Power equipment fault detection method and system based on deep learning network

Abstract

The invention 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, and the method comprises the steps: collecting and preprocessing multi-modal data, such as a power time sequence and an equipment image; monitoring an edge node resource state in real time, and adjusting configuration parameters of the lightweight feature extraction model according to a preset rule; a bidirectional cross-modal attention mechanism is adopted, dynamic weighted fusion is carried out on projection feature vectors of different modals by calculating query, key and value representation and attention weight, and a unified fault representation vector is generated; calculating a fault probability by using the representation vector, and adaptively optimizing parameters of a probability calculation model through online incremental learning; and finally, calculating a task priority, and according to the priority and a real-time resource state, determining whether to upload the data to a superior system for finer analysis. According to the method, the real-time performance, the accuracy and the self-adaptive capability of edge end power equipment fault detection can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power pattern recognition, and in particular to a method and system for detecting power equipment faults based on a deep learning network. Background Art

[0002] The traditional model of relying on remote dispatch centers or master stations for centralized data analysis struggles to meet the demands of real-time processing of massive amounts of data and rapid, accurate fault location and isolation within smart substations. Data transmission latency and network bandwidth become significant bottlenecks, especially when processing high-frequency transient signals or in applications requiring rapid local decision-making. Edge computing, by deploying computing and analysis capabilities to edge nodes close to data sources within the substation, enables on-site, real-time analysis of locally collected data, significantly improving fault response speed, reducing reliance on communication networks, and enhancing local closed-loop data processing capabilities.

[0003] While edge computing offers numerous advantages for smart substations, deploying highly complex deep learning models in resource-constrained edge nodes still faces significant challenges. First, while edge computing devices may possess more resources than some end devices, they still face resource limitations such as computing power and memory storage compared to cloud platforms. This resource bottleneck is particularly pronounced when multiple complex diagnostic models targeting different devices need to be run simultaneously. Second, the varying feature distributions of multimodal devices, the diversity of abnormal patterns, and environmental noise interference require edge computing models to not only maintain classification accuracy but also possess cross-domain generalization capabilities, which complicates model design. Furthermore, ensuring efficient and secure integration and collaboration between edge intelligent applications and existing station-wide control and protection systems to avoid mutual interference is a key technical issue that urgently needs to be addressed. Existing solutions often struggle to balance the detection accuracy, inference efficiency, and adaptability of deep learning models to diverse devices and complex operating conditions within the limited resources of substation edges.

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

[0005] The present invention aims to provide a method and system for power equipment fault detection based on a deep learning network. This method, executed at edge computing nodes, aims to address the conflict between limited edge computing resources and the need for high-precision, real-time detection. The method includes: using edge nodes to collect and preprocess multimodal data, such as power time series and equipment images; real-time monitoring of edge node resource status; and dynamically adjusting the configuration parameters of a lightweight feature extraction model based on preset rules to achieve efficient, resource-adaptive feature extraction. A bidirectional cross-modal attention mechanism is employed to dynamically weighted-fuse projected feature vectors from different modalities by calculating query, key, and value representations, along with attention weights, to generate a unified fault representation vector. This representation vector is used to calculate the fault probability, and through an online incremental learning mechanism, the parameters of the probability calculation model are adaptively optimized based on newly detected and stored anomaly samples. Finally, combining device importance with the priority of the fault probability calculation task, the system intelligently schedules tasks based on this priority and real-time resource status, deciding whether to upload the data to a higher-level system for more detailed analysis. This method can effectively improve the real-time, accuracy, and adaptability of power equipment fault detection at the edge computing end.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting faults in power equipment based on a deep learning network, comprising:

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

[0009] Monitoring the system resource status of the edge node, adjusting the configuration parameters of the lightweight image feature extraction model and the time series feature extraction model through a dynamic pruning controller, and obtaining a feature vector set;

[0010] Mapping the feature vector set to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; calculating the attention weight matrix through a bidirectional cross-modal attention mechanism to generate an attention-weighted feature vector; concatenating the attention-weighted feature vector with the projected feature vector set to generate a fault representation vector;

[0011] Calculating the fault probability of the fault characterization vector using a CNN-LSTM network; detecting and recording potential new anomalies based on the fault probability, and verifying and outputting the fault probability using an incremental learning algorithm;

[0012] Calculate the device importance and the priority weighted by the failure probability; and adaptively determine whether to send the information to a higher-level system for analysis based on the priority and the system resource status.

[0013] Furthermore, the dynamic pruning controller determines the resource mode based on the comparison results of the system resource status with the first threshold and the second threshold, including: resource-sufficient mode, if the system resource status does not exceed the first threshold, the complete model calculation is performed; resource-constrained mode, if the system resource status exceeds the first threshold and does not exceed the second threshold, skip the specified intermediate calculation layer and adjust the calculation path; resource-strained mode, if the system resource status exceeds the second threshold, execute the simplified calculation path that only includes the core feature extraction layer.

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

[0015] For the first projected feature vector and the second projected feature vector in the projected feature vector set, the corresponding first query representation, second query representation, first key representation and second key representation are generated respectively; based on the interaction between the first query representation and the second key representation, a first cross-modal attention score is calculated; based on the interaction between the second query representation and the first key representation, a second cross-modal attention score is calculated; the first cross-modal attention score and the second cross-modal attention score are normalized 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 projected feature vector and the second projected feature vector in the projected feature vector set, a corresponding first value representation and a second value representation are generated respectively; based on the weight value of the first projected feature vector corresponding to the second projected feature vector, the second value representation is weighted summed to obtain a first attention weighted feature; based on the weight value of the second projected feature vector for the first projected feature vector, the first value representation is weighted 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 characterization vector corresponding to the fault probability and the abnormal sample feature vector stored in the edge sample library is lower than the preset similarity threshold, it is determined to be a potential new abnormality; the feature vector and fault label information of the potential new abnormality are stored in the edge sample library; when the number of samples stored in the edge sample library reaches the preset upper limit, the FIFO strategy is adopted to replace the earliest abnormal sample feature vector.

[0020] Furthermore, the process of verifying and outputting the fault probability using an incremental learning algorithm is based on the MAML algorithm, including: randomly extracting 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; based on the support set, calculating the current task loss including the cross entropy loss and the L2 regularization term; using a first learning rate to obtain temporary model parameters through gradient descent; using the query set, calculating the meta-loss under the temporary model parameters, and using a second learning rate to generate global parameters through gradient descent.

[0021] Furthermore, the method further comprises:

[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 time series feature extraction model deployed on the edge node to perform fault diagnosis on the uploaded data, and returns the diagnosis result to the edge node for verifying the fault probability, and / or uses the uploaded data to perform offline retraining of the lightweight image feature extraction model and the time series feature extraction model.

[0023] A power equipment fault detection system based on a deep learning network, comprising:

[0024] A data acquisition module, used to collect and pre-process power time series data and equipment image data using edge nodes;

[0025] A vector generation module is configured to monitor the system resource status of the edge node, adjust the configuration parameters of the lightweight image feature extraction model and the time series feature extraction model through a dynamic pruning controller, and 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 an attention-weighted feature vector; and concatenate the attention-weighted feature vector with the set of projected feature vectors to generate a fault representation vector.

[0026] The fault detection module is configured to calculate the fault probability of the fault characterization 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 device importance and priority weighted by the fault probability; and adaptively determine whether to send the data to a higher-level system for analysis based on the priority and the system resource status.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. The present invention integrates real-time resource monitoring and dynamic model adjustment mechanisms for edge nodes. The dynamic pruning controller adaptively matches resource-sufficient, resource-constrained, and resource-constrained modes based on the real-time status of resources such as CPU and memory, and dynamically adjusts the computational configuration of the preset lightweight feature extraction model. Compared to deploying static models of fixed complexity at the edge or relying entirely on the cloud, the present invention achieves on-demand allocation and efficient utilization of edge computing resources, while ensuring low-latency response and improving the real-time, accuracy, and adaptability of edge power equipment fault detection.

[0029] 2. This invention integrates multimodal data processing and online learning technologies. Through a bidirectional cross-modal attention mechanism, the correlation weights between different sensor data features are dynamically calculated and utilized to generate a more discriminative unified fault representation vector. This representation is further used to calculate the fault probability. Combined with a MAML-based online incremental learning algorithm, this method performs low-sample and rapid adaptive model optimization based on new abnormal samples discovered and stored locally at the edge, thereby improving the real-time, accuracy, and adaptability of edge power equipment fault detection.

[0030] 3. The present invention constructs an intelligent edge and upper-level hierarchical collaborative reasoning architecture. By calculating the priority of the fusion device importance level and the real-time fault probability, and combining the current resource load status of the edge node for adaptive task scheduling, it is decided whether to send high-priority or edge-doubtful fault analysis tasks to the upper-level system with stronger computing power and complex models. This ensures that critical faults are diagnosed in a timely manner, avoids the blocking of edge resources by non-urgent tasks, and realizes the complementary advantages of rapid edge response and in-depth analysis at the center, thereby improving the real-time, accuracy and adaptability of edge-end power equipment fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The present invention provides a flow chart of a method for detecting faults in power equipment based on a deep learning network;

[0032] Figure 2 Provides a schematic flow chart of forming a feature vector set for the present invention;

[0033] Figure 3 The present invention provides a structural schematic diagram of an electric power equipment fault detection system based on a deep learning network. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] See also Figures 1 to 3 The present invention provides a method and system for detecting power equipment faults based on a deep learning network. The technical solution 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 the edge computing nodes installed in the substation (for example, intelligent terminal IEDs and edge servers with certain computing capabilities). The core goal is to achieve real-time monitoring of the operating status of key power equipment in the station (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 the condition of limited edge computing resources.

[0038] like Figure 1 As shown, a method and system for detecting power equipment faults based on a deep learning network include:

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

[0040] The data collection step of this embodiment is intended to provide comprehensive, real-time, multimodal input data for the edge fault detection system based on a deep learning network. The target devices for collection include a key 220kV main power transformer (labeled as T1) and its associated 220kV high-voltage circuit breaker (labeled as CB1) in the smart substation. The types of data collected include electrical quantities, chemical quantities, mechanical quantities, heat, and visual information. Among them, electrical quantities, chemical quantities, and mechanical quantities constitute the power time series data, and heat and visual information constitute the device image data.

[0041] Specifically, T1 collects data including partial discharge (PD) signals, dissolved gas in oil (DGA), vibration signals, and temperature distribution. CB1 collects data including operating mechanism status, main circuit current, auxiliary contact status, gas status, and external appearance. These raw data streams are aggregated in real time or near real time via optical fiber or Ethernet transmission methods to edge computing nodes deployed within the substation. They are then stored in designated input interfaces or memory buffers to support subsequent fault detection and analysis.

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

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

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

[0045] In addition, if Figure 2 As shown, in this embodiment, IFE adopts a lightweight model based on the EfficientNet-Lite0 architecture, which has been INT8 quantized and optimized by static structure pruning. Its input is a preprocessed 112×112 pixel image tensor, and the output is a 128-dimensional feature vector. TFE adopts a lightweight variant based on the Temporal Convolutional Network (TCN), which has also been INT8 quantized. The input of the model is a preprocessed time series data window with a length of 1024 sampling points, and the output is a 64-dimensional feature vector.

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

[0047] Resource sufficient mode, if the system resource status does not exceed the first threshold, then perform full model calculation; 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 percentage of available memory, IFE Active_Blocks Indicates the number of main computational phase 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 - limited 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); or, adjust the calculation path, including: instruct it to use a preset smaller convolution kernel (such as switching 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 preset smaller convolution kernel.

[0050] [[ID=十七]]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 computing 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 - limited environments, thus enhancing the real - time performance, accuracy, and adaptive ability of power equipment fault detection at the edge. [[ID=三十]]

[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 representation vector;

[0053] Furthermore, it is assumed that for a power device (such as main transformer T1) at a certain timestamp t, the feature vector set includes the following feature vectors:

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

[0055] Time series feature vector: dimension 64 (TCN, vibration signal).

[0056] Time series feature vector: dimension 64 (TCN, partial discharge signal).

[0057] Time series feature vector: dimension 32 (TCN, DGA data).

[0058] Set the target dimension to 128 and construct a separate linear projection layer for the feature vector of each input modality. The linear projection layer is essentially a fully connected layer (without or with a linear activation function). Each projection layer contains a weight matrix W and an optional bias vector b. For example, the image projection layer is represented as:

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

[0060] Among them, img emb is the image projection feature vector; Activation() is an optional activation function, such as ReLU activation function parameters; ife features is the image feature vector, @ represents matrix multiplication, W img is the image weight matrix, b img is the image bias vector; (W img ,b img ) are learnable, and their values ​​are optimized through backpropagation during the training of the deep learning network. This process solves the dimensionality mismatch problem when vector interactions are required later (such as dot products in attention calculations).

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

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

[0063] Specifically, the projected feature vector set takes two modalities as examples: the image projected feature vector (dimension 128) is used as the first projected feature vector, and the vibration projected feature vector (dimension 128) is used as the second projected feature vector.

[0064] In order to calculate attention, it is necessary to generate a query (for initiating attention) and a key (for being followed) representation for each modality participating in the interaction. The first query is represented as Q1 = img emb @W q1 ; The first key is represented by K1=img emb @W k1 ; The second query is expressed as Q2 = vib emb @W q2 ; The second key is represented by K2=vib emb @W k2 Among them, W q1 、W q2 、W k1 and W k2 are the corresponding weights, vib emb is the vibration projection eigenvector;

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

[0066] Score 12 and Score 21 Use the Softmax function for normalization, and calculate the first weight value W corresponding to the attention weight matrix 12 and W 21 Second weight value. These weight matrices will be passed to the next sub-step to generate the final attention weighted feature vector.

[0067] Attention weights are dynamically adjusted based on the input data rather than being fixed. This means the model can flexibly capture correlations between different modalities based on the content of real-time signals, such as determining whether localized overheating in an image is associated with a specific vibration mode or enhanced PD signal. This dynamic attention mechanism enables the model to more intelligently integrate multimodal information, improving the effectiveness and stability of power equipment fault diagnosis, especially when dealing with complex faults, early-stage weak faults, or multi-factor coupled faults.

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

[0069] For a first projected eigenvector and a second projected eigenvector in the projected eigenvector set, respectively generating a corresponding first value representation and a second value representation;

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

[0071] Based on the weight value 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 weighted_feature 12 , expressed as:

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

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

[0074] Similarly, based on the weight value of the second projected feature vector on the first projected feature vector, the first value representation is 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.

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

[0076] The attention-weighted feature vector is no longer simply vibration information; it is instead a vibration signature closely related to the current image state after being "harmonized" with image information. Similarly, image features are also influenced by vibration information. This interaction imbues the feature vector with cross-modal correlation information, allowing the final weighted feature to better highlight joint information related to the fault. Irrelevant or noisy components have a lower weight, reducing their impact. This lays the foundation for the subsequent CNN-LSTM output of power equipment failure probabilities, thereby improving the real-time, accuracy, and adaptability of edge power equipment fault detection.

[0077] Furthermore, a CNN-LSTM network is used to calculate the fault probability of the fault characterization vector; potential new anomalies are detected and recorded based on the fault probability, and an incremental learning algorithm is used to verify and output the fault probability.

[0078] Furthermore, the CNN-LSTM network receives a unified fault representation vector. To fully leverage the LSTM's ability to process sequence information, the CNN-LSTM network does not process the representation vector of the current time point alone. Instead, it combines it with the representation vectors of 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 sigmoid activation function is used to output a probability vector, 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 characterization vector corresponding to the fault probability and the abnormal sample feature vector stored in the edge sample library is lower than the preset similarity threshold, it is determined to be a potential new abnormality; the feature vector and fault label information of the potential new abnormality are stored in the edge sample library; when the number of samples stored in the edge sample library reaches the preset upper limit, the FIFO strategy is adopted to replace the earliest abnormal sample feature vector.

[0081] Among them, the edge sample library represents a database or memory structure with limited capacity (for example, upper limit Capacity = 50) stored locally in the edge node, which contains previously identified and confirmed abnormal sample feature vectors and their corresponding fault labels. For example, the fault judgment threshold is 0.8 and the similarity threshold is 0.9. Retrieve all stored abnormal sample feature vectors from the edge sample library, calculate the cosine similarity between the current feature vector and each stored vector in the library, and find the maximum value of 0.75 among all calculated similarities, indicating that it is not similar enough to all known abnormal patterns in the sample library, and then it is determined to be a potential new abnormality. It needs to be stored in the edge sample library for subsequent learning. If the sample library size becomes 51 after the addition, the earliest record stored in the sample library will be deleted according to the first-in-first-out (FIFO) strategy to keep the library size at 50.

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

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

[0084] Randomly extract a support set (vector) containing K samples from the marginal sample library s1 ,label s1 ) and querysets (vector q1 ,label q1 );

[0085] The support set is a small set of labeled training samples used to simulate the model's rapid learning process on new tasks. The query set is another set of samples with the same labels but different from the support set, used to evaluate the model's new performance after rapid adaptation through the support set. s1 and vector q1 is the feature vector, label s1 and label q1 is the corresponding fault label.

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

[0087] Based on the support set, calculate the current task loss L including the cross entropy loss and L2 regularization term S (μ), expressed as:

[0088]

[0089] Among them, 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 is the failure probability predicted based on μ, i is the index variable, ||μ|| 2 is the L2 norm squared of the parameter μ.

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

[0091]

[0092] in, is the gradient operator, and the first learning rate is the internal learning rate with a value of 0.001, which is 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 by gradient descent new :

[0094]

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

[0096] Through incremental learning, the CNN-LSTM network can 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 a small number of samples, the computational and storage costs of online learning are reduced, enabling it to realize fault detection of power equipment on resource-constrained edge devices and improve its real-time performance and accuracy.

[0097] Furthermore, the importance of the equipment and the priority weighted by the failure probability are calculated; and according to the priority and the system resource status, a decision is adaptively made whether to send the information to the upper-level system for analysis.

[0098] The device importance score is the importance score of the i-th power device, indicating its criticality. A predefined priority threshold is then defined. If the priority is greater than the threshold, the data is considered high priority and is always sent, regardless of resource availability. Otherwise, if resources are sufficient, data is not sent, processed locally, or queued for low-priority data. If resources are limited, data is not sent or queued for resources.

[0099] Furthermore, the method further comprises:

[0100] When the adaptive decision is to send the data to an 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 time series feature extraction model deployed at the edge node to perform fault diagnosis on the uploaded data, and returns the diagnosis result to the edge node for verifying the fault probability, and / or the upper-level system uses the uploaded data to perform offline retraining of the lightweight image feature extraction model and the time series feature extraction model.

[0101] The uploaded data includes the fault characterization vector and the feature vector set;

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

[0103] The present invention achieves a balance between computing efficiency and accuracy at the edge through real-time resource monitoring and dynamic model optimization (such as dynamic pruning and calculation path adjustment). Combined with a bidirectional cross-modal attention mechanism, it deeply mines the correlation between multi-source data and introduces online incremental learning to enhance the model's adaptability to new fault modes. At the same time, an intelligent scheduling mechanism based on task priority and an edge-cloud collaborative architecture are constructed to improve the power equipment fault detection accuracy, resource utilization efficiency, and system generalization capabilities while ensuring rapid response, forming an edge intelligent diagnosis solution that integrates real-time, robustness, and adaptability.

[0104] Example 2:

[0105] Based on the first embodiment, this embodiment further elaborates a power equipment fault detection system based on a deep learning network. Figure 3 Shown, including:

[0106] refer to Figure 3 A data acquisition module for collecting and preprocessing power time series data and equipment image data using edge nodes;

[0107] refer to Figure 3 A vector generation module is configured to monitor the system resource status of the edge node, 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 feature vector set; map the feature vector set to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; calculate the attention weight matrix through a bidirectional cross-modal attention mechanism to generate an attention weighted feature vector; and concatenate the attention weighted feature vector with the projected feature vector set to generate a fault representation vector.

[0108] refer to Figure 3 The fault detection module is used to calculate the fault probability of the fault characterization vector using a CNN-LSTM network; detect and record potential new anomalies based on the fault probability, and use an incremental learning algorithm to verify and output the fault probability; calculate the device importance and priority weighted by the 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 further comprises:

[0110] When the adaptive decision is to send the data to an 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 time series feature extraction model deployed at the edge node to perform fault diagnosis on the uploaded data, and returns the diagnosis result to the edge node for verifying the fault probability, and / or the upper-level system uses the uploaded data to perform offline retraining of the lightweight image feature extraction model and the time series 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 detailed diagnostic results, such as the specific fault type, severity, and confidence score. The generated detailed diagnostic results are packaged and encrypted and sent back to the edge node that initiated the request via the communication link. The edge node receives and decrypts the diagnostic results of the upper-level system, compares and records them with the locally calculated fault probability, and evaluates the accuracy of the edge model. Based on the detailed diagnostic results of the upper-level system, the local device status record is updated, and more specific and accurate alarm information may be triggered for operation and maintenance personnel.

[0113] Alternatively, the upper-level system archives and stores the received data, forming a valuable fault case library. This data is used to offline retrain the lightweight feature extraction models (IFE, TFE) and the basic version of the CNN-LSTM deployed at the edge. The new models are then compressed and optimized, and the updated lightweight models are ultimately pushed back to the edge nodes. This forms a long-term closed loop of model adaptation and performance improvement.

[0114] To verify the effectiveness of the hierarchical reasoning approach proposed in this paper, which includes initial edge screening, adaptive scheduling, and refined diagnosis by the upper-level system, we compared its performance with the following two benchmark methods:

[0115] Benchmark method 1 (pure edge detection system): All detection logic is executed at the edge node, using the same lightweight feature extraction, fusion, and probability calculation model as the edge end of the present invention, directly outputting the locally calculated fault probability without sending data up.

[0116] Benchmark method 2 (pure cloud / station-based detection system): The edge node is only responsible for basic data collection and forwarding, uploading the pre-processed feature data (or raw data) to the upper-level system (simulated as a cloud server), which runs a high-complexity and high-precision deep learning model for comprehensive fault diagnosis.

[0117] As shown in Tables 1 and 2, while Baseline Method 1 is highly efficient, it suffers from a low recall rate when handling complex faults, and its overall performance is inferior to that of the system of the present invention. Baseline Method 2 achieves the highest accuracy, but this comes at the expense of real-time performance and significant network bandwidth consumption. The present invention achieves an overall accuracy very close to that of a pure cloud-based system and strikes a good balance in computational efficiency, improving power equipment fault detection accuracy, resource utilization efficiency, and system generalization capabilities.

[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 2 99.2％ 0.96 97.5％ The present invention 98.8％ 0.95 96.8％

[0120] Table 2 System efficiency comparison

[0121] Detection system Average edge CPU usage Average detection delay Average network bandwidth usage Benchmark Method 1 24％ 50ms Less than 0.1Mbps Benchmark Method 2 5％ 800ms About 15.0Mbps The present invention 38％ 120ms About 1.2Mbps

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

Claims

1. A method for detecting power equipment faults based on a deep learning network, characterized in that: include: Edge nodes collect and pre-process power time series data and equipment image data; Monitoring the system resource status of the edge node, adjusting the configuration parameters of the lightweight image feature extraction model and the time series feature extraction model through a dynamic pruning controller, and obtaining a feature vector set; Mapping the feature vector set to a unified dimensional space through an independent linear projection layer to obtain a projected feature vector set; calculating the attention weight matrix through a bidirectional cross-modal attention mechanism to generate an attention-weighted feature vector; concatenating the attention-weighted feature vector with the projected feature vector set to generate a fault representation vector; Calculating the fault probability of the fault characterization vector using a CNN-LSTM network; detecting and recording potential new anomalies based on the fault probability, and verifying and outputting the fault probability using an incremental learning algorithm; Calculate the device importance and the priority weighted by the failure probability; and adaptively determine whether to send the information to a higher-level system for analysis based on the priority and the system resource status.

2. The method for detecting power equipment faults based on a deep learning network according to claim 1, characterized in that: The dynamic pruning controller determines the resource mode based on the comparison results of the system resource status with the first threshold and the second threshold, including: resource-sufficient mode, if the system resource status does not exceed the first threshold, the complete model calculation is performed; resource-constrained mode, if the system resource status exceeds the first threshold but does not exceed the second threshold, skipping the specified intermediate calculation layer and adjusting the calculation path; resource-scarce mode, if the system resource status exceeds the second threshold, executing the simplified calculation path that only includes the core feature extraction layer.

3. The method for detecting power equipment faults based on a deep learning network according to claim 1, characterized in that: The calculation process of the attention weight matrix includes: For the first projected feature vector and the second projected feature vector in the projected feature vector set, the corresponding first query representation, second query representation, first key representation and second key representation are generated respectively; based on the interaction between the first query representation and the second key representation, a first cross-modal attention score is calculated; based on the interaction between the second query representation and the first key representation, a second cross-modal attention score is calculated; the first cross-modal attention score and the second cross-modal attention score are normalized to obtain the corresponding weight values ​​in the attention weight matrix.

4. The method for detecting power equipment faults based on a deep learning network according to claim 1, wherein: The generation of the attention weighted feature vector includes: For the first projected feature vector and the second projected feature vector in the projected feature vector set, a corresponding first value representation and a second value representation are generated respectively; based on the weight value of the first projected feature vector corresponding to the second projected feature vector, the second value representation is weighted summed to obtain a first attention weighted feature; based on the weight value of the second projected feature vector for the first projected feature vector, the first value representation is weighted 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.

5. The method for detecting power equipment faults based on a deep learning network according to claim 1, characterized in that: The process for detecting and recording potential new anomalies includes: If the fault probability exceeds the fault threshold, and the cosine similarity between the fault characterization vector corresponding to the fault probability and the abnormal sample feature vector stored in the edge sample library is lower than the preset similarity threshold, it is determined to be a potential new abnormality; the feature vector and fault label information of the potential new abnormality are stored in the edge sample library; when the number of samples stored in the edge sample library reaches the preset upper limit, the FIFO strategy is adopted to replace the earliest abnormal sample feature vector.

6. The method for detecting power equipment faults based on a deep learning network 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, including: randomly extracting a support set and a query set containing K samples from an edge sample library; obtaining current parameters of the layer to be optimized in the CNN-LSTM network; based on the support set, calculating the current task loss including the cross entropy loss and the L2 regularization term; using a first learning rate to obtain temporary model parameters through gradient descent; using the query set, calculating the meta-loss under the temporary model parameters, and generating global parameters through gradient descent using a second learning rate.

7. The method for detecting power equipment faults based on a deep learning network according to claim 1, characterized in that: The method further comprises: 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 time series feature extraction model deployed on the edge node to perform fault diagnosis on the uploaded data, and returns the diagnosis result to the edge node for verifying the fault probability, and / or uses the uploaded data to perform offline retraining of the lightweight image feature extraction model and the time series feature extraction model.

8. A power equipment fault detection system based on deep learning network, characterized in that: include: A data acquisition module, used to collect and pre-process power time series data and equipment image data using edge nodes; A vector generation module is configured to monitor the system resource status of the edge node, adjust the configuration parameters of the lightweight image feature extraction model and the time series feature extraction model through a dynamic pruning controller, and 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 an attention-weighted feature vector; and concatenate the attention-weighted feature vector with the set of projected feature vectors to generate a fault representation vector. The fault detection module is configured to calculate the fault probability of the fault characterization 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 device importance and priority weighted by the fault probability; and adaptively determine whether to send the data to a higher-level system for analysis based on the priority and the system resource status.

Images