Intelligent early warning system and method for power equipment state based on multi-source data fusion

Abstract

The application discloses a power equipment state intelligent early warning system and method based on multi-source data fusion, and the early warning system comprises a data acquisition module, which is used for collecting multi-dimensional operation data of power equipment, including equipment operation parameter data, environment sensing data, image detection data and historical operation and maintenance data; a data fusion module, which receives the multi-dimensional operation data for preprocessing, generates a standardized processing data set, and performs fusion processing on the data set. The application belongs to the technical field of power equipment state monitoring and early warning, can cover full-dimensional data of equipment operation, environment, image and operation and maintenance, eliminates the limitation of a single data source, provides complete and reliable data support for subsequent analysis, is efficient and intelligent in fusion processing, extracts core features of multiple types of data through CNN and LSTM, combines genetic and reinforcement learning to optimize fusion weights, greatly improves the relevance and effectiveness of fused data, and avoids heterogeneous data fusion deviation.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and early warning technology, specifically to a power equipment condition intelligent early warning system and method based on multi-source data fusion. Background Technology

[0002] Power equipment is the core carrier for the safe and stable operation of the power system. Its operating status directly determines the reliability, security and continuity of power supply. As the power system upgrades to intelligence and scale, the operating conditions of power equipment such as transformers, switchgear, and wind turbines are becoming increasingly complex. Various potential faults and hidden dangers can easily lead to serious consequences such as power outages and equipment damage. Therefore, accurate monitoring, quantitative assessment and intelligent early warning of the operating status of power equipment have become the core requirements of the power operation and maintenance field. Existing power equipment status early warning technologies have significant limitations. They have a single data collection dimension, focusing only on equipment operating parameters and failing to fully integrate multi-source data such as environmental perception, image detection, and historical operation and maintenance. This results in incomplete data coverage, making it difficult to reflect the overall operating status of the equipment. The fusion effect of multi-source data is poor, and various types of data have heterogeneity and redundancy problems, making it difficult to achieve accurate quantitative assessment of equipment status. In summary, existing power equipment status early warning technologies suffer from incomplete data collection, low fusion accuracy, inaccurate assessment, unintelligent early warning, and poor human-machine interaction, failing to meet the high-quality requirements of modern power systems for intelligent and precise equipment operation and maintenance. Therefore, there is an urgent need to develop intelligent early warning systems and methods for power equipment status based on multi-source data fusion to solve the problems in existing technologies. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent early warning system and method for power equipment status based on multi-source data fusion. This invention aims to solve the problems in the prior art where inaccurate status assessment is caused by single data acquisition and the poor effect of multi-source data fusion affects the accuracy of early warning.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent early warning system for power equipment status based on multi-source data fusion, the early warning system comprising: The data acquisition module is used to collect multi-dimensional operational data of power equipment, including equipment operating parameter data, environmental perception data, image detection data, and historical operation and maintenance data; it realizes multi-dimensional and full-element data collection, covering operating parameters, environment, images, and historical operation and maintenance data, providing more comprehensive information; The data fusion module receives multi-dimensional operational data, preprocesses it to generate a standardized dataset, and then fuses the dataset. It extracts image data features using a Convolutional Neural Network (CNN) and extracts temporal operational parameter features using a Long Short-Term Memory (LSTM) network. Based on a genetic algorithm combined with reinforcement learning, it dynamically optimizes the fusion weights of each data source, performs feature-level fusion of the extracted multi-dimensional features, obtains a fused feature vector, and outputs the fused data. It also standardizes, denoises, and normalizes multi-source data to improve data quality and usability. The module uses CNN to extract image features and LSTM to extract temporal parameter features, simultaneously capturing spatial visual information and temporal series variation patterns. The state assessment module is used to construct a power equipment state assessment model based on an improved BP neural network. This improved BP neural network introduces a sparse training mechanism and uses L1 regularization to constrain the network's internal parameters, eliminating redundant parameters and training and optimizing the assessment model. It receives fused feature vectors, inputs them into the state assessment model, outputs the current power equipment's health status index, and, based on a preset health status classification standard, outputs the corresponding power equipment's operating status assessment level, achieving a quantitative assessment of the power equipment's state. The improved BP neural network, incorporating sparse training and L1 regularization, eliminates redundant parameters, reduces model size, and accelerates inference speed. The early warning module generates corresponding early warning information for different assessment levels based on preset early warning thresholds and performs different early warning operations for different early warning information; it automatically calibrates the early warning thresholds based on changes in the equipment operating environment and historical false alarm data; and it provides graded early warnings to avoid the numbness and missed judgments caused by uniform alarms. The human-machine interaction module is used to receive early warning information, allowing users to view the current operating status of power equipment in real time; it intuitively displays equipment status and early warning information, facilitating real-time monitoring and rapid response by operation and maintenance personnel.

[0005] As a further aspect of the present invention: the device operating parameter data in the data acquisition module is collected through multi-protocol acquisition via the connection between the measurement and control terminal and the intelligent sensor; the environmental perception data is based on the deployment of distributed low-power IoT sensor nodes and is collected periodically according to environmental scenarios; the image detection data is collected in real time based on the intelligent camera and thermal imager terminal; and the historical operation and maintenance data is obtained synchronously through online system integration, offline file import, and manual supplementation.

[0006] As a further aspect of the present invention: the preprocessing within the data fusion module includes data cleaning and data standardization; the preprocessed data is classified and summarized to obtain a standardized dataset; before the dataset is fused, multi-source data under the same power equipment and the same time dimension are selected based on data association identifiers to obtain a fused data group, and multi-dimensional feature extraction is performed on the fused data group.

[0007] As a further aspect of this invention: the image data feature extraction within the data fusion module takes preprocessed standardized image data as input, uses a convolutional neural network (CNN) to build a feature extraction model, and gradually extracts low-level, mid-level, and high-level semantic features of the image through alternating operations of convolutional and pooling layers; the extracted multi-layer features are integrated and dimensionality reduced through a fully connected layer, and the output is a fixed-dimensional image feature vector, retaining the feature confidence, thus completing the transformation of image data from unstructured to feature vector; Time-series runtime parameter feature extraction involves inputting a standardized dataset, setting a time window, and dividing the input sequence into a feature sequence. A time series feature extraction model is built using a Long Short-Term Memory (LSTM) network. Its gating mechanism is used to capture the long-term dependencies and dynamic changes in time series data. Overfitting is suppressed by using a dropout layer, and a time-series running parameter feature vector with the same dimension as the image feature vector is output through the output layer. The time-series feature stability index is output simultaneously to complete the feature extraction of time-series data.

[0008] As a further aspect of the present invention: the data fusion module performs feature-level fusion on the extracted multi-dimensional features to obtain a fused feature vector, the steps of which are as follows: Auxiliary features are extracted from environmental perception and historical operation and maintenance data, quantized and encoded, and then integrated into an auxiliary feature vector. A hybrid optimization strategy combining genetic algorithm and reinforcement learning is adopted to dynamically adjust the fusion weights of image feature vector, temporal running parameter features and auxiliary feature vector; Set initial fusion weights for the three types of feature vectors, and define the weight constraint range based on historical operation and maintenance data; Using the accuracy and stability of the fused features as the fitness function, the initial weights are globally optimized through selection, crossover, and mutation operations to select the optimal weight solution set as the initial action space for reinforcement learning. Using the equipment operating status as environmental feedback, weight adjustment as action, and minimization of fusion error as reward function, the optimal weight solution set selected by the genetic algorithm is locally fine-tuned to dynamically adapt to changes in the real-time operating status of the equipment. The process is iteratively optimized until convergence, with the convergence condition being: the fusion error is lower than a set threshold. The optimal fusion weights of each feature vector are then output to form a dynamic weight matrix. Based on the optimized dynamic weight matrix, feature-level fusion operations are performed on the image feature vector extracted by CNN, the time-series running parameter feature vector extracted by LSTM, and the auxiliary feature vector. The three types of feature vectors are multiplied by their corresponding optimal weights to obtain the weighted feature vector. The weighted feature vectors are fused, spliced, and dimension-normalized to transform them into standardized fused data.

[0009] As a further aspect of the present invention: the steps for constructing a state assessment model within the state assessment module are as follows: The network hierarchy was determined, adopting a three-layer structure of input layer-hidden layer-output layer, and the number of neurons in each layer was determined. The number of neurons in the input layer was consistent with the feature dimension after the fused data was filtered. The hidden layer was set to 2-3 layers, and the initial number of neurons was set according to an empirical formula, reserving space for pruning during sparse training. The number of neurons in the output layer was 4, corresponding to the four types of device status labels. The input layer uses a linear activation function, the hidden layer uses the ReLU (Modified Linear Unit) activation function, and the output layer uses the Softmax activation function. By embedding an L1 regularization term into the network loss function, a new loss function is constructed to constrain the internal connection weight parameters of the network. Its expression is: Loss = Original loss function + λ × Σ|w| Where λ is the regularization coefficient and w is the internal weight parameter of the network, and weight sparsity is achieved through L1 regularization; By setting a sparsity threshold, the output of hidden layer neurons is constrained during network training to achieve network sparsity. Model training and optimization involve inputting the training set into the network, obtaining the device state probability prediction value through forward propagation, recording the hidden layer output, calculating the loss function with L1 regularization term in combination with the real label, and adding a penalty term if the sparsity threshold is not met; the process is repeated for iterative training.

[0010] As a further aspect of the present invention: the evaluation steps within the state evaluation module are as follows: The system receives the standardized fusion feature vector output by the data fusion module in real time, acquires the vector association identifier simultaneously, and inputs the preprocessed fusion feature vector into the state assessment model. Through forward propagation operation, the fusion feature vector is analyzed in multiple dimensions, and combined with the equipment state feature patterns learned during training, the system outputs the current health status index of the power equipment. The health status index is set to 0-100 points. The preset power equipment health status classification standard is called, and the calculated health status index is compared with the preset classification standard to match the corresponding operation status assessment level and clarify the current operation status and potential risks of the equipment. Integrate the evaluation results, form a standardized evaluation report, and then output it.

[0011] As a further aspect of the present invention: the early warning module generates differentiated early warning information based on preset graded early warning thresholds, corresponding to different assessment levels, and performs corresponding differentiated operations, from basic log recording to emergency sound and light warnings and emergency linkage, while retaining complete operation logs. The module combines dynamic data of the equipment's operating environment and historical false alarm records, and uses a combination of periodic calibration and real-time calibration to adaptively and automatically calibrate the warning threshold.

[0012] As a further aspect of the present invention: the human-computer interaction module adopts an intuitive and visual interface design, which is compatible with both computer terminals and mobile APPs; the core display area of ​​the interface includes three major sections: overall device operation overview, hierarchical warning list, and detailed status of individual devices; users can quickly locate target devices through filtering and search functions and view their integrated data, evaluation reports and historical operation records.

[0013] The intelligent early warning method for power equipment status based on multi-source data fusion comprises the following steps: S1. Data Acquisition: Collect multi-dimensional operational data of power equipment, specifically including four types of data: equipment operating parameters, environmental perception, image detection, and historical operation and maintenance data. S2. Data Fusion: Receive multi-dimensional data, preprocess it to generate a standardized dataset; extract image data features through CNN, extract time-series running parameter features through LSTM, combine genetic algorithm and reinforcement learning to dynamically optimize the fusion weights of each data source, perform feature-level fusion of multi-dimensional features, obtain fused feature vectors and output fused data. S3. Status Assessment: Construct an improved BP neural network model, introduce a sparsity training mechanism, and use L1 regularization to constrain parameters and eliminate redundancy to complete model training optimization; receive fused feature vectors as input to the model, output the device health status index, and output the operating status assessment level based on preset division standards. S4. Early Warning Handling: Based on preset thresholds, generate corresponding early warning information for different assessment levels and execute differentiated early warning operations; combine changes in equipment operating environment and historical false alarm data to automatically calibrate early warning thresholds and reduce false alarm and missed alarm rates; S5. Human-computer interaction: Receives early warning information and provides a visual operation interface, allowing users to view the current operating status of the device and early warning details in real time, realizing the linkage between the system and the user.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The data acquisition module of this invention collects data from multiple sources, including measurement and control terminals, smart sensors, IoT sensing nodes, and smart cameras. Compared with single data sources in existing technologies, it can provide more complete and reliable data support and reduce data limitations. The data fusion module uses CNN to extract image features and LSTM to extract temporal features, and combines genetic algorithms and reinforcement learning to dynamically optimize the fusion weights. Compared with fixed-weight fusion methods in existing technologies, it can improve the correlation and effectiveness of fused data and avoid fusion bias of heterogeneous data. The state evaluation module introduces sparse training and L1 regularization through an improved BP neural network. Compared with unoptimized neural networks in existing technologies, it can eliminate redundant parameters, improve model accuracy, and achieve more accurate quantitative evaluation and level classification.

[0015] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0016] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the steps of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] refer to Figure 1 As shown in the embodiment of the present invention, the intelligent early warning system for power equipment status based on multi-source data fusion includes: In this embodiment, the power equipment status assessment and early warning system is mainly applied to core power equipment such as main transformers, circuit breakers, and disconnect switches in 110kV substations. The system adopts a three-layer architecture of "terminal data acquisition - edge processing - cloud management and control", as detailed below: Terminal acquisition layer: Deploy various acquisition terminals to implement the functions of the corresponding data acquisition modules, including power parameter acquisition devices, environmental sensors, high-definition cameras, and operation and maintenance data acquisition terminals. All terminals establish communication with the edge computing gateway through industrial Ethernet. The communication protocol adopts the IEC 61850 (International Electrotechnical Commission 61850 standard) standard to ensure the real-time performance and reliability of data transmission, and control the transmission delay within 500ms. Edge processing layer: Deploys edge computing gateways and data processing servers, integrates the core algorithms of data fusion module and status assessment module, and is responsible for real-time preprocessing, fusion and status assessment of collected multi-dimensional data, avoiding bandwidth occupation caused by a large amount of raw data being transmitted to the cloud. The edge processing server uses Intel Core i7-12700H CPU and 32GB of memory to ensure algorithm running efficiency. Cloud-based control layer: Deploys cloud servers and a human-computer interaction platform, integrating the functions of the early warning module and the human-computer interaction module. It is responsible for receiving fused data and status assessment results transmitted from the edge processing layer, generating early warning information, and allowing users to view and operate in real time through the human-computer interaction platform. The cloud server adopts a distributed deployment to ensure the stability and scalability of the system.

[0019] The data acquisition module collects equipment operating parameter data through multi-protocol acquisition via the connection between the measurement and control terminal and the smart sensor; environmental perception data is based on the deployment of distributed low-power IoT sensor nodes and is collected periodically according to environmental scenarios; image detection data is collected in real time based on the smart camera and thermal imager terminal; historical operation and maintenance data is obtained synchronously through online system integration, offline file import and manual supplementation.

[0020] The data acquisition module is used to collect multi-dimensional operational data from power equipment. It adopts a "multi-terminal collaborative acquisition + on-demand sampling + data caching" approach. The specific acquisition content, acquisition terminals, and acquisition process are as follows: Content to be collected and corresponding collection terminals Equipment operating parameter data: Collect electrical and mechanical parameters during the operation of power equipment, specifically including the winding temperature, oil level, no-load loss, and load loss of the main transformer; the opening and closing coil current, operating time, and contact resistance of the circuit breaker; and the contact temperature and operating torque of the disconnecting switch. The data acquisition terminal adopts a high-precision power parameter acquisition device with a sampling accuracy of 0.5 class. The sampling frequency can be dynamically adjusted according to the equipment type, ranging from 1Hz to 10Hz. The sampling frequency of key parameters (such as winding temperature and opening and closing coil current) is set to 10Hz, while the sampling frequency of non-key parameters is set to 1Hz.

[0021] Environmental sensing data: The system collects data on temperature, humidity, dust concentration, vibration amplitude, and electromagnetic field strength of the operating environment of the power equipment. The data acquisition terminal uses a multi-functional environmental sensor with a temperature acquisition range of -40℃ to 85℃ and an accuracy of ±0.5℃; a humidity acquisition range of 0%RH to 100%RH and an accuracy of ±3%RH; and a vibration amplitude acquisition range of 0.1mm to 10mm with a sampling frequency of 5Hz, ensuring accurate capture of the impact of environmental factors on the operating status of the power equipment.

[0022] Image inspection data: Collect images of the external appearance and internal components of the power equipment, including whether the equipment casing is damaged, whether the insulators are stained or damaged, and whether the terminals are loose or show signs of overheating. The acquisition terminal uses a high-definition infrared camera and a visible light camera with a resolution of 1920×1080 and a frame rate of 25fps. The camera is installed 1.5m-3m in front of the key parts of the equipment, and the angle is adjustable to ensure the integrity and clarity of the image acquisition. The infrared camera can capture images of the surface temperature distribution of the equipment, providing support for subsequent fault identification.

[0023] Historical Operation and Maintenance Data: Collect historical maintenance records, fault records, calibration records, and replacement component records of power equipment. The data collection terminal adopts the operation and maintenance data collection terminal, which automatically captures historical operation and maintenance data by connecting to the existing operation and maintenance management system (MIS system) of the substation. It also supports manual supplementation of missing data. The data format is uniformly JSON (JavaScript Object Notation) to ensure data standardization. The scope of historical data collection covers at least the operation and maintenance records of the equipment for the past 3 years, providing data support for the training of the condition assessment model and the calibration of early warning thresholds.

[0024] Specific implementation of the data fusion module The data fusion module receives multi-dimensional operational data transmitted from the data acquisition module and generates a fused feature vector through a process of "preprocessing - feature extraction - weight optimization - feature fusion". The specific implementation details are as follows: Data preprocessing The purpose of data preprocessing is to eliminate the heterogeneity and redundancy of multi-dimensional data and generate a standardized processed dataset. Specific steps include: Data cleaning: Remove outliers, missing values, and duplicate values ​​from the collected data. Outliers are identified and removed using the 3σ criterion, missing values ​​are supplemented using linear interpolation, and duplicate values ​​are removed directly to ensure data accuracy. Data normalization: Since different types of data have different units (such as temperature in °C and resistance in Ω), the Min-Max normalization algorithm is used to map all data to the [0,1] interval. Data standardization: The normalized dataset is standardized to make the mean 0 and the variance 1. The Z-Score standardization algorithm is used to generate a standardized dataset to ensure the consistency of data in all dimensions.

[0025] Feature extraction Different feature extraction algorithms are used to extract image data features and time-series runtime parameter features for different types of data in the standardized processing dataset: Image data feature extraction: Convolutional Neural Networks (CNNs) are used to extract features from image detection data. The CNN model adopts the LeNet-5 architecture, specifically including an input layer, convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, fully connected layer 1, fully connected layer 2, and output layer. The input layer takes image data (visible light image) with a size of 1920×1080×3 or infrared image data with a size of 1920×1080×1. Convolutional layer 1 uses 32 3×3 convolutional kernels with a stride of 1 and padding mode Sa. The activation function is ReLU; pooling layer 1 uses a 2×2 max pooling kernel with a stride of 2; convolutional layer 2 uses 64 3×3 convolutional kernels with a stride of 1 and padding method Same. The activation function is ReLU; pooling layer 2 uses a 2×2 max pooling kernel with a stride of 2; the number of neurons in fully connected layer 1 is 1024, the number of neurons in fully connected layer 2 is 256, and the output layer outputs an image feature vector with a dimension of 128, which contains key features of the image such as texture, contour, and temperature distribution.

[0026] Temporal Operation Parameter Feature Extraction: A Long Short-Term Memory (LSTM) network is used to extract temporal features from equipment operation parameter data and environmental perception data. The LSTM model includes an input layer, a hidden layer, and an output layer. The input layer takes standardized temporal data as input (dimension N×M, where N is the time step, set to 60, and M is the number of temporal parameters, determined according to the equipment type, ranging from 8 to 12). The hidden layer contains two LSTM units, each with 64 units. The activation function is tanh (hyperbolic tangent function), and the activation functions for the forget gate, input gate, and output gate are sigmoid (S-shaped function). The output layer outputs a temporal feature vector with a dimension of 128, which contains key features of the trend changes and fluctuation patterns of the temporal data.

[0027] Fusion weight optimization A genetic algorithm combined with reinforcement learning is used to dynamically optimize the fusion weights of image data features, time-series operational parameter features, and historical operation and maintenance data (preprocessed as auxiliary features, with a dimension of 64) to ensure the accuracy of the fused feature vector. Specific implementation steps are as follows: Initialize the weight population: Set the initial population size of the fusion weight to 50, the weight value range to [0,1], and the sum of the weights of the three data sources to 1. Each individual corresponds to a set of weight parameters (w1,w2,w3), where w1 is the weight of image data features, w2 is the weight of time series operation parameter features, and w3 is the weight of historical operation and maintenance data features. Genetic algorithm optimization: The number of iterations of the genetic algorithm is set to 100, the crossover probability is 0.8, and the mutation probability is 0.05. The classification accuracy of the fused feature vectors (based on historical data verification) is used as the fitness function. Through selection, crossover, and mutation operations, individuals with higher fitness are selected. Reinforcement learning fine-tuning: The optimal weights optimized by the genetic algorithm are used as initial values ​​to construct a Q-learning reinforcement learning model. The error of state evaluation is used as the reward signal. When the evaluation error decreases, a positive reward is given, and vice versa. By continuously iterating and fine-tuning the weight parameters, the fused feature vector can more accurately reflect the device's operating status, and finally the optimal fused weights (w1*, w2*, w3*) are obtained. Dynamic weight update: During system operation, the above optimization process is repeated every 24 hours based on the latest collected data and status evaluation results to dynamically update the fusion weights and ensure the adaptability of the weights.

[0028] Based on the optimized fusion weights, feature-level fusion is performed on the extracted image data feature vector (128-dimensional), time-series operation parameter feature vector (128-dimensional), and historical operation and maintenance data feature vector (64-dimensional). The weighted summation method is used to obtain a fused feature vector with a dimension of 128, which is then output to the state evaluation module.

[0029] Construction of an improved BP neural network model The improved BP neural network, based on the traditional BP neural network, introduces a sparse training mechanism and uses the L1 regularization algorithm to constrain the network's internal parameters and eliminate redundant parameters. The model structure and improvement methods are as follows: Model Structure: The improved BP neural network includes an input layer, hidden layer 1, hidden layer 2, and an output layer. The input layer receives the fused feature vector (128-dimensional) output by the data fusion module, so the number of neurons in the input layer is 128. Hidden layer 1 has 64 neurons, and hidden layer 2 has 32 neurons. The activation function for both is ReLU. The output layer has 1 neuron and outputs the health status index of the power equipment (range 0-100). The activation function is Sigmoid (mapped to the [0,1] interval and multiplied by 100 to obtain the health status index).

[0030] Sparsity training mechanism: A sparsity training mechanism is introduced. By setting the activation threshold of hidden layer neurons (set to 0.2), the output of most hidden layer neurons is close to 0, and only the output of a few key neurons is retained, reducing redundant connections between neurons. During training, if the output of a hidden layer neuron is less than the activation threshold, the weight of that neuron is temporarily set to 0, and only the weights of neurons whose output is greater than the activation threshold are updated, thus achieving sparsity of the network.

[0031] L1 regularization constraint: The L1 regularization algorithm is used to constrain the weight parameters inside the network. An L1 regularization term is added to the loss function. Through the L1 regularization constraint, some redundant weight parameters can be made close to 0. After training, neurons and connections with absolute values ​​of weight parameters less than a threshold (set to 1e-5) are removed to simplify the model structure and improve the model's generalization ability and running efficiency.

[0032] Model training and optimization The improved BP neural network is trained and optimized using a labeled historical dataset. The specific steps are as follows: Dataset preparation: Collect multi-dimensional operation data, historical maintenance data, and corresponding equipment health status labels for the past three years of power equipment (health status index 0-100 is labeled by professional maintenance personnel based on the actual operating status of the equipment and the test results) to construct training and test datasets, with the training dataset accounting for 80% and the test dataset accounting for 20%. Model initialization: Initialize the network's weight parameters and bias parameters. The weight parameters are randomly initialized (value range [-0.1, 0.1]), and the bias parameters are set to 0.1. Set the training batch size to 32, the learning rate to 0.001, and the number of iterations to 200. Model Training: The fused feature vector of the training dataset is input into the improved BP neural network. The output health status index is calculated through forward propagation and compared with the labeled value to calculate the loss function value. The gradient descent method (Adam adaptive moment estimator optimizer) is used to backpropagate and update the weight and bias parameters of the network, while performing sparsity training and L1 regularization constraints. Every 10 iterations, the accuracy of the model on the validation set (the proportion of samples where the error between the predicted health status index value and the labeled value is less than 5) is calculated. If the accuracy on the validation set does not improve after 20 consecutive iterations, training is stopped and the current optimal model parameters are saved. Model optimization: After training, remove redundant neurons and connections (those with absolute values ​​of weight parameters less than the threshold) from the model and prune the model to simplify its structure. Input the test dataset into the optimized model to verify its performance. The model's prediction accuracy should be no less than 95%. If the accuracy does not meet the requirements, adjust parameters such as the regularization coefficient λ and the learning rate, and retrain until the performance requirements are met.

[0033] Equipment status quantitative assessment The status assessment module receives the fused feature vector output by the data fusion module and inputs it into the improved BP neural network model after training and optimization. Through forward propagation calculation, it outputs the current health status index (0-100) of the power equipment. At the same time, it presets the power equipment health status classification standard and outputs the operation status assessment level according to the corresponding health status index. The specific classification standard is as follows: Excellent level: Health status index 90-100, indicating that the power equipment is operating well, all parameters are normal, and there is no potential risk of failure; Good level: Health status index 70-89, indicating that the power equipment is operating normally, with slight parameter fluctuations and no obvious faults, and requires regular monitoring; Pass rating: Health status index 50-69 indicates that the power equipment is in general operating condition, with some parameters abnormal and potential faults, and the frequency of inspections needs to be increased. Unqualified level: Health status index 0-49, indicating that the power equipment is in an abnormal operating state, with obvious potential faults that may affect the normal operation of the equipment, and requires immediate shutdown for maintenance.

[0034] The status assessment module will output the health status index and the corresponding operating status assessment level, and simultaneously transmit them to the early warning module and the human-machine interaction module.

[0035] The early warning module generates corresponding early warning information and executes early warning operations based on the operating status assessment level output by the status assessment module and the preset early warning threshold. At the same time, it automatically calibrates the early warning threshold. The early warning module receives the health status index and assessment level output by the status assessment module in real time, compares them with the preset early warning threshold, and automatically generates early warning information of the corresponding level. The early warning information includes the device name, assessment level, health status index, early warning content, and generation time.

[0036] Execution of early warning operations Different alert actions are performed for different levels of alert information to ensure that the alert information can be promptly transmitted to relevant personnel. The specific actions are as follows: Blue alert: The alert information is only displayed in the alert list of the human-computer interaction module, without sound or light prompts. Staff can check and handle it periodically. Yellow alert: The alert information is displayed on the human-computer interaction module, and the yellow light of the edge computing gateway is triggered (the light frequency is 1 time / second). A push notification is also sent to the staff's mobile APP to remind the staff to strengthen the inspection. Red Alert: The alert information is displayed on the human-machine interaction module, triggering a red light indicator (light frequency of 2 times / second) and an audible alarm (the alarm sound is a buzzer tone with adjustable volume) on the edge computing gateway. At the same time, a notification is sent to the staff's mobile APP and office phone, and an alarm signal is sent to the substation operation and maintenance management platform. If no response is received from the staff within 10 minutes, the notification is automatically resent until a response is received.

[0037] To improve the accuracy of early warnings and reduce false alarms and missed alarms, the early warning module automatically calibrates the early warning threshold based on changes in the equipment operating environment and historical false alarm data. The calibration cycle is 7 days, and the specific calibration steps are as follows: Data collection: Collect equipment operating environment data (temperature, humidity, dust concentration, etc.), status assessment results, early warning information, and staff handling records for the past 7 days (to confirm whether they are false alarms), and count the number of false alarms and the corresponding health status index ranges for false alarms; Environmental Impact Analysis: Analyze the impact of changes in the equipment operating environment (such as a sudden increase in ambient temperature or excessive dust concentration) on the equipment health status index, and determine the threshold adjustment coefficients (range 0.9-1.1) corresponding to environmental factors. If the environment is harsh (such as temperature exceeding 60℃), the adjustment coefficient is set to 0.9 to lower the warning threshold and improve the warning sensitivity; if the environment is favorable, the adjustment coefficient is set to 1.1 to raise the warning threshold and reduce false alarms. False alarm correction: Based on historical false alarm data, if the false alarm rate of a certain level of warning exceeds 5%, the warning threshold of the corresponding level will be adjusted. For example, if the false alarm rate of yellow warning is too high (exceeding 5%), and the false alarms are mostly concentrated between the health status index of 50-55, the yellow warning threshold will be adjusted to 55 to reduce false alarms. Threshold update: Based on the comprehensive environmental impact coefficient and false alarm correction results, calculate the new warning threshold, update the preset warning threshold, and ensure that the warning threshold can adapt to changes in the equipment operating environment, thereby improving the accuracy and reliability of the warning.

[0038] The human-machine interaction module adopts a dual deployment approach of "cloud platform + local terminal". The hardware includes an industrial monitor, operating keyboard, mouse, and mobile APP for staff in the substation monitoring center. The software adopts a B / S architecture. The cloud platform is developed based on Web and supports browser access. The local terminal is developed based on Windows (Microsoft operating system). The mobile APP supports Android (Android operating system) and iOS (Apple operating system) systems, ensuring that staff can check the equipment operating status anytime and anywhere.

[0039] refer to Figure 2 As shown, the overall operation flow of the power equipment status assessment and early warning system of the present invention is as follows, ensuring that all modules work together to achieve real-time monitoring, assessment and early warning of equipment status: Data Acquisition: Each acquisition terminal of the data acquisition module collects equipment operation parameter data, environmental perception data, and image detection data of power equipment in real time, captures historical operation and maintenance data, and transmits it to the edge computing gateway after preliminary noise reduction and caching. Data fusion: The data fusion module receives the collected data, performs cleaning, normalization, and standardization preprocessing to generate a standardized processing dataset; it uses CNN to extract image data features and LSTM to extract time-series running parameter features, optimizes the fusion weights through genetic algorithm combined with reinforcement learning, performs feature-level fusion of multi-dimensional features, generates a fused feature vector, and outputs it to the state evaluation module. Status Assessment: The status assessment module will input the fused feature vector into the improved BP neural network model, output the equipment health status index, determine the operating status assessment level according to the preset health status classification standard, and transmit the results to the early warning module and the human-machine interaction module. Early warning processing: The early warning module compares the health status index and assessment level with the preset early warning threshold, generates early warning information of the corresponding level, and executes the corresponding early warning operation; at the same time, it automatically calibrates the early warning threshold on a regular basis based on environmental changes and historical false alarm data. Human-computer interaction: The human-computer interaction module receives early warning information and status assessment results, and displays them to users in an intuitive form for real-time viewing, querying, and operation. Users process the early warning information and fill in the processing records to achieve closed-loop operation of the system.

[0040] System effect verification To verify the effectiveness of the system of this invention, it was applied to three main transformers in a 110kV substation for a three-month operation period. The evaluation results were compared with those of the traditional condition assessment method. The specific verification results are as follows: Evaluation accuracy: The state evaluation accuracy of this system is 96.8%, while the accuracy of the traditional BP neural network evaluation method is 88.2%. This system significantly improves the evaluation accuracy by improving the BP neural network (sparse training + L1 regularization) and multi-dimensional feature fusion. Early warning accuracy: The false alarm rate of this system is 3.2% and the false alarm rate is 0.5%, while the false alarm rate of traditional early warning methods is 10.8% and the false alarm rate is 4.3%. This system effectively reduces the false alarm rate and false alarm rate by dynamically optimizing the fusion weight and automatically calibrating the early warning threshold. Real-time performance: The total latency of this system from data acquisition to status assessment and early warning information generation is 380ms, which meets the requirements for real-time monitoring of power equipment (requirement latency ≤500ms). Operation and maintenance efficiency: After adopting this system, the average time for troubleshooting substation equipment is reduced by 60%, and the average operation and maintenance cost is reduced by 35%, effectively improving the operation and maintenance efficiency of power equipment and reducing failure losses.

[0041] The above verification results show that, compared with traditional methods, the system of the present invention has improvements in accuracy, early warning precision, and real-time performance.

[0042] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0043] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A power equipment status intelligent early warning system based on multi-source data fusion, characterized in that, The early warning system includes: The data acquisition module is used to collect multi-dimensional operating data of power equipment, including equipment operating parameter data, environmental perception data, image detection data, and historical operation and maintenance data. The data fusion module receives multi-dimensional running data, preprocesses it to generate a standardized processing dataset, performs fusion processing on the dataset, extracts image data features through a convolutional neural network (CNN), extracts temporal running parameter features through a long short-term memory network (LSTM), dynamically optimizes the fusion weights of each data source based on a genetic algorithm combined with reinforcement learning, performs feature-level fusion on the extracted multi-dimensional features to obtain a fused feature vector, and outputs the fused data. The status assessment module is used to construct a power equipment status assessment model based on an improved BP neural network. The improved BP neural network introduces a sparsity training mechanism and uses an L1 regularization algorithm to constrain the network's internal parameters, eliminate redundant parameters, and train and optimize the assessment model. It receives fused feature vectors, inputs them into the status assessment model, outputs the current power equipment health status index, and outputs the corresponding power equipment operating status assessment level based on a preset health status classification standard, thereby realizing a quantitative assessment of the power equipment status. The early warning module generates corresponding early warning information for different assessment levels based on preset early warning thresholds, and performs different early warning operations for different early warning information; it also automatically calibrates the early warning thresholds based on changes in the equipment operating environment and historical false alarm data. The human-computer interaction module is used to receive early warning information, allowing users to view the current operating status of power equipment in real time.

2. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The data acquisition module collects equipment operating parameter data through multi-protocol acquisition via the connection between the measurement and control terminal and the smart sensor; environmental perception data is based on the deployment of distributed low-power IoT sensor nodes and is collected periodically according to environmental scenarios; image detection data is collected in real time based on the smart camera and thermal imager terminal; historical operation and maintenance data is obtained synchronously through online system integration, offline file import and manual supplementation.

3. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, Preprocessing within the data fusion module includes data cleaning and data standardization. The preprocessed data is classified and summarized to obtain a standardized dataset. Before the dataset is fused, multi-source data under the same power equipment and the same time dimension are selected based on the data association identifier to obtain a fused data group. Multi-dimensional feature extraction is performed on the fused data group.

4. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The image data feature extraction module takes preprocessed standardized image data as input and uses a convolutional neural network (CNN) to build a feature extraction model. Through alternating operations of convolutional and pooling layers, it gradually extracts low-level, mid-level, and high-level semantic features of the image. The extracted multi-layer features are integrated and dimensionality reduced through a fully connected layer, and the output is a fixed-dimensional image feature vector, which retains the feature confidence and completes the transformation of image data from unstructured to feature vector. Time-series runtime parameter feature extraction involves inputting a standardized dataset, setting a time window, and dividing the input sequence into a feature sequence. A time series feature extraction model is built using a Long Short-Term Memory (LSTM) network. Its gating mechanism is used to capture the long-term dependencies and dynamic changes in time series data. Overfitting is suppressed by using a dropout layer, and a time-series running parameter feature vector with the same dimension as the image feature vector is output through the output layer. The time-series feature stability index is output simultaneously to complete the feature extraction of time-series data.

5. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The data fusion module performs feature-level fusion on the extracted multi-dimensional features to obtain the fused feature vector. The steps are as follows: Auxiliary features are extracted from environmental perception and historical operation and maintenance data, quantized and encoded, and then integrated into an auxiliary feature vector. A hybrid optimization strategy combining genetic algorithm and reinforcement learning is adopted to dynamically adjust the fusion weights of image feature vector, temporal running parameter features and auxiliary feature vector; Set initial fusion weights for the three types of feature vectors, and define the weight constraint range based on historical operation and maintenance data; Using the accuracy and stability of the fused features as the fitness function, the initial weights are globally optimized through selection, crossover, and mutation operations to select the optimal weight solution set as the initial action space for reinforcement learning. Using the equipment operating status as environmental feedback, weight adjustment as action, and minimization of fusion error as reward function, the optimal weight solution set selected by the genetic algorithm is locally fine-tuned to dynamically adapt to changes in the real-time operating status of the equipment. The process is iteratively optimized until convergence, with the convergence condition being: the fusion error is lower than a set threshold. The optimal fusion weights of each feature vector are then output to form a dynamic weight matrix. Based on the optimized dynamic weight matrix, feature-level fusion operations are performed on the image feature vector extracted by CNN, the time-series running parameter feature vector extracted by LSTM, and the auxiliary feature vector. The three types of feature vectors are multiplied by their corresponding optimal weights to obtain the weighted feature vector. The weighted feature vectors are fused, spliced, and dimension-normalized to transform them into standardized fused data.

6. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The steps for building a state assessment model within the state assessment module are as follows: The network hierarchy was determined, adopting a three-layer structure of input layer-hidden layer-output layer, and the number of neurons in each layer was determined. The number of neurons in the input layer was consistent with the feature dimension after the fused data was filtered. The hidden layer was set to 2-3 layers, and the initial number of neurons was set according to an empirical formula, reserving space for pruning during sparse training. The number of neurons in the output layer was 4, corresponding to the four types of device status labels. The input layer uses a linear activation function, the hidden layer uses a ReLU activation function, and the output layer uses a Softmax activation function. By embedding an L1 regularization term into the network loss function, a new loss function is constructed to constrain the internal connection weight parameters of the network. Its expression is: Loss = Original loss function + λ × Σ|w| Where λ is the regularization coefficient and w is the internal weight parameter of the network, and weight sparsity is achieved through L1 regularization; By setting a sparsity threshold, the output of hidden layer neurons is constrained during network training to achieve network sparsity. Model training and optimization involve inputting the training set into the network, obtaining the device state probability prediction value through forward propagation, recording the hidden layer output, calculating the loss function with L1 regularization term in combination with the real label, and adding a penalty term if the sparsity threshold is not met; the process is repeated for iterative training.

7. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The evaluation steps within the status evaluation module are as follows: The system receives the standardized fusion feature vector output by the data fusion module in real time, acquires the vector association identifier simultaneously, and inputs the preprocessed fusion feature vector into the state assessment model. Through forward propagation operation, the fusion feature vector is analyzed in multiple dimensions, and combined with the equipment state feature patterns learned during training, the system outputs the current health status index of the power equipment. The health status index is set to 0-100 points. The preset power equipment health status classification standard is called, and the calculated health status index is compared with the preset classification standard to match the corresponding operation status assessment level and clarify the current operation status and potential risks of the equipment. Integrate the evaluation results, form a standardized evaluation report, and then output it.

8. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The early warning module generates differentiated early warning information based on preset graded early warning thresholds, and performs corresponding differentiated operations, from basic log recording to emergency audible and visual warnings and emergency linkage, while retaining complete operation logs. The module combines dynamic data of the equipment's operating environment and historical false alarm records, and uses a combination of periodic calibration and real-time calibration to adaptively and automatically calibrate the warning threshold.

9. The intelligent early warning system for power equipment status based on multi-source data fusion according to claim 1, characterized in that, The human-computer interaction module adopts an intuitive and visual interface design, which is compatible with both computer terminals and mobile APPs. The core display area of ​​the interface includes three major sections: overall equipment operation overview, hierarchical warning list, and detailed status of individual devices. Users can quickly locate target devices and view their integrated data, evaluation reports, and historical operation records through filtering and search functions.

10. A method for intelligent early warning of power equipment status based on multi-source data fusion, applied to the intelligent early warning system for power equipment status based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that, The steps are as follows: S1. Data Acquisition: Collect multi-dimensional operational data of power equipment, specifically including four types of data: equipment operating parameters, environmental perception, image detection, and historical operation and maintenance data. S2. Data Fusion: Receive multi-dimensional data, preprocess it to generate a standardized dataset; extract image data features through CNN, extract time-series running parameter features through LSTM, combine genetic algorithm and reinforcement learning to dynamically optimize the fusion weights of each data source, perform feature-level fusion of multi-dimensional features, obtain fused feature vectors and output fused data. S3. Status Assessment: Construct an improved BP neural network model, introduce a sparsity training mechanism, and use L1 regularization to constrain parameters and eliminate redundancy to complete model training optimization; receive fused feature vectors as input to the model, output the device health status index, and output the operating status assessment level based on preset division standards. S4. Early Warning Handling: Based on preset thresholds, generate corresponding early warning information for different assessment levels and execute differentiated early warning operations; By combining changes in the equipment operating environment with historical false alarm data, the warning threshold is automatically calibrated to reduce the false alarm and missed alarm rates. S5. Human-computer interaction: Receives warning information and provides a visual operation interface, allowing users to view the current operating status of the device and warning details in real time, realizing the linkage between the system and the user.

Images