Multi-view electrical load type identification method based on load joint embedding prediction architecture

Abstract

The application discloses a multi-view electrical load type identification method based on a load joint embedding prediction architecture, comprising: obtaining multi-source electrical data of historical loads, constructing multi-view features, performing electrical segment division and occlusion construction on the features, and generating unoccluded context segments, occluded target segments and occlusion position markers; performing self-supervised pre-training based on a load joint embedding prediction architecture, and taking a completed context encoder as a load representation extractor; extracting multi-view electrical representations of loads to be identified, combining a view time window embedding and a mask attention fusion mechanism to fuse each view representation, inputting global representation into a classifier to output an identification result, and discriminating unknown loads according to classification confidence and representation distance. The application solves the technical problems that the prior art is difficult to stably and accurately identify the load type when the labeled samples are limited, the data collection is incomplete and unknown loads are connected.

Description

Technical Field

[0001] This invention relates to the fields of power load identification and artificial intelligence technology, and in particular to a multi-perspective electrical load type identification method based on a load joint embedding prediction architecture. Background Technology

[0002] With the development of smart distribution networks, user-side energy efficiency management, shared power equipment, park energy management, and non-intrusive load monitoring technologies, accurately identifying the types of user-connected loads has become a crucial foundation for electricity safety monitoring, abnormal electricity consumption identification, refined load management, and energy consumption optimization control. By identifying load types, the operating status and electricity consumption attributes of user-side connected equipment can be determined, providing data support for power equipment management, abnormal load detection, sub-item energy consumption statistics, and demand response control.

[0003] The existing electrical load type identification methods mainly have the following technical bottlenecks: (1) Manual feature methods are limited by the quality of feature design and are difficult to adapt to complex and ever-changing actual working conditions; (2) Supervised deep learning methods rely on a large number of labeled samples and have too high requirements for data integrity and data distribution consistency; (3) Reconstruction-based self-supervised methods take the reconstruction of the original signal as the training target, which is prone to overfitting noise and disturbance, and do not fully utilize the fusion of electrical information from multiple perspectives; (4) Existing methods generally lack the ability to effectively distinguish unknown loads and are difficult to adapt to the actual scenarios where new loads are constantly emerging.

[0004] Therefore, there is an urgent need for a new electrical load identification method that can fully utilize the stable correlation between electrical features from multiple perspectives under the condition of a small number of labeled samples, learn discriminative representations related to the essential attributes of load types, and at the same time have the ability to adapt to data missing, incomplete perspectives and unknown loads, thereby improving the accuracy, robustness and engineering applicability of load identification in real complex power consumption scenarios. Summary of the Invention

[0005] This invention provides a multi-perspective electrical load type identification method based on a load joint embedded prediction architecture. The method solves the technical problem that existing technologies struggle to stably and accurately identify load types when there are limited labeled samples, incomplete or missing data collection, or unknown load access.

[0006] A multi-perspective electrical load type identification method based on a load joint embedding prediction architecture includes: S1: Acquire multi-source electrical data of historical loads and construct multi-view electrical features that characterize the multi-dimensional operating characteristics of the loads; S2: Perform electrical segmentation and occlusion construction on electrical features to form an unoccluded context segment set and an occluded target segment set for self-supervised training, and generate occlusion position markers; S3: Based on the load joint embedding prediction architecture, self-supervised pre-training is performed using the set of unoccluded context segments, the set of occluded target segments, and the occlusion location markers. The latent space predictor predicts the target potential representation of the occluded electrical segment and calculates the prediction loss. The pre-trained context encoder obtained after training is used as the load representation extractor. S4: Extract multi-view electrical representations of the load to be identified based on the load representation extractor, and fuse the electrical representations of each view by combining the view time window embedding and mask attention fusion mechanism to obtain the global representation of the load to be identified. S5: Input the global representation of the load to be identified into the classifier, output the load type identification result, and judge the unknown load based on the classification confidence and representation distance.

[0007] Furthermore, in S1, the viewpoint type of the electrical feature includes time-domain waveform viewpoint features. Power perspective characteristics Frequency domain perspective features Voltage-current trajectory characteristics .

[0008] Further, in step S2, the multi-view electrical features are divided according to a preset time window, and the divided electrical segments are subjected to time alignment, viewpoint combination, and occlusion processing to form a set of context segments and a set of occluded target segments for self-supervised training of the load joint embedding prediction architecture. The specific process is as follows: S21: Set the time window length based on the load sampling frequency, power frequency cycle, start-up transient duration, and steady-state operating characteristics. and sliding step size The electrical segments are obtained by dividing the circuit. Among them, for the first Class-based electrical characteristics In the Electrical segments within a time window Represented as: ; Total number of electrical segments for: ; in, This represents the total length of the original sampling sequence; This indicates the floor function; Multi-view electrical segments within the same time window are represented as follows: ; in, For the first Multiple electrical segments corresponding to each time window; This is an electrical segment viewed from a time-domain waveform perspective; This is an electrical segment from a power perspective; This is an electrical segment from a frequency domain perspective; An electrical segment from the perspective of voltage-current trajectory; S22: Arrange all multi-view electrical segments within the time window in chronological order to form an electrical segment sequence; wherein, the electrical segment sequence is represented as: ; in, For the reason A multi-view electrical segment sequence composed of time windows; S23: For electrical segment sequences Occlusion is performed to obtain a set of unoccluded context fragments. , set of occluded target segments and marking of the obstructed position; Among them, electrical segment sequence A collection of unobstructed context fragments With the set of occluded target segments The relationship between them is represented as follows: .

[0009] Furthermore, in S23, the occlusion of electrical segments is divided into time window occlusion, feature channel occlusion, and viewpoint occlusion according to the object being occluded; Time window occlusion refers to randomly selecting one or more continuous or discrete time window indices according to a preset time occlusion ratio, and occluding all electrical segments within the selected time window. Feature channel occlusion refers to randomly selecting one or more feature channel indices within an electrical segment that is not occluded by a time window, according to a preset channel occlusion ratio, and occluding the selected feature channels within the corresponding time window and viewpoint. Viewpoint occlusion refers to randomly selecting at least one viewpoint type according to a preset viewpoint occlusion ratio and occluding all feature channels of the selected viewpoint within one or more time windows. ; ; in, For the first The amount of occlusion indication for each time window. For the first Occlusion indicator for view-like perspectives For the first The first time window, the first Within the class perspective The occlusion indicator value of each feature channel; when the indicator value is 1, it means that the corresponding object is selected and occluded, and when it is 0, it means that it is not selected and occluded; This is the final occlusion indication formed by the combination of the three types of occlusion. Electrical characteristics before shielding, The occlusion vector is a learnable vector; the original electrical features of the occluded location are preserved in the target fragment set and are input only into the target encoder to generate the target latent representation.

[0010] By setting occlusion position markers to indicate the corresponding time window position, viewpoint type, and feature channel of the occluded segment, the latent space predictor can locate the occluded segment according to the time window position, viewpoint type, and feature channel, and predict its target latent representation; wherein, the occlusion position marker is represented as: ; in, A set of occlusion location markers for the occluded segment; Index of the time window position of the occluded segment; Index of the viewpoint feature category to which the occluded segment belongs; The index is for the occluded feature channels; the viewpoint type is used to distinguish feature sources such as time-domain waveforms, power, frequency domain, and voltage-current trajectories; the feature channels are used to distinguish the specific measurement dimensions within each viewpoint, such as voltage and current channels in the time-domain viewpoint, active power and reactive power channels in the power viewpoint, and harmonic channels in the frequency-domain viewpoint.

[0011] Furthermore, in S3, during the self-supervised pre-training phase, a load representation model is constructed based on the load joint embedding prediction architecture, and the set of unoccluded context fragments is utilized. , set of occluded target segments and occlusion location markers Self-supervised pre-training of the load representation model; The load joint embedding prediction architecture includes a context encoder. Latent space predictor and exponential moving average target encoder ;in, Represents the trainable parameters of the context encoder; This represents the trainable parameters of the latent space predictor; Parameters representing the target encoder; context encoder Used to extract the contextual latent representation of unobstructed electrical segments; target encoder Used to generate latent target representations of occluded target fragments; latent space predictor It is used to predict the latent representation of the occluded target segment in the latent space of the target encoder based on the contextual latent representation and the occlusion location marker; The collection of unoccluded context fragments Input Context Encoder We obtain the contextual latent representation: ; in, This represents the contextual potential representation of the unobstructed electrical segment. contextual latent representation and occlusion position markers Input to latent space predictor Predicting the potential representation of the occluded target fragment : ; At the same time, the set of target segments that will be obscured Input target encoder The latent representation of the target corresponding to the occluded target segment is obtained. : ; Based on context encoder parameters Update the target encoder parameters using an exponential moving average method. : ; in, This is the exponential moving average coefficient, used to control the update rate of the target encoder parameters, and ; For assignment / update operations; Based on the latent representation and prediction of the occluded target fragment, a latent space prediction loss function is constructed to obtain the latent representation of the corresponding target fragment. ; in, Predicting losses for potential spaces; This is a function to stop gradient operations, used to prevent the gradient of the loss function from being passed to the operation within the parentheses; L1 norm represents the sum of the absolute values ​​of the vector components.

[0012] Furthermore, the specific process of S4 is as follows: S41: The first The first time window Electrical fragments from a class perspective Input load characterization extractor to obtain corresponding latent characterizations : ; in, For the first The first time window Electrical fragments from a class perspective Corresponding latent representations; S42: Based on each potential representation, combined with the corresponding viewpoint embedding and time window embedding, the electrical representation after incorporating viewpoint embedding and time window embedding is obtained: ; ; in, This represents the electrical characterization after incorporating viewpoint embedding and time window embedding; Indicates the first The view embedding corresponding to the view electrical features of the class (the view embedding is obtained by looking up a table according to the view category index of the learnable view embedding matrix); Indicates the first The time window embeddings corresponding to each time window (the time window embeddings are obtained by looking up a table according to the time window index of the learnable time embedding matrix). Embedding matrices for learnable perspectives, For the first The one-hot vector corresponding to the class perspective; Embedding matrices for learnable time windows, For the first The one-hot vector corresponding to each time window; This represents the potential representation dimension. A one-hot vector is an encoded vector that takes a value of 1 only at the index position corresponding to the view category or time window, and a value of 0 at all other positions.

[0013] S43: Arrange all embedded electrical representations according to time window and perspective order to form a set of representations to be fused: ; in, The set of representations to be fused; Number of time windows; The total number of perspective types for electrical features; S44: Based on the time window and whether there are missing electrical features of each viewpoint type in the set of representations to be fused, construct a mask matrix to determine whether to mask the corresponding missing areas. Among them, for the first in the mask matrix The first time window Missing electrical features from a class perspective It is expressed as follows: ; S45: Set of representations to be fused and mask matrix Inputting the multi-head attention fusion module yields the fused intermediate representation. : ; in, This is the intermediate representation obtained after multi-head attention fusion; This is a multi-head attention fusion function used to perform weighted fusion of the output representations to be fused based on the mask matrix; S46: Introducing Learnable Query Vectors Cross-attention fusion is performed on the fused intermediate representations to obtain the global representation of the payload to be identified. : ; in, This is a global representation of the load to be identified; The query vector is a learnable vector, serving as the query endpoint across attention mechanisms. The fused intermediate representation serves as the key and value. By calculating the matching weights between the query vector and the representations of each effective time window and each effective viewpoint, the effective representations are weighted and aggregated to obtain a global representation with fixed dimensions. This is a cross-attention fusion function.

[0014] Furthermore, based on the attention weights generated by the multi-head attention fusion module, the contribution of each feature perspective to the load recognition result is calculated: ; in, For the first The average contribution of electrical features from a class perspective to the recognition results; For the first The first time window Attention weight values ​​corresponding to electrical features from a class perspective.

[0015] Furthermore, in S5, the global representation of the load to be identified is input into the load classifier, the probability distribution of the electricity object to be identified belonging to various types of loads is output, and whether it belongs to a known load category is determined based on the classification confidence and representation distance. Among them, the global representation of the load to be identified Input the load classifier to obtain the load type probability distribution: ; in, The predicted probability that the load to be identified belongs to each known category; This is the weight matrix of the classifier; The bias of the classifier; This is the activation function, used to convert the classifier output into a probability distribution; When training the classifier, the classification loss function is expressed as: ; in, This is the classification loss value; The total number of labeled training samples; This is the index of the training sample number; Indexed by the sequence number of the load category; The total number of known load categories; For the first Does the sample belong to the ? A binary indicator variable for load-like conditions; For the first The sample was predicted as the first... The probability of class load.

[0016] Furthermore, for the payload to be identified, the category with the highest predicted probability is selected as the candidate identification result: ; in, The index for the load category with the highest predicted probability; To select the variable that maximizes the objective function The possible values ​​of ; The load belongs to the first The predicted probability of a class; When the maximum class probability is not lower than the preset confidence threshold When, the load to be identified is determined to be the first Class load, where, A preset probability confidence threshold is used to determine whether the recognition result is reliable. ; in, The maximum value in the predicted probability vector; When the maximum class probability is lower than the preset confidence threshold, calculate the global representation of the load to be identified. The distance between the known class centers; where the first class center is the distance between the known class centers. The category center of the class load is represented as: ; in, For the first The class center vector of the class load; For the first Number of labeled samples for class load; For the first The first in the class Global characterization of the load of each labeled sample; The criteria for identifying unknown loads are expressed as follows: ; in, This is a distance metric function used to calculate the distance between two vectors; For all load categories Take the minimum value; The preset distance threshold is used to determine whether the load representation deviates from the known category distribution.

[0017] Furthermore, in S5, when the load to be identified is determined to be an unknown load, it is added to the sample library to be confirmed. After confirmation, it is added to the training sample library as a new load category sample or an abnormal load sample for subsequent model updates. During model updates, the pre-trained encoder is kept unchanged while only the classifier is updated; or the encoder parameters are incrementally fine-tuned after the number of new samples reaches a preset scale.

[0018] A multi-perspective electrical load type identification method based on a load joint embedding prediction architecture has the following advantages:

[0019] First, at the representation learning level, this invention abandons the traditional self-supervised method's approach of using the reconstruction of original voltage, current, or power waveforms as the training objective, and instead adopts a latent space representation prediction strategy. By extracting latent representations of visible electrical segments through a context encoder, and using a latent space predictor to directly predict the target representation of occluded segments in the latent space, the model focuses on stable structural features related to the essential attributes of the load type, including startup transient response, steady-state power level, harmonic distortion degree, and voltage-current trajectory shape, while consciously ignoring redundant information such as sampling noise, voltage fluctuations, transformer errors, and transient disturbances. Compared to reconstruction-based autoencoder methods, the representations learned in this invention have stronger discriminative and anti-interference capabilities, significantly improving the stability of load identification under complex acquisition conditions.

[0020] Secondly, at the data utilization level, this invention constructs self-supervised pre-training samples through three strategies: time segment occlusion, feature channel occlusion, and viewpoint occlusion. This effectively simulates common non-ideal operating conditions in actual engineering, such as communication packet loss, abnormal acquisition terminals, missing local data, unavailable harmonic data, and insufficient voltage-current trajectory sampling. During the pre-training phase, the model has already learned the potential correlation ability to infer the load operating state from incomplete electrical data. Therefore, even if some feature viewpoints are missing or some time segments are of poor quality, it can still maintain a high recognition accuracy during the inference phase.

[0021] Third, regarding the dependence on labeled samples, this invention adopts a two-stage paradigm of pre-training followed by lightweight fine-tuning. Self-supervised pre-training requires only unlabeled electrical data, eliminating the need for manual labeling of load categories; the downstream classification task requires only a small number of labeled samples to train a lightweight classifier, significantly reducing the cost of large-scale manual labeling. This characteristic makes this invention highly valuable for practical engineering applications in scenarios where labeled samples are scarce.

[0022] Fourth, in terms of unknown load discrimination, this invention abandons the traditional closed set classification paradigm and effectively identifies novel loads, temporary access loads, or abnormal loads not included in the training set by using dual criteria of classification confidence threshold and representation spatial distance threshold. This avoids the risk of misjudgment caused by forcibly classifying unknown loads into known categories and improves the scalability and security of the method.

[0023] Fifth, at the multi-view fusion level, this invention uses a multi-head masking attention mechanism to adaptively weight and fuse electrical representations from different time windows and feature perspectives. This dynamically adjusts the contribution weights of each perspective and segment based on actual data availability, ensuring the fusion process is unaffected by missing perspectives or abnormal segments. Simultaneously, learnable perspective embedding and time window embedding endow the model with the ability to distinguish different feature sources and temporal locations, further enhancing the semantic richness of load representation.

[0024] In summary, this invention significantly improves upon the accuracy of identification, robustness against noise, adaptability to small samples, tolerance to missing data, and ability to discover unknown loads, effectively meeting the needs of load type identification in complex real-world power consumption scenarios. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating the multi-perspective electrical load type identification method based on a load joint embedding prediction architecture provided in an embodiment of the present invention. Figure 2 This is a comparison chart of the accuracy of load type identification under different missing proportions from different perspectives provided in the embodiments of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] Example 1

[0029] like Figure 1 As shown, this invention provides a multi-perspective electrical load type identification method based on a load joint embedding prediction architecture, including: S1: Acquire multi-source electrical data of historical loads and construct multi-view electrical features that characterize the multi-dimensional operating characteristics of the loads.

[0030] Specifically, the acquired multi-source electrical data undergoes preprocessing, including time synchronization, outlier removal, missing value imputation, filtering and denoising, and normalization. This aims to reduce the impact of time deviations between different acquisition channels, sensor errors, communication packet loss, and amplitude scale differences on subsequent load identification.

[0031] The viewpoint types of electrical features include time-domain waveform viewpoint features. Power perspective characteristics Frequency domain perspective features Voltage-current trajectory characteristics .

[0032] Time-domain waveform characteristics include voltage waveforms, current waveforms, voltage change rate, current change rate, peak starting current, and transient current variation characteristics, used to characterize the dynamic response during load addition, removal, and operation state switching. Power-domain characteristics include active power, reactive power, apparent power, and power factor, used to characterize load energy consumption, reactive power demand, and power variation characteristics. Frequency-domain harmonic characteristics include the fundamental component, harmonic amplitude, harmonic phase, and total harmonic distortion rate, used to characterize the harmonic distortion characteristics of nonlinear loads and power electronic loads. Voltage-current trajectory characteristics consist of voltage and current signals within the same time window, used to characterize the load's phase relationship, nonlinearity, and differences in operating states.

[0033] Constructing a multi-level set of electrical features to characterize the multi-dimensional operating characteristics of loads .

[0034] S2: Perform electrical segmentation and occlusion construction on electrical features to form an unoccluded context segment set and an occluded target segment set for self-supervised training, and generate occlusion position markers.

[0035] Specifically, the multi-view electrical features are divided according to a preset time window, and the divided electrical segments are subjected to time alignment, viewpoint combination, and occlusion processing to form a set of context segments and a set of occluded target segments for self-supervised training of the load joint embedding prediction architecture. The specific process is as follows: S21: Set the time window length based on the load sampling frequency, power frequency cycle, start-up transient duration, and steady-state operating characteristics. and sliding step size The electrical segments are obtained by dividing the circuit.

[0036] More specifically, the time window length is used to determine the number of sampling points included in each electrical segment, and the sliding step size is used to determine the degree of overlap between adjacent electrical segments. For loads with obvious start-up transients, the time window length is set to cover the time range of load connection, disconnection, or state switching processes; for steady-state operating loads, the time window length is set to cover several complete power frequency cycles to ensure that the segments contain stable voltage, current, power, and harmonic information.

[0037] Among them, for the first Class-based electrical characteristics In the Electrical segments within a time window Represented as: ; Total number of electrical segments for: ; in, This represents the total length of the original sampling sequence; This indicates a floor operation. Thus, the original continuous electrical signal is converted into several electrical segments with local timing characteristics, enabling subsequent models to learn the load's operational variation patterns across different time windows using these segments as basic units.

[0038] Time-domain waveform segments, power segments, frequency-domain harmonic segments, and voltage-current trajectory segments within the same time window are time-aligned so that characteristics from different perspectives correspond to the same load operating state. Multi-perspective electrical segments within the same time window are represented as follows: ; in, For the first Multiple electrical segments corresponding to each time window; This is an electrical segment viewed from a time-domain waveform perspective; This is an electrical segment from a power perspective; This is an electrical segment from a frequency domain perspective; An electrical segment from the perspective of voltage-current trajectory; S22: Arrange all multi-view electrical segments within the time window in chronological order to form an electrical segment sequence; wherein, the electrical segment sequence is represented as: ; in, For the reason A sequence of multi-view electrical segments composed of time windows. When a time window contains missing views, missing channels, or data quality below a preset threshold, the time window is not directly deleted. Instead, a corresponding missing marker is set for masking in subsequent occlusion construction and attention fusion modules. This approach preserves the temporal continuity of the load operation process and enhances the model's adaptability to incomplete data acquisition.

[0039] S23: For electrical segment sequences Occlusion is performed to obtain a set of unoccluded context fragments. , set of occluded target segments And marking the occlusion position.

[0040] Among them, electrical segment sequence A collection of unobstructed context fragments With the set of occluded target segments The relationship between them is represented as follows: .

[0041] Temporal occlusion refers to occluding continuous or discrete time window segments to simulate situations such as partial sampling loss, communication interruption, incomplete transient processes, or short-term acquisition failures during load operation. For example, occluding several consecutive time windows allows the model to predict the potential operational characteristics of the load within that time period based on the context segments before and after the occlusion. Feature channel occlusion refers to occluding partial features in voltage, current, active power, reactive power, power factor, or harmonic channels within a certain time window to simulate partial measurement anomalies in voltage transformers, current transformers, power quality monitoring devices, or acquisition terminals. Viewpoint occlusion refers to occluding partial views in the time domain waveform view, power view, frequency domain harmonic view, or voltage-current trajectory view to simulate situations where certain features are unavailable or the data quality is insufficient in actual engineering. For example, in low-frequency sampling scenarios, a complete voltage-current trajectory may not be obtainable; in ordinary smart meter scenarios, high-precision harmonic components may not be obtainable. In these cases, viewpoint occlusion allows the model to learn the load characterization under conditions of partial viewpoint loss.

[0042] Therefore, in S23, the occlusion of electrical segments is divided into time window occlusion, feature channel occlusion, and viewpoint occlusion according to the object being occluded. In specific implementation, the corresponding time window index, feature channel index, or viewpoint index is randomly selected according to the preset occlusion ratio. The three types of occlusion can be executed individually or in combination.

[0043] Time window occlusion refers to randomly selecting one or more continuous or discrete time window indices according to a preset time occlusion ratio, and occluding all electrical segments within the selected time window. Feature channel occlusion refers to randomly selecting one or more feature channel indices within an electrical segment that is not occluded by a time window, according to a preset channel occlusion ratio, and occluding the selected feature channels within the corresponding time window and viewpoint. Viewpoint occlusion refers to randomly selecting at least one viewpoint type according to a preset viewpoint occlusion ratio and occluding all feature channels of the selected viewpoint within one or more time windows. ; ; in, For the first The amount of occlusion indication for each time window. For the first Occlusion indicator for view-like perspectives For the first The first time window, the first Within the class perspective The occlusion indicator of each feature channel; when the indicator value is 1, it means that the corresponding object is selected and occluded, and when it is 0, it means that it is not selected and occluded; This is the final occlusion indication formed by the combination of the three types of occlusion. Electrical characteristics before shielding, The occlusion vector is a learnable vector; the original electrical features of the occluded location are preserved in the target fragment set and are input only into the target encoder to generate the target latent representation.

[0044] To enable the model to identify the location, viewpoint, and feature channel of an occluded segment, an occlusion location marker is set to indicate the corresponding time window location, viewpoint type, and feature channel of the occluded segment. This allows the latent space predictor to locate the occluded segment according to the time window location, viewpoint type, and feature channel, and predict its target latent representation. The occlusion location marker is represented as follows: ; in, A set of occlusion location markers for the occluded segment; Index of the time window position of the occluded segment; Index of the viewpoint feature category to which the occluded segment belongs; This indexes the occluded feature channels. The viewpoint type is used to distinguish feature sources such as time-domain waveforms, power, frequency-domain waveforms, and voltage-current trajectories; the feature channel is used to distinguish the specific measurement dimensions within each viewpoint, such as the voltage and current channels in the time-domain viewpoint, the active power and reactive power channels in the power viewpoint, and the harmonic channels in the frequency-domain viewpoint.

[0045] In specific implementation, the occlusion location marker is only used to provide the latent space predictor with the location information of the occluded segment during the self-supervised pre-training stage, and does not participate in the calculation during the inference and recognition stage; the set of unoccluded context segments is used to input the load joint embedding prediction architecture into the context encoder during the self-supervised pre-training stage to extract the context latent representation, and the set of occluded target segments is used to input the load joint embedding prediction architecture into the target encoder during the self-supervised pre-training stage to generate the target latent representation.

[0046] S3: Based on the load joint embedding prediction architecture, it uses the set of unoccluded context segments, the set of occluded target segments, and the occlusion location markers for self-supervised pre-training. It predicts the target potential representation of the occluded electrical segment through the latent space predictor and calculates the prediction loss. The pre-trained context encoder obtained after training is used as the load representation extractor.

[0047] Specifically, in the self-supervised pre-training phase, a load representation model is constructed based on the load joint embedding prediction architecture, and an unoccluded set of context fragments is utilized. , set of occluded target segments and occlusion location markers The load representation model is pre-trained under self-supervised conditions.

[0048] The load joint embedding prediction architecture includes a context encoder. Latent space predictor and exponential moving average target encoder ;in, Represents the trainable parameters of the context encoder; This represents the trainable parameters of the latent space predictor; Parameters representing the target encoder; context encoder Used to extract the contextual latent representation of unobstructed electrical segments; target encoder Used to generate latent target representations of occluded target fragments; latent space predictor It is used to predict the latent representation of an occluded target fragment in the latent space of the target encoder based on the contextual latent representation and the occlusion location marker.

[0049] Unlike autoencoders that directly reconstruct the original voltage and current waveforms, this method does not use the reconstruction of the original electrical signal as the training objective. Instead, the latent space predictor predicts the target representation of the obscured segment in the latent space based on the load context information contained in the unobscured electrical segment, as well as the time position, feature viewpoint, and feature channel information corresponding to the obscured electrical segment. This allows the model to focus on learning stable structural features related to the load type (including load start-up transient features, steady-state power change features, harmonic distortion features, and voltage-current trajectory features), rather than predicting the original voltage, current, or power values ​​of the obscured segment.

[0050] In practical implementation, the context encoder can employ a Transformer encoder, a temporal convolutional encoder, a gated cyclic encoder, or a combination thereof. The context encoder maps time-domain waveform information, power variation information, harmonic information, and voltage-current trajectory information from electrical segments into a unified latent space, enabling load features from different perspectives to be correlated and learned within the same representation space. The context encoder includes a segment embedding layer, a multi-head self-attention layer, a feedforward neural network layer, and a normalization layer. The segment embedding layer converts multi-view electrical segments into token representations; the multi-head self-attention layer learns the correlation between different time windows and different electrical perspectives; the feedforward neural network layer enhances nonlinear representation capabilities; and the normalization layer stabilizes the model training process. Through the context encoder, the model can extract contextual information about the load operating state from unobstructed electrical segments, providing a basis for subsequent prediction of the latent representation of obstructed segments.

[0051] The collection of unoccluded context fragments Input Context Encoder We obtain the contextual latent representation: ; in, This represents the contextual potential representation of the unshielded electrical segment.

[0052] The latent space predictor predicts the representation of a blocked target segment in the latent space based on the contextual latent representation of the unblocked electrical segment and the blocking location markers, rather than predicting the raw voltage, current, or power values ​​of the blocked segment. This approach reduces the model's dependence on transient noise, occasional disturbances, and sampling errors, allowing the model to focus more on the stability discrimination characteristics of the load type itself.

[0053] contextual latent representation and occlusion position markers Input to latent space predictor Predicting the potential representation of the occluded target fragment : ; At the same time, the set of target segments that will be obscured Input target encoder The latent representation of the target corresponding to the occluded target segment is obtained. : ; The target encoder and the context encoder have the same or similar network structure, but the target encoder parameters are not updated directly through backpropagation, but are based on the context encoder parameters. Update the target encoder parameters using an exponential moving average method. : ; in, This is the exponential moving average coefficient, used to control the update rate of the target encoder parameters, and ; This is for the assignment update operation. Through exponential moving average updates, the target encoder can provide a relatively stable latent representation of the target, avoiding drastic fluctuations in the target representation during training, thereby improving the stability of self-supervised pre-training and enabling the learning of potential relationships between electrical segments under conditions of no class labels or a small number of labels.

[0054] Based on the latent representation and prediction of the occluded target fragment, a latent space prediction loss function is constructed to obtain the latent representation of the corresponding target fragment. ; in, Predicting losses for potential spaces; This is a function to stop gradient operations, used to prevent the gradient of the loss function from being passed to the operation within the parentheses; This represents the L1 norm, which is the sum of the absolute values ​​of the vector components. The stop gradient operation prevents the gradient of the loss function from being directly passed to the target encoder, ensuring that the target encoder is updated only through an exponential moving average. This avoids representation collapse or training instability caused by simultaneous rapid changes in both the context encoder and the target encoder.

[0055] In practice, the context encoder and latent space predictor are updated by minimizing the latent space prediction loss, so that the predicted latent representation gradually approaches the target latent representation. After pre-training, the context encoder is retained as the load representation extractor, while the target encoder and latent space predictor are no longer used in the inference and recognition stage.

[0056] Therefore, it can be seen that the self-supervised training is completed by extracting the representation of unobstructed electrical segments through the context encoder, predicting the representation of obstructed electrical segments through the latent space predictor, and then providing stable target representations through the exponential moving average target encoder. This does not depend on the original electrical signal reconstruction process, thus enabling the model to learn stable latent features related to the load type.

[0057] S4: Based on the load characterization extractor, extract the multi-view electrical characterization of the load to be identified, and combine the viewpoint time window embedding and mask attention fusion mechanism to fuse the electrical characterization of each viewpoint to obtain the global characterization of the load to be identified.

[0058] In the inference and recognition stage, the pre-trained context encoder is used as the load representation extractor to encode electrical segments of various time windows and electrical features of the load to be identified. Occlusion construction is no longer performed, nor are target encoders and latent space predictors used. For electrical segments of the load to be identified, training targets are no longer generated through the target encoder, nor are decoders used to reconstruct the original electrical signals. Instead, the context encoder is used directly to extract the electrical latent representations corresponding to each time window and feature viewpoint. After self-supervised pre-training, the pre-trained context encoder is used as the load representation extractor to encode the multi-view electrical segments of the load to be identified, obtaining the electrical latent representations corresponding to each time window and feature viewpoint. Then, a viewpoint embedding, time window embedding, and mask attention fusion mechanism are introduced to uniformly fuse electrical representations from different sources and time locations, obtaining a global load representation for load type identification.

[0059] S41: The first The first time window Electrical fragments from a class perspective Input load characterization extractor to obtain corresponding latent characterizations : ; in, For the first The first time window Electrical fragments from a class perspective The corresponding latent representations. The context encoder is used to extract load operation features from different electrical perspectives, including time-domain waveform variation features, power variation features, harmonic distortion features, and voltage-current trajectory morphology features. Since the encoder has been pre-trained through a latent space prediction task, its output representation tends to retain stable features related to load type, while weakening the impact of sampling noise, transient disturbances, and locally missing data.

[0060] S42: To enable the model to distinguish between different feature sources and temporal locations, based on each latent representation, the corresponding viewpoint embedding and temporal window embedding are combined to obtain the electrical representation after incorporating viewpoint embedding and temporal window embedding: ; ; in, This represents the electrical characterization after incorporating viewpoint embedding and time window embedding; Indicates the first View embedding corresponding to class-view electrical features; Indicates the first Embedding of time windows corresponding to each time window; Embedding matrices for learnable perspectives, For the first The one-hot vector corresponding to the class perspective; Embedding matrices for learnable time windows, For the first The one-hot vector corresponding to each time window; This represents the potential representation dimension. A one-hot vector is an encoded vector that takes a value of 1 only at the index position corresponding to the view category or time window, and a value of 0 at all other positions. For example, Corresponding to the time-domain waveform perspective, power perspective, frequency-domain perspective, and voltage-current trajectory perspective, respectively, the model is from... Obtain the corresponding embedding vector from the data.

[0061] Perspective embedding is used to identify whether the current representation originates from a time-domain waveform perspective, a power perspective, a frequency-domain harmonic perspective, or a voltage-current trajectory perspective; time window embedding is used to identify whether the current representation corresponds to a load startup phase, a steady-state operation phase, a state transition phase, or other time locations. By incorporating perspective embedding and time window embedding, the model can distinguish the physical meaning of different features during subsequent fusion. For example, the frequency-domain harmonic perspective is more suitable for distinguishing power electronic loads, the power perspective is more suitable for distinguishing resistive and inductive loads, and the time-domain waveform perspective and transient segments are more suitable for distinguishing motor-type and refrigeration-type loads.

[0062] S43: Arrange all embedded electrical representations according to time window and perspective order to form a set of representations to be fused: ; in, The set of representations to be fused; Number of time windows; The total number of perspective types for electrical features; by , For example, the permutation result formed according to the above rules is as follows: ; S44: Based on the time window and whether there are missing electrical features of each viewpoint type in the set of representations to be fused, construct a mask matrix to determine whether to mask the corresponding missing areas. Among them, for the first in the mask matrix The first time window Missing electrical features from a class perspective It is expressed as follows: ; S45: Set of representations to be fused and mask matrix Inputting the multi-head attention fusion module yields the fused intermediate representation. : ; in, This is the intermediate representation obtained after multi-head attention fusion; This is a multi-head attention fusion function used to perform weighted fusion of the output representations to be fused based on the mask matrix.

[0063] The multi-head attention fusion module is used to learn the relationships between different time windows, different feature perspectives, and different operational stages. Through a mask matrix, the model ignores unusable or low-quality segments when calculating attention weights, so that the fusion process is not affected by missing perspectives or abnormal segments.

[0064] S46: To obtain the overall features for load type identification, a learnable query vector is introduced. Cross-attention fusion is performed on the fused intermediate representations to obtain the global representation of the payload to be identified. : ; in, This is a global representation of the load to be identified; The query vector is a learnable vector, serving as the query endpoint across attention mechanisms. The fused intermediate representation serves as the key and value. By calculating the matching weights between the query vector and the representations of each effective time window and each effective viewpoint, the effective representations are weighted and aggregated to obtain a global representation with fixed dimensions. This is a cross-attention fusion function. The global representation of the load to be identified comprehensively reflects the load operation characteristics under different time windows and feature perspectives.

[0065] More specifically, based on the attention weights generated by the multi-head attention fusion module, the contribution of each feature perspective to the load recognition result is calculated: ; in, For the first The average contribution of electrical features from a class perspective to the recognition results; For the first The first time window Attention weight values ​​corresponding to electrical features from different perspectives. The perspective contribution provides interpretable evidence for load type identification results.

[0066] S5: Input the global representation of the load to be identified into the classifier, output the load type identification result, and judge the unknown load based on the classification confidence and representation distance.

[0067] The global characterization of the load to be identified includes stability discrimination information such as load transient response, steady-state power change, harmonic distortion, and voltage-current trajectory morphology.

[0068] The global representation of the load to be identified is input into the load classifier, which outputs the probability distribution of the electricity consumption object belonging to each load category. The classifier determines whether the load belongs to a known load category based on the classification confidence and representation distance. The load classifier is lightweight; in practice, it can employ any of the following: a linear classifier, a multilayer perceptron classifier, or an attention classifier. Preferably, the context encoder obtained through self-supervised pre-training in S3 is frozen during the load identification stage, and only the lightweight classifier is trained. This reduces reliance on a large number of manually labeled samples and avoids disrupting the stable load representation obtained through pre-training under small sample conditions.

[0069] Among them, the global representation of the load to be identified Input the load classifier to obtain the load type probability distribution: ; in, The predicted probability that the load to be identified belongs to each known category; This is the weight matrix of the classifier; The bias of the classifier; This is the activation function, used to convert the classifier output into a probability distribution; When training the classifier, the classification loss function is expressed as: ; in, This is the classification loss value; The total number of labeled training samples; This is the index of the training sample number; Indexed by the sequence number of the load category; The total number of known load categories; For the first Does the sample belong to the ? A binary indicator variable for load-like conditions; For the first The sample was predicted as the first... The probability of classifying loads. Since S3 has learned stable potential associations between electrical segments of loads through self-supervised pre-training, only a small number of labeled samples are needed to complete the training of the load classifier, thereby improving the load identification capability in small sample scenarios.

[0070] More specifically, for the payload to be identified, the category with the highest predicted probability is selected as the candidate identification result: ; in, The index for the load category with the highest predicted probability; To select the variable that maximizes the objective function The possible values ​​of ; The load belongs to the first The predicted probability of a class; When the maximum class probability is not lower than the preset confidence threshold When, the load to be identified is determined to be the first Class load, where, A preset probability confidence threshold is used to determine whether the recognition result is reliable. ; in, This is the maximum value in the predicted probability vector.

[0071] In practical implementation, load categories include at least one of the following: resistive loads, inductive loads, motor loads, power electronic loads, lighting loads, charging loads, cooling loads, agricultural irrigation and drainage loads, and commercial mixed loads. In practical applications, load categories can be set according to different scenarios. For example, in a shared power equipment scenario, they can be set as lighting loads, charging loads, electric heating loads, motor loads, agricultural irrigation and drainage loads, night market stall mixed loads, and abnormal loads; in a park energy efficiency management scenario, they can be set as air conditioning and cooling loads, power equipment loads, office equipment loads, lighting loads, and charging facility loads.

[0072] When the maximum class probability is lower than the preset confidence threshold, calculate the global representation of the load to be identified. The distance between the known class centers; where the first class center is the distance between the known class centers. The category center of the class load is represented as: ; in, For the first The class center vector of the class load; For the first Number of labeled samples for class load; For the first The first in the class Global characterization of the load of each labeled sample; The criteria for identifying unknown loads are expressed as follows: ; in, This is a distance metric function used to calculate the distance between two vectors; For all load categories Take the minimum value; The preset distance threshold is used to determine whether the load representation deviates from the known category distribution.

[0073] When any of the above conditions are met, the electrical object to be identified is marked as an unknown load, and its original electrical data, multi-view electrical characteristics, global load characterization and identification time are recorded for subsequent manual verification or incremental training.

[0074] Output the identification results of the load to be identified, including the load category, category probability, unknown load marker, main contributing perspective, and identification time.

[0075] Preferably, when the load to be identified is determined to be an unknown load, it is added to the sample library to be confirmed. After confirmation, it is added to the training sample library as a new load category sample or an abnormal load sample for subsequent model updates. During the model update process, the pre-trained encoder is kept unchanged and only the classifier is updated. Alternatively, after the number of new samples reaches a preset scale, the parameters of the encoder are incrementally fine-tuned.

[0076] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0077] To verify the identification stability of this invention under conditions of incomplete electrical data from multiple perspectives, different proportions of missing perspectives were set in Example 1 to simulate the situation where some perspectives were unavailable in the time-domain waveform perspective, power perspective, frequency-domain harmonic perspective, and voltage-current trajectory perspective. CNN, GNN, Transformer, and the LoadJEPA method described in this invention were then used for load type identification and comparison. Figure 2As shown, the recognition accuracy of all methods decreases as the proportion of missing viewpoints increases from 0% to 40%. Among them, the CNN, GNN, and Transformer methods show a more significant decrease as the proportion of missing viewpoints increases, indicating their strong dependence on complete input features; while the LoadJEPA method described in this invention maintains a high recognition accuracy under different proportions of missing viewpoints, and the decrease is relatively small.

[0078] The above results demonstrate that this invention constructs self-supervised learning samples through temporal segment occlusion, feature channel occlusion, and viewpoint occlusion, enabling the model to learn potential load correlations under missing viewpoint conditions during the training phase. Simultaneously, it masks unusable viewpoints through a masked attention fusion mechanism and extracts effective load discrimination information from the remaining available viewpoints. Therefore, this invention maintains good load type identification capabilities even under conditions such as missing harmonic data, unavailable voltage-current trajectories, or abnormal acquisition channels.

[0079] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-perspective electrical load type identification method based on a load joint embedded prediction architecture, characterized in that, include: S1: Acquire multi-source electrical data of historical loads and construct multi-view electrical features that characterize the multi-dimensional operating characteristics of the loads; S2: Perform electrical segmentation and occlusion construction on electrical features to form an unoccluded context segment set and an occluded target segment set for self-supervised training, and generate occlusion position markers; S3: Based on the load joint embedding prediction architecture, self-supervised pre-training is performed using the set of unoccluded context segments, the set of occluded target segments, and the occlusion location markers. The latent space predictor predicts the target potential representation of the occluded electrical segment and calculates the prediction loss. The pre-trained context encoder obtained after training is used as the load representation extractor. S4: Extract multi-view electrical representations of the load to be identified based on the load representation extractor, and fuse the electrical representations of each view by combining the view time window embedding and mask attention fusion mechanism to obtain the global representation of the load to be identified. S5: Input the global representation of the load to be identified into the classifier, output the load type identification result, and judge the unknown load based on the classification confidence and representation distance.

2. The method according to claim 1, characterized in that, In S1, the perspective types of electrical features include time-domain waveform perspective features, power perspective features, frequency-domain perspective features, and voltage-current trajectory perspective features.

3. The method according to claim 1, characterized in that, In step S2, the multi-view electrical features are divided according to a preset time window, and the divided electrical segments are subjected to time alignment, viewpoint combination, and occlusion processing to form a set of context segments and a set of occluded target segments for self-supervised training of the load joint embedding prediction architecture. The specific process is as follows: S21: Set the time window length based on the load sampling frequency, power frequency cycle, start-up transient duration, and steady-state operating characteristics. and sliding step size The electrical segments are obtained by dividing the circuit. Among them, for the first Class-based electrical characteristics In the Electrical segments within a time window Represented as: ； Total number of electrical segments for: ； in, This represents the total length of the original sampling sequence; This indicates the floor function; Multi-view electrical segments within the same time window are represented as follows: ； in, For the first Multiple electrical segments corresponding to each time window; This is an electrical segment viewed from a time-domain waveform perspective; This is an electrical segment from a power perspective; This is an electrical segment from a frequency domain perspective; An electrical segment from the perspective of voltage-current trajectory; S22: Arrange all multi-view electrical segments within the time window in chronological order to form an electrical segment sequence; wherein, the electrical segment sequence is represented as: ； in, For the reason A multi-view electrical segment sequence composed of time windows; S23: For electrical segment sequences Occlusion is performed to obtain a set of unoccluded context fragments. , set of occluded target segments and marking of the obstructed position; Among them, electrical segment sequence A collection of unobstructed context fragments With the set of occluded target segments The relationship between them is represented as follows: 。 4. The method according to claim 3, characterized in that, In S23, the occlusion of electrical segments is divided into time window occlusion, feature channel occlusion, and viewpoint occlusion according to the object being occluded; Time window occlusion refers to randomly selecting one or more continuous or discrete time window indices according to a preset time occlusion ratio, and occluding all electrical segments within the selected time window. Feature channel occlusion refers to randomly selecting one or more feature channel indices within an electrical segment that is not occluded by a time window, according to a preset channel occlusion ratio, and occluding the selected feature channels within the corresponding time window and viewpoint. Viewpoint occlusion refers to randomly selecting at least one viewpoint type according to a preset viewpoint occlusion ratio and occluding all feature channels of the selected viewpoint within one or more time windows. ； ； in, For the first The amount of occlusion indication for each time window. For the first Occlusion indicator for view-like perspectives For the first The first time window, the first Within the class perspective Occlusion indication of each feature channel; This is the final occlusion indication formed by the combination of the three types of occlusion. Electrical characteristics before shielding, The occlusion vector is a learnable vector. By setting occlusion position markers to indicate the corresponding time window position, viewpoint type, and feature channel of the occluded segment, the latent space predictor can locate the occluded segment according to the time window position, viewpoint type, and feature channel, and predict its target latent representation; wherein, the occlusion position marker is represented as: ； in, A set of occlusion location markers for the occluded segment; Index of the time window position of the occluded segment; Index of the viewpoint feature category to which the occluded segment belongs; Index of the occluded feature channels.

5. The method according to claim 1, characterized in that, In S3, during the self-supervised pre-training phase, a load representation model is constructed based on the load joint embedding prediction architecture, and an unoccluded set of context fragments is utilized. , set of occluded target segments and occlusion location markers Self-supervised pre-training of the load representation model; The load joint embedding prediction architecture includes a context encoder. Latent space predictor and exponential moving average target encoder ;in, Represents the trainable parameters of the context encoder; This represents the trainable parameters of the latent space predictor; Parameters representing the target encoder; context encoder Used to extract the contextual latent representation of unobstructed electrical segments; target encoder Used to generate latent target representations of occluded target fragments; latent space predictor It is used to predict the latent representation of the occluded target segment in the latent space of the target encoder based on the contextual latent representation and the occlusion location marker; The collection of unoccluded context fragments Input Context Encoder We obtain the contextual latent representation: ； in, This represents the contextual potential representation of the unobstructed electrical segment. contextual latent representation and occlusion position markers Input to latent space predictor Predicting the potential representation of the occluded target fragment : ； At the same time, the set of target segments that will be obscured Input target encoder The latent representation of the target corresponding to the occluded target segment is obtained. : ； Based on context encoder parameters Update the target encoder parameters using an exponential moving average method. : ； in, This is the exponential moving average coefficient, used to control the update rate of the target encoder parameters, and ; For assignment / update operations; Based on the latent representation and prediction of the occluded target fragment, a latent space prediction loss function is constructed to obtain the latent representation of the corresponding target fragment. ； in, Predicting losses for potential spaces; This is a function to stop gradient operations, used to prevent the gradient of the loss function from being passed to the operation within the parentheses; L1 norm represents the sum of the absolute values ​​of the vector components.

6. The method according to claim 1, characterized in that, The specific process in S4 is as follows: S41: The first The first time window Electrical fragments from a class perspective Input load characterization extractor to obtain corresponding latent characterizations : ； in, For the first The first time window Electrical fragments from a class perspective Corresponding latent representations; S42: Based on each potential representation, combined with the corresponding viewpoint embedding and time window embedding, the electrical representation after incorporating viewpoint embedding and time window embedding is obtained: ； ； in, This represents the electrical characterization after incorporating viewpoint embedding and time window embedding; Indicates the first View embedding corresponding to class-view electrical features; Indicates the first Embedding of time windows corresponding to each time window; Embedding matrices for learnable perspectives, For the first The one-hot vector corresponding to the class perspective; Embedding matrices for learnable time windows, For the first The one-hot vector corresponding to each time window; As a potential representation dimension; S43: Arrange all embedded electrical representations according to time window and perspective order to form a set of representations to be fused: ； in, The set of representations to be fused; Number of time windows; The total number of perspective types for electrical features; S44: Based on the time window and whether there are missing electrical features of each viewpoint type in the set of representations to be fused, construct a mask matrix to determine whether to mask the corresponding missing areas. Among them, for the first in the mask matrix The first time window Missing electrical features from a class perspective It is expressed as follows: ； S45: Set of representations to be fused and mask matrix Inputting the multi-head attention fusion module yields the fused intermediate representation. : ； in, This is the intermediate representation obtained after multi-head attention fusion; This is a multi-head attention fusion function used to perform weighted fusion of the output representations to be fused based on the mask matrix; S46: Introducing Learnable Query Vectors Cross-attention fusion is performed on the fused intermediate representations to obtain the global representation of the payload to be identified. : ； in, This is a global representation of the load to be identified; A learnable query vector; This is a cross-attention fusion function.

7. The method according to claim 6, characterized in that, Based on the attention weights generated by the multi-head attention fusion module, the contribution of each feature perspective to the load recognition result is calculated: ； in, For the first The average contribution of class-specific features to the recognition results; For the first Attention weight values ​​corresponding to the first type of viewpoint features within each time window.

8. The method according to claim 1, characterized in that, In step S5, the global representation of the load to be identified is input into the load classifier, which outputs the probability distribution of the electrical object to be identified belonging to various types of loads, and determines whether it belongs to a known load category based on the classification confidence and representation distance. Among them, the global representation of the load to be identified Input the load classifier to obtain the load type probability distribution: ； in, The predicted probability that the load to be identified belongs to each known category; This is the weight matrix of the classifier; The bias of the classifier; This is the activation function, used to convert the classifier output into a probability distribution; When training the classifier, the classification loss function is expressed as: ； in, This is the classification loss value; The total number of labeled training samples; This is the index of the training sample number; Indexed by the sequence number of the load category; The total number of known load categories; For the first Does the sample belong to the ? A binary indicator variable for load-like conditions; For the first The sample was predicted as the first... The probability of class load.

9. The method according to claim 8, characterized in that, For the payload to be identified, the category with the highest predicted probability is selected as the candidate identification result: ； in, The index for the load category with the highest predicted probability; To select the variable that maximizes the objective function The possible values ​​of ; The load belongs to the first The predicted probability of a class; When the maximum class probability is not lower than the preset confidence threshold When, the load to be identified is determined to be the first Class load, where, A preset probability confidence threshold is used to determine whether the recognition result is reliable. ； in, The maximum value in the predicted probability vector; When the maximum class probability is lower than the preset confidence threshold, calculate the global representation of the load to be identified. The distance between the known class centers; where the first class center is the distance between the known class centers. The category center of the class load is represented as: ； in, For the first The class center vector of the class load; For the first Number of labeled samples for class load; For the first The first in the class Global characterization of the load of each labeled sample; The criteria for identifying unknown loads are expressed as follows: ； in, This is a distance metric function used to calculate the distance between two vectors; For all load categories Take the minimum value; The preset distance threshold is used to determine whether the load representation deviates from the known category distribution.

10. The method according to claim 1, characterized in that, In S5, when the load to be identified is determined to be an unknown load, it is added to the sample library to be confirmed. After confirmation, it is added to the training sample library as a new load category sample or an abnormal load sample for subsequent model updates. During model updates, the pre-trained encoder is kept unchanged while only the classifier is updated; or the encoder parameters are incrementally fine-tuned after the number of new samples reaches a preset scale.

Images