A power equipment fault detection method and system based on a deep learning network

Abstract

The application relates to the technical field of power mode recognition, in particular to a power equipment fault detection method and system based on a deep learning network, which comprises the following steps: collecting and preprocessing multimodal data such as power time series and equipment images; monitoring the resource state of an edge node in real time, and adjusting the configuration parameters of a lightweight feature extraction model according to a preset rule; adopting a bidirectional cross-modal attention mechanism, calculating query, key and value representation and attention weight, dynamically weighting and fusing projection feature vectors of different modes, and generating a unified fault representation vector; calculating a fault probability by using the representation vector, and adaptively optimizing the parameters of a probability calculation model through online incremental learning; finally, calculating a task priority, and determining whether data is uploaded to a higher-level system for more detailed analysis according to the priority and a real-time resource state. The application can effectively improve the real-time performance, accuracy and adaptive capacity of edge-end power equipment fault detection.

Description

Technical Field

[0001] This invention relates to the field of power pattern recognition technology, specifically to a method and system for detecting power equipment faults based on deep learning networks. Background Technology

[0002] Traditional centralized data analysis relying on remote dispatch centers or master stations is insufficient to meet the demands of real-time processing of massive amounts of data and rapid, accurate fault location and isolation within smart substations. This is especially true in applications requiring high-frequency transient signals or rapid local decision-making, where data transmission latency and network bandwidth become significant bottlenecks. Edge computing, by deploying computing and analytical capabilities to edge nodes close to the data sources within the substation, enables real-time on-site analysis of locally collected data. This significantly improves fault response speed, reduces reliance on communication networks, and enhances the local closed-loop capability of data processing.

[0003] While edge computing offers numerous advantages to smart substations, deploying highly complex deep learning models on resource-constrained edge nodes still presents significant challenges. First, although edge computing devices may have more resources than some terminal devices, they still face limitations in computing power and memory storage compared to cloud platforms, especially when multiple complex diagnostic models targeting different devices need to be run simultaneously, highlighting resource bottlenecks. Second, the differences in feature distribution, anomaly patterns, and environmental noise interference of multimodal devices require edge computing models to not only maintain classification accuracy but also possess cross-domain generalization capabilities, resulting in high model design complexity. Furthermore, ensuring efficient and secure integration and collaboration between edge intelligent applications and existing substation control and protection systems, avoiding mutual interference, is a critical technical issue that urgently needs to be addressed. Existing solutions often struggle to balance the detection accuracy, inference efficiency, and adaptability to diverse equipment and complex operating conditions of deep learning models within the limited resources of the substation edge.

[0004] To address this, a method and system for power equipment fault detection based on deep learning networks are proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for power equipment fault detection based on deep learning networks. Executed on edge computing nodes, it aims to resolve the contradiction between limited edge computing resources and the need for high-precision real-time detection. The method includes: collecting and preprocessing multimodal data such as power time series and equipment images using edge nodes; monitoring the resource status of edge nodes in real time and dynamically adjusting the configuration parameters of a lightweight feature extraction model according to preset rules to achieve resource-adaptive and efficient feature extraction; employing a bidirectional cross-modal attention mechanism to dynamically weight and fuse projected feature vectors from different modalities by calculating query, key-value representations, and attention weights to generate a unified fault representation vector; calculating the fault probability using this representation vector and adaptively optimizing the parameters of the probability calculation model based on locally detected and stored new abnormal samples through an online incremental learning mechanism; finally, combining equipment importance and fault probability calculation task priority, and intelligently scheduling tasks based on this priority and real-time resource status to determine whether to send data to a higher-level system for more refined analysis. This effectively improves the real-time performance, accuracy, and adaptability of power equipment fault detection at the edge computing end.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A power equipment fault detection method based on deep learning networks includes:

[0008] Use edge nodes to collect and preprocess power time-series data and equipment image data;

[0009] The system resource status of the edge nodes is monitored, and the configuration parameters of the lightweight image feature extraction model and the temporal feature extraction model are adjusted through a dynamic pruning controller to obtain a set of feature vectors.

[0010] The feature vector set is mapped to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; an attention weight matrix is ​​calculated through a bidirectional cross-modal attention mechanism to generate an attention-weighted feature vector; the attention-weighted feature vector is concatenated with the projected feature vector set to generate a fault characterization vector.

[0011] The fault probability of the fault representation vector is calculated using a CNN-LSTM network; potential new anomalies are detected and recorded based on the fault probability, and the fault probability is verified and output using an incremental learning algorithm.

[0012] The system calculates the importance of the equipment and prioritizes it based on the weighted average of the failure probability; then, based on the priority and the system resource status, it adaptively decides whether to send the data to the higher-level system for analysis.

[0013] Furthermore, the resource mode determined by the dynamic pruning controller based on the comparison result between the system resource status and the first threshold and the second threshold includes: a resource-sufficient mode, in which the complete model calculation is executed if the system resource status does not exceed the first threshold; a resource-constrained mode, in which either the specified intermediate calculation layer is skipped or the calculation path is adjusted if the system resource status exceeds the first threshold but does not exceed the second threshold; and a resource-scarce mode, in which the simplified calculation path including only the core feature extraction layer is executed if the system resource status exceeds the second threshold.

[0014] Furthermore, the calculation process of the attention weight matrix includes:

[0015] For the first and second projected feature vectors in the projected feature vector set, generate corresponding first query representation, second query representation, first key representation, and second key representation, respectively; calculate a first cross-modal attention score based on the interaction between the first query representation and the second key representation; calculate a second cross-modal attention score based on the interaction between the second query representation and the first key representation; normalize the first cross-modal attention score and the second cross-modal attention score to obtain the corresponding weight values ​​in the attention weight matrix.

[0016] Furthermore, the generation of the attention-weighted feature vector includes:

[0017] For the first and second projected feature vectors in the set of projected feature vectors, corresponding first and second value representations are generated respectively; based on the weight values ​​of the first and second projected feature vectors corresponding to the second projected feature vectors, the second value representations are weighted and summed to obtain a first attention-weighted feature; based on the weight values ​​of the second and first projected feature vectors on the first projected feature vectors, the first value representations are weighted and summed to obtain a second attention-weighted feature; the attention-weighted feature vector includes the first attention-weighted feature and the second attention-weighted feature.

[0018] Furthermore, the process of detecting and recording potential new anomalies includes:

[0019] If the fault probability exceeds the fault threshold, and the cosine similarity between the fault representation vector corresponding to the fault probability and the feature vector of the abnormal sample already stored in the edge sample library is lower than a preset similarity threshold, then it is determined to be a potential new anomaly; the feature vector and fault label information of the potential new anomaly are stored in the edge sample library; when the number of samples stored in the edge sample library reaches a preset upper limit, the earliest abnormal sample feature vector is replaced by a FIFO strategy.

[0020] Furthermore, the process of verifying and outputting the fault probability using an incremental learning algorithm is based on the MAML algorithm, including: randomly selecting a support set and a query set containing K samples from the edge sample library; obtaining the current parameters of the layer to be optimized in the CNN-LSTM network; calculating the current task loss including cross-entropy loss and L2 regularization term based on the support set; calculating temporary model parameters using gradient descent with a first learning rate; calculating the meta-loss under the temporary model parameters using the query set, and generating global parameters using gradient descent with a second learning rate.

[0021] Furthermore, the method also includes:

[0022] When the adaptive decision is to send the data to the upper-level system for analysis, the upper-level system uses a deep learning model with higher complexity than the lightweight image feature extraction model and the temporal feature extraction model deployed on the edge node to perform fault diagnosis on the sent data, and returns the diagnosis results to the edge node to verify the fault probability, and / or uses the sent data to perform offline retraining of the lightweight image feature extraction model and the temporal feature extraction model.

[0023] A power equipment fault detection system based on deep learning networks includes:

[0024] The data acquisition module is used to collect and preprocess power time-series data and equipment image data using edge nodes;

[0025] The vector generation module monitors the system resource status of the edge nodes, adjusts the configuration parameters of the lightweight image feature extraction model and the temporal feature extraction model through a dynamic pruning controller to obtain a set of feature vectors; maps the set of feature vectors to a unified dimensional space through an independent linear projection layer to obtain a set of projected feature vectors; calculates the attention weight matrix through a bidirectional cross-modal attention mechanism to generate attention-weighted feature vectors; and concatenates the attention-weighted feature vectors with the set of projected feature vectors to generate a fault characterization vector.

[0026] The fault detection module is used to calculate the fault probability of the fault representation vector using a CNN-LSTM network; detect and record potential new anomalies based on the fault probability, verify and output the fault probability using an incremental learning algorithm; calculate the importance of the equipment and the priority under the weighted fault probability; and adaptively decide whether to send the data to the upper-level system for analysis based on the priority and the system resource status.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] 1. This invention integrates a real-time resource monitoring and dynamic model adjustment mechanism for edge nodes. Through a dynamic pruning controller, it adaptively matches resource availability (sufficient, limited, and strained) modes based on real-time resource status such as CPU and memory, and dynamically adjusts the computational configuration of a pre-set lightweight feature extraction model. Compared to deploying static models of fixed complexity at the edge or relying entirely on the cloud, this invention achieves on-demand allocation and efficient utilization of edge computing resources, improving the real-time performance, accuracy, and adaptability of fault detection for edge power equipment while ensuring low-latency response.

[0029] 2. This invention integrates multimodal data processing and online learning technologies. Through a bidirectional cross-modal attention mechanism, it dynamically calculates and utilizes the correlation weights between features from different sensor data to generate a more discriminative unified fault representation vector. This representation is then used to calculate the fault probability, and combined with a MAML-based online incremental learning algorithm, it performs small-sample and rapid adaptive model optimization based on new anomaly samples discovered and stored locally at the edge. This improves the real-time performance, accuracy, and adaptability of fault detection in edge-end power equipment.

[0030] 3. This invention constructs an intelligent edge and upper-level hierarchical collaborative reasoning architecture. By calculating the priority of the importance level of fusion devices and the real-time failure probability, and combining it with the current resource load status of edge nodes for adaptive task scheduling, it determines whether to send high-priority or edge-problem fault analysis tasks to the upper-level system with stronger computing power and more complex models. This ensures that critical faults are diagnosed in a timely manner, avoids edge resources being blocked by non-urgent tasks, and achieves the complementary advantages of rapid edge response and in-depth central analysis, thereby improving the real-time performance, accuracy, and adaptability of fault detection for power equipment at the edge. Attached Figure Description

[0031] Figure 1 This invention provides a flowchart of a power equipment fault detection method based on deep learning networks.

[0032] Figure 2 A schematic diagram illustrating the process of forming a feature vector set is provided for this invention;

[0033] Figure 3 This invention provides a schematic diagram of the structure of a power equipment fault detection system based on a deep learning network. Detailed Implementation

[0034] 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 embodiments of the present invention, and not all embodiments. 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.

[0035] Please see Figures 1 to 3 This invention provides a method and system for fault detection of power equipment based on deep learning networks, the technical solution of which is as follows:

[0036] Example 1:

[0037] This embodiment provides and elaborates on the specific implementation process of the method according to the present invention in a smart substation. This process is mainly deployed and executed on edge computing nodes installed within the substation (e.g., intelligent terminal devices (IEDs) and edge servers with certain computing capabilities). The core objective is to achieve real-time monitoring of the operating status of key power equipment within the substation (in this embodiment, the main power transformer and high-voltage circuit breaker are examples), and to achieve high-precision, low-latency, and adaptive fault detection and preliminary diagnosis under conditions of limited edge computing resources.

[0038] like Figure 1 As shown, a method and system for fault detection of power equipment based on deep learning networks includes:

[0039] Use edge nodes to collect and preprocess power time-series data and equipment image data;

[0040] The data acquisition steps in this embodiment aim to provide comprehensive, real-time multimodal input data for an edge fault detection system based on deep learning networks. The target equipment for acquisition includes a critical 220kV main power transformer (labeled T1) and its associated 220kV high-voltage circuit breaker (labeled CB1) within a smart substation. The types of data acquired cover electrical quantities, chemical quantities, mechanical quantities, thermal quantities, and visual information. Among these, electrical, chemical, and mechanical quantities constitute power time-series data, while thermal and visual information constitute equipment image data.

[0041] Specifically, the data acquired by T1 includes partial discharge (PD) signals, dissolved gas in oil (DGA), vibration signals, and temperature distribution. The data acquired by CB1 includes the status of operating mechanisms, main circuit current, auxiliary contact status, gas status, and appearance. These raw data streams are transmitted in real-time or near real-time via fiber optics or Ethernet to edge computing nodes deployed within the substation, and stored in designated input interfaces or memory buffers to support subsequent fault detection and analysis.

[0042] Furthermore, the system resource status of the edge nodes is monitored, and the configuration parameters of the lightweight image feature extraction model (IFE) and the temporal feature extraction model (TFE) are adjusted through the dynamic pruning controller (DPC) to obtain a set of feature vectors.

[0043] Among them, the system resource status reflects the operation of edge computing nodes at a specific point in time, and mainly focuses on two important indicators: one is the average CPU load, which represents the average busyness of the CPU in the past minute, expressed as a percentage; the other is the available memory percentage, which represents the proportion of currently unused physical memory to the total physical memory.

[0044] DPC refers to a software control module deployed on edge nodes that generates configuration parameters in real time based on the current resource status. This guides the IFE and TFE on how to adjust their computation methods during inference, thereby ensuring that the system can still run efficiently with limited resources.

[0045] In addition, such as Figure 2 As shown, in this embodiment, IFE employs a lightweight model based on the EfficientNet-Lite0 architecture, which undergoes INT8 quantization and static structure pruning optimization. Its input is a preprocessed 112×112 pixel image tensor, and its output is a 128-dimensional feature vector. TFE employs a lightweight variant based on a Temporal Convolutional Network (TCN), also undergoing INT8 quantization. This model's input is a preprocessed temporal data window with a length of 1024 sampling points, and its output is a 64-dimensional feature vector.

[0046] Furthermore, for example Figure 2 As shown, DPC continuously monitors system resource status, including average CPU load C and unavailable memory percentage Mem. Two resource thresholds are set: a first threshold Th1 (e.g., C = 50%, Mem = 20%) and a second threshold Th2 (e.g., C = 85%, Mem = 40%). DPC compares the current resource status (C, Mem) with the thresholds to determine the resource pattern, including:

[0047] In the resource-sufficient mode, if the system resource status does not exceed the first threshold, then the full model calculation is performed; for example, if C≤50% and Mem≤20%, then IFE... Active_Blocks =7, TFE Active_Blocks =5; where |Mem| represents the absolute value of the available memory percentage, IFE Active_Blocks Indicates the number of major computational stage blocks activated when the specified IFE model performs inference (range 4-7), TFE Active_BlocksIndicates the number of residual blocks activated when the specified TFE model performs inference (range 3 - 5).

[0048] Resource - constrained mode: If the system resource status exceeds the first threshold and does not exceed the second threshold, skip either the specified intermediate calculation layer or adjust the calculation path. For example, if (50% < C ≤ 85% or 20% < Mem ≤ 40%), then set IFE Active_Blocks = 5 (skip the last 2 MBConv blocks), TFE Active_Blocks = 4 (skip the last 1 residual block); alternatively, adjust the calculation path, including: instruct it to use a pre - set smaller convolution kernel (e.g., switch from 5×5 to 3×3).

[0049] Resource - tight mode: If the system resource status exceeds the second threshold, execute the most simplified calculation path that only includes the core feature extraction layer. For example, if C > 85% or Mem > 40%, set IFE Active_Blocks = 4 (only execute the first 4 MBConv blocks), TFE Active_Blocks = 3 (only execute the first 3 residual blocks), and for the convolution layers that are not skipped, the controller can instruct it to use a pre - set smaller convolution kernel.

[0050] When new pre - processed power time - series data and device image data arrive, first query the current configuration parameters (IFE Active_Blocks and TFE Active_Blocks ) through DPC. Subsequently, execute the IFE model and pass the IFE Active_Blocks parameters to the inference engine or the internal logic of the model, enabling it to only calculate the specified number of front - end blocks and output a 128 - dimensional image feature vector. Similarly, execute the TFE model, pass the TFE Active_Blocks parameters, enabling it to only calculate the specified number of front - end blocks and output a 64 - dimensional time - series feature vector. Finally, combine these two feature vectors with meta - information such as timestamps and device IDs to form a feature vector set and pass it to the next stage (feature fusion).

[0051] By introducing a dynamic pruning controller and a configurable lightweight model, it is possible to dynamically adjust the computational amount of the model according to the real - time fluctuating computational resources of the edge node, avoiding system crashes or response delays caused by resource exhaustion, while ensuring that basic detection capabilities can still be provided under extreme resource constraints rather than complete failure. This not only improves the robustness of the system but also enables it to have the ability to continuously operate in resource - constrained environments, thus enhancing the real - time performance, accuracy, and adaptive ability of power equipment fault detection at the edge.

[0052] Furthermore, the feature vector set is mapped to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; an attention weight matrix is ​​calculated through a bidirectional cross-modal attention mechanism to generate an attention-weighted feature vector; the attention-weighted feature vector is concatenated with the projected feature vector set to generate a fault characterization vector.

[0053] Furthermore, suppose that for a power device (such as main transformer T1) at a certain timestamp t, this set of feature vectors contains the following feature vectors:

[0054] Image feature vector: 128 dimensions (EfficientNet-Lite0).

[0055] Temporal feature vector: 64-dimensional (TCN, vibration signal).

[0056] Temporal feature vector: 64-dimensional (TCN, partial discharge signal).

[0057] Temporal feature vector: 32 dimensions (TCN, DGA data).

[0058] Setting the target dimension to 128, an independent linear projection layer is constructed for the feature vectors of each input modality. A linear projection layer is essentially a fully connected layer (containing or without a linear activation function), and each projection layer contains a weight matrix W and an optional bias vector b. For example, an image projection layer is represented as:

[0059] img emb =Activation(ife features @W img +b img );

[0060] Among them, img emb It is the image projection feature vector; Activation() is an optional activation function, such as the ReLU activation function parameter; ife features It is an image feature vector, @ represents matrix multiplication, W img It is the image weight matrix, b img It is the image bias vector; (W) img ,b img These values ​​are learnable; they are optimized through backpropagation during the training of the deep learning network. This process resolves the dimensionality mismatch issue that arises when subsequent vector interactions are required (such as the dot product in attention calculations).

[0061] Furthermore, the calculation process of the attention weight matrix includes:

[0062] For the first and second projected feature vectors in the projected feature vector set, generate corresponding first query representation, second query representation, first key representation, and second key representation, respectively; calculate a first cross-modal attention score based on the interaction between the first query representation and the second key representation; calculate a second cross-modal attention score based on the interaction between the second query representation and the first key representation; normalize the first cross-modal attention score and the second cross-modal attention score to obtain the corresponding weight values ​​in the attention weight matrix.

[0063] Specifically, the set of projection feature vectors, taking two modes as examples: the image projection feature vector (dimension 128), as the first projection feature vector; and the vibration projection feature vector (dimension 128), as the second projection feature vector.

[0064] To calculate attention, a query (for initiating attention) and a key (for matching the attention received) need to be generated for each modality participating in the interaction. The first query is represented as Q1 = img emb @W q1 The first key is represented as K1 = img emb @W k1 The second query is represented as Q2 = vib emb @W q2 The second key is represented as K2 = vib emb @W k2 Among them, W q1 W q2 W k1 and W k2 These are the corresponding weights, vib emb The vibration projection feature vector;

[0065] Next, the attention score is calculated using a scaled dot product attention mechanism, specifically the first cross-modal attention score. Second crossmodal attention score in, For the transpose of K1, d is the transpose of K2 qk It is a scaling factor used to stabilize the gradient.

[0066] For Score 12 and Score 21 The attention weight matrix is ​​calculated by normalizing using the Softmax function, resulting in the first weight value W. 12 and W 21 The second set of weights. These weight matrices will be passed to the next sub-step to generate the final attention-weighted feature vector.

[0067] The attention weights are dynamically adjusted based on the input data, rather than being fixed. This means that the model can flexibly capture the correlation between different modes based on the content of real-time signals, such as determining whether local overheating in an image is related to a specific vibration mode or PD signal enhancement. Through this dynamic attention mechanism, the model can more intelligently integrate multimodal information, especially when dealing with complex faults, early weak faults, or multi-factor coupled faults, which can improve the fault diagnosis effect and stability of power equipment.

[0068] Furthermore, the generation of the attention-weighted feature vector includes:

[0069] For the first projection feature vector and the second projection feature vector in the projection feature vector set, generate corresponding first value representations and second value representations respectively;

[0070] Among them, the projected feature vector can be directly used as the value representation;

[0071] Based on the weight values ​​of the first projected feature vector corresponding to the second projected feature vector, the second value representation is weighted and summed to obtain the first attention-weighted feature. 12 , represented as:

[0072] weighted_feature 12 =matmul(W 12 ,V2);

[0073] Here, matmul() is the attention-weighted summation operation, and V2 is the second value representation.

[0074] Similarly, based on the weight values ​​of the first projection feature vector to the second projection feature vector, the first value representation is weighted and summed to obtain the second attention-weighted feature; the attention-weighted feature vector includes the first attention-weighted feature and the second attention-weighted feature.

[0075] If more modal pairs exist (such as image-partial discharge, vibration-partial discharge), the weighted feature vectors between them are repeatedly calculated, and the final output is a set containing all the calculated attention-weighted feature vectors.

[0076] Attention-weighted feature vectors are no longer simply vibration information, but vibration features closely related to the current image state after being "harmonized" with image information. Similarly, image features are also influenced by vibration information. This interaction gives the feature vectors cross-modal correlation information, making the final weighted features more prominent in conjunction with fault-related information. Irrelevant or noisy components are weakened in impact due to their low weight. This lays the foundation for the subsequent output of fault probabilities for power equipment by CNN-LSTM, thereby improving the real-time performance, accuracy, and adaptability of fault detection in edge-based power equipment.

[0077] Furthermore, the fault probability of the fault representation vector is calculated using a CNN-LSTM network; potential new anomalies are detected and recorded based on the fault probability, and the fault probability is verified and output using an incremental learning algorithm.

[0078] Furthermore, the CNN-LSTM network receives a unified fault representation vector. To fully utilize the LSTM's ability to process sequence information, the CNN-LSTM network does not process the representation vector at the current time point alone, but combines it with the representation vectors from several past time points to form a fixed-length time series as input. The 1D CNN layer first processes this sequence along the time dimension, extracting local spatiotemporal features. Subsequently, the LSTM layer further processes the CNN output to capture long-term temporal dependencies in the sequence. Finally, a probability vector is output through the sigmoid activation function, where each element represents the probability of belonging to the corresponding category, and the sum of all elements is 1.

[0079] Furthermore, the process of detecting and recording potential new anomalies includes:

[0080] If the fault probability exceeds the fault threshold, and the cosine similarity between the fault representation vector corresponding to the fault probability and the feature vector of the abnormal sample already stored in the edge sample library is lower than a preset similarity threshold, then it is determined to be a potential new anomaly; the feature vector and fault label information of the potential new anomaly are stored in the edge sample library; when the number of samples stored in the edge sample library reaches a preset upper limit, the earliest abnormal sample feature vector is replaced by a FIFO strategy.

[0081] The edge sample library refers to a limited-capacity (e.g., maximum Capacity = 50) database or memory structure stored locally on the edge node. It contains feature vectors of previously identified and confirmed anomalous samples and their corresponding fault labels. For example, the fault judgment threshold is 0.8, and the similarity threshold is 0.9. All stored anomalous sample feature vectors are retrieved from the edge sample library. The cosine similarity between the current feature vector and each existing vector in the library is calculated. The maximum similarity among all calculated similarities (0.75) indicates insufficient similarity to any known anomalous patterns in the sample library, and is therefore identified as a potential new anomalous sample. This new anomalous sample needs to be added to the edge sample library for subsequent learning. If the sample library size becomes 51 after adding the new anomalous sample vector, the earliest record added to the library is deleted according to a first-in, first-out (FIFO) strategy, maintaining the library size at 50.

[0082] By comparing similarity, only significantly different anomaly patterns are included in the learning sample library, avoiding repeated learning of known or similar patterns and improving the efficiency and targeting of online learning. Edge nodes can automatically discover and record anomaly patterns occurring in their specific monitored objects and operating environments. These patterns may not be fully reflected in the initial training data, thus gradually building a personalized local fault knowledge base. This mechanism provides a filtered and information-valuable source of small-sample data for subsequent efficient and stable online incremental learning (such as MAML), thereby improving the real-time performance, accuracy, and adaptability of fault detection for edge power equipment.

[0083] Furthermore, the process of verifying and outputting the fault probability using an incremental learning algorithm, based on the MAML algorithm, includes:

[0084] A support set (vector) containing K samples is randomly drawn from the edge sample library. s1 ,label s1 ) and query set (vector) q1 ,label q1 );

[0085] The support set consists of a small subset of labeled training samples used to simulate the model's rapid learning process on new tasks. The query set, on the other hand, is another subset of labeled samples, distinct from the support set, used to evaluate the model's performance after rapid adaptation using the support set; vector s1 and vector q1 For feature vectors, label s1 and label q1 This corresponds to the fault label.

[0086] Obtain the current parameters of the layer to be optimized in the CNN-LSTM network, i.e., the parameters μ of the current probability calculation layer;

[0087] Based on the support set, calculate the current task loss L, which includes cross-entropy loss and L2 regularization. S (μ), represented as:

[0088]

[0089] Where CrossEntropy() is the cross-entropy loss function, CNN-LSTM() is the CNN-LSTM network, k is the number of samples in the support set, λ is the L2 regularization strength, and pred si Let ||μ|| be the failure probability predicted based on μ, where i is the index variable. 2 Let μ be the square of the L2 norm of the parameter μ.

[0090] The temporary model parameters μ′ are calculated using gradient descent with the first learning rate:

[0091]

[0092] in, This is the gradient operator. The first learning rate is the internal learning rate, with a value of 0.001, used in the inner loop of MAML.

[0093] Using the query set, calculate the meta-loss L under the temporary model parameters. Q (μ′), and use the second learning rate to generate the global parameter μ through gradient descent. new :

[0094]

[0095] Among them, Q size The number of samples in the query set is denoted by , and the second learning rate is the inter-task learning rate with a value of 0.01, used in the outer loop of MAML.

[0096] Incremental learning enables CNN-LSTM networks to quickly adapt to new tasks (such as identifying new potential anomalies) with a small number of samples (K support set samples) and a small number of gradient updates. By updating only some layers (such as the classification head) and utilizing fewer samples, the computational and storage costs of online learning are reduced, enabling fault detection of power equipment on resource-constrained edge devices and improving its real-time performance and accuracy.

[0097] Furthermore, the importance of the calculated device is weighted by the probability of failure; based on the priority and the system resource status, an adaptive decision is made as to whether to send the data to the higher-level system for analysis.

[0098] Here, the equipment importance is the importance score of the i-th power equipment, representing its criticality. Then, a priority threshold is predefined. If the priority is greater than the priority threshold, it is a high priority and the data is always sent, regardless of the resource status; otherwise, if resources are sufficient, the data is not sent / processed locally / queued for low-priority sending; if resources are scarce, the data is not sent / queued for waiting for resources.

[0099] Furthermore, the method also includes:

[0100] When the adaptive decision is to send the data to the upper-level system for analysis, the upper-level system uses a deep learning model with higher complexity than the lightweight image feature extraction model and the temporal feature extraction model deployed on the edge node to perform fault diagnosis on the sent data, and returns the diagnosis results to the edge node to verify the fault probability, and / or the upper-level system uses the sent data to perform offline retraining of the lightweight image feature extraction model and the temporal feature extraction model.

[0101] The uploaded data includes the fault characterization vector and the feature vector set, etc.

[0102] The edge is responsible for rapid initial screening, dynamic extraction, and intelligent triage, while the upper-level system (cloud / site domain) focuses on detailed diagnosis and global analysis. This collaborative mechanism fully leverages the advantages of each component, flexibly adapting to different operation and maintenance strategies, risk preferences, and edge hardware configurations, thereby improving the real-time performance, accuracy, and adaptability of power equipment fault detection.

[0103] This invention achieves a balance between computational efficiency and accuracy at the edge by real-time resource monitoring and dynamic model optimization (such as dynamic pruning and computation path adjustment). Combined with a bidirectional cross-modal attention mechanism, it deeply mines the correlations between multi-source data and introduces online incremental learning to enhance the model's adaptability to new fault modes. Simultaneously, it constructs an intelligent scheduling mechanism based on task priority and an edge-cloud collaborative architecture, improving the accuracy of power equipment fault detection, resource utilization efficiency, and system generalization capability while ensuring rapid response. This forms an edge intelligent diagnostic solution integrating real-time performance, robustness, and adaptability.

[0104] Example 2:

[0105] This embodiment, based on Embodiment 1, further elaborates on a power equipment fault detection system based on deep learning networks, such as... Figure 3 As shown, it includes:

[0106] refer to Figure 3 The data acquisition module is used to collect and preprocess power time-series data and equipment image data using edge nodes;

[0107] refer to Figure 3 The vector generation module is used to monitor the system resource status of the edge nodes, adjust the configuration parameters of the lightweight image feature extraction model and the temporal feature extraction model through a dynamic pruning controller to obtain a set of feature vectors; map the set of feature vectors to a unified dimensional space through an independent linear projection layer to obtain a set of projected feature vectors; calculate the attention weight matrix through a bidirectional cross-modal attention mechanism to generate attention-weighted feature vectors; and concatenate the attention-weighted feature vectors with the set of projected feature vectors to generate a fault characterization vector.

[0108] refer to Figure 3 The fault detection module is used to calculate the fault probability of the fault representation vector using a CNN-LSTM network; detect and record potential new anomalies based on the fault probability, verify and output the fault probability using an incremental learning algorithm; calculate the importance of the equipment and the priority under the weighted fault probability; and adaptively decide whether to send the data to the upper-level system for analysis based on the priority and the system resource status.

[0109] Furthermore, the system also includes:

[0110] When the adaptive decision is to send the data to the upper-level system for analysis, the upper-level system uses a deep learning model with higher complexity than the lightweight image feature extraction model and the temporal feature extraction model deployed on the edge node to perform fault diagnosis on the sent data, and returns the diagnosis results to the edge node to verify the fault probability, and / or the upper-level system uses the sent data to perform offline retraining of the lightweight image feature extraction model and the temporal feature extraction model.

[0111] Specifically, the fault characterization vector, feature vector set, and metadata to be uploaded are packaged into a predefined format and encrypted. The encrypted data packet is then sent to the designated receiving endpoint of the upper-level system via a configured communication interface. The upper-level system receives and decrypts the data packet, extracting the feature vector and metadata.

[0112] The upper-level system loads the uncompressed full version of EfficientNet (or other advanced models such as ResNet and ViT), performs in-depth analysis of these features, and outputs more refined diagnostic results, such as specific fault types, severity, and confidence scores. The generated refined diagnostic results are packaged and encrypted, and sent back to the requesting edge node via the communication link. The edge node receives and decrypts the diagnostic results from the upper-level system, compares and records them with locally calculated fault probabilities to evaluate the accuracy of the edge model. Based on the detailed diagnostic results from the upper-level system, the local device status record is updated, and more specific and accurate alarm information may be triggered for maintenance personnel.

[0113] Alternatively, the upper-level system can archive and store the received data to form a valuable failure case library. Using this data, the lightweight feature extraction model (IFE, TFE) and the base version of CNN-LSTM deployed at the edge are retrained offline. Then, the new model is compressed and optimized, and finally, the updated lightweight model is pushed back to the edge nodes. This constitutes a long-term closed loop of model adaptation and performance improvement.

[0114] To verify the effectiveness of the hierarchical inference method proposed in this invention, which includes edge screening, adaptive scheduling, and fine-grained diagnosis of the upper-level system, its performance was compared with the following two benchmark methods:

[0115] Baseline Method 1 (Pure Edge Detection System): All detection logic is executed at the edge node. It adopts the same lightweight feature extraction, fusion, and probability calculation model as the edge end of this invention, and directly outputs the locally calculated fault probability without uploading data.

[0116] Benchmark Method 2 (Pure Cloud / Site Domain Detection System): Edge nodes are only responsible for basic data collection and forwarding, uploading preprocessed feature data (or raw data) to the upper-level system (simulated as a cloud server), where the upper-level system runs a highly complex and accurate deep learning model for comprehensive fault diagnosis.

[0117] As shown in Tables 1 and 2, while the baseline method 1 is highly efficient, it has a low recall rate when handling complex faults, and its overall performance is inferior to the system of this invention. The baseline method 2 has the highest accuracy, but this comes at the cost of sacrificing real-time performance and consuming a large amount of network bandwidth. This invention achieves an overall accuracy very close to that of a pure cloud system and a good balance in computational efficiency, thereby improving the accuracy of power equipment fault detection, resource utilization efficiency, and system generalization ability.

[0118] Table 1 Comparison of Fault Detection Performance

[0119] Detection system Overall accuracy Overall F1 score Critical / Complex Fault Recall Rate Benchmark Method 1 96.5％ 0.88 75.2％ Benchmark Method Two 99.2％ 0.96 97.5％ This invention 98.8％ 0.95 96.8％

[0120] Table 2 Comparison of System Efficiency

[0121] Detection system Average edge CPU utilization Average detection delay Average network bandwidth usage Benchmark Method 1 24％ 50ms Less than 0.1Mbps Benchmark Method Two 5％ 800ms Approximately 15.0 Mbps This invention 38％ 120ms Approximately 1.2 Mbps

[0122] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for fault detection of power equipment based on deep learning networks, characterized in that, include: Edge nodes collect and preprocess power time-series data and equipment image data; monitor the system resource status of the edge nodes, and adjust the configuration parameters of the lightweight image feature extraction model and the time-series feature extraction model through a dynamic pruning controller to obtain a set of feature vectors; the dynamic pruning controller determines the resource mode based on the comparison result of the system resource status with a first threshold and a second threshold, including: a resource-sufficient mode, in which the full model calculation is performed if the system resource status does not exceed the first threshold; In resource-constrained mode, if the system resource status exceeds the first threshold but does not exceed the second threshold, either skip the specified intermediate computing layer or adjust the computing path; in resource-scarce mode, if the system resource status exceeds the second threshold, execute the simplified computing path that only includes the core feature extraction layer. The feature vector set is mapped to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; an attention weight matrix is ​​calculated through a bidirectional cross-modal attention mechanism to generate an attention-weighted feature vector; the attention-weighted feature vector is concatenated with the projected feature vector set to generate a fault characterization vector. The fault probability of the fault representation vector is calculated using a CNN-LSTM network; potential new anomalies are detected and recorded based on the fault probability, and the fault probability is verified and output using an incremental learning algorithm; the importance of the equipment and the priority under the weighted fault probability are calculated; and the system resource status is adaptively determined to send the data to the upper-level system for analysis based on the priority and the system resource status.

2. The power equipment fault detection method based on deep learning networks according to claim 1, characterized in that, The calculation process of the attention weight matrix includes: generating corresponding first query representation, second query representation, first key representation, and second key representation for the first and second projected feature vectors in the projected feature vector set, respectively; calculating a first cross-modal attention score based on the interaction between the first query representation and the second key representation; calculating a second cross-modal attention score based on the interaction between the second query representation and the first key representation; and normalizing the first cross-modal attention score and the second cross-modal attention score to obtain the corresponding weight values ​​in the attention weight matrix.

3. The power equipment fault detection method based on deep learning networks according to claim 1, characterized in that, The generation of the attention-weighted feature vector includes: generating a first value representation and a second value representation for the first projected feature vector and the second projected feature vector in the projected feature vector set, respectively; performing a weighted summation on the second value representation based on the weight value of the first projected feature vector corresponding to the second projected feature vector to obtain a first attention-weighted feature; and performing a weighted summation on the first value representation based on the weight value of the second projected feature vector relative to the first projected feature vector to obtain a second attention-weighted feature; the attention-weighted feature vector includes the first attention-weighted feature and the second attention-weighted feature.

4. The power equipment fault detection method based on deep learning networks according to claim 1, characterized in that, The process of detecting and recording potential new anomalies includes: if the fault probability exceeds a fault threshold, and the cosine similarity between the fault representation vector corresponding to the fault probability and the feature vector of the anomaly sample already stored in the edge sample library is lower than a preset similarity threshold, then it is determined to be a potential new anomaly; the feature vector and fault label information of the potential new anomaly are stored in the edge sample library; when the number of samples stored in the edge sample library reaches a preset upper limit, the earliest anomaly sample feature vector is replaced using a FIFO strategy.

5. The power equipment fault detection method based on deep learning networks according to claim 1, characterized in that, The process of verifying and outputting the fault probability using an incremental learning algorithm is based on the MAML algorithm and includes: randomly selecting a support set and a query set containing K samples from the edge sample library; obtaining the current parameters of the layer to be optimized in the CNN-LSTM network; calculating the current task loss including cross-entropy loss and L2 regularization term based on the support set; calculating temporary model parameters using gradient descent with a first learning rate; calculating the meta-loss under the temporary model parameters using the query set, and generating global parameters using gradient descent with a second learning rate.

6. The power equipment fault detection method based on deep learning networks according to claim 1, characterized in that, The method further includes: when the adaptive decision is to send the data to the upper-level system for analysis, the upper-level system uses a deep learning model with higher complexity than the lightweight image feature extraction model and the temporal feature extraction model deployed on the edge node to perform fault diagnosis on the sent data, and returns the diagnosis results to the edge node to verify the fault probability, and / or uses the sent data to perform offline retraining of the lightweight image feature extraction model and the temporal feature extraction model.

7. A power equipment fault detection system based on deep learning networks, characterized in that, include: The data acquisition module is used to collect and preprocess power time-series data and equipment image data using edge nodes; The vector generation module is used to monitor the system resource status of the edge nodes and adjust the configuration parameters of the lightweight image feature extraction model and the temporal feature extraction model through the dynamic pruning controller to obtain a set of feature vectors. The dynamic pruning controller determines resource modes based on the comparison results of the system resource status with a first threshold and a second threshold, including: a resource-sufficient mode, where if the system resource status does not exceed the first threshold, the complete model calculation is performed; a resource-constrained mode, where if the system resource status exceeds the first threshold but does not exceed the second threshold, either the specified intermediate calculation layer is skipped or the calculation path is adjusted; and a resource-scarce mode, where if the system resource status exceeds the second threshold, the simplified calculation path including only the core feature extraction layer is executed. The feature vector set is mapped to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; an attention weight matrix is ​​calculated through a bidirectional cross-modal attention mechanism to generate attention-weighted feature vectors; and the attention-weighted feature vectors are concatenated with the projected feature vector set to generate a fault characterization vector. The fault detection module is used to calculate the fault probability of the fault representation vector using a CNN-LSTM network; detect and record potential new anomalies based on the fault probability, verify and output the fault probability using an incremental learning algorithm; calculate the importance of the equipment and the priority under the weighted fault probability; and adaptively decide whether to send the data to the upper-level system for analysis based on the priority and the system resource status.

Images