback, merges the feedback information with the information in the knowledge graph by taking the union of the feedback information, and dynamically updates the knowledge graph. The remaining lifespan of the equipment and the corresponding explanatory information are combined to generate equipment maintenance strategy recommendations.

2. The predictive assessment and maintenance method for nuclear power plant equipment based on self-attention as described in claim 1, characterized in that: Inference using a Transformer model on time window series data includes the following steps: The time window sequence data is input into the attention mechanism module, and query vector, key vector and value vector are generated through three sets of linear projection weight matrices respectively. The attention weights at each time step are calculated using a multi-head attention mechanism and then normalized using a softmax function to obtain the attention weight vector at each time step. The attention weight vector and the value vector are weighted and fused to obtain the state vector of the fused context information at the current time point. : Perform multi-head attention concatenation and feature integration to obtain the state vectors from all attention heads. By concatenating along the feature dimensions, a global comprehensive feature is formed; The concatenated global comprehensive features are then mapped using a linear mapping. Perform dimension alignment and output feature vectors in a uniform format. ; Perform residual connections to connect the feature vectors. With the original input Add them together and then normalize them to obtain the global state vector. ; The obtained global state vector After feature extraction via a feedforward network, and subsequent operations including fully connected layers, activation operations, residual connections, and layer normalization, the final state vector is obtained. .

3. The predictive assessment and maintenance method for nuclear power plant equipment based on self-attention as described in claim 1, characterized in that: During the training of the Transformer model, the loss function constructed includes a prediction loss term, a long-term error penalty term, and a confidence modeling term.

4. The predictive assessment and maintenance method for nuclear power plant equipment based on self-attention as described in claim 1, characterized in that: The method for visualizing the attention matrix obtained during the inference process of the Transformer model's attention mechanism to obtain a device health status interpretation map of the attention weights and generate explanatory information includes the following steps: Obtain the attention matrix output by the Transformer model. Based on the combination of time steps and feature channels with high weight values ​​in the attention matrix, the input data regions where the Transformer model's attention is concentrated are identified as high attention regions. Based on the feature channels and time windows of high-interest regions, abnormal features of key equipment are extracted, abnormal behavior patterns with physical significance are identified, and equipment degradation features and their time-series information are extracted to construct a degradation feature set. By fusing abnormal features from the degradation feature set with historical operational data, a device health status interpretation map of the weights that the Transformer model focuses on is constructed.

5. The predictive assessment and maintenance method for nuclear power plant equipment based on self-attention as described in claim 4, characterized in that: By fusing abnormal features from the degradation feature set with historical operational data, a device health status interpretation map is constructed based on the weights of the Transformer model. This process includes the following steps: The generated degradation feature set is aligned with the historical operating data in terms of time dimension, and a mapping index is established according to equipment, operating parameters and timestamps to form a dataset that associates the current state with the historical evolution path. For each key equipment feature in the degradation feature set, its evolution curve over time is fitted by combining the corresponding historical data, and the current abnormal location is superimposed to generate an equipment degradation trend map. Based on historical data, the frequency of occurrence of various abnormal events in different time periods and under different working conditions is statistically analyzed to construct an abnormal intensity heat map or a fault frequency map. By combining information on abnormal feature type, severity, and duration, the health score or risk level of the device in each time period is calculated, and a risk level distribution map is constructed. By merging and integrating the equipment degradation trend map, failure frequency map, and risk level distribution map based on the same time axis and equipment identification, an equipment health status interpretation map is constructed. The global state vector output by the Transformer model By comparing with the health baseline, the health status information of the equipment at each moment is obtained, the degree of health deviation is determined, and it is correlated with each data point of the nuclear power equipment health status interpretation map to obtain the final nuclear power equipment health status interpretation map.

6. A predictive assessment and maintenance system for nuclear power plant equipment based on self-attention, characterized in that, include: The sliding window unit is configured to acquire multi-source sensor data from nuclear power plant equipment using a sliding window mechanism, and then preprocess it to construct an input time window sequence. A sliding window mechanism is used to acquire multi-source sensor data from nuclear power plant equipment. After preprocessing, the data is constructed into an input time window sequence, including the following steps: Synchronously monitor the received multi-source sensor data, DCS operating status signals, and interlocking signals to identify events; Based on the identified event occurrence point, the time period before and after the event occurrence point is set as a transition window according to the event type. The time period after the event occurs is divided into a steady state window. The transition window is smaller than the steady state window. The size of the transition window and the steady state window are set to different sizes according to the event type. The time window without events is set as a fixed window, resulting in a dynamically changing time window. The acquired multi-source sensor data is divided into multiple time window data based on dynamically updated time windows. Each time window data is preprocessed and then used as the input time window sequence. The inference unit is configured to infer the time window sequence data through the Transformer model, obtain the state vector, perform regression analysis, and generate the remaining service life of the equipment. The explanation unit is configured to visualize the attention matrix obtained by the attention mechanism of the Transformer model during inference to obtain a device health status explanation graph with attention weights; and to construct a dynamic knowledge graph to generate explanation information for abnormal deviations in the device health status explanation graph. For abnormal deviations in the device health status interpretation graph, a dynamic knowledge graph is constructed based on these abnormal deviations to generate explanatory information, including the following steps: Using equipment nodes, component nodes, fault mode nodes, and operating condition nodes as nodes, and the temporal correlation weights between nodes as edges, a directed graph is constructed as a dynamic knowledge graph. Based on the constructed directed graph, an explanation path with time evolution characteristics is generated as a health status explanation information with textual description, which is used to characterize the correspondence between the current state of the device and potential failure modes. The system obtains feedback data and operation annotation results from operation and maintenance personnel as feedback, merges the feedback information with the information in the knowledge graph by taking the union of the feedback information, and dynamically updates the knowledge graph. The fusion unit is configured to fuse the obtained remaining lifespan of the equipment and its corresponding explanatory information to generate equipment maintenance strategy recommendations.

7. The self-attention-based predictive assessment and maintenance system for nuclear power plant equipment as described in claim 6, characterized in that: The method for visualizing the attention matrix obtained during the inference process of the Transformer model's attention mechanism to obtain a device health status interpretation map of the attention weights and generate explanatory information includes the following steps: Obtain the attention matrix output by the Transformer model. Based on the combination of time steps and feature channels with high weight values ​​in the attention matrix, the input data regions where the Transformer model's attention is concentrated are identified as high attention regions. Based on the feature channels and time windows of high-interest regions, abnormal features of key equipment are extracted, abnormal behavior patterns with physical significance are identified, and equipment degradation features and their time-series information are extracted to construct a degradation feature set. By fusing abnormal features from the degradation feature set with historical operational data, a device health status interpretation map of the weights that the Transformer model focuses on is constructed.

Images