Unknown load identification and incremental learning method based on feature space multi-objective optimization

Abstract

This invention discloses an unknown load identification and incremental learning method based on feature space multi-objective optimization. Step S1: Acquire data and convert it into V-I trajectory grayscale images to construct a sample library after normalization. Step S2: Construct a feature extraction backbone network to obtain feature vectors. Step S3: Construct a dynamic multi-prototype classifier. Step S4: Optimize the network using a multi-objective joint loss function. Step S5: During the online monitoring phase, calculate the distance between the test sample and each known class prototype, and set a rejection threshold to effectively intercept unknown loads. Step S106: When the class incremental learning mechanism is triggered, use the cluster center of the new load features as the new class prototype, and perform incremental optimization and closed-loop evolution without data replay using an asymmetric learning rate and multi-objective joint loss. This invention effectively solves the problem of rejection of unknown devices and the catastrophic forgetting problem during incremental updates in traditional models, significantly reducing computational and storage overhead.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of intelligent electrical measurement, non-intrusive load monitoring (NILM), and artificial intelligence, and particularly to a method for identifying unknown loads and incremental learning based on multi-objective optimization of feature space. This method primarily addresses the problem of accurate identification of electrical loads and continuous model evolution in open-set environments. By mapping samples to a feature metric space with strict objective constraints, and combining this with an incremental learning strategy without data replay, it achieves efficient identification of known and unknown electrical loads and dynamic expansion of new load categories. It is applicable to scenarios such as edge computing for smart meters, refined energy management in homes, and load situation awareness in new power systems. Background Technology

[0002] With the rapid development of smart grids and new power systems, demand-side response and refined electricity management are playing an increasingly crucial role in improving grid operating efficiency and promoting energy conservation and emission reduction. Non-Intrusive Load Monitoring (NILM) technology, with its advantages of low cost and easy deployment, can identify the operating status and energy consumption information of various electrical devices simply by collecting total voltage and current signals at the main power input, becoming a key technological link connecting the power grid and end users. In recent years, deep learning technology has made significant progress in the field of NILM, greatly improving the accuracy of identifying known devices. However, existing deep learning load identification methods still face extremely severe technical bottlenecks and application limitations when dealing with real and highly dynamic power consumption environments.

[0003] First, most traditional deep learning load identification models are based on the "closed set" assumption, meaning they assume all electrical device categories encountered during the testing phase are already included in the training set. In real-world household or industrial power scenarios, as user equipment is added or upgraded, "unknown loads" never seen in the training library inevitably connect to the power grid bus. Existing models typically use fully connected layers and the Softmax activation function for classification, which forces any input sample to be classified into one of the known categories. When faced with unknown new devices, traditional closed-set models often give high-confidence false predictions, failing to effectively intercept them, leading to a sharp increase in the false positive rate. Therefore, how to accurately identify and separate unknown loads in open environments is the primary challenge for NILM technology to become practical.

[0004] Secondly, existing methods for characterizing and measuring non-intrusive load signals often lack effective modeling of complex intra-class polymorphism and strict constraints on decision boundaries. Many household appliances have multiple complex operating modes or multi-level power states, resulting in highly non-convex and multimodal distributions in the feature space. Traditional single-class center methods struggle to fully cover this multimodal distribution, easily leading to intra-class feature divergence. Simultaneously, most existing loss functions lack geometric constraints on intra-class compactness and inter-class repulsion in the feature space, causing blurred boundaries and even overlap between different load categories. When the feature space is crowded by known categories, it cannot reserve clear, independent space for potentially unknown loads, making it easy for unknown samples to fall within the decision boundaries of known classes, thus significantly reducing the sensitivity of open-set identification.

[0005] Finally, after the system successfully detects unknown loads and subsequently obtains labels, how to enable the model to learn this new knowledge becomes another key challenge. The traditional approach is to collect all historical data and retrain the entire network along with the new data. However, this approach incurs extremely high computational costs, making it completely unsuitable for edge devices such as smart meters, which are constrained by both computing power and storage space. To reduce overhead, existing incremental learning methods have been introduced into the NILM field. However, conventional fine-tuning strategies can lead to severe "catastrophic forgetting" of the model, meaning that while learning new load features, the ability to identify old loads deteriorates significantly or is completely lost. Some advanced incremental learning algorithms attempt to mitigate the forgetting problem through data replay strategies—that is, retaining some old category samples for joint training. However, in actual electricity consumption data collection, long-term storage of users' high-frequency raw electricity consumption data not only consumes a large amount of solid-state storage resources on terminal devices but also violates the red line of user data privacy and security, making the practical implementation of incremental learning methods based on data replay in the NILM field extremely difficult.

[0006] In summary, existing non-intrusive load monitoring technologies urgently need a novel identification method that can overcome the limitations of the closed-set assumption. This method needs to construct a feature metric space with clear topological boundaries and high discriminative power, without relying on historical data replay or infringing on user privacy. It should accurately solve the challenge of characterizing complex polymorphic loads, accurately intercept unknown loads in open environments, and achieve adaptive incremental evolution of new load categories with extremely low computational and storage costs. This has significant engineering value and practical implications for promoting the large-scale application of next-generation smart electricity sensing terminals. Summary of the Invention

[0007] The purpose of this invention is to propose an unknown load identification and incremental learning method based on multi-objective optimization of feature space, to solve the problem of traditional models rejecting unknown devices and the catastrophic forgetting problem during incremental updates. Furthermore, this invention aims to address a series of key technical challenges faced by existing non-intrusive load monitoring (NILM) technologies in practical engineering applications. First, it addresses the poor robustness of existing load identification models in open-set environments. Traditional identification models based on the closed-set assumption typically use a Softmax output layer, which essentially compares probabilities between known categories and lacks the ability to intercept unknown categories. When an unknown load, never before seen, is connected to the power grid bus, the model will force it to be mapped to a known category and give a false high-confidence prediction. This invention aims to accurately intercept unknown loads by constructing a constrained feature metric space, solving the problem of misjudgment in open-set environments. Second, it addresses the problem of complex distribution of electrical load characteristics and the difficulty in characterizing multimodal features. Many electrical devices have multiple sub-modes during steady-state operation, and their features exhibit a non-convex, multi-peak distribution in space. Traditional single-class center identification algorithms struggle to accurately characterize the decision boundaries of such complex electrical devices. This invention constructs a dynamic multi-prototype classifier, allowing each category to have multiple learnable local centers, aiming to achieve accurate coverage of complex load manifold structures. Finally, it addresses the contradiction between catastrophic forgetting and data replay in the incremental learning process. Existing incremental learning algorithms often rely on replaying historical data to maintain the recognition accuracy of old categories, but this faces the dual risks of limited storage resources and leakage of user electricity privacy at the smart meter edge. This invention designs a multi-objective joint loss function, optimizing the feature space and shaping feature topology without storing any historical raw data, aiming to solve the feature drift and forgetting problems during load category expansion and achieve closed-loop evolution of model recognition capabilities.

[0008] An unknown load identification and incremental learning method based on feature space multi-objective optimization includes the following steps: Step S1: Extract data and construct a sample database; In a non-intrusive load monitoring system, the single-phase AC voltage and current sequences (sampling frequency 6.4kHz, single-phase AC) of ordinary household electrical equipment (equipment refers to ordinary household electrical equipment used in the non-intrusive load monitoring system) under steady-state operating conditions are extracted. The voltage and current sequences within each 50Hz power frequency cycle are normalized and mapped to a VI trajectory grayscale image to construct a sample database. The cycle refers to the next cycle of the 50Hz power grid. The voltage and current signals here are collected at the main power inlet, but the collected signals are voltage and current signal sequences of each device running independently. Then, the small sequences of each power frequency cycle in the voltage and current sequence are processed into a grayscale image of the vi trajectory. In step S1, a measurement terminal (such as a high-precision smart meter) installed at the power inlet is used to synchronously collect the high-frequency voltage sequence of the target device during its steady-state operating cycle in real time. With current sequence To capture the subtle nonlinear physical characteristics of different electrical appliances, the sampling frequency was set to be no less than 6.4 kHz; Step S2: Construct a feature extraction backbone network to obtain feature vectors; A lightweight convolutional neural network containing convolutional blocks, batch normalization layers, activation functions, and pooling layers is constructed as the feature extraction backbone network. A grayscale image of the VI trajectory is input into this network, and the extracted high-dimensional spatial features (here, features are high-dimensional spatial features output by the deep learning network from the image; these are high-dimensional abstractions without specific physical meaning) are then processed. Standardize the constraints, map them onto the unit hypersphere, and output the feature vector; Step S3: Construct a multi-prototype classifier; Construct a multi-prototype classifier, assigning a specific number of learnable prototype vectors to each known category in the sample database; define the distance between a sample and any category as the minimum of the squared Euclidean distance between the sample's feature vector and all prototype vectors of that category; introduce a temperature coefficient to scale this distance and transform it into a log probability distribution to establish the classification mapping relationship; Step S4: Perform basic training on the model; A multi-objective joint loss function consisting of cross-entropy loss (based on distance metric), intra-class / inter-class boundary loss, and prototype loss is constructed to train the model in the basic stage. The model includes a feature extraction backbone network and a multi-prototype classifier built in the previous steps, which are used to perform feature extraction and classification functions. Among them, cross-entropy loss is used to optimize the global probability distribution; intra-class-inter-class boundary loss makes the features of samples of the same class closer to their corresponding prototypes in the feature space to reduce intra-class differences and pushes the prototypes of different classes apart to increase inter-class distance; prototype loss is used to minimize the absolute distance between sample features and the true class. Step S5: Online load monitoring, performing category matching on online test samples, and detecting unknown loads; During the online load monitoring phase, feature vectors of online test samples are extracted, and their distances to all known class prototypes are calculated to obtain the optimal matching known class. Based on the mean and standard deviation of the distances to correctly classified samples in the training set, independent rejection thresholds are set for each known class. If the global minimum Euclidean distance squared is greater than the threshold of all known classes, the sample is determined as an unknown load and temporarily stored in the unknown sample buffer. (This step is the online load monitoring phase of the system; step S4 is the design of the training phase.) Step S6: Incremental learning is performed on the model; When the unknown load samples in the unknown sample buffer acquire labels, the class increment mechanism is triggered to enter the incremental learning stage. The features of the new load samples are extracted and cluster centers are found in the feature space. After standardization, they are used as the initial prototype vector of the new class and added to the classifier to achieve dynamic expansion. The backbone network and prototype are trained and updated using a multi-objective joint loss function, where a learning rate is assigned to the old prototype and the backbone network. A learning rate is assigned to the new class prototype (here referring to the new class prototype, there is no need to add "backbone network", the backbone network is unique, and the prototype is divided into old class prototype and new class prototype). This allows for incremental optimization of the feature space; Much smaller than the learning rate ; After training converges, the adaptive rejection thresholds for all categories are recalculated; this completes one incremental learning phase. The system then returns to the online monitoring phase in step S5 to continue identifying load categories. If an unknown load is detected, it will enter the next incremental learning phase, and so on.

[0009] In step S1, the specific process of data acquisition and VI trajectory grayscale image construction is as follows: high-frequency voltage sequence within the steady-state operating cycle of the target device is acquired through a smart meter. With current sequence The sampling frequency is 6.4kHz; to eliminate the influence of amplitude, the signal is normalized using the Min-Max normalization method. The basic calculation process is as follows: ;(Formula 1) in, It is the normalized output voltage or current data value; Voltage sequence at a certain moment or current sequence The original value; and These correspond to the maximum and minimum values ​​within the period of the original data sequence; subsequently, the normalized voltage-current pairs are mapped to... In the grid matrix, a single-channel VI trajectory grayscale image is generated based on the distribution of sampling points within the grid. This process transforms the original one-dimensional time-series signal into two-dimensional image features with spatial structure. The physical significance of this step lies in converting the one-dimensional time-series signal into two-dimensional image features with spatial topological structure, thereby effectively eliminating the influence of the initial sampling phase and highlighting the nonlinear shape characteristics of the internal impedance properties of electrical equipment.

[0010] In step S2, the specific process of the lightweight convolutional neural network and feature mapping is as follows: Construct a backbone network consisting of cascaded basic convolutional blocks, each basic convolutional block consisting of... The backbone network consists of standard convolutional layers, batch normalization (BN) layers, ReLU activation function layers, and max pooling layers; the backbone network processes the input VI trajectory grayscale image. Feature extraction is performed, and the convolutional layer is calculated as follows: ;(Formula 2) in, These are the weight coefficients of the filters in this convolutional layer (the weight coefficients are dynamically changing values ​​during training, initialized using the Kaiming normal distribution, and automatically updated as the network trains). This represents the feature data input to this layer; It is a convolution operation; The bias coefficients for the convolution operation (initialized to 0, automatically updated during network training); These are the feature data extracted after convolution; It is the ReLU activation function for convolution operations; the feature map output by the convolutional layer undergoes spatial dimensionality compression through an adaptive average pooling layer, compressing it from 16×16×256 dimensions to 1×1×256 dimensions, and is then mapped by a fully connected layer to... dimensional feature space, where The value is set to 128, resulting in the original feature vector. To achieve a stable distance metric in subsequent multi-prototype classifiers, the original feature vector is processed... After standardization, it is forced to be constrained to a unit hypersphere, and the calculation is as follows: ;(Formula 3) In the formula, The final standardized VI trajectory feature vector is generated. express Norm (this standardization process effectively transforms the classification problem into a metric problem on a hypersphere).

[0011] In step S3, the specific calculation process of the construction of the dynamic multi-prototype classifier and the distance metric logic is as follows: establish a dynamic multi-prototype classification architecture based on metric learning, and for each known category in the basic database (including common electrical load categories such as air conditioners, microwave ovens, and washing machines; a specific load category table is provided in Example 1 below), Pre-set a set of learnable prototype vectors ,in This represents the number of prototypes assigned to each category. These prototype vectors act as class centers for that category in the hyperspherical feature space, capturing the feature diversity of the same household appliance under different modes, and are iteratively updated synchronously during training. The classification metric is the minimum squared Euclidean distance between the sample features and the prototype set of each category, calculated as follows: ;(Formula 4) In the formula, It represents the minimum Euclidean distance squared between a sample feature and the prototype set of a certain category; This represents the set of prototype vectors assigned to this category; This represents the square of the L2 norm (i.e., Euclidean distance) of the solution vector. This multi-prototype design allows the model to form a non-convex decision region in the feature space, greatly enhancing the accuracy of representing complex loads; to achieve end-to-end discrimination, a temperature coefficient is introduced. The minimum Euclidean distance is non-linearly scaled and converted into class log odds. : ;(Formula 5) in: The preset temperature scalar coefficient (typically 5.0) controls the clarity of the decision boundary, ensuring that the gradient can be effectively backpropagated to the backbone network and each prototype vector during training. The dynamic multi-prototype classifier effectively constructs a refined decision boundary by learning multiple local centroids rather than a single class centroid.

[0012] In step S4, the specific process and calculation logic of the multi-objective joint optimization based on feature topological space constraints are as follows: In the basic training and incremental update phases of the model, the traditional fully connected layer and linear classification structure are abandoned. Instead, a multi-objective loss function is used to jointly optimize the network parameters. The aim is to construct a feature metric space that is highly compact within classes, forced to be far apart between classes, and has topological smoothness. The multi-objective joint optimization shapes the feature space through three losses, ensuring that the feature space still has anti-drift capability and nonlinear discriminative ability even without historical data replay. First, a distance-based cross-entropy loss is calculated to optimize the global probability distribution in the feature space. This loss function replaces the traditional inner product operation with the negative Euclidean distance between sample features and each class prototype as the logarithm of the classification, fundamentally transforming the model into a metric-based discriminative model. For any input sample, the model calculates the distance distribution between its feature vector and the set of all class prototypes in the current stage, and obtains the predicted probability through Softmax normalization. First calculate the sample feature vector With the Each category prototype Squared Euclidean distance between : ;(Formula 6) Negate and exponentialize the distance to calculate the true class label of the sample. Predicted probability : ;(Formula 7) Constructing a cross-entropy loss based on distance metric : ;(Formula 8) in: This represents the total number of training samples in the current batch. The samples were extracted and standardized by the backbone network. 3D feature vector; True label for the sample The corresponding prototype vector; The total number of currently known categories; This represents the squared operation of the Euclidean distance; this loss term establishes the basic probabilistic classification boundary of the feature space. Secondly, the intra-class and inter-class boundary loss is calculated to explicitly constrain the geometric topological relationship of the feature space. In order to establish a hard exclusion mechanism in the feature space and prevent feature overlap between different classes, the loss function explicitly optimizes the geometric relationship between the sample and its class prototype and the nearest error class prototype by introducing adaptive boundary constraints. First, calculate the intra-class distance from the sample features to their true class prototypes. : ;(Formula 9) And calculate the inter-class distance from the sample features to the nearest non-class prototype (i.e., the most easily confused class). : ;(Formula 10) The difference between the two values ​​is then nonlinearly smoothed using the Sigmoid function to calculate the intra-class vs. inter-class boundary loss. The formula is as follows: ;(Formula 11) in: For all non-class prototypes, the distance from the current feature vector The most recent prototype vector; The scaling hyperparameter (typically 0.1) controls the gradient response sensitivity; the difference term... Right now It describes the relative position of the sample at the decision boundary; by minimizing this loss, the model is forced to expand the geometric margin between the true class and the nearest neighboring perturbation class, thereby forming a non-intrusive cluster manifold on the hypersphere, which effectively enhances the model's discrimination accuracy for samples with abnormal position distribution and reserves blank feature storage space for subsequent incremental learning.

[0013] To further lock onto the feature distribution of known categories and prevent cluster divergence due to a lack of old data during the incremental update phase, prototype loss is introduced. As a global geometric anchor point, by directly constraining the absolute spatial displacement of samples and their corresponding class prototypes, the nonlinear robustness of the decision boundary is enhanced, ensuring that the feature points of each load category are closely distributed around their corresponding prototype vectors, and their prototype loss... The calculation is as follows: ;(Formula 12) Finally, by weighting and fusing the above losses, a global total loss function for end-to-end training is constructed. The calculation is as follows: ;(Formula 13) in: This represents the total global loss value. and These are the dynamic weighting coefficients that adjust the inter-class repulsion strength and the intra-class compactness constraint strength, respectively. The value is 0.1. The value is 0.001; The multi-objective joint optimization scheme drives the classification logic through cross-entropy loss, shapes the topological spacing through boundary loss, and locks the spatial location through prototype loss. The three work together to enable the model to learn new load characteristics without storing any historical electricity consumption data. At the same time, the model can rely on the originally constructed high-rigidity topological boundary to resist the collapse of the feature space, thus fundamentally solving the catastrophic forgetting problem of non-intrusive load identification in incremental scenarios.

[0014] In step S5, the specific process and discrimination logic of the unknown load detection based on adaptive dynamic threshold are as follows: During the online monitoring phase, for any input load sample to be identified... First, the corresponding standardized feature vector is extracted through the lightweight feature extraction network. Subsequently, the feature vector is compared with all known categories. The optimal matching category for a sample is determined by finding the global minimum distance between it and its corresponding prototype set. and its corresponding metric span The calculation is as follows: ;(Formula 14) in, The known load category closest to this test sample; To accurately intercept unseen payload classes in an open-set environment, for each known class... Establish independent distance thresholds That is, in the category After the initial training and convergence, the feature distribution of the current training batch for this category is used to determine: the Euclidean distance between samples of this category and their corresponding prototypes is calculated, and the mean is determined. with standard deviation The calculation is as follows: ;(Formula 15) in, The preset rejection sensitivity coefficient (typically set to 3.0) is used to adjust the model's rejection strength for unknown samples and its tolerance for fluctuations in known samples; ultimately, the global minimum distance of the online test samples is used to... Its optimal matching category exclusive threshold Real-time comparison is performed, and the following discrimination logic is executed to obtain the final output result. : ;(Formula 16); This threshold is not a static value, but is dynamically calculated using Formula 15 (i.e., the mean and standard deviation of feature distances based on the category training data); If the judgment result is If the load is unknown, the current load is an unknown new device. It is rejected from being incorrectly classified as a known load, and its feature data and the original trajectory grayscale image are automatically stored in the unknown sample buffer. When the number of samples reaches a certain scale and labels are obtained (generally, 50 or more samples of the device are collected), it will automatically enter the incremental learning process to realize the closed-loop evolution of the model's recognition capability.

[0015] In step S6, the specific calculation process and strategy of incremental learning and closed-loop evolution are as follows: When the unknown sample buffer triggers incremental learning, while keeping the front-end lightweight CNN feature extractor network structure completely fixed, the model parameters are updated hierarchically using an asymmetric learning rate to suppress catastrophic forgetting under the condition of no old data replay; the feature extractor parameters are set. and old class prototype The learning rate is Set a new class prototype The learning rate is And satisfy By minimizing the incremental total loss function Update the parameters as follows: This step involves updating the feature extractor parameters. inverted triangle This represents the gradient operator, i.e., the differentiation operation; This step involves updating the old class prototype. The arrow indicates assignment; This step is to update the prototype of the new class. ;(Formula 17) in, This is the gradient operator (differentiation operation). This indicates assignment. Using step S4 A multi-objective joint loss strategy is employed; after training converges, the rejection threshold for newly added class samples is calculated. : ;(Formula 18) in, The mean and standard deviation of the feature distribution of the new category samples. Standard deviation, The preset rejection sensitivity coefficient (value 3.0) is used to ultimately achieve continuous evolution and closed-loop operation of the model in an open set environment.

[0016] Beneficial effects: The unknown load identification and incremental learning method based on feature space multi-objective optimization of the present invention, This invention discloses an unknown load identification and incremental learning method based on multi-objective optimization of feature space. Step S1: Extract steady-state high-frequency voltage and current sequences from the equipment, normalize them, and convert them into VI trajectory grayscale images to construct a sample library. Step S2: Construct a lightweight convolutional neural network to extract features and map them onto a unit hypersphere to output feature vectors. Step S3: Construct a dynamic multi-prototype classifier, assign learnable prototype vectors to known classes, and define distance metric classification criteria. Step S4: Optimize the network using a multi-objective joint loss function to shape the feature topology metric space. Step S5: During the online monitoring phase, calculate the distance between the test sample and each known class prototype, and set a rejection threshold to effectively intercept unknown loads. Step S106: When the class incremental learning mechanism is triggered, use the cluster center of the new load features as the new class prototype, and use an asymmetric learning rate and multi-objective joint loss to perform incremental optimization and closed-loop evolution without data replay. This invention effectively solves the problems of traditional models refusing to identify unknown devices and catastrophic forgetting during incremental updates, significantly reducing computational and storage overhead and possessing significant value for edge computing applications. Specifically, compared with existing technologies, this invention overcomes the closed-set limitations of traditional load identification models and solves the technical problem of existing deep learning methods easily generating high-confidence misjudgments when facing unknown new devices. By innovatively transforming non-intrusive load identification from probabilistic classification into a metric problem subject to strict geometric and topological constraints, and combining it with a threshold interception mechanism, this invention can accurately identify and isolate unseen electrical devices in dynamic open power environments, greatly improving the reliability and practicality of non-intrusive load monitoring systems. Simultaneously, this invention effectively overcomes the technical bottleneck of inaccurate representation of polymorphic complex load features. The innovatively introduced dynamic multi-prototype classifier endows a single category with multiple local cluster centers, making its decision boundary more flexible and enhancing its ability to identify complex power loads. Furthermore, this invention eliminates the dependence of traditional incremental learning on historical data replay, resolving the core contradiction between model catastrophic forgetting and user privacy protection. By employing multi-objective joint optimization and asymmetric learning strategies, this invention resists model parameter collapse by relying on the topology and boundaries of the feature space without the constraint of storing historical data. This reduces the computational and storage overhead of model updates, enabling this lightweight framework to be perfectly adapted to resource-constrained edge terminals such as smart meters, possessing extremely high engineering application value and broad market prospects. Attached Figure Description

[0017] Figure 1 This is a basic flowchart of the present invention.

[0018] Figure 2 This is a diagram of the model structure based on multi-objective optimization in feature space.

[0019] Figure 3 A flowchart for the overall process of identifying and incrementally updating unknown loads.

[0020] Figure 4 The confusion matrix diagram is used to represent the model in the basic stage.

[0021] Figure 5 The confusion matrix diagram for the model in increment 1.

[0022] Figure 6 The confusion matrix diagram for the model in incremental stage 2.

[0023] Figure 7 The confusion matrix diagram for the model in the incremental 3 stage. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Example 1

[0025] This invention addresses three core technical challenges faced by non-intrusive load monitoring (NILM) systems in real-world open power consumption environments: First, traditional closed-set models cannot effectively reject unknown new devices, forcibly classifying them into known categories and generating high-confidence erroneous predictions. Second, the multi-mode operation of household appliances makes it difficult for a single class center to fully represent such complex features. Third, conventional incremental learning methods rely on the playback of historical data to mitigate catastrophic forgetting, which faces dual constraints of limited storage resources and user privacy protection on edge terminals such as smart meters. This invention innovatively transforms the load identification problem from a traditional probabilistic classification problem into a feature metric learning problem subject to strict geometric topological constraints. It introduces a dynamic multi-prototype classification architecture and multi-objective joint optimization loss, while combining unknown load identification and a data-free incremental strategy. This achieves accurate identification of known loads, effective interception of unknown loads, and closed-loop evolution of new load categories in open-set environments, effectively solving the core problems in existing technologies.

[0026] An unknown load identification and incremental learning method based on feature space multi-objective optimization, such as... Figure 1 As shown, firstly, high-frequency voltage and current sequences during the steady-state operating cycle of the smart meter are synchronously collected. The original signals are then processed using Min-Max normalization and mapped to a grid matrix to generate VI trajectory grayscale images, constructing a labeled sample database. Subsequently, a lightweight backbone network composed of multiple cascaded convolutional blocks is constructed to extract features from the samples. L2 normalization is used to constrain the feature vectors to a unit hypersphere, achieving stable metric space mapping. Based on this, a dynamic multi-prototype classification architecture is established, assigning multiple learnable local prototype vectors to each known load category, using the minimum Euclidean distance squared between sample features and the prototype sets of each category. As a classification criterion, it is transformed into a classification log probability that can be used for end-to-end training through a temperature coefficient. Then, a multi-objective joint loss function consisting of cross-entropy loss (distance metric), intra-class / inter-class boundary loss, and prototype loss is used to globally optimize the network, simultaneously shaping a feature distribution that is highly compact within classes and forcibly separated between classes in the feature space. After the system enters the online monitoring stage, an independent rejection threshold is set for each known class, and unknown loads are identified in real time and temporarily stored in a buffer. After the buffer data is labeled, incremental optimization without data replay is performed to achieve dynamic expansion of the classifier and closed-loop evolution of the model's recognition capability.

[0027] The specific implementation steps of the unknown load identification and incremental learning method based on feature space multi-objective optimization are as follows: In step S1, the voltage sequence of the target device during its steady-state operating cycle is acquired by a smart meter installed at the main power inlet at a sampling frequency of 6.4 kHz. With current sequence To eliminate the effects of voltage fluctuations and amplitude differences between devices, the original signal is subjected to Min-Max normalization. The basic calculation process is as follows: ;(Formula 1) in, It is the normalized output voltage or current data value; Voltage sequence at a certain moment or current sequence The original value; and These correspond to the maximum and minimum values ​​within the period of the original data sequence. Subsequently, the normalized voltage-current pairs are mapped to... In the grid matrix, a single-channel VI trajectory grayscale image is generated based on the distribution of sampling points within the grid. This transforms the original one-dimensional time-series signal into two-dimensional image features with spatial structure. The types of acquisition equipment are shown in the table below: ; In step S2, a lightweight backbone network composed of multiple cascaded basic convolutional blocks is constructed to adapt to the limited computing power constraints at the edge of the smart meter. Each basic convolutional block sequentially contains a standard convolutional layer, a batch normalization (BN) layer, a ReLU activation function, and a max pooling layer, with its basic convolution operation being... ;(Formula 2) in, For filter weights, This is to account for the bias. The original feature vector is output from the fully connected layer. Later, it was introduced Standardization constraints: The eigenvectors are normalized. ;(Formula 3) Force all load characteristics to be mapped to one On a hypersphere of dimensionality, the classification problem is effectively transformed into a metric problem on the hypersphere. The specific layer structure and dimensional transformation of the backbone network are shown in Table 2. ; In step S3, a preset is made for each known load category. A learnable prototype vector , These prototype vectors act as class centers for the category in the hyperspherical feature space, capturing the feature diversity of the same appliance under different modes, and are iteratively updated synchronously during training. The minimum Euclidean distance squared between the sample features and the prototype sets of each category is used as the classification metric, calculated as follows: ;(Formula 4) This multi-prototype design allows the model to form a non-convex decision region in the feature space, greatly enhancing the accuracy of representing complex loads. To achieve end-to-end discrimination, a temperature coefficient is introduced. The minimum Euclidean distance is non-linearly scaled and converted into class log odds. : ;(Formula 5) By temperature coefficient Adjusting the gradient can control the clarity of the decision boundary, ensuring that the gradient can be effectively backpropagated to the backbone network and each prototype vector during training.

[0028] In step S4, a multi-objective joint loss function is used to optimize the network, aiming to create a feature metric space on the hypersphere that is highly compact within classes, forces distancing between classes, and possesses topological smoothness. One approach is cross-entropy loss based on distance metrics. By replacing the traditional inner product operation with the negative Euclidean distance as the classification log odds, the feature space forms the basic probability classification boundary, which is calculated as follows: ;(Formula 6) in For sample feature vectors, As a category prototype, This represents the total number of currently known categories. Secondly, intra-class / inter-class boundary loss By introducing the Sigmoid function to measure intra-class distance Inter-class distance The difference is used to constrain the categories, making them more compact and expanding the space between categories, reserving a clear area in the feature space for unknown load categories that may appear in the future. The specific calculation is as follows: ; ; ;(Formula 7) Third, prototype loss By directly minimizing the absolute displacement between sample features and their true class prototypes, spatial drift of known class clusters can be resisted in the absence of old class data during the incremental learning stage, enhancing the nonlinear robustness of the decision boundary for known classes. The specific calculation is as follows: ;(Formula 8) Finally, a global total loss function is constructed through weighted fusion; ;(Formula 9) The three loss terms work together to finely optimize the feature space from three dimensions: global probability distribution, topological spacing constraint, and spatial location anchoring, fundamentally constructing a highly rigid feature topology structure that is conducive to incremental updates.

[0029] In step S5, after the system enters the online monitoring phase, the load samples to be identified are... First, the standardized feature vector is obtained through a lightweight feature extraction network that has been trained. Then calculate The global minimum distance between the prototype set of all known categories. And determine the optimal matching category The calculation is as follows: ;(Formula 10) For each known category Establish independent distance thresholds After training convergence for each category, this threshold is calculated by taking the mean Euclidean distance between the features of the training samples for that category and their corresponding prototypes. with standard deviation Confirmed, the calculation is as follows: ;(Formula 11) in The preset rejection sensitivity coefficient is used to adjust the model's rejection strength for unknown samples and its tolerance for fluctuations in known samples; ultimately, the global minimum distance of the online test samples is used to... Its optimal matching category The system performs a real-time comparison with the specific threshold and executes the following judgment logic: ;(Formula 12) If the judgment result is If the current load is an unknown new device, it will be rejected from being incorrectly classified as a known load, and its feature data and the original trajectory grayscale image will be automatically stored in the unknown sample buffer. When the number of samples reaches a certain scale and labels are obtained, it will automatically enter the incremental learning process to realize the closed-loop evolution of the model's recognition capability.

[0030] In step S106, after the data in the unknown sample buffer is labeled, the class incremental learning mechanism is triggered. The overall process is as follows: Figure 3 As shown, firstly, a lightweight CNN feature extractor is used to extract features from the new class samples in the buffer. K-Means clustering is then performed in the feature space to find K initial cluster centers. After standardization, these centers are appended to the classifier as the initial prototype vectors for the new class, achieving dynamic scaling. Subsequently, an asymmetric learning rate strategy is used for incremental optimization: setting the feature extractor parameters θ and the old class prototypes... The learning rate is Set a new class prototype The learning rate is And satisfy By minimizing the multi-objective joint loss function The parameters are updated hierarchically, allowing the new class prototypes to quickly converge to their feature distribution region under a high learning rate, while the old class prototypes and backbone network parameters remain stable under a low learning rate constraint, avoiding feature space collapse and drift. After training convergence, an adaptive rejection threshold is recalculated based on the feature distribution of the new class samples and applied. ;(Formula 13) Ultimately, this achieves a closed-loop, continuous evolution of the model's recognition capabilities in open-set environments. The complete training hyperparameter configuration for the model is shown in Table 3. ; To objectively and comprehensively evaluate the effectiveness of the method described in this invention in open set load identification and incremental class learning tasks, this embodiment constructs an experimental verification environment based on actual user data collected from electricity meters, and strictly follows the evaluation specifications for open set identification scenarios in the experimental design. The experimental hardware consists of an Intel(R) Core(TM) i9-7920X CPU @ 2.90GHz, an NVIDIA GeForce RTX 2080 Ti graphics card, and 64GB of memory. The program is implemented using Python 3.9 and the PyTorch 2.0 framework. The dataset contains 6.4kHz high-frequency data of 11 categories of common household appliances from multiple brands, divided into training, validation, and test sets in a 7:2:1 ratio. Eight categories of appliances are randomly selected as known categories for the basic training phase, while the remaining three categories appear as unknown categories in the test phase to evaluate the model's rejection capability. In the incremental learning phase, the three unknown categories of appliances are gradually expanded by introducing one category at a time to verify the method's ability to retain old knowledge and quickly learn new categories under conditions without data replay.

[0031] This embodiment comprehensively evaluates the method of the present invention from three dimensions: the accuracy of known category recognition in the basic stage, the ability to intercept and detect unknown loads in open sets, and the overall maintenance performance after multiple rounds of incremental learning. It also compares the method with traditional closed-set CNN networks (direct fine-tuning for incremental learning), the data replay-based iCaRL model, and the knowledge distillation-based LwF model without data replay. The main evaluation indicators include the recognition accuracy in the basic stage, the detection rate of unknown loads, the recognition accuracy in each of the three incremental stages, and the final average recognition accuracy. Specific comparison results are shown in Table 4. ; As can be seen from the experimental results in Table 4, the proposed method achieves optimal performance across all core metrics. In terms of the accuracy of known category recognition in the basic stage, the method achieves an accuracy of 99.32%. Regarding the ability to detect unknown loads in open sets, the method achieves an accuracy of 98.07%, which is attributed to the fact that the multi-target loss method reserves a clear blank area in the feature space for unknown samples, thus accurately distinguishing between known and unknown loads. In terms of multi-round incremental learning, traditional closed-set CNNs suffer from catastrophic forgetting due to direct model fine-tuning; the LwF model without data replay introduces knowledge distillation, but the lack of old data replay still leads to forgetting of old knowledge, reducing accuracy to 63.85%. Our method, without relying on any historical data replay, achieves an average recognition accuracy of 97.71%, surpassing the iCaRL model which relies on old data retention mechanisms. This fully verifies the effectiveness of the feature space constructed in this invention in resisting catastrophic forgetting under conditions without data replay. The above experimental results strongly demonstrate that the unknown load identification and incremental learning method based on feature space multi-objective optimization proposed in this embodiment fully realizes efficient adaptation and closed-loop evolution in open power consumption environments under the premise of terminal device storage pressure and user privacy leakage risks, and has outstanding engineering practical value and broad prospects for promotion and application.

[0032] The foregoing has shown and described the basic principles, main features, and advantages of this embodiment. Those skilled in the art should understand that this embodiment is not limited to the specific embodiments described above. The specific embodiments and descriptions in the specification are merely for further illustrating the principles of this embodiment. Various changes and modifications can be made to this embodiment without departing from the spirit and scope of this embodiment, and all such changes and modifications fall within the scope of this embodiment as claimed. The scope of protection of this embodiment is defined by the claims and their equivalents.

Claims

1. A method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space, characterized in that, Includes the following steps: Step S1: Extract data and construct a sample database; In a non-intrusive load monitoring system, the single-phase AC voltage and current sequences of ordinary household electrical equipment under steady-state operation are extracted; the voltage and current sequences within each 50Hz power frequency cycle are normalized and mapped to VI trajectory grayscale images to construct a sample database. Step S2: Construct a feature extraction backbone network to obtain feature vectors; A lightweight convolutional neural network containing convolutional blocks, batch normalization layers, activation functions, and pooling layers is constructed as the feature extraction backbone network; the grayscale image of the VI trajectory is input into this network, and the extracted high-dimensional spatial features are processed. Standardize the constraints, map them onto the unit hypersphere, and output the feature vector; Step S3: Construct a multi-prototype classifier; Construct a multi-prototype classifier, assigning a specific number of learnable prototype vectors to each known category in the sample database; define the distance between a sample and any category as the minimum of the squared Euclidean distance between the sample's feature vector and all prototype vectors of that category; introduce a temperature coefficient to scale this distance and transform it into a log probability distribution to establish the classification mapping relationship; Step S4: Perform basic training on the model; A multi-objective joint loss function consisting of cross-entropy loss (based on distance metric), intra-class / inter-class boundary loss, and prototype loss is constructed to train the model in the basic stage. The model includes a feature extraction backbone network and a multi-prototype classifier built in the previous steps, which are used to perform feature extraction and classification functions. Step S5: Online load monitoring, performing category matching on online test samples, and detecting unknown loads; During the online load monitoring phase, feature vectors of online test samples are extracted, and their distances to all known category prototypes are calculated to obtain the optimal matching known category. Based on the mean and standard deviation of the distances to correctly classified samples in the training set, an independent rejection threshold is set for each known class; if the squared minimum global Euclidean distance is greater than the threshold of all known classes, the sample is judged as an unknown load and temporarily stored in the unknown sample buffer. Step S6: Incremental learning is performed on the model; When the unknown load samples in the unknown sample buffer acquire labels, the class increment mechanism is triggered to enter the incremental learning stage. The features of the new load samples are extracted and cluster centers are found in the feature space. After standardization, they are used as the initial prototype vector of the new class and added to the classifier to achieve dynamic expansion. The backbone network and prototype are trained and updated using a multi-objective joint loss function, where a learning rate is assigned to the old prototype and the backbone network. Assign a learning rate to the prototype of the new class. This allows for incremental optimization of the feature space; Much smaller than the learning rate ; After training converges, the adaptive rejection thresholds for all categories are recalculated; this completes one incremental learning phase. The system then returns to the online monitoring phase in step S5 to continue identifying load categories. If an unknown load is detected, it will enter the next incremental learning phase, and so on.

2. The method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space according to claim 1, characterized in that, In step S1, the specific process of data acquisition and VI trajectory grayscale image construction is as follows: high-frequency voltage sequence within the steady-state operating cycle of the target device is acquired through a smart meter. With current sequence The sampling frequency is 6.4kHz; to eliminate the influence of amplitude, the signal is normalized using the Min-Max normalization method. The basic calculation process is as follows: ； in, It is the normalized output voltage or current data value; Voltage sequence at a certain moment or current sequence The original value; and These correspond to the maximum and minimum values ​​within the period of the original data sequence; subsequently, the normalized voltage-current pairs are mapped to... In the grid matrix, a single-channel VI trajectory grayscale image is generated based on the distribution of sampling points within the grid. This transforms the original one-dimensional time-series signal into two-dimensional image features with spatial structure.

3. The method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space according to claim 1, characterized in that, In step S2, the specific process of the lightweight convolutional neural network and feature mapping is as follows: Construct a backbone network consisting of cascaded basic convolutional blocks, each basic convolutional block consisting of... The backbone network consists of standard convolutional layers, batch normalization layers, ReLU activation function layers, and max pooling layers; the backbone network processes the input VI trajectory grayscale image. Feature extraction is performed, and the convolutional layer is calculated as follows: ； in, These are the weight coefficients of the filters in this convolutional layer; This represents the feature data input to this layer; It is a convolution operation; The deviation coefficient for the convolution operation; These are the feature data extracted after convolution; It is the ReLU activation function for convolution operations; the feature map output by the convolutional layer undergoes spatial dimensionality compression through an adaptive average pooling layer, compressing it from 16×16×256 dimensions to 1×1×256 dimensions, and is then mapped by a fully connected layer to... 3D feature space, where The value is set to 128, resulting in the original feature vector. To achieve a stable distance metric in subsequent multi-prototype classifiers, the original feature vector is processed... After standardization, it is forced to be constrained to a unit hypersphere, and the calculation is as follows: ； In the formula, The final standardized VI trajectory feature vector is generated. express Norm.

4. The method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space according to claim 1, characterized in that, In step S3, the specific process of constructing the dynamic multi-prototype classifier and calculating the distance metric logic is as follows: Establish a dynamic multi-prototype classification architecture based on metric learning, and calculate the distance metric logic for each known category in the basic database. Pre-set a set of learnable prototype vectors ,in This represents the number of prototypes assigned to each category. These prototype vectors act as class centers for that category in the hyperspherical feature space, capturing the feature diversity of the same household appliance under different modes, and are iteratively updated synchronously during training. The minimum Euclidean distance squared between the sample features and the prototype set of each category is used as the classification metric, and its calculation is as follows: ； In the formula, It represents the minimum Euclidean distance squared between a sample feature and the prototype set of a certain category; This represents the set of prototype vectors assigned to this category; This represents the L2 norm of the solution vector, i.e., the square of the Euclidean distance; this multi-prototype design allows the model to form a non-convex decision region in the feature space, greatly enhancing the accuracy of representing complex loads; to achieve end-to-end discrimination, a temperature coefficient is introduced. The minimum Euclidean distance is non-linearly scaled and converted into class log odds. : ； in: The preset temperature scalar coefficient can control the clarity of the decision boundary, ensuring that the gradient can be effectively backpropagated to the backbone network and each prototype vector during training.

5. The method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space according to claim 4, characterized in that, In step S4, the cross-entropy loss based on the distance metric is first calculated: First calculate the sample feature vector With the Each category prototype Squared Euclidean distance between : ； Negate and exponentialize the distance to calculate the true class label of the sample. Predicted probability : ； Constructing a cross-entropy loss based on distance metric : ； in: This represents the total number of training samples in the current batch. The samples were extracted and standardized by the backbone network. 3D feature vector; True label for the sample The corresponding prototype vector; The total number of currently known categories; This represents the squared operation of the Euclidean distance; this loss term establishes the basic probabilistic classification boundary of the feature space. Secondly, calculate the intra-class and inter-class boundary loss to explicitly constrain the geometric and topological relationships of the feature space; First, calculate the intra-class distance from the sample features to their true class prototypes. : ； And calculate the inter-class distance from the sample features to the nearest non-class prototype. : ； The difference between the two values ​​is then nonlinearly smoothed using the Sigmoid function to calculate the intra-class vs. inter-class boundary loss. The formula is as follows: ； in: For all non-class prototypes, the distance from the current feature vector The most recent prototype vector; The scaling hyperparameter is used to control the gradient response sensitivity; Prototype loss The calculation is as follows: ； Finally, by weighting and fusing the above losses, a global total loss function for end-to-end training is constructed. The calculation is as follows: ； in: This represents the total global loss value. and These are the dynamic weighting coefficients that adjust the inter-class repulsion strength and the intra-class compactness constraint strength, respectively.

6. The method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space according to claim 5, characterized in that, In step S5, the specific process and discrimination logic of the unknown load detection based on adaptive dynamic threshold are as follows: During the online monitoring phase, for any input load sample to be identified... First, a lightweight feature extraction network is used to extract the corresponding standardized feature vectors. Subsequently, the feature vector is compared with all known categories. The optimal matching category for a sample is determined by finding the global minimum distance between it and its corresponding prototype set. and its corresponding metric span The calculation is as follows: ； in, The closest known load category to the test sample; To accurately intercept unseen payload classes in an open-set environment, for each known class... Establish independent distance thresholds That is, in the category After the initial training and convergence, the feature distribution of the current training batch for this category is used to determine: the Euclidean distance between samples of this category and their corresponding prototypes is calculated, and the mean is determined. with standard deviation The calculation is as follows: ； in, The preset rejection sensitivity coefficient is used to adjust the model's rejection strength for unknown samples and its tolerance for fluctuations in known samples; ultimately, the global minimum distance of the online test samples is used. Its optimal matching category exclusive threshold Real-time comparison is performed, and the following discrimination logic is executed to obtain the final output result. : ； If the judgment result is If the load is unknown, the current load is considered an unknown new device. It is rejected from being incorrectly classified as a known load, and its feature data and the original trajectory grayscale image are automatically stored in the unknown sample buffer. When the number of samples reaches a certain scale and labels are obtained, it will automatically enter the incremental learning process to realize the closed-loop evolution of the model's recognition capability.

7. The method for identifying and incrementally learning unknown loads based on multi-objective optimization of feature space according to any one of claims 1-6, characterized in that, In step S6, the specific calculation process and strategy of incremental learning and closed-loop evolution are as follows: When the unknown sample buffer triggers incremental learning, under the premise of keeping the front-end lightweight CNN feature extractor network structure completely fixed, the model parameters are updated hierarchically using an asymmetric learning rate to suppress catastrophic forgetting under the condition of no old data replay. Setting feature extractor parameters and old class prototype The learning rate is Set a new class prototype The learning rate is And satisfy By minimizing the incremental total loss function Update the parameters as follows: ； ； ； in, For gradient operators, This indicates assignment. Using step S4 A multi-objective joint loss strategy is employed; after training converges, the rejection threshold for newly added class samples is calculated. : ； in, The mean and standard deviation of the feature distribution of the new category samples. Standard deviation, The preset rejection sensitivity coefficient is used to ultimately achieve continuous evolution and closed-loop operation of the model in an open set environment.

Images