Improved YOLOv8 power grid equipment operation state fault monitoring system

Abstract

The invention discloses an improved YOLOv8 power grid equipment operation state fault monitoring system. The system comprises a multispectral image acquisition module, a data fusion module, an improved YOLOv8 target detection module, a multi-modal feature analysis module, a fault diagnosis module, a self-adaptive information updating module, an alarm module and a remote interaction module. The problems of multispectral data redundancy, insufficient target detection precision, weak model generalization ability, poor fault diagnosis real-time performance and lack of dynamic updating ability existing in a traditional power grid equipment monitoring system are solved. According to the invention, fusion processing is carried out on multispectral image data through an improved weighted principal component analysis IWPCA algorithm, data redundancy is eliminated, and the information utilization rate is improved; through the improved YOLOv8 network structure and the loss function, high-precision detection is performed on the type, the position and the operation state of the power grid equipment, the small target identification capability is significantly improved, and background interference is reduced to ensure stable operation and efficient maintenance of the power grid equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring and fault diagnosis of power equipment, and more particularly to a power grid equipment operation state fault monitoring system based on an improved YOLOv8 network architecture. BACKGROUND

[0002] With the exponential growth of the construction scale and real-time requirements of the power grid equipment intelligent monitoring system, there is a lack of multi-source data fusion efficiency and dynamic fault diagnosis adaptability in power grid equipment state monitoring, and the implementation of the corresponding technology is not perfect.

[0003] Under the existing technical conditions, power grid equipment monitoring has formed a dual-spectrum detection method based on visible light-infrared image analysis and a traditional PCA data fusion strategy, which has to some extent alleviated the problem of high fault omission rate caused by the limitations of single sensor data. However, it still faces the problems of low computational efficiency caused by multispectral data redundancy, insufficient detection accuracy of small sample fault categories, and the inability of the static model updating mechanism to adapt to dynamic new fault modes of equipment.

[0004] In order to effectively solve the above problems, a comprehensive and perfect system needs to be constructed. By deeply mining the multi-modal spectral characteristics and dynamic operating parameters of power grid equipment, and using the improved YOLOv8 network's weighted principal component fusion module, hybrid attention mechanism and incremental online learning framework to realize real-time perception of equipment operating state and adaptive updating of model parameters, the bottleneck of the existing technology in the sustainable evolution ability of complex scene generalization system is broken through. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application discloses a multispectral fusion adaptive fault monitoring system based on improved YOLOv8. The dynamic characteristic weight of multi-source spectral data is optimized and evaluated by the improved weighted principal component analysis IWPCA algorithm, the multi-modal characteristics and long-tail distribution characteristics of the equipment target are detected and evaluated with high precision by the YOLOv8 network fused with attention mechanism and hybrid loss function, and the feature analysis results of the new fault mode are fed back to the target detection and diagnosis module and the model parameter adaptive updating is performed by means of the incremental online learning iCaRL framework, so as to ensure that the power grid equipment can realize high-precision fault recognition under real-time fusion of multispectral data, the detection accuracy is improved by 8.7%, the low omission rate in dynamic incremental scene is reduced by 3.2%, and the false alarm rate in complex working conditions is 1.5%.

[0006] The present application adopts the following technical solutions: The improved YOLOv8 power grid equipment operation state fault monitoring system comprises a multi-spectral image acquisition module, a data fusion module, an improved YOLOv8 target detection module, a multi-modal feature analysis module, a fault diagnosis module, an adaptive information updating module, an alarm module and a remote interaction module. The multi-source data acquisition module acquires operation images of power grid equipment under different spectral bands through an optical imaging module and a spectral analysis module. The data fusion module fuses multi-spectral image data at the output end of the multi-spectral image acquisition module based on an improved weighted principal component analysis (IWPCA) algorithm to eliminate data redundancy. The improved YOLOv8 target detection module detects the type, position and operation state of power grid equipment through an improved network structure module and an improved loss function module. The multi-modal feature analysis module extracts and analyzes features based on an improved ResNet-50 convolutional neural network, mines multi-modal features of the target detection result, and mines potential features of the equipment operation state. The adaptive information updating module updates parameters of the improved YOLOv8 target detection module, the multi-modal feature analysis module and the fault diagnosis module based on an incremental online learning (iCaRL) algorithm. The fault diagnosis module diagnoses the operation state of the power grid equipment based on an improved long short-term memory (LSTM) network model to determine whether the equipment has a fault and the type of the fault. The alarm module issues corresponding alarm information when the fault diagnosis module determines that the equipment has a fault based on fuzzy logic warning rules, and the alarm modes include audible and visual alarms, SMS alarms and email alarms. The remote interaction module enables an operator to remotely view the equipment operation state, fault diagnosis result and alarm information, and set and control the system through interaction with a remote monitoring center or the operator. The multi-source data acquisition module is connected to the input end of the data fusion module, the data fusion module is connected to the input end of the improved YOLOv8 target detection module, the improved YOLOv8 target detection module is connected to the input end of the multi-modal feature analysis module, the multi-modal feature analysis module and the improved YOLOv8 target detection module are connected to the input end of the fault diagnosis module, the fault diagnosis module is connected to the input end of the alarm module and the input end of the remote interaction module, and the remote interaction module interacts with the improved YOLOv8 target detection module, the multi-modal feature analysis module, the fault diagnosis module and the adaptive information updating module.

[0007] As a further technical solution of the present invention, the optical imaging module includes a FLIR SC640 thermal imager and a Phase One iXM 100MP multispectral camera to collect visible light, infrared, and multispectral images of the device; the spectral analysis module uses ASD FieldSpec Hi-Res hyperspectral imaging technology to extract characteristic information of the device in different spectral bands.

[0008] The working method of the improved weighted principal component analysis IWPCA algorithm is as follows: Step 1: Data preprocessing; Normalize the multispectral image data output by the multispectral image acquisition module to ensure that the data are within the same scale range; use singular spectrum analysis (SSA) to denoise the data to improve the data quality Step 2: Weight calculation; Assign a weight to each data point based on the feature relevance and importance of the data. The weight is calculated based on the variance and covariance matrix of the data. A Gaussian kernel function is introduced to weight the similarity between data points. Step 3: principal component extraction; Decompose the weighted data matrix through singular value decomposition (SVD) to extract the main principal components; select the first k principal components; Step 4: Data fusion; The extracted principal components are linearly combined to generate a fused feature map; the fused feature map is further processed using the Kalman filter algorithm.

[0009] As a further technical solution of the present invention, the implementation steps of the improved YOLOv8 target detection module are as follows: Step (1) network structure initialization; Initialize the improved YOLOv8 network structure, which includes a feature extraction module, an attention mechanism module, and a target detection module. The feature extraction module uses the EfficientFormerv2 network to extract the features of the fused data. The attention mechanism module is used to calculate the weight of the feature map, highlight important features, and suppress irrelevant features. The target detection module is used to output the type, location, and operating status of the power grid equipment. Step (2) Attention mechanism calculation: The weight of the feature map is calculated by the module to suppress irrelevant features. The calculation formula of the improved attention mechanism is: In formula (1) ，A is the attention weight, X is the input feature map, is the weight matrix, is the bias vector, is the activation function, is the normalization function, C represents the number of channels of the feature map; Step (3) loss function calculation; The improved loss function combines Focal Loss and Dice Loss to improve the model's detection performance for targets of different categories and sizes. The Focal Loss is used to deal with the imbalance problem of positive and negative samples, and the Dice Loss focuses on the segmentation of the target area. The improved loss function expression is: In formula (2), To adjust the parameters, is the probability predicted by the model, A and B are the areas of the predicted results and the true labels respectively, is a smoothing term to prevent division by zero errors; Step (4) target detection and output; The fused data is input into the improved YOLOv8 network. After feature extraction, attention mechanism and loss function optimization, the type, location and operating status of the power grid equipment are output to ensure the accuracy and robustness of the detection results. As a further technical solution of the present invention, the working method of the improved ResNet-50 convolutional neural network is as follows: 1) Convolutional feature enhancement; By performing sliding calculations on the convolution kernel of the fusion attention mechanism, local features are associated with the global channel. The expression of the convolution operation is: In formula (3), For the l The convolutional layer is at position ( i , j ), For the Layer at position ( ), For the first The convolution kernel of the layer is at position ( m , n ), For the The bias of the layer; a represents the channel attention coefficient, Indicates the The attention weight of the kth channel of the layer, C represents the number of channels; 2) Adaptive pooling compression; A hybrid pooling operation is used to balance the spatial information preservation and noise suppression of the feature map. The pooling calculation expression is: In formula (4), For the The pooling layer is at position The output of , S is the pooling window, s is the pooling step size, denotes the size of the pooling window, denotes the size of the pooling window, denotes the size of the pooling window, denotes the size of the pooling window, denotes the size of the pooling window, optimizes the feature expression through normalization and enhancement of the feature map; 4) multi-modal feature fusion; The optimized feature map is subjected to multi-modal feature fusion to generate a comprehensive feature vector. As a further technical solution of the application, the working method of the incremental online learning iCaRL algorithm is as follows: 1) model and data initialization; Initialize the model parameters of the improved YOLOv8 target detection module, multi-modal feature analysis module and fault diagnosis module; 2) old knowledge learning and storage; The model of each module is trained using the initial training data set; during the training process, the model parameters are continuously adjusted through the back propagation algorithm to minimize the loss function; after the training is completed, the current state of the model is stored, and representative samples are selected from the training data set as old knowledge samples and stored in the sample buffer; 3) new knowledge acquisition and processing; When new data samples arrive, the new data is preprocessed and the power grid equipment information contained in the new data is analyzed to determine whether there are new equipment types, location characteristics and operating state modes; 4) mixed data set construction; The old knowledge samples are taken out from the sample buffer and merged with the new data samples to construct a mixed data set; during the merging process, the old knowledge samples and the new data samples can be weighted according to different strategies to balance the proportion of new and old knowledge in the training; 5) model incremental update; The improved YOLOv8 target detection module, multi-modal feature analysis module and fault diagnosis module are incrementally trained using the mixed data set; during the training process, the iCaRL algorithm utilizes the information of the old knowledge samples, and in the back propagation process, the gradient information of the old knowledge samples and the gradient information of the new knowledge samples are combined to update the model parameters; 6) sample buffer update; After the training is completed, the sample buffer is updated according to the rules; a representative sampling method is adopted to select a part of samples from the mixed data set to remain in the buffer, while removing obsolete samples to ensure the relative stability of the number and quality of samples in the buffer; 7) repeated updating and optimization; Based on the update of new data, repeat steps 3)-6) to continuously update and optimize the model incrementally so that the model can adapt to changes in the operating status of power grid equipment in a timely manner.

[0010] As a further technical solution of the present invention, the improved neural network model adopts a multi-layer perceptron MLP model for fault diagnosis, wherein the multi-layer perceptron MLP model includes a hidden layer and an output layer, and the MLP model includes one or more hidden layers, each hidden layer consisting of multiple neurons; The hidden layer performs nonlinear transformation on the input data to extract complex features in the data. The calculation formula of the hidden layer is: In formula (5), For the The output of the jth neuron in the hidden layer, For the Layer to The connection weights of the layer, For the The input of the i-th neuron in the layer, For the The bias of the layer, is the activation function; a represents the channel attention coefficient, Indicates the The attention weight of the kth channel of the layer, C represents the number of channels; the number of neurons in the output layer is usually consistent with the number of fault categories, and each neuron corresponds to a fault category; the calculation formula of the output layer is: In formula (6), is the output of the kth neuron in the output layer, is the connection weight from the hidden layer to the output layer, is the output of the jth neuron in the hidden layer, is the bias of the output layer, As a further technical solution of the present invention, the working method of the alarm module is: First, the feature map output by the multimodal feature analysis module is converted into a one-dimensional feature vector sequence and normalized. An attention mechanism is introduced into the LSTM unit to enhance the ability to capture key time step features and optimize the structures of the input gate, forget gate, and output gate. Then, the input feature vector sequence is processed by the improved LSTM unit to extract time series features, and feature fusion is performed to generate a comprehensive feature vector. The extracted features are then classified to determine whether the equipment has a fault and the fault type, and the fault is graded according to its severity. Finally, the fault status and fault type of the equipment are output, and the confidence of the diagnosis is expressed by a probability value. At the same time, the diagnosis results are fed back to the model, and the model parameters are updated in real time through an online learning algorithm to adapt to changes in the operating status of power grid equipment.

[0011] Compared with the prior art, the present invention has the following positive effects: The present invention uses an improved weighted principal component analysis (IWPCA) algorithm to evaluate the dynamic feature weights and redundant information of multi-source spectral data, and uses the YOLOv8 network with a hybrid attention mechanism to evaluate the morphological characteristics, thermodynamic properties, and small sample failure modes of equipment targets. At the same time, it optimizes the configuration of data fusion, target detection, and incremental learning modules. Through the adaptive information update module iCaRL framework, the dynamic new fault feature analysis results are fed back to the target detection and diagnosis model and prompted to perform online iterative parameter updates, thereby ensuring that the power grid equipment monitoring system can operate stably and efficiently, achieving significant technical breakthroughs in improving multi-spectral fusion efficiency by 43%, achieving a fault detection accuracy of 97.3% under extreme lighting conditions, and reducing model update time by 63% in incremental learning scenarios. In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which: Figure 1 This is a flow chart of a method for power system fault diagnosis and analysis using an improved deep learning algorithm according to the present invention; Figure 2 This is a schematic diagram of the improved YOLOV8 deep learning algorithm model architecture of a power system fault diagnosis and analysis method using an improved deep learning algorithm of the present invention; Figure 3 This is an architectural diagram of a full life cycle assessment model for a power system fault diagnosis and analysis method using an improved deep learning algorithm according to the present invention; Figure 4A signal connection diagram of a central processing unit of an improved deep learning algorithm power system fault diagnosis analysis method. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments will be clearly and completely described below with reference to the drawings in the embodiments. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. It should be understood that the description is only exemplary, but not to limit the scope of the present application. In addition, in the description, the description of the known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present application.

[0013] An improved power grid equipment operation state fault monitoring system based on YOLOv8, comprising a multi-spectral image acquisition module, a data fusion module, an improved YOLOv8 target detection module, a multi-modal feature analysis module, a fault diagnosis module, an adaptive information updating module, an alarm module and a remote interaction module; wherein: The multi-source data acquisition module acquires operation images of power grid equipment under different spectral bands through an optical imaging module and a spectral analysis module; The data fusion module fuses the multi-spectral image data output from the multi-spectral image acquisition module based on an improved weighted principal component analysis (IWPCA) algorithm to eliminate data redundancy; The improved YOLOv8 target detection module detects and identifies the type, position and operation state of the power grid equipment through an improved network structure module and an improved loss function module based on the fused data; The multi-modal feature analysis module extracts and analyzes features based on an improved ResNet-50 convolutional neural network, mines multi-modal features of the target detection results, and mines potential features of the equipment operation state; The adaptive information updating module updates the parameters of the improved YOLOv8 target detection module, the multi-modal feature analysis module and the fault diagnosis module based on an incremental online learning iCaRL algorithm; The fault diagnosis module diagnoses the operation state of the power grid equipment based on an improved long short-term memory (LSTM) network model to determine whether the equipment has a fault and the type of the fault; The alarm module issues corresponding alarm information when the fault diagnosis module determines that the equipment has a fault based on fuzzy logic warning rules, and the alarm modes include audible and visual alarms, SMS alarms and email alarms; The remote interaction module allows the operation and maintenance personnel to remotely view the equipment operation state, fault diagnosis results and alarm information through the module, and set and control the parameters of the system through interaction with the remote monitoring center or the operation and maintenance personnel; The output end of the multi-source data acquisition module is connected to the input end of the data fusion module, the output end of the data fusion module is connected to the input end of the improved YOLOv8 target detection module, the output end of the improved YOLOv8 target detection module is connected to the input end of the multi-modal feature analysis module, the output ends of the multi-modal feature analysis module and the improved YOLOv8 target detection module are connected to the input end of the fault diagnosis module, the output end of the fault diagnosis module is connected to the input end of the alarm module and the input end of the remote interaction module, and the remote interaction module interacts with the improved YOLOv8 target detection module, the multi-modal feature analysis module, the fault diagnosis module and the adaptive information update module.

[0014] Further, the optical imaging module includes a FLIR SC640 thermal imager and a Phase One iXM 100MP multispectral camera to collect visible light, infrared and multispectral images of the equipment; the spectral analysis module uses ASDFieldSpec Hi-Res hyperspectral imaging technology to extract feature information of the equipment under different spectral bands.

[0015] In a specific embodiment, in a specific embodiment, The FLIR SC640 thermal imager obtains the temperature field distribution of the equipment surface through non-contact infrared radiation detection technology, the detector receives the 8-14 μm long-wave infrared radiation emitted by the target object, converts the radiation signal into an electrical signal through a HgCdTe focal plane array, generates raw thermal image data after 16-bit ADC analog-digital conversion, and finally generates a temperature pseudo-color image combined with blackbody calibration parameters, with a temperature resolution of 0.03℃; the Phase One iXM 100MP multispectral camera uses a combination of a light splitting prism and a filter array to realize multi-band synchronous imaging, including 5 independent spectral channels of 450nm blue light, 550nm green light, 650nm red light, 750nm near-infrared and 850nm short-wave infrared. The optical system decomposes the incident light into different waveband sensors through a dichroic mirror, and cooperates with a 100MP CMOS chip to capture multispectral images with a spatial resolution of 11608x8708 pixels at a rate of 15 frames per second, which can accurately identify the fine defects of equipment surface oxidation spots and electric arc ablation. The ASDFieldSpec Hi-Res hyperspectral imaging module is based on a push-broom imaging spectrometer architecture, with an InGaAs detector array for synchronous acquisition of 256 continuous waveband data, combined with a GPS / INS integrated navigation system to realize construction of a space-spectrum three-dimensional data cube, which can extract the absorption peak information of the C-H / O-H chemical bond characteristic waveband of the equipment surface material, effectively detecting potential faults of insulation material aging and metal component corrosion.

[0016] Furthermore, the working method of the improved weighted principal component analysis IWPCA algorithm is: Step 1: Data preprocessing; Normalize the multispectral image data output by the multispectral image acquisition module to ensure that the data are within the same scale range; use singular spectrum analysis (SSA) to denoise the data to improve the data quality Step 2: Weight calculation; Assign a weight to each data point based on the feature relevance and importance of the data. The weight is calculated based on the variance and covariance matrix of the data. A Gaussian kernel function is introduced to weight the similarity between data points. Step 3: principal component extraction; Decompose the weighted data matrix through singular value decomposition (SVD) to extract the main principal components; select the first k principal components; Step 4: Data fusion; The extracted principal components are linearly combined to generate a fused feature map; and a Kalman filtering algorithm is used to further process the fused feature map. In a specific implementation example, in the data preprocessing stage, first, the visible light image is normalized by using a minimum-maximum normalization method to map its data range to [0, 1], the infrared thermal imaging data retains the temperature physical quantity, and the hyperspectral data is normalized by using spectral reflectivity to ensure that all channel data are in the same scale range. Then, singular spectrum analysis (SSA) is used to denoise each spectral channel. The spectral channel is decomposed by using a sliding window with a length of 64, and the main components are retained and the noise components are removed. In the weight calculation stage, first, the covariance matrix of each spectral channel is calculated based on the feature weighting method of variance-covariance, and the diagonal elements are extracted as the variances of the channels. According to the variances, initial weights are assigned to the channels, and higher weight gains are manually assigned to the infrared channels and some key bands in the hyperspectral data. Then, the similarity between data points is calculated by using a Gaussian kernel function to construct a similarity matrix by using a Gaussian kernel similarity weighting method. The final weight combines the initial weight based on variance-covariance and the similarity between data points to form a comprehensive weight. In the principal component extraction stage, first, the pixel-level weight is embedded into the data matrix to construct a weighted matrix. Then, the weighted matrix is subjected to singular value decomposition (SVD) to extract the first 15 principal components. In the data fusion stage, first, the extracted principal components are linearly combined by using adaptive weight coefficients to obtain a fused feature map. These adaptive weight coefficients are obtained by grid search optimization. Then, the fused feature map is optimized by using Kalman filtering. The state equation and the observation equation are established, the fused feature map is taken as the state vector, and the noise covariance matrix is obtained by training the first 10 frames of data, so as to realize smooth tracking of the dynamic change characteristics of the equipment. The root mean square error of the time series of the filtered feature map is significantly reduced. The implementation results show that after the improved weighted principal component analysis (IWPCA) processing, the dimension of the multispectral data is greatly reduced, the data redundancy is reduced by 78%, and the key features related to the equipment defects are retained.

[0017] Further, the implementation steps of the improved YOLOv8 target detection module are as follows: Step (one) network structure initialization The improved YOLOv8 network structure is initialized, and the network structure includes a feature extraction module, an attention mechanism module and a target detection module; the feature extraction module adopts an efficientformerv2 network to extract the features of the fused data; the attention mechanism module is used to calculate the weight of the feature map, highlight important features and suppress irrelevant features; and the target detection module is used to output the type, position and running state of the power grid equipment Step (two) attention mechanism calculation The weights of the feature maps are calculated by the module to suppress irrelevant features. The calculation formula of the improved attention mechanism is: In formula (1) ，A is the attention weight, X is the input feature map, is the weight matrix, is the bias vector, is the activation function, is a normalization function, and C represents the number of channels in the feature map. Multi-source data for electrical equipment (such as current, voltage, temperature, and vibration) contains numerous features, but the contributions of different features to fault diagnosis vary significantly. For example, a short circuit fault may be highly correlated with sudden changes in current and voltage, while an insulation fault may be more dependent on temperature and partial discharge signals. Noise (such as ambient vibration and sensor errors) can cause irrelevant features to interfere with diagnosis.

[0018] By calculating the weight of the feature map (A in the formula), the attention mechanism allows the model to automatically identify and enhance the feature channels most relevant to the current fault category (such as the current channel), and suppress irrelevant channels (such as the vibration channel during normal operation), thereby highlighting key information.

[0019] By visualizing weights (such as channel attention heatmaps), the attention mechanism can intuitively display the key features the model focuses on (for example, a high weight on a channel indicates that the sensor data is more important for the current fault diagnosis), improving the transparency of the diagnostic process. Furthermore, suppressing noisy features can reduce the model's overfitting to irrelevant information and enhance generalization capabilities. This is particularly suitable for fault diagnosis in environments with unbalanced samples or noisy environments. Table 1 shows the differences and advantages of this mechanism over traditional neural networks. By strengthening fault-sensitive features (such as current mutation signals during short circuits), missed diagnoses or misdiagnoses can be reduced. When fusing multi-sensor data such as current, voltage, and temperature, the importance of different modes is automatically assigned (for example, the temperature channel has a higher weight in high-temperature faults). When the equipment load changes or the environment fluctuates, the feature weights are dynamically adjusted to avoid misjudgments caused by fixed weights in traditional models. Through channel pruning (channels with weights approaching 0 can be ignored), the amount of calculation is reduced without sacrificing performance, which is suitable for embedded diagnostic equipment. Step (3) Loss function calculation; The improved loss function combines Focal Loss and Dice Loss to improve the model's detection performance for targets of different categories and sizes. The Focal Loss is used to deal with the imbalance problem of positive and negative samples, and the Dice Loss focuses on the segmentation of the target area. The improved loss function expression is: In formula (2), To adjust the parameters, is the probability of model prediction, A and B are the area of prediction result and true label respectively, is the smoothing term to prevent division by zero error; the loss function is used to measure the difference between the model prediction result and the true label. During the neural network training process, the goal of the model is to continuously adjust its parameters so that the value of the loss function is as small as possible. The loss function is like a "compass" that points the direction of the model's training, telling the model which direction to adjust the parameters to make the prediction result closer to the true situation. By minimizing the loss function, the model can learn the mapping relationship between the input data and the output result, thereby improving the prediction accuracy and generalization ability of the model. In power grid equipment fault diagnosis, the imbalance between positive and negative samples is a common problem. For example, the number of normal operation state samples is usually much larger than that of fault state samples. Traditional loss functions treat each sample equally, which leads the model to be more inclined to fit the larger number of normal samples during the training process, while ignoring the smaller number of fault samples, thereby making the model's detection performance on fault samples poor. The focal loss reduces the weight of easy-to-classify samples by introducing a modulation factor, and increases the attention to difficult-to-classify samples (usually minority class fault samples), thereby effectively solving the problem of imbalance between positive and negative samples, and making the model pay more attention to the learning of fault samples. Power grid equipment fault diagnosis sometimes not only needs to determine whether a fault has occurred, but also needs to accurately locate the area where the fault occurred, such as determining which part of the equipment has failed. The dice loss focuses on the segmentation of the target area, and it measures the performance of the model by calculating the overlap between the predicted result and the true label area. In fault diagnosis, using the dice loss can encourage the model to more accurately segment the fault area and improve the accuracy of fault location.

[0020] Step (four) target detection and output; The fused data is input into the improved YOLOv8 network, and after feature extraction, attention mechanism and loss function optimization, the type, location and running state of the power grid equipment are output to ensure the accuracy and robustness of the detection result. In a specific implementation example, the input feature map X is multiplied by the weight matrix and the bias vector is added, and then processed by the activation function , and then multiplied by the weight matrix and summed and averaged on the channel number C, and finally normalized by the normalization function to get the attention weight A, which highlights important features and suppresses irrelevant features. The target detection module is responsible for outputting the type, location and running state of the power grid equipment; then calculate the focal loss part according to the model prediction probability and the modulation parameter ​, focal loss is used to deal with the imbalance problem of positive and negative samples; finally, the dice loss part is calculated by the area A and B of the predicted results and the true label and the smoothing term calculate The dice loss focuses on segmenting the target area, and the two are summed to obtain the final loss function value. Finally, the fused data is input into the improved YOLOv8 network. After feature extraction, attention mechanism, and loss function optimization, the type, location, and operating status of the power grid equipment are output. The output data is used for actual power grid operation monitoring, and the model parameters are modified based on the feedback results. Table 2 shows the performance comparison of the improved YOLOv8. Table 2 shows that the improved YOLOv8 target detection module reduces the false positive rate from 4.7% to 3.8% in equipment-dense areas. Its adaptive threshold mechanism dynamically adjusts the IoU threshold between 0.5 and 0.7 based on the spatial density of the target, avoiding the over-suppression problem of traditional NMS in overlapping target scenarios. Experiments have shown that when the surface defect area of ​​the equipment accounts for less than 5%, the model's detection rate for subtle features such as arc burns and metal corrosion can reach 93.6%, and the false positive rate is controlled within 1.2%, which is significantly better than the original model. Furthermore, the working method of the improved ResNet-50 convolutional neural network is as follows: 1) Convolutional feature enhancement; By performing sliding calculations on the convolution kernel of the fusion attention mechanism, local features are associated with the global channel. The expression of the convolution operation is: In formula (3), For the l The convolutional layer is at position ( i , j ), For the Layer at position ( ), For the first The convolution kernel of the layer is at position ( m , n ), For the The bias of the layer; a represents the channel attention coefficient, Indicates the The attention weight of the kth channel of the layer, C represents the number of channels; 2) Adaptive pooling compression; A hybrid pooling operation is used to balance the spatial information preservation and noise suppression of the feature map. The pooling calculation expression is: In formula (4), For the The pooling layer is at position The output of , S is the pooling window, s is the pooling step size, Represents the pooling strategy weight, which is used to balance the contribution of maximum pooling and average pooling. represents the size of the pooling window, Indicates the Layer feature map at position ;3) Feature map optimization; Optimize feature expression through normalization and enhancement of feature maps; 4) Multimodal feature fusion; The optimized feature map is subjected to multimodal feature fusion to generate a comprehensive feature vector. In the specific implementation example, the convolution sum of the convolution kernel and the input feature is first calculated, that is, Plus the first Layer bias , and then calculate the channel attention part, that is , the three are added together to get The convolutional layer is at position Output features Yi , jl , so that local features are associated with global channels. First calculate the maximum pooling part , and then calculate the average pooling part , adding the two together gives The pooling layer is at position Output , thereby balancing the preservation of spatial information and noise suppression in the feature map. Feature map optimization is performed by normalizing and enhancing the feature map to optimize feature expression. Finally, multimodal feature fusion is performed to fuse the optimized feature maps to generate a comprehensive feature vector. The calculated data undergoes iterative calculations of parameters such as convolution and pooling to further improve network performance. Table 3 shows a data comparison using the improved ResNet-50 convolutional neural network. Table 3 shows that the improved ResNet-50 convolutional neural network offers greater performance improvements than the conventional ResNet-50 convolutional neural network. Given the same number of samples, the improved algorithm achieves higher feature extraction accuracy and noise suppression than the conventional algorithm, and also achieves a better overall performance score. By integrating an attention mechanism and adopting a hybrid pooling strategy, the improved ResNet-50 convolutional neural network balances the association between local features and global channels, retaining spatial information, and suppressing noise. This reduces the algorithm's sensitivity to data noise, resulting in improved stability, higher computational efficiency, faster convergence, and stronger feature extraction capabilities.

[0021] Furthermore, the incremental online learning iCaRL algorithm works as follows: 1) Model and data initialization; Initialize the model parameters of the improved YOLOv8 target detection module, multi-modal feature analysis module and fault diagnosis module; 2) Old knowledge learning and storage; Train the model of each module using the initial training data set; During the training process, the model parameters are continuously adjusted through the backpropagation algorithm to minimize the loss function; After the training is completed, the current state of the model is stored, and representative samples are selected from the training data set as old knowledge samples and stored in the sample buffer; 3) New knowledge acquisition and processing; When new data samples arrive, pre-process the new data and analyze the power grid equipment information contained in the new data to determine whether there are new equipment types, location characteristics and operating state modes; 4) Construction of mixed data set; Merge the old knowledge samples from the sample buffer with the new data samples to construct a mixed data set; During the merging process, the old knowledge samples and new data samples can be weighted according to different strategies to balance the proportion of new and old knowledge in training; 5) Incremental model update; Use the mixed data set to perform incremental training on the improved YOLOv8 target detection module, multi-modal feature analysis module and fault diagnosis module; During the training process, the iCaRL algorithm uses the information of the old knowledge samples, and in the backpropagation process, the gradient information of the old knowledge samples and the gradient information of the new knowledge samples are combined to update the model parameters; 6) Sample buffer update; After training, update the sample buffer according to the rules; Use a representative sampling method to select a portion of samples from the mixed data set to keep in the buffer, while removing outdated samples to ensure the relative stability of the number and quality of samples in the buffer; 7) Repeat the update and optimization; Based on the update of new data, repeat steps 3) - 6) to continuously update and optimize the model incrementally, so that the model can adapt to changes in the operating state of the power grid equipment in a timely manner.

[0022] In order to improve the above technical capabilities, the following experimental scheme is adopted: Hardware configuration: NVIDIA RTX 4090 GPU (24GB VRAM), Intel Xeon W-2245 CPU (8 cores 3.9GHz), 64GB RAM, 2TB NVMe SSD; Software platform: Python 3.9 + PyTorch 2.1 + CUDA 12.1; Framework tools: YOLOv8 official implementation (Ultralytics library), Scikit-learn (feature processing), TensorBoard (training monitoring).

[0023] Data source: A provincial power company of State Grid, 110kV-500kV substations, containing multi-modal data of 3 types of equipment (transformers, circuit breakers, insulators): Visual data: Equipment infrared thermal images (2000, resolution 640×480), visible light inspection images (5000, resolution 1280×720) Sensor data: Vibration acceleration (10kHz sampling rate), oil temperature (accuracy ±0.5℃), leakage current (resolution 1mA) Operating parameters: Voltage / current effective value, load rate, power factor (collected every second) Fault types: Covering 7 typical fault modes (winding deformation, bushing crack, insulator contamination, etc.), label using One-Hot encoding Data division: Initial training set: 5000 samples (new and old knowledge ratio 7:3, containing historical fault data of 3 types of equipment) Online test set: Real-time incoming new data (200-300 new samples per day, including 20% new equipment types / state modes) YOLOv8 detection module: Input size 640×640, backbone network using CSPDarknet, detection head supporting multi-scale feature fusion Multi-modal feature module: Visual features (extracted by ResNet50) and sensor features (mapped by fully connected layers) are fused through Concatenate, outputting 256-dimensional feature vectors Fault diagnosis module: 3-layer fully connected network (256→128→64), Softmax output 7-class fault probability The verification process of incremental learning is as follows. Old knowledge learning stage: Initial training is 200 epochs, learning rate is 1e-4, batch size is 32, and the loss function uses cross entropy + DiceLoss (for unbalanced fault data) Select the top 30% representative samples (based on the principle of minimum intra-class distance) and store them in the buffer to build the initial knowledge base Online learning phase: New knowledge processing: Detect new device types through the DBSCAN clustering algorithm (threshold: feature distance > 2σ), triggering dynamic expansion of the model structure (such as adding new detection categories) Mixed dataset construction: New and old samples are mixed in a 1:1 ratio, using class-balanced weighting (faulty sample weight × 2, normal sample weight × 0.5) Model update: incremental training for 50 epochs, freezing the first 3 layers of the backbone network (preserving old knowledge), learning rate 1e-5, adding elastic weight consolidation (EWC) regularization term (λ=0.1) Buffer Update: After each round of training, the class representativeness of the samples is calculated (using k-NN density estimation) and the 20% samples with the lowest density in the buffer are replaced. Core indicators: Category Incremental Accuracy (CIA): Accuracy of newly added category recognition Old Knowledge Retention Rate (OKR): The accuracy retention of the initial category after incremental training Mean Average Precision (mAP@0.5): Object Detection Task Performance Fault diagnosis F1-score: classification performance after multimodal fusion The comparative test is shown in Table 4. The object detection performance is shown in Table 5. The fault diagnosis accuracy is shown in Table 6. When a new 500kV reactor (not included in initial training) was connected, this solution detected new feature distributions using DBSCAN and automatically expanded the detection categories, achieving a first-time recognition accuracy of 78% (compared to a traditional method's >50% probability of misclassifying it as a transformer). In a scenario with an imbalance between new and old samples (80% of new samples), class balancing and weighting improved the accuracy of the old knowledge categories from 65% to 82%. Compared to single-visual modality, the integration of vibration and oil temperature data improved the recognition rate of early faults (such as minor insulator contamination) from 68% to 81%. The dynamic iCaRL algorithm outperformed traditional solutions in both object detection and fault diagnosis tasks, particularly when handling new equipment types and fault modes, achieving improvements of 10.5% in mAP and 11.3% in F1-score, respectively, demonstrating the adaptability of the incremental learning framework. A representative sampling approach improved the quality of buffer samples by 25.8%, effectively alleviating the problem of old knowledge forgetting and achieving a 92.4% old knowledge retention rate.

[0024] In a specific implementation, the system uses an improved YOLOv8 object detection module to identify device type, location, and operating status. A multimodal feature analysis module based on an improved ResNet-50 convolutional neural network mines potential features of device operating status. A fault diagnosis module based on an improved Long Short-Term Memory (LSTM) model determines the presence and type of fault. During initialization, the model parameters of these modules are initialized using pretrained weights, using weights pretrained on a large-scale image dataset to accelerate model convergence. The initial training dataset consists of multispectral image data from various devices in the regional power grid under different operating conditions over the past month. This data, processed by the data fusion module, is then used for subsequent training. The models of each module are trained using this initial training dataset. During training of the improved YOLOv8 object detection module, the network weights and biases are continuously adjusted using a backpropagation algorithm to minimize the improved loss function. Backpropagation is also used to optimize the model parameters of the multimodal feature analysis and fault diagnosis modules. After training is complete, the current state of each module's model is stored, for example, as a model file. From the initial training dataset, representative samples are selected using cluster analysis as old knowledge samples. For transformers, clusters are formed based on their operating parameters, such as temperature and voltage. Several samples are selected from each cluster and stored in a sample buffer. When new data arrives, preprocessing is performed, including image normalization and cropping, to ensure data consistency. The new data is then analyzed for grid equipment information to determine whether new equipment types, location characteristics, and operating status patterns exist. Old knowledge samples are then extracted from the sample buffer and merged with the new data to construct a hybrid dataset. New data samples are weighted based on their degree of difference from the old knowledge samples. New equipment type data with significant differences receive a weight of 0.8, while new data with less significant differences receive a weight of 0.5. For old knowledge samples, the overall weight is 0.2, with the weight appropriately increased to 0.3 for key, representative old knowledge samples. Assuming there are 200 new data samples and 1000 old knowledge samples in the sample buffer, the effective number of samples in the hybrid dataset, after weighting, is equivalent to 500. The improved YOLOv8 object detection module, multimodal feature analysis module, and fault diagnosis module are incrementally trained using a mixed dataset. During training, the iCaRL algorithm leverages information from previous knowledge samples. During backpropagation, the gradient information from previous and new knowledge samples is weighted and summed to update the model parameters. After training, the sample buffer is updated according to the rules. A representative sampling method is used to select a subset of samples from the mixed dataset and retain them in the buffer, while removing outdated samples.Suppose that, after calculation, 150 samples were selected from the mixed data set to be added to the buffer. At the same time, the 150 oldest samples with significant distribution differences from the current data were removed to ensure a relatively stable sample quantity and quality in the buffer. The system continuously collects new data and incrementally updates and optimizes the model. The system's detection accuracy for new smart meters has increased from an initial 71% to 89%. The system's diagnostic accuracy for various equipment faults has also significantly improved, enabling it to promptly adapt to changes in the operating status of power grid equipment. Furthermore, the improved neural network model adopts a multi-layer perceptron MLP model for fault diagnosis, wherein the multi-layer perceptron MLP model includes a hidden layer and an output layer, and the MLP model includes one or more hidden layers, each hidden layer consisting of multiple neurons; The hidden layer performs nonlinear transformation on the input data to extract complex features in the data. The calculation formula of the hidden layer is: In formula (5), For the The output of the jth neuron in the hidden layer, For the Layer to The connection weights of the layer, For the The input of the i-th neuron in the layer, For the The bias of the layer, is the activation function; a represents the channel attention coefficient, Indicates the The attention weight of the kth channel of the layer, C represents the number of channels; the number of neurons in the output layer is usually consistent with the number of fault categories, and each neuron corresponds to a fault category; the calculation formula of the output layer is: In formula (6), is the output of the kth neuron in the output layer, is the connection weight from the hidden layer to the output layer, is the output of the jth neuron in the hidden layer, is the bias of the output layer, Represents the adaptive normalization coefficient. In this specific implementation example, multi-source data from power equipment under different operating conditions is first collected, including sensor data such as current, voltage, temperature, and vibration, totaling 1,000 sets of samples. This data is normalized to eliminate the dimensional effects between different features, with a mean of 0 and a standard deviation of 1. Based on the equipment's historical fault records, the fault types are categorized into four categories: short circuit fault, overload fault, insulation fault, and ground fault, and the samples are labeled accordingly.

[0025] The difference and superiority of the improved MLP model compared with the neural network model in the prior art are as follows: The channel attention mechanism is introduced in the hidden layer of the improved MLP model and the channel attention weight , which can adaptively weight the features of different channels and enhance the model's attention to important features. The traditional MLP model usually does not have such an attention mechanism. Adaptive normalization: adaptive normalization coefficients are introduced in the output layer , which can adaptively normalize the output and help improve the stability and generalization ability of the model. The MLP model in the prior art may use a fixed normalization method or not perform normalization. Stronger feature extraction capability: the channel attention mechanism enables the model to automatically learn the importance of different channel features, thereby more effectively extracting complex features in the data and improving the model's ability to capture fault features. For example, in power equipment fault diagnosis, certain sensor data (such as temperature, vibration) may be more critical to a specific fault type, and the channel attention mechanism can highlight these important features to improve the accuracy of fault diagnosis.

[0026] Better generalization capability: the adaptive normalization coefficients can automatically adjust the scale of the output according to the distribution of the input data, reducing the model's overfitting to the training data and improving the model's generalization capability in different data sets and actual application scenarios.

[0027] Better fault diagnosis performance: by combining the advantages of the channel attention mechanism and adaptive normalization, the improved MLP model may outperform the neural network model in the prior art in terms of accuracy, recall rate, and other evaluation indicators in fault diagnosis, enabling more accurate diagnosis of the fault type of power equipment and providing more reliable protection for the safe operation of the power system.

[0028] The operation of electrical equipment in power grids generates a large amount of heterogeneous data from multiple sources, such as sensor data for current, voltage, temperature, vibration, and sound. This data varies in type, dimension, and time scale, making it difficult for traditional methods to effectively process such complex data. However, improved neural network models, through a multi-layered structure and nonlinear transformations, can automatically learn complex features and patterns in this data, adapting to its diversity and complexity. Electrical equipment can experience a variety of fault types, such as short circuits, overloads, insulation faults, and ground faults, each with varying manifestations and severity. Improved neural network models possess powerful classification capabilities and can accurately distinguish different fault types by learning from a large amount of fault sample data, providing accurate fault diagnosis results. In power systems, electrical equipment failures can have serious consequences, necessitating their timely detection and diagnosis. After training, improved neural network models can rapidly process and analyze real-time data, providing fault diagnosis results that meet real-time requirements. Improved neural network models may incorporate attention mechanisms, such as channel attention coefficients and channel attention weights. This mechanism allows the model to automatically focus on important feature channels in the data, enhancing its ability to extract key features. For example, in electrical equipment fault diagnosis, certain sensor data (such as sudden changes in temperature or current) may be more critical for fault diagnosis. Attention mechanisms can highlight these important features and improve fault diagnosis accuracy. The hidden layers of a multilayer perceptron (MLP) model can perform multiple nonlinear transformations on the input data, capturing complex features and underlying patterns within the data. Compared to traditional linear models, it can better fit the nonlinear relationships in electrical equipment fault data. Improved models may employ adaptive normalization methods, such as adaptive normalization coefficients in the output layer. This method automatically adjusts the output scale based on the distribution of the input data, reducing overfitting of the model to the training data and improving the model's generalization across different datasets and practical application scenarios. Even when operating under new conditions or with changes in the data distribution, the model maintains good fault diagnosis performance. Improved models may also employ regularization techniques, such as L1 and L2 regularization, to constrain model parameters, prevent overfitting, and further improve generalization. Improved neural network models can dynamically adjust their structure based on actual needs, such as increasing or decreasing the number of hidden layers or neurons. In electrical equipment fault diagnosis, if new fault types or changes in data characteristics are encountered, the model structure can be adjusted to adapt to these changes, improving the model's flexibility and adaptability. Integrating different types of sensor data (such as visual, auditory, and tactile) into the model allows for comprehensive utilization of multiple information for fault diagnosis, improving diagnostic accuracy and reliability.

[0029] In specific embodiments, by constructing an improved MLP model, 2 hidden layers are set, each containing 50 neurons. The activation function is selected as ReLU function, the channel attention coefficient a is initialized as 0.5, and the adaptive normalization coefficient b is initialized as 0.3. 1000 groups of samples are divided into training set and test set according to the ratio of 8:2. In the training process, the stochastic gradient descent (SGD) optimization algorithm is used, the learning rate is set to 0.01, and the iteration number is 100. Then, through the ReLU activation function for nonlinear transformation, combined with the channel attention part, the final output is calculated. The trained model is evaluated using the test set, and the accuracy, recall rate, F1 value and other indicators are calculated. After testing, the improved MLP model has a fault diagnosis accuracy of 92%, a recall rate of 90%, and an F1 value of 0.91, which is significantly improved compared with the traditional MLP model. Further, the working method of the alarm module is: Firstly, the feature maps output by the multi-modal feature analysis module are converted into one-dimensional feature vector sequences and normalized; an attention mechanism is introduced into the LSTM unit to enhance the ability to capture key time step features and optimize the structure of the input gate, forget gate and output gate; then, the input feature vector sequence is processed by the improved LSTM unit to extract time series features and perform feature fusion to generate a comprehensive feature vector; the extracted features are then classified to determine whether the equipment has a fault and the fault type, and are graded according to the severity of the fault; finally, the fault state and fault type of the equipment are output, and the confidence of the diagnosis is represented by a probability value, while the diagnosis results are fed back to the model to update the model parameters in real time through an online learning algorithm to adapt to changes in the operating state of the power grid equipment.

[0030] In specific embodiments, the feature maps output by the multi-modal feature analysis module contain various information about the equipment, such as infrared thermal imaging, vibration and current voltage characteristics, and the alarm module converts these feature maps into one-dimensional feature vector sequences and performs normalization to unify the numerical range to [0, 1] and eliminate the dimensional differences between different features. An attention mechanism is introduced into the LSTM unit to enhance the ability to capture key time step features and optimize the structure of the input gate, forget gate and output gate to more accurately control the inflow, retention and output of information. The normalized feature vector sequence is processed by the improved LSTM unit to extract time series features and fuse different features to generate a comprehensive feature vector. The comprehensive features are classified using the trained classification model to determine whether the equipment has a fault and the fault type, and are graded according to the severity of the fault. Finally, the fault state and fault type of the equipment are output, and the confidence of the diagnosis is represented by a probability value, while the diagnosis results are fed back to the model to update the model parameters in real time using an online learning algorithm to improve the accuracy and reliability of fault diagnosis.

[0031] While specific embodiments of the application have been described above, it will be appreciated that those skilled in the art within the scope of the present application can make modifications, substitutions and changes in the form and details of the method and system described above without departing from the spirit and scope of the present application. For example, it is well within the scope and spirit of the present application to combine various of the illustrative steps described above to perform substantially the same functions in substantially the same way to achieve the same results. Therefore, the scope of the application shall be limited only by the claims below.

Claims

1. A power grid equipment operating status fault monitoring system based on improved YOLOv8, characterized by: It includes a multispectral image acquisition module, a data fusion module, an improved YOLOv8 target detection module, a multimodal feature analysis module, a fault diagnosis module, an adaptive information update module, an alarm module, and a remote interaction module; among which: The multi-source data acquisition module collects operating images of power grid equipment in different spectral bands through the optical imaging module and the spectral analysis module; The data fusion module fuses the multispectral image data output by the multispectral image acquisition module based on the improved weighted principal component analysis (IWPCA) algorithm to eliminate data redundancy. The improved YOLOv8 target detection module uses the improved network structure module and the improved loss function module to detect targets on the fused data and identify the type, location, and operating status of power grid equipment; The multimodal feature analysis module extracts and analyzes features based on an improved ResNet-50 convolutional neural network, mines multimodal features of target detection results, and discovers potential features of the device's operating status. The adaptive information update module adaptively updates the parameters of the improved YOLOv8 target detection module, multimodal feature analysis module, and fault diagnosis module based on the incremental online learning iCaRL algorithm; The fault diagnosis module diagnoses the operating status of power grid equipment based on the improved long short-term memory network (LSTM) model to determine whether the equipment has a fault and the type of fault. The alarm module uses fuzzy logic early warning rules. When the fault diagnosis module determines that the equipment has a fault, it will issue a corresponding alarm message. The alarm methods include sound and light alarm, SMS alarm and email alarm. Remote interaction module: Through interaction with the remote monitoring center or operation and maintenance personnel, the operation and maintenance personnel can remotely view the equipment operating status, fault diagnosis results and alarm information, and set and control system parameters; The output end of the multi-source data acquisition module is connected to the input end of the data fusion module, the output end of the data fusion module is connected to the input end of the improved YOLOv8 target detection module, the output end of the improved YOLOv8 target detection module is connected to the input end of the multimodal feature analysis module, the output ends of the multimodal feature analysis module and the improved YOLOv8 target detection module are connected to the input end of the fault diagnosis module, the output end of the fault diagnosis module is connected to the input end of the alarm module and the input end of the remote interaction module, and the remote interaction module interacts with the improved YOLOv8 target detection module, the multimodal feature analysis module, the fault diagnosis module and the adaptive information update module.

2. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1 is characterized in that: The optical imaging module includes a FLIR SC640 thermal imager and a Phase One iXM 100MP multispectral camera to capture visible light, infrared, and multispectral images of the device. The spectral analysis module uses ASD FieldSpec Hi-Res hyperspectral imaging technology to extract characteristic information of the device in different spectral bands.

3. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1, characterized in that: The working method of the improved weighted principal component analysis IWPCA algorithm is as follows: Step 1: Data preprocessing; Normalize the multispectral image data output by the multispectral image acquisition module to ensure that the data are within the same scale range; use singular spectrum analysis (SSA) to denoise the data to improve the data quality; Step 2: Weight calculation; Assign a weight to each data point based on the feature relevance and importance of the data. The weight is calculated based on the variance and covariance matrix of the data. A Gaussian kernel function is introduced to weight the similarity between data points. Step 3: principal component extraction; Decompose the weighted data matrix through singular value decomposition (SVD) to extract the main principal components; select the first k principal components; Step 4: Data fusion; The extracted principal components are linearly combined to generate a fused feature map; the fused feature map is further processed using the Kalman filter algorithm.

4. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1, characterized in that: The implementation steps of the improved YOLOv8 target detection module are as follows: Step (1) network structure initialization; Initialize an improved YOLOv8 network structure, which includes a feature extraction module, an attention mechanism module, and a target detection module. The feature extraction module uses an EfficientFormerv2 network to extract features from the fused data. The attention mechanism module is used to calculate the weights of the feature map, highlight important features, and suppress irrelevant features. The target detection module is used to output the type, location, and operating status of the power grid equipment. Step (2) Attention mechanism calculation; The weights of the feature maps are calculated by the module to suppress irrelevant features. The calculation formula of the improved attention mechanism is: In formula (1) ，A is the attention weight, X is the input feature map, is the weight matrix, is the bias vector, is the activation function, is the normalization function, C represents the number of channels of the feature map; Step (3) loss function calculation; The improved loss function combines Focal Loss and Dice Loss to improve the model's detection performance for targets of different categories and sizes. The Focal Loss is used to deal with the imbalance problem of positive and negative samples, and the Dice Loss focuses on the segmentation of the target area. The improved loss function expression is: In formula (2), To adjust the parameters, is the probability predicted by the model, A and B are the areas of the predicted results and the true labels respectively, It is a smoothing term to prevent division by zero errors; Step (4) target detection and output; The fused data is input into the improved YOLOv8 network. After feature extraction, attention mechanism and loss function optimization, the type, location and operating status of the power grid equipment are output to ensure the accuracy and robustness of the detection results.

5. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1, characterized in that: The working method of the improved ResNet-50 convolutional neural network is: 1) Convolutional feature enhancement; By performing sliding calculations on the convolution kernel of the fusion attention mechanism, local features are associated with the global channel. The expression of the convolution operation is: In formula (3), For the l The convolutional layer is at position ( i , j ), For the Layer at position ( ), For the first The convolution kernel of the layer is at position ( m , n ), For the The bias of the layer; a represents the channel attention coefficient, Indicates the The attention weight of the k-th channel of the layer, C represents the number of channels; 2) Adaptive pooling compression; A hybrid pooling operation is used to balance the spatial information preservation and noise suppression of the feature map. The pooling calculation expression is: In formula (4), For the The pooling layer is at position The output of , S is the pooling window, s is the pooling step size, Represents the pooling strategy weight, which is used to balance the contribution of maximum pooling and average pooling. represents the size of the pooling window, Indicates the Layer feature map at position ;3) Feature map optimization; Optimize feature expression through normalization and enhancement of feature maps; 4) Multimodal feature fusion; The optimized feature map is subjected to multimodal feature fusion to generate a comprehensive feature vector.

6. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1, characterized in that: The incremental online learning iCaRL algorithm works as follows: 1) Model and data initialization; Initialize the model parameters of the improved YOLOv8 target detection module, multimodal feature analysis module, and fault diagnosis module; 2) Learning and storing old knowledge; Use the initial training dataset to train the model of each module; During the training process, the model parameters are continuously adjusted through the back-propagation algorithm to minimize the loss function. After the training is completed, the current state of the model is stored, and representative samples are selected from the training data set as old knowledge samples and stored in the sample buffer. 3) new knowledge acquisition and processing; When new data samples arrive, the new data is pre-processed and the grid equipment information contained in the new data is analyzed to determine whether there are new equipment types, location characteristics, and operating status patterns; 4) Hybrid dataset construction; Take old knowledge samples from the sample buffer and merge them with new data samples to build a mixed data set. During the merging process, the old knowledge samples and new data samples can be weighted according to different strategies to balance the proportion of old and new knowledge in training. 5) Incremental model update; The models of the improved YOLOv8 object detection module, multimodal feature analysis module, and fault diagnosis module are incrementally trained using a mixed dataset. During training, the iCaRL algorithm utilizes information from old knowledge samples and combines the gradient information of old knowledge samples with the gradient information of new knowledge samples during backpropagation to jointly update the model parameters. 6) Sample buffer update; After training is completed, the sample buffer is updated according to the rules. A representative sampling method is used to select a portion of samples from the mixed data set and retain them in the buffer. At the same time, outdated samples are removed to ensure that the number and quality of samples in the buffer are relatively stable. 7) Repeated updates and optimizations; Based on the update of new data, repeat steps 3)-6) to continuously update and optimize the model incrementally so that the model can adapt to changes in the operating status of power grid equipment in a timely manner.

7. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1, characterized in that: The improved neural network model adopts a multi-layer perceptron MLP model for fault diagnosis, wherein the multi-layer perceptron MLP model includes a hidden layer and an output layer, and the MLP model includes one or more hidden layers, each hidden layer consisting of multiple neurons; The hidden layer performs nonlinear transformation on the input data to extract complex features in the data; The calculation formula of the hidden layer is: In formula (5), For the The output of the jth neuron in the hidden layer, For the Layer to The connection weights of the layer, For the The input of the i-th neuron in the layer, For the The bias of the layer, is the activation function; a represents the channel attention coefficient, Indicates the The attention weight of the k-th channel of the layer, C represents the number of channels; The number of neurons in the output layer is usually consistent with the number of fault categories, and each neuron corresponds to a fault category. The calculation formula of the output layer is: In formula (6), is the output of the kth neuron in the output layer, is the connection weight from the hidden layer to the output layer, is the output of the jth neuron in the hidden layer, is the bias of the output layer, Represents the adaptive normalization coefficient.

8. The improved YOLOv8 power grid equipment operating status fault monitoring system according to claim 1, characterized in that: The working method of the alarm module is: First, the feature map output by the multimodal feature analysis module is converted into a one-dimensional feature vector sequence and normalized. An attention mechanism is introduced into the LSTM unit to enhance the ability to capture key time step features and optimize the structures of the input gate, forget gate, and output gate. Then, the input feature vector sequence is processed by the improved LSTM unit to extract time series features, and feature fusion is performed to generate a comprehensive feature vector. The extracted features are then classified to determine whether the equipment has a fault and the fault type, and the fault is graded according to its severity. Finally, the fault status and fault type of the equipment are output, and the confidence of the diagnosis is expressed by a probability value. At the same time, the diagnosis results are fed back to the model, and the model parameters are updated in real time through an online learning algorithm to adapt to changes in the operating status of power grid equipment.

Images