State evaluation method, system, equipment and medium

Abstract

The invention discloses a state evaluation method, system and device and a medium, and belongs to the field of machine learning, and the method comprises the steps: obtaining multi-modal sensor data, carrying out the preprocessing and fusion, and obtaining a fusion feature vector; detecting the fusion feature vector based on a mixed architecture model constructed by a spline enhanced multi-layer sensing network, a convolutional neural network and a Transform attention mechanism to obtain an anomaly type and confidence; when the detection result is abnormal, generating an abnormal event according to an event driving mechanism; responding and acquiring time sequence data from historical sensor data, and performing simulation prediction according to the physical information neural network model to obtain a prediction result; the prediction result is fed back to the detection model to update the confidence degree, the target detection result is obtained, the state evaluation result is obtained in combination with the prediction result, and therefore the state evaluation accuracy and reliability can be improved.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more particularly to a state assessment method, system, device, and medium. Background Technology

[0002] In high-voltage direct current (HVDC) transmission systems, smoothing reactors, as core components, are used to smooth DC current ripple, suppress harmonics, limit short-circuit current, and maintain DC link voltage stability. These functions have a decisive impact on the reliable operation of HVDC systems. However, during long-term operation, smoothing reactors face various power faults and degradation mechanisms, such as overheating, partial discharge (PD), winding deformation, inter-turn short circuits, and commutation failures and DC line grounding faults unique to HVDC systems. Therefore, accurate and timely condition assessment and fault detection of core power equipment such as smoothing reactors are particularly important.

[0003] Current condition assessment technologies mainly rely on rule-driven methods and traditional machine learning models. Both methods have significant drawbacks and are insufficient to meet the accurate assessment requirements under complex operating conditions. Firstly, existing rule-driven methods rely on preset fixed thresholds (such as upper limits for winding temperature and vibration amplitude) and simple logical judgments for condition classification. However, the operating parameters of smoothing reactors are affected by factors such as grid load fluctuations, the randomness of new energy output, and changes in ambient temperature and humidity. The normal fluctuation range far exceeds the fixed threshold range, and the thresholds cannot be adaptively adjusted. This easily leads to misjudging normal parameter fluctuations as faults or missing weak signals of early minor faults. First, existing static models based on fixed rules have insufficient generalization ability and cannot adapt to changes in different operating conditions and external environments, thus failing to meet the need for accurate identification of the state of smoothing reactors under complex operating conditions. Second, although traditional machine learning methods can handle some nonlinear mapping relationships (such as the simple correlation between current and temperature), the performance of machine learning models heavily relies on manual feature engineering, which does not fully utilize the physical mechanism. Furthermore, it is difficult for humans to extract deep coupling features and capture long-term time-series dependencies, ultimately limiting the accuracy of state identification and making it difficult to achieve effective detection of early minor faults and accurate location of fault roots. Summary of the Invention

[0004] This invention provides a state assessment method, system, device, and medium that can solve the problems of insufficient generalization ability and inadequate utilization of physical mechanisms in existing machine learning models, which lead to the inability of state assessment to adapt to complex operating condition fluctuations and insufficient accuracy and reliability of assessment.

[0005] This invention provides a state assessment method, comprising: Acquire multimodal sensor data, preprocess and fuse the sensor data to obtain a fused feature vector; The fused feature vector is detected based on an anomaly detection model to obtain detection results. The anomaly detection model is a hybrid architecture model built on spline-enhanced multilayer perceptron, convolutional neural network and Transformer attention mechanism. The detection results include anomaly type and confidence level. When the detection result is abnormal, an abnormal event is generated according to the event-driven mechanism; In response to the abnormal event, time-series data corresponding to the abnormal type and the time window of the abnormal event are obtained from historical sensor data, and the time-series data are simulated and predicted according to the physical information neural network model to obtain the prediction result; The prediction results are fed back to the anomaly detection model, and the confidence level in the detection results is updated according to the prediction results to obtain the target detection results. The state assessment results are then determined based on the target detection results and the prediction results.

[0006] This invention, through preprocessing and fusion of multimodal sensor data (such as temperature, vibration, and partial discharge), eliminates differences in data format and dimensions, integrating scattered single-modal data into a unified fused feature vector, thus avoiding evaluation bias caused by the one-sidedness of single-modal data. In the hybrid architecture anomaly detection model, a convolutional neural network (CNN) extracts local and short-term temporal features of the fused features (such as the instantaneous signal of a sudden temperature rise), a spline-enhanced multilayer sensing network (SEMLP) enhances nonlinear discrimination through spline functions and Fourier series (amplifying the difference between fault and normal features), and a Transformer attention mechanism captures global long-distance dependencies (such as the temporal correlation between partial discharge and subsequent temperature anomalies). These three mechanisms work together to take into account both local / short-term and global / long-term temporal features, improving the accuracy of anomaly identification and the precision of anomaly type judgment. The event-driven mechanism triggers the Physical Information Neural Network (PINN) simulation only when an anomaly is detected, avoiding the loss of real-time performance caused by continuous operation of the physical model. At the same time, the multi-physics simulation of PINN (such as temperature field and electromagnetic field) provides a physical basis for anomaly detection, balancing "real-time monitoring" and "precise source tracing". Requirements: Upgrade the data-driven preliminary detection (statistical probability) to a target detection result with dual verification of "data + physics" by updating the confidence level through PINN prediction results (e.g., increasing the confidence level of thermal faults from 60% to 92%), solving the "black box" problem of pure data-driven models, improving detection reliability, and achieving interpretable assessment; finally, combine the target detection results with the physical quantities predicted by PINN (such as fault location and temperature distribution) to generate a complete assessment report containing fault type, confidence level, and physical root cause, meeting the core requirements of operation and maintenance personnel for "accuracy + interpretability".

[0007] Furthermore, the preprocessing and fusion of the sensor data specifically includes: Wavelet denoising algorithm is used to denoise the sensor data of each mode, and the denoised sensor data is normalized to obtain the first sensor data corresponding to each mode. The target proximity information recovery filtering technique is used to perform data interpolation on the missing data in each of the first sensor data by combining spatial and temporal dimensions to obtain the second sensor data. The second sensor data is processed by a convolutional neural network to obtain modal feature vectors corresponding to each modality. A fusion weight corresponding to each modal feature vector is determined by an attention mechanism. The modal feature vectors are fused according to the fusion weights to obtain the fused feature vector.

[0008] This approach utilizes wavelet denoising algorithms to selectively filter out high-frequency noise in sensor data (such as electromagnetic interference from partial discharge signals and environmental noise from vibration signals). Normalization (such as Z-Score and Min-Max) eliminates dimensional differences between different modal data (such as temperature "°C" versus vibration "g"), preventing feature distortion caused by noise and dimensional interference and providing high-quality "first sensor data" for subsequent modeling. Target proximity information recovery filtering combines time dimension (data from adjacent time points) and spatial dimension (data from adjacent sensors) to interpolate missing values, avoiding the bias of single-dimensional interpolation (such as using only the time mean), solving the problem of data loss caused by sensor failure and transmission delay, ensuring the integrity of "second sensor data," and preventing missing data from affecting model training and detection accuracy. CNNs extract exclusive modal feature vectors for each modality (such as high-frequency pulse features for vibration modes and gradient features for temperature modes), and then dynamically allocate fusion weights based on an attention mechanism (e.g., higher weights for partial discharge modes that contribute significantly to inter-turn short-circuit faults), avoiding the weakening of key features due to "one-size-fits-all" fusion and improving the sensitivity of the fused feature vectors to faults.

[0009] Furthermore, the method employs target proximity information recovery filtering technology, combining spatial and temporal dimensions to perform data interpolation on missing data in each of the first sensor data to obtain second sensor data, specifically as follows: The data from each of the first sensors are used as a multi-dimensional data matrix to determine the target time and target sensor corresponding to the missing data. A first time set within a preset time range is determined, and a first data set of the target sensor within the first time set is obtained through the multidimensional data matrix. Based on the time distance between each first data in the first data set and the missing data, the time weight corresponding to each first data is determined. A first set of sensors within a preset correlation range is determined for the target sensors. A second set of data for the first sensor set at the target time is obtained through the multidimensional data matrix. The correlation weight corresponding to each second data is determined based on the correlation between each second data in the second set and the missing data. The third data set of the first sensor set within the first time set is obtained through the multidimensional data matrix. Based on the time distance and correlation between each third data in the third data set and the missing data, the third weight corresponding to each third data is determined. Obtain a first weight coefficient, a second weight coefficient, and a third weight coefficient. Weight each first data point and each time weight according to the first weight coefficient. Weight each second data point and each related weight according to the second weight coefficient. Weight each third data point and each third weight according to the third weight coefficient. Sum the weighted data to obtain the target data. The missing data is interpolated based on the target data to obtain the second sensor data.

[0010] Compared to traditional single-dimensional or single-dimensional interpolation, this technology employs a three-dimensional collaborative approach to calculate the target value of missing data. This involves "time distance weighting" (the closer the time, the greater the weight; for example, data from time t-1 and t+1 is more reliable than data from time t-10), "sensor correlation weighting" (the higher the correlation, the greater the weight; for example, a temperature sensor in the same winding is more correlated than an ambient temperature sensor), and "spatiotemporal combined third weighting" (considering the dual advantages of proximity and high correlation). This significantly improves interpolation accuracy compared to single-dimensional methods and enhances the reliability of missing data. Since sensors in smoothing reactors are typically deployed in a "spatial distribution + temporal acquisition" configuration (e.g., temperature sensors at different locations in the winding, vibration data collected at different times), the multi-dimensional data matrix and spatiotemporal weighting design of this technology align with this layout, avoiding interpolation deviations caused by neglecting spatial correlations (such as the coordinated changes of adjacent sensors), and ensuring the accuracy of the interpolated "second sensor data." It better matches the actual operating status of the equipment, adapts to the multi-sensor layout characteristics of the smoothing reactor, and reduces data deviation; it eliminates the "feature gaps" caused by missing data (such as the break in time series features caused by missing temperature data at a certain moment) through precise interpolation, ensuring that there are no data gaps when CNN extracts modal features and Transformer captures time series correlations, and provides data integrity guarantee for the stable operation of the anomaly detection model.

[0011] Furthermore, the detection of the fused feature vector based on the anomaly detection model specifically involves: The fused feature vector is input into the convolutional neural network module in the anomaly detection model to extract features from the fused feature vector through a preset convolutional kernel, thereby obtaining the first anomaly feature; The first abnormal feature is input into the spline-enhanced multilayer perceptron module in the abnormality detection model to parameterize the first abnormal feature through spline functions and Fourier series, to perform nonlinear transformation on the first abnormal feature, and to fit a nonlinear abnormal correlation relationship based on the transformed first abnormal feature to enhance the distinguishability between abnormal features and normal features, thereby obtaining the second abnormal feature. The second abnormal feature is arranged in chronological order and then input into the Transformer module in the anomaly detection model. The second abnormal feature is calculated through a multi-head self-attention layer to obtain the attention weights corresponding to each attention head. The second abnormal feature is then weighted and summed according to the attention weights to obtain the third abnormal feature. The classification layer in the anomaly detection model compares the third anomaly feature with each anomaly feature template in the preset anomaly feature template library to determine the anomaly type corresponding to the anomaly feature template with the highest matching degree. The feature weight coefficients are obtained, and the first anomaly probability of the convolutional neural network module, the second anomaly probability of the spline-enhanced multilayer perceptron module, and the third anomaly probability of the Transformer module are weighted and calculated based on the feature weight coefficients to obtain the confidence level of the detection result.

[0012] This approach addresses the problem of "global features ignoring minor local faults" by using CNNs to extract local short-term anomalous features (such as instantaneous temperature spikes or high-frequency vibration pulses) through sliding convolutions. SEMLP uses spline functions and Fourier series to perform nonlinear mapping on the first anomalous feature, amplifying the difference between fault and normal features (e.g., converting weak inter-turn short-circuit signals into easily identifiable nonlinear features), thus solving the problem of "linear features failing to distinguish complex faults." The Transformer multi-head self-attention mechanism arranges the second anomalous feature sequentially and captures global correlations across different time steps and modalities in parallel (e.g., the correlation between partial discharge anomalies at time t and temperature anomalies at time t+5), solving the problem of traditional machine learning's difficulty in capturing long-term temporal dependencies and improving the recognition rate of early minor faults (features need to accumulate and manifest). The classification layer accurately locates the anomaly type by comparing it with a pre-set anomalous feature template library (e.g., typical feature templates for inter-turn short circuits and winding overheating), avoiding the problem of "only recognizing anomalies but failing to determine the type." The limitations of CNN, SEMLP, and Transformer are addressed by calculating confidence levels based on the output probabilities of CNN, SEMLP, and Transformer. This provides a preliminary quantification of the reliability of the detection results and offers a "priority basis" for subsequent PINN verification (e.g., low-confidence anomalies require focused verification).

[0013] Furthermore, the loss function of the physical information neural network model is determined based on the physical information loss term, the boundary and initial condition loss term, and the data loss term, specifically as follows: The physical equation residuals are calculated based on the prediction results and the physical equations, and the mean square error of the physical equation residuals is calculated to obtain the physical information loss term. The constraint error is calculated based on the prediction results and physical constraints, and the mean square error of the constraint error is calculated to obtain the boundary and initial condition loss terms. Calculate the mean square error between the prediction result and the sensor data to obtain the data loss term; The loss weight parameters are dynamically determined based on the gradient norm normalization algorithm, and the physical information loss term, the boundary and initial condition loss term, and the data loss term are weighted and summed according to the loss weight parameters to obtain the loss value corresponding to the loss function.

[0014] The physical information loss term calculates the mean square error of the residuals between the predicted results and physical equations (such as the heat conduction equation and Maxwell's equations), forcing PINN predictions to conform to the core physical rules of power equipment operation, such as smoothing reactors (e.g., temperature distribution must meet the heat conduction law, electromagnetic field must meet the electromagnetic coupling principle). This solves the problem of pure data-driven models "potentially outputting physically impossible results" (e.g., a sudden temperature rise without a heat source). The boundary and initial condition loss term calculates the mean square error between the predicted results and equipment constraints (e.g., initial temperature = ambient temperature, winding surface temperature ≤ insulation tolerance threshold), preventing PINN predictions from exceeding the actual operating boundaries of the equipment (e.g., predicted temperature exceeding the insulation material tolerance value). This ensures that the predictions meet equipment constraints, closely match actual operating boundaries, and improves the engineering practicality of the prediction results. The data loss term ensures that PINN predictions closely match sensor measured data, avoiding the problem of pure physical models "having large deviations from actual operation." The GradNorm algorithm dynamically adjusts the weights of the three loss terms, solving the problem of "single loss term dominating training" (e.g., excessive data loss leading to the neglect of physical rules), enabling PINN to simultaneously ensure "physical consistency." By combining "data fit", it outputs high-fidelity multiphysics prediction results (such as temperature distribution at the fault location and electromagnetic field distortion range), balances physical rules with data fitting, improves simulation accuracy, and provides accurate physical basis for fault tracing.

[0015] Furthermore, the feature processing of each second sensor data using a convolutional neural network to obtain the modality feature vector corresponding to each modality, wherein the second sensor data includes a one-dimensional time series of multiple modalities, and the convolutional neural network includes convolutional layers, activation functions, and pooling layers, specifically: For each modality, the one-dimensional time series is processed by sliding convolution through the convolution kernel in the convolutional layer; The feature vector obtained by sliding convolution is nonlinearly processed by the activation function to obtain a feature map; The feature map is downsampled by the pooling layer to obtain the feature vectors of each modality.

[0016] For multimodal one-dimensional time series (such as temperature changes over time, vibration signal time series), convolutional layers automatically extract local features (such as short-term temperature gradient changes, specific frequency pulses of vibration) through sliding convolution kernels, eliminating the need for manual feature design and solving the problem that "manual feature engineering is difficult to cover all local features of the fault." Activation functions (such as ReLU, GELU) perform nonlinear processing on the linear features of the convolution output, transforming the "straightforward" local features into features that fit the nonlinear characteristics of the fault (such as the abrupt change in temperature before and after the fault threshold), solving the problem that "linear features cannot characterize complex fault modes," improving the sensitivity of modal features to faults, enhancing the nonlinear expression of features, and adapting to the complex characteristics of faults. Pooling layers downsample the nonlinearly processed feature maps, reducing the feature dimension (e.g., compressing 1000-dimensional features to 100 dimensions) while retaining key features (such as fault peak features), reducing the computational load of subsequent attention fusion and anomaly detection models, avoiding the problem of "high-dimensional features leading to slow model training and poor real-time performance," achieving real-time monitoring, dimensionality reduction and computational efficiency, and balancing feature quality and computational efficiency.

[0017] Further, the second abnormal feature is calculated using a multi-head self-attention layer to obtain the attention weights corresponding to each attention head. The second abnormal feature is then weighted and summed according to these attention weights to obtain the third abnormal feature. Specifically: The second abnormal feature is projected by multiple projection matrices of each attention head to obtain the query vector, key vector and value vector corresponding to each second abnormal feature; Calculate the dot product of the query vector and the key vector to obtain the attention score corresponding to each attention head; The attention scores are processed by a normalization function to obtain the attention weights. The value vectors of each attention head are weighted and summed according to the attention weights to obtain an initial vector. The initial vector is then projected through a linear layer in the multi-head self-attention layer to obtain the third abnormal feature.

[0018] This approach generates Q (query vector), K (key vector), and V (value vector) for each attention head through an independent projection matrix, capturing feature associations across different dimensions in parallel (e.g., head 1 captures the association between temperature and vibration, head 2 captures the association between partial discharge and electromagnetic fields). This addresses the problem of "single attention head being unable to cover multiple types of associations," comprehensively exploring the multi-dimensional temporal dependencies of power faults, capturing feature associations dimensionally, and improving global dependency recognition capabilities. The dot product of Q and K quantifies the similarity of different features (e.g., the similarity between temperature features at a certain moment and historical fault features). A normalization function (e.g., Softmax) converts the scores into 0-1 weights, ensuring that highly relevant features (e.g., key fault features) receive high attention, while low-relevance features (e.g., environmental interference) are weakened, avoiding the problem of "irrelevant features interfering with global association judgment." By weighted summing of the V vectors of each attention head, multi-dimensional association information is integrated, and then the feature dimensions are unified through linear layer projection. The resulting third anomaly feature contains both local key information and covers global long-distance dependencies, providing "comprehensive + accurate" support for subsequent anomaly type judgment. The feature support improves the Transformer module's accuracy in identifying complex faults (multi-feature collaborative anomalies).

[0019] Another embodiment of the present invention provides a state assessment system, including: a feature processing module, an anomaly detection module, an event-driven module, a simulation prediction module, and a confidence update and assessment module; The feature processing module is used to acquire multimodal sensor data, preprocess and fuse the sensor data to obtain a fused feature vector; The anomaly detection module is used to detect the fused feature vector based on the anomaly detection model to obtain the detection result. The anomaly detection model is a hybrid architecture model built based on spline-enhanced multilayer perceptron, convolutional neural network and Transformer attention mechanism. The detection result includes anomaly type and confidence level. The event-driven module is used to generate an abnormal event according to the event-driven mechanism when the detection result is abnormal. The simulation prediction module is used to respond to the abnormal event by acquiring time-series data corresponding to the abnormal type and the time window of the abnormal event from historical sensor data, and performing simulation prediction on the time-series data according to the physical information neural network model to obtain the prediction result. The confidence update and evaluation module is used to feed the prediction result back to the anomaly detection model, update the confidence level in the detection result according to the prediction result to obtain the target detection result, and determine the state evaluation result according to the target detection result and the prediction result.

[0020] This invention, through preprocessing and fusion of multimodal sensor data (such as temperature, vibration, and partial discharge), eliminates differences in data format and dimensions, integrating scattered single-modal data into a unified fusion feature vector, thus avoiding evaluation bias caused by the one-sidedness of single-modal data. In the hybrid architecture anomaly detection model, a convolutional neural network (CNN) extracts local and short-term temporal features of the fused features (such as the instantaneous signal of a sudden temperature rise), a spline-enhanced multilayer perceptron (SEMLP) enhances nonlinear discrimination through spline functions and Fourier series (amplifying the difference between fault and normal features), and a Transformer attention mechanism captures global long-distance dependencies (such as the temporal correlation between partial discharge and subsequent temperature anomalies). These three mechanisms work together to consider both local / short-term and global / long-term temporal features, improving the accuracy of anomaly identification and the precision of anomaly type judgment. The event-driven mechanism only... When an anomaly is detected, the Physical Information Neural Network (PINN) simulation is triggered to avoid the loss of real-time performance caused by continuous operation of the physical model. At the same time, PINN's multi-physics simulation (such as temperature field and electromagnetic field) provides physical evidence for anomaly detection, taking into account both the needs of "real-time monitoring" and "precise source tracing". The confidence level is updated by PINN prediction results, upgrading the data-driven preliminary detection (statistical probability) to the target detection result with "data + physics" dual verification (such as the confidence level of thermal faults increasing from 60% to 92%), solving the "black box" problem of pure data-driven models, improving detection reliability, and achieving interpretable assessment. Finally, the target detection results and the physical quantities predicted by PINN (such as fault location and temperature distribution) are combined to generate a complete assessment report containing fault type, confidence level, and physical root cause, meeting the core needs of operation and maintenance personnel for "accuracy + interpretability".

[0021] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the state assessment method of the present invention.

[0022] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform steps as described in the state assessment method of the present invention. Attached Figure Description

[0023] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1This is a schematic flowchart of a state assessment method provided in an embodiment of the present invention; Figure 2 This is a flowchart of an event-driven co-simulation algorithm provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a CNN-SEMLP-Attention model provided in an embodiment of the present invention; Figure 4 A schematic diagram of the structure of a PINN model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a smoothing reactor condition assessment system provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of another state assessment system provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0027] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0030] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0031] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0032] See Figure 1 To address the problem that existing rule-driven assessment methods cannot adapt to complex operating condition fluctuations, resulting in insufficient accuracy and reliability of state assessment, an embodiment of the present invention provides a state assessment method, including steps S101-S105, comprising: Step S101: Obtain multimodal sensor data, preprocess and fuse the sensor data to obtain a fused feature vector.

[0033] In this embodiment, the state assessment method of a smoothing reactor is used as an example. Raw heterogeneous sensor data of different modes are collected. The raw heterogeneous sensor data is preprocessed, including cleaning, normalization, and imputation of missing data. For data fusion, a feature-level fusion strategy can be adopted. This strategy includes using an encoder-combiner specific architecture to process the heterogeneous sensor inputs. For each type of sensor data, the sampling frequency and internal structure may differ. An independent dedicated encoder is designed. Each encoder encodes the raw heterogeneous data into mode-specific feature vectors. The mode-specific vectors from different encoders are concatenated and input into a combiner to fuse these features to generate a unified fused feature vector. The sensor data includes electrical quantities (current, voltage), temperature quantities (winding hot spot temperature, surface temperature), vibration quantities, acoustic quantities, and partial discharge signals. High-precision current transformers (CTs) and voltage transformers (PTs) are used to acquire current and voltage signals at a sampling frequency of 10 kHz. Thermocouples or resistance temperature detectors (RTDs) are used to monitor winding hot spot and surface temperatures at a sampling frequency of 5 kHz. An accelerometer is used to acquire vibration signals, with a sampling frequency of 15 kHz for high-frequency vibration characteristics. An acoustic emission sensor is used to acquire transient elastic wave signals caused by material deformation at a sampling frequency of 100 kHz. A very high frequency (UHF) sensor is used to acquire partial discharge signals at a sampling frequency of 10 MHz. In this embodiment, the following implementation steps are performed by inputting the acquired data into a pre-trained model. The model training data comes from historical operating data under various working conditions.

[0034] Step S102: Detect the fused feature vector based on the anomaly detection model to obtain the detection result. The anomaly detection model is a hybrid architecture model built based on spline-enhanced multilayer perceptron, convolutional neural network and Transformer attention mechanism. The detection result includes anomaly type and confidence level.

[0035] In this embodiment, a hybrid architecture anomaly detection model (CNN-SEMLP-Attention) based on a spline-enhanced multilayer perceptron (SEMLP), a convolutional neural network (CNN), and a Transformer attention mechanism is used. The fused feature vector is fed into the anomaly detection model. The CNN layer extracts local and short-term temporal features from the fused feature vector. The SEMLP layer takes the output of the CNN as input and uses its learnable spline function to perform nonlinear mapping on the local features extracted by the CNN, learning complex nonlinear relationships. The Transformer attention layer follows the SEMLP layer, receives the features processed by the SEMLP, and establishes global correlations and long-distance dependencies between features through a self-attention mechanism, compensating for the feature decay problem that may occur in SEMLP. This ensures that the model comprehensively considers the relationship between all time steps and different sensor data. Finally, a fully connected layer and a Softmax activation function are used for final classification to obtain the anomaly detection result.

[0036] Step S103: When the detection result is abnormal, an abnormal event is generated according to the event-driven mechanism.

[0037] In this embodiment, an event-driven co-simulation framework can be used to dynamically exchange data and feedback between the data-driven model (i.e., the anomaly detection model) and the physical model. When the anomaly detection module detects an anomaly or a change in the system's operating status, it generates an anomaly event signal according to the event-driven mechanism and sends it to the physical model.

[0038] Step S104: In response to the abnormal event, obtain time-series data corresponding to the abnormal type and the time window of the abnormal event from historical sensor data, and perform simulation prediction on the time-series data according to the physical information neural network model to obtain the prediction result.

[0039] In this embodiment, in response to the abnormal event, a physical model based on an adaptive weighted physical information neural network (AW-PINN) is triggered to perform high-fidelity prediction of the temperature field or electromagnetic field of the smoothing reactor, obtaining the prediction result. The loss function of PINN includes: a physical information loss term, used to evaluate the residual of the model prediction value to the multi-physics control partial differential equation of the smoothing reactor; a boundary and initial condition loss term, used to evaluate the error between the model prediction value and the specified boundary and initial condition values; and a data loss term, used to evaluate the error between the model prediction value and the available labeled measurement data. The control partial differential equations embedded in the physical information loss term include, but are not limited to, Maxwell's equations and the thermal equation. In the PINN-based physical model prediction step, PINN employs an adaptive weighting method based on the gradient norm normalization (GradNorm) algorithm to solve the problem of balancing the weights of multiple loss functions during training, ensuring that the model can simultaneously learn physical equations and data information.

[0040] Step S105: Feed the prediction result back to the anomaly detection model, update the confidence level in the detection result according to the prediction result to obtain the target detection result, and determine the state evaluation result according to the target detection result and the prediction result.

[0041] In this embodiment, under the hybrid modeling architecture of anomaly detection model and physical model, the state evaluation result is generated based on the fault detection result of the CNN-SEMLP-Attention anomaly detection model and the prediction result of the PINN physical model. The state evaluation result includes the fault detection result, fault prediction result, and real-time state warning of the smoothing reactor; finally, the state evaluation result is presented. Specifically, during the generation of the fault detection result, the CNN-SEMLP-Attention module updates its fault detection confidence after receiving physical feedback. Specifically, when the CNN-SEMLP-Attention model initially determines it to be a thermal fault, and the PINN model predicts an increase in winding hotspot temperature, the CNN-SEMLP-Attention module will increase its detection probability of the thermal fault and generate a target detection result. The fault detection result is output in the form of fault category and detection confidence. Generating Fault Prediction Results: When a thermal fault is diagnosed, the PINN module provides physical source tracing of the fault. Specifically, the PINN model, through its embedded thermal and electromagnetic equations, can accurately predict the temperature distribution and heat source in the fault location area. The prediction results are output as trends in physical quantity changes and the spatial location of the fault. Real-time Status Warnings: The system generates multi-level warnings based on the confidence and severity of the fault detection results. When a thermal fault is detected and the confidence exceeds a preset threshold, the system will issue a high-level warning; the warning results are output in the form of warning levels.

[0042] It should be noted that, as Figure 2The diagram shows a flowchart of an event-driven co-simulation algorithm provided by an embodiment of the present invention. The event-driven co-simulation framework enables intelligent, on-demand collaborative work between two sub-modules. It is not a static, fixed process, but rather an intelligent control system that dynamically adjusts according to the system state. Specifically, the core of the framework is to implement two bidirectional data and feedback loops. The data sub-module (CNN-SEMLP-Attention) acts as an event listener and primary detector, continuously processing fused feature vectors from sensors for efficient preliminary fault detection. The physical sub-module (PINN) acts as a high-fidelity predictor and fault tracing engine, and is in a standby state under normal circumstances, without performing calculations. The overall collaborative working loop algorithm is as follows: Step 1 Initialization: The data-driven sub-module starts and enters real-time listening mode. The physical submodule is in standby mode; the second step is monitoring and initial detection: at each time step, it receives the fused feature vector sequence and uses its trained model to perform preliminary fault detection; the third step is event judgment: if the output of the data-driven submodule does not trigger any preset event (smoothing reactor body fault or system fault), the algorithm returns to step 2 and continues monitoring; the fourth step is physical model triggering: if an event is detected, the coupling module extracts the corresponding time series window from historical data according to the event type and passes it as input to the physical submodule (PINN); the fifth step is high-fidelity prediction: the physical submodule is activated and uses its model with embedded physical equations to perform high-fidelity prediction of multiphysics based on the received data, generating interpretable physical quantity prediction results; the sixth step is feedback and enhancement: the physical submodule sends the prediction results back to the data-driven submodule. The data-driven submodule uses this physical information as auxiliary features to update its judgment of fault type and confidence level; the seventh step is result output: finally, the system combines the data-driven detection results and the prediction feedback of the physical model to generate the final status assessment report, including fault detection results, fault prediction, and health status warning. The fault prediction steps are performed under the condition of considering the operating status of the HVDC system (normal, HVDC commutation failure, DC line grounding fault).

[0043] Overall, this invention overcomes the limitations of traditional fault diagnosis methods in terms of accuracy, timeliness, interpretability, and handling complex multi-source data by integrating data-driven models and high-fidelity physical models. It achieves high-fidelity prediction of smoothing reactor faults and fault tracing and analysis considering system operating conditions, significantly improving the accuracy, interpretability, and applicability of smoothing reactor condition assessment. This provides an advanced solution for predictive maintenance and safe operation of critical power system equipment. It should be noted that these descriptions represent only preferred embodiments of the invention and do not limit the scope of protection of the invention. Any adjustments, equivalent transformations, and optimizations made under the guidance of the core ideas and basic principles of this invention should be considered part of the protection scope of this invention.

[0044] As an example of an embodiment of the present invention, the preprocessing and fusion of the sensor data specifically includes: using a wavelet denoising algorithm to denoise the sensor data, and normalizing the denoised sensor data to obtain first sensor data corresponding to each modality; using a target proximity information recovery filtering technique, combining spatial and temporal dimensions to perform data interpolation on missing data in each of the first sensor data to obtain second sensor data; using a convolutional neural network to perform feature processing on each of the second sensor data to obtain modal feature vectors corresponding to each modality; using an attention-based mechanism to determine the fusion weights corresponding to each modal feature vector; and fusing the modal feature vectors according to the fusion weights to obtain the fused feature vector.

[0045] In this embodiment, during data processing, electrical quantity signals are processed using digital filters to remove interference. For example, a band-stop filter is used to eliminate power frequency and harmonic interference; or a wavelet denoising algorithm is used to remove transient impacts and random noise by setting a threshold for high-frequency wavelet coefficients, while retaining the effective temporal characteristics of the signal. The Z-Score normalization method is used to process the cleaned electrical quantity data to eliminate dimensional differences. For example, the value of each data point is subtracted from the average value of the data sequence and then divided by the standard deviation of the data sequence. Vibration signals are processed using digital filters to remove environmental and low-frequency interference. Specifically, a high-pass filter is used to eliminate low-frequency equipment shaking or drift, and a low-pass filter is used to smooth high-frequency noise, ensuring the retention of key frequency band information related to the fault; the Z-Score normalization method is used to process the cleaned vibration data, converting signals with different vibration amplitudes to the same standard scale. For abnormal temperature readings, a physical threshold-based method is used for identification and correction; a minimum-maximum normalization method is used to process the cleaned temperature data. Wavelet thresholding denoising is employed to clean acoustic quantities and partial discharge signals. By identifying and removing noise figures in the wavelet domain, signal smoothing is achieved while fully preserving the pulse characteristics associated with discharge events. An improved method based on proximity information recovery (AIR) filtering is used to imputate missing data. This method utilizes both temporal and spatial neighboring data; data from different sensors are treated as a multidimensional matrix, and when a data point is missing, it is imputed using the spatiotemporal neighborhood data of that data point.

[0046] In the data fusion process, a dedicated encoder is designed for each type of sensor data. Each encoder encodes the raw heterogeneous data into modality-specific feature vectors. The modality-specific vectors from different encoders are concatenated and input into a combiner. The combiner, based on an attention mechanism, fuses the feature vectors from different encoders and assigns different weights to each modality. A learnable query vector The input to the combiner is the set of feature vectors from all encoders, denoted as . ; The first step is to calculate the attention score for each modality feature: For each modality feature vector, calculate its attention score with all... Query the dot product of vectors: ;in, It is the first The feature vector of a modality For the first A learnable query vector, Indicates the first The eigenvector of the first mode and the eigenvalue of the second mode The similarity of the query vectors; The second step is to obtain the maximum similarity score: To determine which modality is most important for predicting the current input data, the query vector that best matches each modality is selected. ;in, Indicates the first The maximum contribution of a modality to all potential anomalous modes; The third step involves converting the maximum similarity score into normalized attention weights using the Softmax function. These fusion weights represent the importance of each modality in the final fusion. Finally, the combiner uses these normalized weights to sum the feature vectors of all modalities, generating a unified fused feature vector. : ;in, It is the number of modes. It is modal Attention weights.

[0047] As an example of an embodiment of the present invention, the feature processing of each second sensor data through a convolutional neural network to obtain modal feature vectors corresponding to each modality, wherein the second sensor data includes a one-dimensional time series of multiple modalities, and the convolutional neural network includes a convolutional layer, an activation function, and a pooling layer, specifically: for the one-dimensional time series of each modality, the one-dimensional time series is subjected to sliding convolution processing through the convolutional kernel in the convolutional layer; the feature vector obtained by the sliding convolution processing is subjected to nonlinear processing through the activation function to obtain a feature map; and the feature map is downsampled through the pooling layer to obtain the modal feature vectors.

[0048] In this embodiment, for each modality, the encoder is an independent one-dimensional convolutional neural network (1D-CNN), including multiple one-dimensional convolutional layers, activation functions, and pooling layers stacked together; the input is a preprocessed one-dimensional time series for each modality, such as the vibration signal input to a vibration encoder. Temperature encoder input temperature signal Electrical quantity encoder input electrical quantity signal One-dimensional convolutional layers automatically extract features of a modality by sliding the convolution kernel across the time series. (Formula) ;in, It is the first Input sequences for each modality It is the convolution kernel of the modal encoder. For bias vectors, It is the extracted feature map. It is an activation function; after convolution, the pooling layer downsamples the feature map to reduce its dimensionality. Finally, each encoder compresses the original heterogeneous time series data into a fixed-dimensional feature vector, such as the output of a vibration encoder. Temperature encoder output Electrical quantity encoder output .

[0049] As an example of an embodiment of the present invention, the method of using target proximity information recovery filtering technology, combining spatial and temporal dimensions to perform data interpolation processing on missing data in each of the first sensor data to obtain second sensor data, specifically involves: using each of the first sensor data as a multidimensional data matrix to determine the target time and target sensor corresponding to the missing data; determining a first time set within a preset time range for the target time; obtaining a first data set of the target sensor within the first time set using the multidimensional data matrix; determining the time weight corresponding to each of the first data in the first data set based on the time distance between each of the first data and the missing data; determining a first sensor set within a preset correlation range for the target sensor; obtaining a second data set of the first sensor set at the target time using the multidimensional data matrix; and ... and determining a second data set of the first sensor set within a preset correlation range for the target sensor; and obtaining a second data set of the first sensor set at the target time using the second data. The correlation between each second data point in the dataset and the missing data is used to determine the correlation weight of each second data point. Using the multidimensional data matrix, a third data set of the first sensor set within the first time set is obtained. Based on the time distance and correlation between each third data point in the third data set and the missing data, a third weight is determined for each third data point. A first weight coefficient, a second weight coefficient, and a third weight coefficient are obtained. Each first data point and each time weight are weighted according to the first weight coefficient. Each second data point and each correlation weight are weighted according to the second weight coefficient. Each third data point and each third weight are weighted according to the third weight coefficient. The weighted data are summed to obtain the target data. Data interpolation is performed on the missing data based on the target data to obtain the second sensor data.

[0050] In this embodiment, an improved method based on proximity information recovery (AIR) filtering is used to impute missing data. This method utilizes both temporal and spatial neighboring sensor data; it treats data from different sensors as a multidimensional matrix, and when a data point is missing, it uses the spatiotemporal neighborhood data of that data point for imputation. Specifically, let the sensor data matrix be... ,in, For the number of sensors, For the time step, when a certain data point When missing, its interpolated value for: ; in, For time The set of adjacent time points, For sensors The set of adjacent sensors, The weights are based on distance or correlation; the closer the temporal distance, the greater the weight. Similarly, the higher the sensor correlation, the greater the weight. The weighting coefficients are used to balance the contributions of time, space, and spatiotemporal influences.

[0051] As an example of an embodiment of the present invention, the detection of the fused feature vector based on the anomaly detection model specifically involves: inputting the fused feature vector into the convolutional neural network module of the anomaly detection model to extract features from the fused feature vector using a preset convolutional kernel, thereby obtaining a first anomalous feature; inputting the first anomalous feature into the spline-enhanced multilayer perceptron module of the anomaly detection model to parameterize the first anomalous feature using spline functions and Fourier series, thereby performing a nonlinear transformation on the first anomalous feature, and fitting a nonlinear anomalous correlation relationship based on the transformed first anomalous feature to enhance the distinguishability between the anomalous feature and normal features, thereby obtaining a second anomalous feature; and inputting the second anomalous feature into the anomaly detection model after arranging it in chronological order. The Transformer module calculates the second abnormal feature using a multi-head self-attention layer to obtain the attention weights corresponding to each attention head. It then performs a weighted summation of the second abnormal feature based on these attention weights to obtain a third abnormal feature. The classification layer in the anomaly detection model compares the similarity of the third abnormal feature with each abnormal feature template in a preset anomaly feature template library to determine the anomaly type corresponding to the template with the highest matching degree. Feature weight coefficients are obtained, and these coefficients are used to weight and calculate the confidence level of the detection result by weighting the first anomaly probability of the convolutional neural network module, the second anomaly probability of the spline-enhanced multilayer perceptron module, and the third anomaly probability of the Transformer module.

[0052] In this embodiment, as Figure 3 The diagram shown is a structural schematic of a CNN-SEMLP-Attention model provided in an embodiment of the present invention. The CNN layer is composed of multiple one-dimensional convolutional layers and activation layers stacked together, which fuse the feature vector sequence. Input convolutional layer for convolution: ; ; in, For time step, For feature dimension, This is a one-dimensional convolution operation that slides along the time dimension. For the first Learnable weights of convolutional layers, The kernel size is [size]. For the input feature dimension, The number of convolution kernels, For learnable bias terms, For activation function, For the first Output feature map of layer convolution; The final feature map obtained after multiple convolutions is denoted as... ,Will The flattened input serves as the input to the SEMLP layer. The SEMLP layer is not a single mathematical function, but a network layer composed of multiple neurons. The output of each neuron is obtained by summing the inputs from the previous layer after they have been transformed by independent, learnable univariate functions. The SEMLP layer can be understood as a network that transforms the input vector... To output vector The mapping, input vector Output feature vectors for the flattened CNN layers , It is the dimension of the input features, and the output vector. , This is the dimension output by the SEMLP layer; the first dimension of the SEMLP layer. one neuron Output value for: ; in, Output vector of SEMLP layer The The component (i.e., the first component) (output of each neuron) For input vectors The One portion, For a learnable univariate function, the first variable of the input vector is... Each component After transformation, as the first The contribution of each neuron; among which Parameterization is performed using Fourier series: ; in, For learnable constant terms, The order of the Fourier series determines the complexity and accuracy of the function fitting. For the first The first neuron and the second The first input connection cosine coefficients, For the first The first neuron and the second The first input connection The first-order sine coefficients; the complete SEMLP layer consists of The above neurons are combined to ultimately output a vector. Represented as: ; Each component of the vector All of them are calculated using the formulas mentioned above.

[0053] The input to the Transformer attention layer is the final output vector of the SEMLP layer. The system employs a self-attention mechanism to establish global associations and long-distance dependencies between features. This self-attention mechanism is a multi-head self-attention mechanism, which projects the input sequence onto multiple heads. Each head independently learns different association patterns, and the results are then integrated to obtain the final fault diagnosis probability vector.

[0054] The anomaly detection model provided in this embodiment optimizes all learnable parameters in the model by minimizing a loss function, which is the cross-entropy loss function: ,in, The loss value. Number of fault categories For real labels, This is the predicted probability vector.

[0055] As an example of an embodiment of the present invention, the step of calculating the second abnormal feature through a multi-head self-attention layer to obtain the attention weights corresponding to each attention head, and then performing a weighted summation of the second abnormal feature based on each attention weight to obtain the third abnormal feature, specifically: The second abnormal feature is projected onto multiple projection matrices of each attention head to obtain the query vector, key vector, and value vector corresponding to each second abnormal feature; the dot product of the query vector and the key vector is calculated to obtain the attention score corresponding to each attention head; each attention score is processed by a normalization function to obtain each attention weight; the value vectors of each attention head are weighted and summed according to each attention weight to obtain an initial vector, and the initial vector is projected onto a linear layer in the multi-head self-attention layer to obtain the third abnormal feature.

[0056] In this embodiment, each attention head needs to process the input sequence. Projecting onto three different learnable spaces yields the query. ,key Sum Vector, for the th A person's attention : ; ; ; in, For the number of attention heads, For the first Learnable query projection matrix of size For the first Learnable key projection matrix of size For the first The learnable value projection matrix of the head; Calculate the attention score for each attention head: ;in, This is a scaling factor to prevent the dot product result from becoming too large, which could lead to gradient vanishing. Then calculate the attention weights: ;in, Convert the score to a range of 0 to 1; The outputs of all attention heads are concatenated and projected through a final linear layer to obtain the final output. : ; ;in, Concatenate all the output headers along the feature dimension. The final projection matrix is ​​learnable; Finally, a fully connected layer and a softmax activation function are used for final classification. ; ;in, This is the weight matrix of the fully connected layer. For bias terms of fully connected layers; It is a fractional vector. For activation function, This represents the final fault diagnosis probability vector.

[0057] As an example of an embodiment of the present invention, the loss function of the physical information neural network model is determined based on the physical information loss term, the boundary and initial condition loss term, and the data loss term, specifically as follows: The physical equation residuals are calculated based on the prediction results and the physical equations, and the mean square error of the physical equation residuals is calculated to obtain the physical information loss term. The constraint error is calculated based on the prediction results and the physical constraints, and the mean square error of the constraint error is calculated to obtain the boundary and initial condition loss term. The mean square error between the prediction results and the sensor data is calculated to obtain the data loss term. The loss weight parameters are dynamically determined based on the gradient norm normalization algorithm, and the physical information loss term, the boundary and initial condition loss term, and the data loss term are weighted and summed according to the loss weight parameters to obtain the loss value corresponding to the loss function.

[0058] In this embodiment, as Figure 4The diagram shows the structure of the PINN model. By embedding multiple physical equations into PINN, it generates high-precision predictions for multiphysics, which is crucial for subsequent fault attribution and health assessment. The total loss function of PINN is a weighted sum of various losses, including physical information loss, boundary and initial condition loss, and data loss. The total loss function is expressed as follows: ;in, Let be the total loss function, and the goal of training is to minimize this value; This is the physical information loss term, used to measure the degree of agreement between the PINN prediction results and the physical equations; The boundary and initial condition loss terms are used to ensure that the prediction results meet the constraints of the physical system. This is the data loss term, used to measure the error between the neural network's predictions and the actual measured data. These are learnable weight hyperparameters to balance the importance of each loss term.

[0059] Physical information loss item The forced prediction results must comply with physical rules, namely the governing partial differential equations (PDEs): ;in, The number of physical points randomly sampled in the training domain. For the first Spatial and temporal coordinates of a physical point This is the residual function of the physical equations, used to calculate the error of the physical equations; the governing partial differential equations include, but are not limited to: Maxwell's equations and the heat equations; for the heat equations... ,in, For density, For specific heat capacity, For temperature, Thermal conductivity, Let the volumetric heat source intensity be , and its residual function be . ,in, To predict the output, and Obtained through automatic differentiation techniques; for Maxwell's equations: , , , ,in, For curl operator, For divergence operators, The magnetic field strength, For current density, For electric field strength, Where is the dielectric constant. Let be the charge density, and its residual function be: ; in, , , and To predict the output, , , and Obtained through automatic differentiation techniques; Boundary and initial condition loss terms for: ; in, The number of constraint points randomly sampled at the boundary or initial time. and For the first The spatial and temporal coordinates of several constraint points located at the boundaries of the physical domain or... time, For predicted values, The physical values ​​are known. Data loss items for: ; in, The number of data points measured by the sensor. and For the first Spatial and temporal coordinates of each data point For the predicted values ​​of these data points, The actual physical value measured by the sensor; An adaptive weighting method based on the GradNorm algorithm is adopted. GradNorm dynamically adjusts the weight of each loss term to ensure that the gradient norm of all loss terms converges at a similar speed, thereby avoiding the situation where the gradient of a certain loss term is too large and dominates the entire training process, thus preventing other loss terms from learning effectively.

[0060] The total loss function is composed of ( The components of the sub-loss terms are: ; in, This represents all trainable parameters (weights and biases) in PINN. For the first The learnable weights of each loss term are initialized to 1 and dynamically adjusted during training.

[0061] First, for each loss item Calculate its weights with respect to the penultimate layer of the network (or other selected shared parameter layers). Gradient norm: ; in, Regarding shared parameters gradient, It is the L2 norm. For the first The gradient norm of each loss term; Then calculate the equilibrium loss function. : ; ; ; ; in, The average gradient norm, For averaging operations, It is a relative loss, for each loss item Loss at the start of training The ratio, The relative inverse training rate represents the training difficulty of each task; the better the task is trained, the higher the training difficulty. The smaller, This represents the average relative loss across all tasks. To balance the loss function; The network parameters and loss weights are updated using gradient descent. ; ; in, For learning rate, The learning rate is used to update the weights; after updating, the weights are normalized to obtain the loss weight coefficients. .

[0062] like Figure 5 As shown, a smoothing reactor condition assessment system is provided, comprising: A multimodal sensor data acquisition module is configured to acquire raw heterogeneous sensor data of different modes from the smoothing reactor; A data preprocessing and fusion module, connected to the multimodal sensor data acquisition module, is configured to clean, normalize, imputate missing data, and fuse multi-sensor data on the raw heterogeneous sensor data to generate a fused feature vector. The data preprocessing and fusion module includes a missing data imputation submodule and a multi-sensor data fusion submodule. The missing data imputation submodule is configured to imput missing data using an improved method based on proximity information recovery filtering. The multi-sensor data fusion submodule is configured to employ a feature-level fusion strategy and use a specific encoder-combiner architecture to process the heterogeneous sensor input. The hybrid modeling and analysis module, connected to the data preprocessing and fusion module, includes a data-driven submodule, a physics submodule, and a coupling module. The data-driven submodule is configured to receive the unified fused feature vector and perform anomaly detection on the unified fused feature vector based on a CNN-SEMLP-Attention architecture. The physics submodule is configured to establish a physical model of the smoothing reactor's multi-physics fields (electromagnetic field, thermal field) and perform high-fidelity prediction of the temperature or electromagnetic field of the smoothing reactor based on an adaptive weighted physical information neural network (AW-PINN). The coupling module is configured to dynamically exchange data and feedback between the data-driven submodule and the physics submodule through an event-driven co-simulation framework. The status assessment and decision support module, connected to the hybrid modeling and analysis module, is configured to generate fault diagnosis results, fault source analysis results, and real-time status warnings for the smoothing reactor based on the output of the hybrid modeling and analysis module. The status assessment and decision support module includes a fault diagnosis output head, a fault source analysis output head, and a real-time status warning output head. The fault diagnosis output head is used to identify the fault type and severity of the smoothing reactor. The fault source analysis output head is used to analyze the root cause of the fault, considering system operating conditions (normal, HVDC commutation failure, DC line grounding fault). The real-time status warning output head is used to issue fault warnings based on the prediction results of the physical submodule.

[0063] The reporting module is configured to present the status assessment results.

[0064] like Figure 6 As shown, based on the above method embodiments, corresponding system embodiments are provided; One embodiment of the present invention provides a state assessment system 600, including: a feature processing module 601, an anomaly detection module 602, an event-driven module 603, a simulation prediction module 604, and a confidence update and assessment module 605; The feature processing module 601 is used to acquire multimodal sensor data, preprocess and fuse the sensor data to obtain a fused feature vector; The anomaly detection module 602 is used to detect the fused feature vector based on the anomaly detection model to obtain the detection result. The anomaly detection model is a hybrid architecture model built based on spline-enhanced multilayer perceptron, convolutional neural network and Transformer attention mechanism. The detection result includes anomaly type and confidence level. The event-driven module 603 is used to generate an abnormal event according to the event-driven mechanism when the detection result is abnormal. The simulation prediction module 604 is used to respond to the abnormal event by acquiring time-series data corresponding to the abnormal type and the time window of the abnormal event from historical sensor data, and performing simulation prediction on the time-series data according to the physical information neural network model to obtain the prediction result. The confidence update and evaluation module 605 is used to feed the prediction result back to the anomaly detection model, update the confidence level in the detection result according to the prediction result to obtain the target detection result, and determine the state evaluation result according to the target detection result and the prediction result.

[0065] It is understood that the above system item embodiments correspond to the method item embodiments of the present invention, and can implement the state evaluation method provided by any of the above method item embodiments of the present invention.

[0066] It should be noted that the system embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0067] For ease of description and brevity, the system embodiments of the present invention include all the implementation methods described in the above embodiments of the UAV-based information recognition method, and will not be repeated here.

[0068] Based on the above-described embodiments of the state assessment method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the state assessment method of any embodiment of the present invention.

[0069] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0070] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0071] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0072] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the state assessment method described in any of the above-described method embodiments of the present invention.

[0073] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0074] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A state assessment method, characterized in that, include: Acquire multimodal sensor data, preprocess and fuse the sensor data to obtain a fused feature vector; The fused feature vector is detected based on an anomaly detection model to obtain detection results. The anomaly detection model is a hybrid architecture model built on spline-enhanced multilayer perceptron, convolutional neural network and Transformer attention mechanism. The detection results include anomaly type and confidence level. When the detection result is abnormal, an abnormal event is generated according to the event-driven mechanism; In response to the abnormal event, time-series data corresponding to the abnormal type and the time window of the abnormal event are obtained from historical sensor data, and the time-series data are simulated and predicted according to the physical information neural network model to obtain the prediction result; The prediction results are fed back to the anomaly detection model, and the confidence level in the detection results is updated according to the prediction results to obtain the target detection results. The state assessment results are then determined based on the target detection results and the prediction results.

2. The state assessment method as described in claim 1, characterized in that, The preprocessing and fusion of the sensor data specifically involves: Wavelet denoising algorithm is used to denoise the sensor data of each mode, and the denoised sensor data is normalized to obtain the first sensor data corresponding to each mode. The target proximity information recovery filtering technique is used to perform data interpolation on the missing data in each of the first sensor data by combining spatial and temporal dimensions to obtain the second sensor data. The second sensor data is processed by a convolutional neural network to obtain modal feature vectors corresponding to each modality. A fusion weight corresponding to each modal feature vector is determined by an attention mechanism. The modal feature vectors are fused according to the fusion weights to obtain the fused feature vector.

3. The state assessment method as described in claim 2, characterized in that, The second sensor data is obtained by employing target proximity information recovery filtering technology, which combines spatial and temporal dimensions to perform data interpolation on missing data in each of the first sensor data, specifically as follows: The data from each of the first sensors are used as a multi-dimensional data matrix to determine the target time and target sensor corresponding to the missing data. A first time set within a preset time range is determined, and a first data set of the target sensor within the first time set is obtained through the multidimensional data matrix. Based on the time distance between each first data in the first data set and the missing data, the time weight corresponding to each first data is determined. A first set of sensors within a preset correlation range is determined for the target sensors. A second set of data for the first sensor set at the target time is obtained through the multidimensional data matrix. The correlation weight corresponding to each second data is determined based on the correlation between each second data in the second set and the missing data. The third data set of the first sensor set within the first time set is obtained through the multidimensional data matrix. Based on the time distance and correlation between each third data in the third data set and the missing data, the third weight corresponding to each third data is determined. Obtain a first weight coefficient, a second weight coefficient, and a third weight coefficient. Weight each first data point and each time weight according to the first weight coefficient. Weight each second data point and each related weight according to the second weight coefficient. Weight each third data point and each third weight according to the third weight coefficient. Sum the weighted data to obtain the target data. The missing data is interpolated based on the target data to obtain the second sensor data.

4. The state assessment method as described in claim 1, characterized in that, The detection of the fused feature vector based on the anomaly detection model is specifically as follows: The fused feature vector is input into the convolutional neural network module in the anomaly detection model to extract features from the fused feature vector through a preset convolutional kernel, thereby obtaining the first anomaly feature; The first abnormal feature is input into the spline-enhanced multilayer perceptron module in the abnormality detection model to parameterize the first abnormal feature through spline functions and Fourier series, to perform nonlinear transformation on the first abnormal feature, and to fit a nonlinear abnormal correlation relationship based on the transformed first abnormal feature to enhance the distinguishability between abnormal features and normal features, thereby obtaining the second abnormal feature. The second abnormal feature is arranged in chronological order and then input into the Transformer module in the anomaly detection model. The second abnormal feature is calculated through a multi-head self-attention layer to obtain the attention weights corresponding to each attention head. The second abnormal feature is then weighted and summed according to the attention weights to obtain the third abnormal feature. The classification layer in the anomaly detection model compares the third anomaly feature with each anomaly feature template in the preset anomaly feature template library to determine the anomaly type corresponding to the anomaly feature template with the highest matching degree. The feature weight coefficients are obtained, and the first anomaly probability of the convolutional neural network module, the second anomaly probability of the spline-enhanced multilayer perceptron module, and the third anomaly probability of the Transformer module are weighted and calculated based on the feature weight coefficients to obtain the confidence level of the detection result.

5. The state assessment method as described in claim 1, characterized in that, The loss function of the physical information neural network model is determined based on the physical information loss term, the boundary and initial condition loss term, and the data loss term, specifically as follows: The physical equation residuals are calculated based on the prediction results and the physical equations, and the mean square error of the physical equation residuals is calculated to obtain the physical information loss term. The constraint error is calculated based on the prediction results and physical constraints, and the mean square error of the constraint error is calculated to obtain the boundary and initial condition loss terms. Calculate the mean square error between the prediction result and the sensor data to obtain the data loss term; The loss weight parameters are dynamically determined based on the gradient norm normalization algorithm, and the physical information loss term, the boundary and initial condition loss term, and the data loss term are weighted and summed according to the loss weight parameters to obtain the loss value corresponding to the loss function.

6. The state assessment method as described in claim 2, characterized in that, The second sensor data is processed using a convolutional neural network to obtain modal feature vectors corresponding to each modality. The second sensor data includes a one-dimensional time series of multiple modalities. The convolutional neural network includes convolutional layers, activation functions, and pooling layers, specifically: For each modality, the one-dimensional time series is processed by sliding convolution through the convolution kernel in the convolutional layer; The feature vector obtained by sliding convolution is nonlinearly processed by the activation function to obtain a feature map; The feature map is downsampled by the pooling layer to obtain the feature vectors of each modality.

7. The state assessment method as described in claim 4, characterized in that, The second abnormal feature is calculated using a multi-head self-attention layer to obtain the attention weights corresponding to each attention head. The second abnormal feature is then weighted and summed according to these attention weights to obtain the third abnormal feature. Specifically: The second abnormal feature is projected by multiple projection matrices of each attention head to obtain the query vector, key vector and value vector corresponding to each second abnormal feature; Calculate the dot product of the query vector and the key vector to obtain the attention score corresponding to each attention head; The attention scores are processed by a normalization function to obtain the attention weights. The value vectors of each attention head are weighted and summed according to the attention weights to obtain an initial vector. The initial vector is then projected through a linear layer in the multi-head self-attention layer to obtain the third abnormal feature.

8. A state assessment system, characterized in that, include: The module includes a feature processing module, an anomaly detection module, an event-driven module, a simulation prediction module, and a confidence update and evaluation module. The feature processing module is used to acquire multimodal sensor data, preprocess and fuse the sensor data to obtain a fused feature vector; The anomaly detection module is used to detect the fused feature vector based on the anomaly detection model to obtain the detection result. The anomaly detection model is a hybrid architecture model built based on spline-enhanced multilayer perceptron, convolutional neural network and Transformer attention mechanism. The detection result includes anomaly type and confidence level. The event-driven module is used to generate an abnormal event according to the event-driven mechanism when the detection result is abnormal. The simulation prediction module is used to respond to the abnormal event by acquiring time-series data corresponding to the abnormal type and the time window of the abnormal event from historical sensor data, and performing simulation prediction on the time-series data according to the physical information neural network model to obtain the prediction result. The confidence update and evaluation module is used to feed the prediction result back to the anomaly detection model, update the confidence level in the detection result according to the prediction result to obtain the target detection result, and determine the state evaluation result according to the target detection result and the prediction result.

9. A terminal device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the state assessment method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, include: A stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the state assessment method as described in any one of claims 1-7.

Images