A lightweight power distribution terminal health assessment method fusing expert knowledge and deep learning

Abstract

The present application relates to the technical field of smart grid equipment state monitoring and predictive maintenance, in particular to a lightweight power distribution terminal health evaluation method fusing expert knowledge and deep learning. The present application converts expert knowledge into computable feature indicators and differentiable constraint conditions, uses domain prior knowledge to guide model learning, thereby improving the learning effect of the model under the condition of less labeled samples, and enabling the judgment basis output by the model to be traced back to specific expert rules, thereby reducing the dependence of the model on a large amount of labeled data and improving the transparency of the decision-making process. A lightweight dual-channel hybrid neural network is constructed, deep separable convolution is used to reduce the model parameter quantity, different types of features are respectively accessed by relying on the dual-channel structure, and lightweight strategies such as small convolution kernel and channel compression are combined, so that the model can be deployed on edge embedded devices and meet the timeliness requirements of real-time online evaluation on the premise of maintaining evaluation accuracy.

Description

Technical Field

[0001] This invention belongs to the field of smart grid equipment condition monitoring and predictive maintenance technology, specifically involving a lightweight distribution terminal health assessment method that integrates expert knowledge and deep learning. It is applicable to embedded distribution automation terminals with limited computing power and storage resources, such as feeder terminals (FTU), distribution terminals (DTU), and transformer terminals (TTU). Background Technology

[0002] As the core equipment of the power distribution automation system, the reliability of the distribution terminal directly affects the power supply quality. Currently, the operation and maintenance mode of the distribution terminal is still mainly based on regular inspections and reactive repairs. This passive operation and maintenance mode is not only costly and slow to respond, but also makes it difficult to detect potential faults caused by factors such as component aging and environmental changes in advance.

[0003] The existing technologies and research mainly face the following problems: First, the data collection dimensions are relatively limited. Most existing systems focus on communication status and business data such as telemetry and remote signaling, but the collection of operational data reflecting the health of the terminal hardware itself is not comprehensive enough. This includes data such as CPU load, memory usage, internal temperature, port error rate, and storage media wear. Furthermore, there is a lack of correlation analysis of the above multi-dimensional data. Second, there is a contradiction between model accuracy and edge deployment capabilities. High-precision deep learning models have complex structures and high computational costs, making them unsuitable for direct deployment on resource-constrained power distribution terminals or edge gateways. Raw data typically needs to be uploaded to the cloud for processing, which introduces communication latency and other issues, leading to insufficient real-time performance and high network bandwidth consumption. Traditional machine learning models that can run at the edge (such as linear models and decision trees) have limited accuracy and cannot output unified quantitative health indicators. Third, power distribution terminal hardware has high reliability, but real-world fault data is scarce. Moreover, hardware performance degradation is a continuous and gradual process, making accurate labeling difficult. This makes pure data-driven deep learning models unstable during training in small-sample, unlabeled scenarios, prone to overfitting. Fourth, the expert knowledge fusion method is simple, the model interpretability is poor, and the early warning capability is weak. After long-term development, the power industry has accumulated a large amount of mature expert experience and physical models. However, most existing technologies only use expert rules for data preprocessing or correction after model output, and fail to deeply embed domain knowledge into the model's feature learning and training process. This makes the model lack physical constraints when data is insufficient, and the decision-making process is not interpretable enough.

[0004] In summary, existing technologies cannot simultaneously address the requirements of edge device resource constraints, small-sample training, and model interpretability. Therefore, this invention proposes a power distribution terminal health assessment scheme that balances assessment accuracy, model lightweighting, and decision interpretability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a lightweight power distribution terminal health assessment method that integrates expert knowledge and deep learning. Expert knowledge is converted into computable feature indicators and differentiable constraints. Domain prior knowledge guides model learning, thereby improving the model's learning performance even with limited labeled samples. Furthermore, the model's output judgments are traceable to specific expert rules, reducing its dependence on large amounts of labeled data and enhancing the transparency of the decision-making process. A lightweight dual-channel hybrid neural network is constructed, employing depthwise separable convolutions to reduce the number of model parameters. Different types of features are accessed through the dual-channel structure, and lightweight strategies such as small convolutional kernels and channel compression are combined to maintain assessment accuracy while enabling the model to be deployed on edge embedded devices and meeting the timeliness requirements of real-time online assessment. A ternary composite loss function consisting of supervised loss, consistency regularization loss, and expert rule loss is constructed to build a semi-supervised training framework. This framework utilizes labeled data, unlabeled data, and expert knowledge for joint parameter optimization, reducing the model's dependence on large amounts of labeled samples during training.

[0006] To solve the aforementioned technical problem, the technical solution adopted by the present invention is: a lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning, comprising the following steps: Step 1: Standardization and interpretable feature construction of multi-source data based on expert knowledge.

[0007] This step primarily involves transforming the tacit experience accumulated by operations and maintenance experts into explicit, computable features, which serve as prior inputs to the model. This improves evaluation accuracy with a limited number of labeled samples and ensures the traceability of the model's judgment criteria. This includes: 1.1 Standardized Acquisition of Multi-Source Heterogeneous Data: Based on data type and source, the collected raw data is divided into three categories: Category A (Power Distribution Terminal Body Status Data): Reflects the real-time operating status of the core hardware components of the power distribution terminal, including CPU utilization. U cpu Memory usage M Internal temperature T int Battery voltage V bat Wear and tear of Flash memory cells W flash .

[0008] Category B (Business Application Performance Data): Reflects the reliability of the power distribution terminal's business functions, including the telemetry data transmission qualification rate. R telem Remote execution success rate R ctrl.

[0009] Category C (Environmental and Lifetime Information Data): Describes the external environment and equipment aging status of the power distribution terminal, including ambient temperature. T env ,humidity H env Cumulative operational time t op Historical fault records.

[0010] 1.2 Characterization of Expert Knowledge: Based on the physical operating mechanism of power distribution terminals and on-site operation and maintenance experience, statistical characteristics and trend characteristics with clear physical meaning are constructed, including: (1) Thermal operation health characteristics: defining the temperature-load synergy factor C thermal This quantifies the relationship between CPU load and internal temperature changes, reflecting the performance of the cooling system. The calculation formula is: , in For CPU utilization U cpu With internal temperature T int The covariance is used to calculate the correlation between CPU utilization and internal temperature; The standard deviation of CPU utilization. The standard deviation of the internal temperature of the power distribution terminal represents the degree of fluctuation in CPU utilization and internal temperature. This indicator is calculated based on covariance and standard deviation, and its value ranges from -1 to 1. The closer the value is to 1, the normal the heat dissipation condition is. When the value drops significantly or is negative, it indicates that the heat dissipation of the equipment is abnormal.

[0011] (2) Power system aging characteristics: defining battery health factors H bat The formula for quantifying the performance degradation trend of a battery over time is as follows: , in V bat This is the current measured voltage of the battery. V nom This is the battery's rated voltage. t op The cumulative operating time of the battery. t life Battery expected lifespan H bat ∈[0,1], the closer the value is to 1, the healthier the battery is.

[0012] (3) Business function decay characteristics: Based on the least squares estimation method of univariate linear regression slope, a univariate linear regression is performed on the remote control failure rate within the sliding window, and the regression slope k ctrl The formula for calculating the attenuation rate of remote control function is as follows: , Where: n is the number of sampling points in the sliding window, the sliding window is preset to a fixed length, and the window width is n (n is a positive integer); This is the sampling sequence number. If the sampling interval is uniform, it can be a consecutive natural number 1, 2, ..., n; if the sampling interval is not uniform, it should be replaced with the actual cumulative timestamp. The remote control failure rate is the rate at the i-th sampling point. The slope... k ctrl The meaning is: the average change in remote control failure rate within a unit sampling interval, that is, the instantaneous degradation rate of remote control function. When the... k ctrl If the value remains greater than 0 for multiple consecutive sliding window periods, the remote control function is considered to have experienced systematic degradation; the larger the value, the faster the degradation rate.

[0013] (4) Storage medium wear characteristics: Define the Flash abnormal wear index I flash， The calculation formula is as follows: , in W flash This represents the number of programming / erase cycles that have occurred in the Flash memory. This refers to the maximum number of cycles specified for this type of Flash chip (as given in the datasheet). I flash ∈[0,1], the larger the value, the more severe the wear and tear, and the higher the risk of storage failure.

[0014] Step 2: Construct a lightweight dual-channel hybrid evaluation network.

[0015] To meet the low-latency deployment requirements in resource-constrained edge environments, this invention designs a lightweight dual-channel hybrid network, which is divided into three parts: a shared feature extraction backbone, an expert feature injection channel, and a dual-task output head.

[0016] 2.1 Shared Feature Extraction Backbone: The shared feature extraction backbone consists of multi-channel time-series signals, including but not limited to: CPU utilization. U cpu Memory usage M Internal temperature T int Battery voltage V batWear and tear of Flash memory cells W flash、 Ambient temperature T env ,humidity H env Cumulative operational time t op 、 Remote execution success rate R ctrl At least one or more of the above-mentioned multi-channel time-series signals are used for feature encoding. The input feature matrix is ​​formed by the above-mentioned multi-channel time-series signals and is then encoded by the backbone network. The backbone network adopts a structure of 4 layers of one-dimensional depthwise separable convolution + 2 layers of max pooling + 1 layer of Flatten layer. Compared with standard convolution, this structure can reduce the number of parameters by about 70% and the amount of computation by about 80%, making it suitable for embedded device deployment scenarios.

[0017] 2.2 Expert Feature Injection Channel: Set up an independent expert feature injection channel to inject the scalar expert features constructed in step 1 ( H bat ,、 I flash ,、 C thermal Using expert features as input, a lightweight SE-Net (Squeeze-and-Excitation Network) module is embedded before the fully connected layers. This module first performs global average pooling on the expert features, and then dynamically generates attention weights for each expert feature through two fully connected layers, automatically highlighting factors that have a greater impact on the current health assessment (such as when...). I flash When the value rises abnormally, its weight increases automatically, thus realizing the integration of data-driven models and expert knowledge.

[0018] 2.3 Feature Fusion and Dual-Task Output Head: The abstract feature vector from the shared backbone output is concatenated with the attention-weighted expert feature vector at the feature fusion layer, and then two parallel lightweight output heads are generated to form a multi-task learning structure. The division of labor between the two heads is as follows: (1) Health regression head: Outputs continuous health index HI, with a value range of [0,1], where 1 represents brand new health and 0 represents complete failure.

[0019] (2) Health status classification header: output four categories, with the status classification criteria as follows: healthy (HI≥0.9), attention (0.8≤HI<0.9), abnormal (0.6≤HI<0.8), and serious (HI<0.6).

[0020] Maintenance personnel can directly determine the equipment status level based on the classification results, which facilitates the arrangement of maintenance plans.

[0021] 2.4 Lightweight Model Post-processing: After the model training is completed, multi-level compression is carried out in a fixed order of first structured sparsity, then functional compression, and then numerical compression. This order can effectively avoid the pruning bias caused by weight distortion after quantization. Under the same compression ratio, the accuracy retention rate can be improved by about 10%~15%. The steps are as follows: (1) Channel pruning: Based on the contribution of feature channels, redundant channels are pruned to reduce the number of parameters and computation. (2) Knowledge distillation: The pruned model is used as the teacher model to train the student model with a more compact structure. The model volume is further reduced under the premise of basically no loss of accuracy. (3) Fixed-point quantization: The 32-bit floating-point parameters of the student model are converted into 8-bit integer parameters. After quantization, the model storage volume is compressed to 1 / 4 of the original floating-point model, and the inference speed is significantly improved.

[0022] Step 3: Semi-supervised training method incorporating expert rules: To address the issues of scarce on-site labeled samples and high costs of manual labeling, this invention employs a semi-supervised training scheme, combining a small amount of labeled data with a large amount of unlabeled data to complete model training, and using a ternary composite loss function for overall constraint.

[0023] 3.1 Ternary Composite Loss Function: The total loss function consists of three parts: supervision loss, consistency regularization loss, and expert rule loss, defined as follows: , The supervised loss Lsup is calculated based on labeled samples and is obtained by weighted summation of health regression loss and state classification loss: +(1- ) , Where α is the weighting coefficient, with a value range of [0,1] (preferably α=0.5, i.e., both are equally important), used to balance the contribution ratio of regression and classification tasks.

[0024] Health regression loss Mean squared error (MSE) is used to measure the difference between the health level predicted by the model and the actual health level labeled by humans. The calculation formula is as follows: , in This represents the number of labeled samples. The health score of the i-th sample predicted by the model; The health status of this sample is manually labeled.

[0025] Equipment condition classification loss Cross-entropy (CE) loss is used to measure the difference between the model's predicted probability distributions of the four states and the true class labels. The calculation formula is as follows: , Where the subscript i represents the sequence number of the labeled sample. This represents the total number of labeled samples; the subscript c is the status category index, where c is the category index (1=healthy, 2=attention, 3=abnormal, 4=severe). The one-hot encoding of the true class label of the i-th sample is set to 1 when the sample belongs to class c, and 0 otherwise. This is the predicted probability that the i-th sample belongs to class c, as output by the Softmax function of the model.

[0026] (2) Consistency regularization loss L con This method is applied to unlabeled samples. Two independent random perturbations are applied to the same unlabeled sample (the random perturbation is either adding Gaussian noise to the input data or randomly deactivating features in the intermediate layers of the network). The perturbed samples are then input into the model separately to obtain two health prediction values. and The difference between the two is measured by the mean squared error, and the calculation formula is as follows: , in, The number of unlabeled samples used in the calculation; and Let be the predicted health values ​​output by the model after the i-th unlabeled sample has undergone two sets of independent random perturbations. This loss function is used to force the model to be insensitive to small changes in the input data, thereby leveraging a large amount of unlabeled data to enhance the model's generalization ability and robustness.

[0027] (3) Loss due to expert rules L knowledge The operation and maintenance expert rules are transformed into differentiable constraint terms, and the calculation formula is as follows: , Where N is the total number of samples involved in the calculation (usually all training samples, including labeled and unlabeled data). I (Premise) i ) This is an indicator function. It's introduced to determine whether the expert features meet the rule preconditions. If the preconditions are met, the indicator function is 1; otherwise, it's 0. As can be seen from the expert rule loss, a penalty is applied when the expert rule meets the rule preconditions and the network output health index violates the upper limit requirement, ensuring that the evaluation results conform to the physical operating laws of the equipment.

[0028] The typical expert rules set in this scheme are as follows: Table 1. Example of Rules (Thresholds and upper limits are set based on expert experience) 3.2 Loss weight setting and training process: The weights λ1, λ2, and λ3 of the loss function are determined through cross-validation: the labeled samples are divided into training and validation sets, and the optimal weights are selected by searching within the preset weight range that minimizes the mean squared error of the health regression on the validation set. When there are few labeled samples, the expert rule loss weight λ3 can be appropriately increased.

[0029] During the training phase, labeled and unlabeled data are jointly input into the network to perform end-to-end optimization. After each training round, the model's loss value on the validation set is calculated. When the validation set loss no longer decreases for several consecutive rounds, the model is considered to have converged, and training is stopped. After the model converges, the lightweight processing in step 2.4 is performed to obtain the final deployable model.

[0030] 3.3 Differences between this invention and conventional practices: In traditional technical solutions, expert rules and deep learning are often loosely coupled. Common approaches include data screening and labeling before training, and rule validation of the results after model output. This invention embeds expert knowledge into the model at two levels. At the feature level, expert features such as temperature-load synergy coefficients and battery health factors are set as independent input channels and dynamically weighted by the SE-Net module. When a certain expert feature is abnormally prominent (e.g., increased Flash wear index), its corresponding weight automatically increases, making the model more attentive to these abnormal factors during inference. At the loss level, expert rules are transformed into differentiable loss terms. During training, if the model's output health level violates physical laws (e.g., severely aged battery with a high score), a penalty is applied to the loss term. This design is fundamentally different from existing loosely coupled techniques.

[0031] Step 4: Online evaluation and cloud-edge collaborative adaptive update. This invention adopts a three-level collaborative architecture of "terminal-edge-cloud" to realize closed-loop iteration of the model's entire life cycle.

[0032] 4.1 Model Deployment and Real-time Online Evaluation: The trained and lightweight model is deployed to the power distribution terminal or edge gateway. The terminal-side program continuously collects raw data, constructs hybrid feature vectors, and inputs them into the model, enabling deployment on edge embedded devices and meeting the timeliness requirements of real-time online evaluation. It outputs a health score (HI) and a four-category health status. When the HI falls below a preset threshold, or the device status is determined to be abnormal or severe, the system triggers multi-level warnings and simultaneously outputs key causative factors. These key causative factors are obtained by calculating the partial derivatives of each input feature with respect to the output health score (HI), using the absolute value of the partial derivative as the feature's contribution to the health score, and outputting the top three features with the highest contribution as key causative factors. This provides direction for maintenance and troubleshooting; warning information and on-site operational data are synchronously uploaded to the edge gateway and cloud platform for unified management.

[0033] 4.2 Cloud-edge collaborative adaptive update: The cloud platform continuously aggregates new fault cases and early warning data, and uses incremental learning to fine-tune and optimize the model. The new version of the model is batch-deployed to edge devices and terminals via OTA encryption protocol. The terminal verifies the effectiveness of the new version of the model locally. If the verification is successful, the upgrade is completed automatically. If the verification fails, the abnormal information is sent back to the cloud for re-optimization, forming a closed-loop iteration mechanism that enables the system to continuously adapt to new operating conditions and new fault modes.

[0034] The beneficial effects of this invention are: Deep knowledge fusion, with both accuracy and interpretability: This invention embeds expert knowledge into the feature layer (dynamically weighting expert features through the SE-Net attention mechanism) and the loss function (constrained by differentiable expert rules), forming a two-level deep coupling mechanism. This enables the model to maintain high generalization ability even with a small number of labeled samples. At the same time, each evaluation result can be traced back to a specific expert feature or rule, effectively solving the problem of poor interpretability of pure data-driven models.

[0035] Lightweight design to meet the real-time operation requirements of embedded terminals: This invention optimizes the model from two aspects: model structure design and post-training compression. Structurally, it adopts depthwise separable convolution and dual-channel architecture, which has high parameter efficiency. In terms of compression, it combines compression in a specific order of "pruning → distillation → quantization". The final model meets the deployment requirements of real-time inference on a single-chip microcontroller.

[0036] A semi-supervised training mechanism reduces reliance on labeled data: This invention constructs a ternary composite loss function comprising supervision loss, consistency regularization loss, and expert rule loss, enabling joint optimization using a small amount of labeled data, a large amount of unlabeled data, and domain expert knowledge. This mechanism effectively alleviates the problems of scarce and difficult-to-label fault samples in power distribution terminals, making it more suitable for real-world scenarios with limited power distribution terminal fault samples.

[0037] With cloud-edge collaboration and OTA updates, the system has self-evolution capabilities: the model can be upgraded online to adapt to terminal upgrades and new fault modes, avoiding a decrease in accuracy after long-term model operation and reducing manual maintenance costs.

[0038] Promoting the transformation of operation and maintenance model to predictive maintenance: This invention outputs a quantifiable health index, multi-level risk warnings, and key causal factor analysis, enabling operation and maintenance personnel to identify equipment degradation trends and locate the main sources of risk in advance, thereby transforming the operation and maintenance model from the traditional "passive maintenance" to "proactive predictive maintenance". Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the overall workflow of the method of the present invention; Figure 2 This is a schematic diagram of the lightweight dual-channel hybrid evaluation model of the present invention; Figure 3 This is a diagram of the "end-edge-cloud" collaborative deployment architecture of this invention. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0041] Example 1 The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. In this embodiment, 100 FTU terminals in a distribution network in a certain urban area are selected as the evaluation objects. The verification is carried out based on the field operation data of the terminals for about 18 consecutive months. The dataset covers typical operating conditions such as normal operation of equipment, slow performance degradation, and sudden failure.

[0042] This embodiment discloses a lightweight power distribution terminal health assessment method that integrates expert knowledge and deep learning, such as... Figure 1 As shown, it includes the following steps: 1. Data and Feature Preparation: Collect three types of raw operational data from each device, including terminal status data (CPU utilization) U cpu Memory usage M Internal temperature T int Battery voltage V bat Wear and tear of Flash memory cells W flash etc.), business performance (data telemetry pass rate, pass rate R) telem Remote control execution success rate R ctrl ), environmental and lifespan information data (environmental temperature and humidity, commissioning time).

[0043] Six types of interpretable expert features are constructed: temperature-load synergy coefficient C. thermal Battery health factor H bat Remote control function attenuation rate k ctrl Flash Abnormal Wear Index I flash Power supply ripple RMS value V ripple_rms High load ratio during idle periods (R) idle_high In addition, based on the experience of on-site maintenance personnel, two expert rules are introduced and incorporated into the model loss function for constraint. The rules are as follows: R5: If the power ripple is continuously greater than the set threshold (e.g., 50mV) for more than a specified time (e.g., 1 hour), the health HI should not be higher than the set upper limit (e.g., 0.85). R6: When the idle high load ratio is consistently greater than the set threshold (e.g., 30%), the health HI should not be higher than the set upper limit (e.g., 0.8).

[0044] 2. Model building and training.

[0045] 2.1 Network Structure: Construct a lightweight dual-channel hybrid health assessment network. The hybrid health assessment network structure corresponds to... Figure 2 The parameters for each layer are shown in Table 2.

[0046] Table 2 Lightweight Dual-Channel Hybrid Evaluation Network Structure Parameters 2.2 Semi-supervised training: This embodiment simulates a scenario where field labels are scarce. A small number of fault samples and normal / early warning samples are manually labeled from the full dataset to form a labeled dataset; the remaining large amount of data is used as an unlabeled sample set.

[0047] The model uses a ternary composite loss function, and the weights are determined through cross-validation. The model is trained on a cloud server, employing piecewise learning rate decay and early stopping mechanisms to prevent overfitting.

[0048] Experimental comparisons show that: Under normal labeled sample conditions, the dual-channel model of this invention, which integrates time series data and expert knowledge, has lower regression error and higher classification accuracy. Its overall performance is better than that of a pure data-driven model that does not introduce expert knowledge and traditional rule post-processing methods. In scenarios with small sample sizes, this invention relies on expert knowledge constraints, resulting in significantly stronger generalization ability. It can effectively suppress model overfitting and is more suitable for the current application situation where field-labeled samples for power distribution terminals are scarce.

[0049] 3. Lightweight model and embedded deployment.

[0050] Following step 2.4, multi-stage lightweight compression was performed on the trained model: redundant convolutional channels were pruned, a simplified student model was obtained through knowledge distillation, and then the floating-point parameters were quantized into integer parameters. After lightweight compression, the total number of model parameters was significantly reduced, and the model file size was compressed to approximately 1 / 8 of the original size. The lightweight model was deployed to the FTU embedded platform for testing, and the single-round inference time met the real-time requirements of the field, with inference efficiency significantly improved compared to before optimization.

[0051] Figure 3 This is a diagram of the "end-edge-cloud" collaborative deployment architecture of this invention. The system is divided into three layers: the power distribution terminal layer is responsible for multi-source data acquisition, local feature construction, and lightweight model inference; the edge gateway layer deploys inference and scheduling units, performs health assessment, early warning judgment, and data caching and uploading; the cloud platform layer collects fault samples and expert rules, completes incremental model learning and version management, and securely distributes updated models and rules to the edge and terminals through OTA encryption protocol. This architecture forms a closed-loop evolution mechanism of "on-site data acquisition - cloud optimization - terminal iteration".

[0052] 4. Online verification and closed-loop iteration of the model.

[0053] A typical abnormal terminal was selected as a verification case: the device's initial health level (HI) was high, but it gradually decreased during operation. The system promptly triggered the corresponding level of warning and located the cause of the anomaly (such as exceeding the high load ratio during idle periods). After the maintenance personnel investigated, confirmed, and repaired the problem on-site, the health level gradually recovered.

[0054] After the data from this abnormal case was sent back to the cloud platform, incremental learning was used to fine-tune the model, and the new version was distributed to all FTUs of the same model across the network via OTA encryption protocol, realizing a closed-loop evolution of the model from on-site data collection to cloud optimization to terminal iteration.

[0055] The above embodiments verify that the present invention can achieve accurate, real-time, interpretable and continuously evolving health assessment in a real power distribution terminal environment, and has good engineering practical value.

[0056] The above description is merely the basic principle and preferred embodiment of the present invention. Improvements and substitutions made by those skilled in the art based on the present invention are within the scope of protection of the present invention.

Claims

1. A lightweight method for assessing the health of power distribution terminals that integrates expert knowledge and deep learning, characterized in that: Includes the following steps: S01. Data Acquisition and Expert Feature Construction: The acquired data includes terminal status data, business application performance data, and environmental and lifespan information data. Expert knowledge features are constructed by combining the physical operation mechanism of the power distribution terminal with on-site operation and maintenance experience. These features have clear physical meaning, including thermal operation health features, power system aging features, business function degradation features, and storage medium loss features. S02. Construct a lightweight dual-channel hybrid evaluation network, including a shared feature extraction backbone, an expert feature injection channel, and a dual-task output head. The shared feature extraction backbone extracts features from the input multi-channel time-series signal to obtain an abstract feature vector. The expert feature injection channel dynamically generates attention weights for each expert feature through the SE-Net module, highlighting expert features that have a greater impact on the current health assessment; the dual-task output head concatenates the abstract feature vector with the attention-weighted expert feature vector at the feature fusion layer, and then outputs two parallel output heads to output the health index and health status classification of the power distribution terminal respectively. S03. Perform semi-supervised training on the dual-channel hybrid evaluation network, which integrates expert rules. The loss function includes supervision loss, consistency regularization loss, and expert rule loss. The supervision loss is calculated based on labeled samples and is obtained by weighted summation of health regression loss and state classification loss. The consistency regularization loss is applied to unlabeled samples. Two independent random perturbations are applied to the same unlabeled sample, and then the consistency regularization loss is calculated. The expert rule loss transforms the operation and maintenance expert rules into differentiable constraint terms and is calculated based on the indicator function and health status classification. The dual-channel hybrid evaluation network is jointly input with labeled and unlabeled data to complete end-to-end optimization. After each training round, the loss value of the model on the validation set is calculated. When the loss on the validation set no longer decreases for several consecutive rounds, the model is considered to have converged and training is stopped. S04. Lightweight post-processing of the model: Multi-level compression is performed on the trained dual-channel hybrid evaluation network to achieve lightweight post-processing of the model. S05. Deploy the trained and lightweight dual-channel hybrid evaluation network to the power distribution terminal or edge gateway. The terminal continuously collects raw data, constructs hybrid feature vectors, and inputs them into the dual-channel hybrid evaluation network. After calculation, it outputs the health index and health status classification.

2. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: Thermal operating health characteristics are measured by the temperature-load synergy factor C. thermal Characterization, C thermal The calculation formula is: ， in Let σ be the covariance between the CPU utilization rate of the power distribution terminal and the internal temperature of the power distribution terminal, and let σ be the standard deviation of the CPU utilization rate of the power distribution terminal. The standard deviation of the internal temperature of the power distribution terminal; Power system aging characteristics are assessed through battery health factors. H bat Characterization, H bat The calculation formula is: ， in V bat This represents the current measured voltage of the battery in the power distribution terminal. V nom The rated voltage of the battery in the power distribution terminal. t op The cumulative operating time of the batteries in the power distribution terminal. t life The expected lifespan of the battery in the power distribution terminal; Service function attenuation characteristics are measured by remote control function attenuation rate. k ctrl Characterization, k ctrl The calculation formula is: ， Where n is the number of sampling points within the sliding window. For sampling sequence number, The remote control failure rate for the i-th sampling point; Storage media wear characteristics are identified by the Flash abnormal wear index I. flash Characterization, I flash The calculation formula is: , Among them W flash This represents the number of programming / erase cycles that have occurred in the Flash memory of the power distribution terminal. This refers to the maximum number of cycles specified for this type of Flash chip.

3. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: The formula for calculating the loss function of the dual-channel hybrid evaluation network is as follows: ， in For the total loss function, To monitor losses, For consistency regularization loss, For the loss of experts, , , These are the corresponding weighting coefficients; Consistency regularization loss L con The calculation process is as follows: Two independent random perturbations are applied to the same unlabeled sample, and the perturbed samples are then input into a dual-channel hybrid evaluation network to obtain two health prediction values. and The difference between the two is measured by the mean squared error, and the calculation formula is as follows: ， in The number of unlabeled samples used in the calculation. and These are the health prediction values ​​output by the dual-channel hybrid evaluation network after the i-th unlabeled sample has undergone two sets of independent random perturbations; Expert Rule Loss L knowledge The calculation formula is: ， Where N is the total number of expert features involved in the calculation, I(Premise) i ) is an indicator function. This represents the predicted health score. This indicates the maximum health level.

4. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 3, characterized in that: The indicator function is used to determine whether the expert feature meets the rule preconditions. When the expert feature meets the rule preconditions, the indicator function is 1, and when the expert feature does not meet the rule preconditions, the indicator function is 0. These include battery aging rules, temperature-load coordination rules, idle load rules, and storage wear rules.

5. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: Step S04 follows a fixed order of first structured sparsity, then functional compression, and finally numerical compression to perform multi-level compression. Structured sparsity is channel pruning, which removes redundant channels based on the contribution of feature channels, thereby reducing the number of parameters and computational load. Functional compression is knowledge distillation, which uses the pruned model as the teacher model to train a more compact student model. Numerical compression is fixed-point quantization, which converts the 32-bit floating-point parameters of the student model into 8-bit integer parameters.

6. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: The shared feature extraction backbone consists of four one-dimensional depthwise separable convolutional layers, two max pooling layers, and one Flatten layer.

7. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: When the health index is lower than the preset threshold or the health status is classified as abnormal or severe, a multi-level warning is triggered, and key causative factors are output simultaneously. The key causative factors are obtained in the following way: calculate the partial derivative of each input feature of the evaluation network with respect to the output health index, use the absolute value of the partial derivative as the contribution of the feature to the health index, and output the top 3 features with the highest contribution as key causative factors.

8. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: Terminal status data includes the CPU utilization rate of the power distribution terminal. U cpu Memory usage M Internal temperature T int Battery voltage V bat Wear and tear of Flash memory cells W flash Business application performance data includes the telemetry data transmission pass rate R. telem Remote control execution success rate R ctrl Environmental and lifespan information data includes ambient temperature. T env ,humidity H env Cumulative operational time t op Historical fault records.

9. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: The health index is a continuous value, ranging from [0,1], where 1 represents brand new health and 0 represents complete failure. The classification criteria for health status are as follows: HI≥0.9, healthy status; 0.8≤HI<0.9, attentive status; 0.6≤HI<0.8, abnormal status; HI<0.6, severe status. HI is the health index.

10. The lightweight power distribution terminal health assessment method integrating expert knowledge and deep learning according to claim 1, characterized in that: This method adopts a three-layer deployment architecture of terminal-edge-cloud, with the power distribution terminal layer responsible for multi-source data acquisition, local feature construction, and lightweight model inference. The edge gateway layer deploys inference and scheduling units to perform health assessments, early warning judgments, and data caching and uploading; the cloud platform layer gathers fault samples and expert rules, completes incremental model learning and version management, and securely distributes the updated model and rules to the edge and terminals through OTA encryption protocol.

Images