Method for monitoring the state of use of a device based on the electrical signals of a multi-energy complementary energy supply system

Abstract

The application provides a kind of based on the energy state monitoring method of equipment of multi-energy complementary energy supply system electric signal, steps include: obtaining each energy equipment timing electric signal, extracts multi-dimensional feature after pre-processing and constructs feature vector, and obtains original sample set in set;Known state original sample is applied to enhance, and abnormal behavior agent sample set is generated;Based on the two-stage model training mechanism, the Encoder-Classifier architecture model is constructed;Real-time acquisition of electric signal and output according to fixed window, pre-treatment is extracted into the model after feature and the monitoring result is obtained.The application combines the timing characteristics of electric signal and the operation law of equipment, can automatically generate abnormal behavior agent sample and does not need abnormal label, adapts to complex equipment scene;Through two-stage training of data enhancement and similarity learning, the model identifies equipment state and abnormality, and provides intelligent support for safe operation and energy efficiency improvement of multi-energy system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of measuring electrical variables, and in particular to a device energy state monitoring method based on electrical signals of a multi-energy complementary energy supply system. BACKGROUND

[0002] With the continuous advancement of carbon reduction goals and the acceleration of energy transformation, multi-energy complementary energy supply systems are gradually widely used in industrial parks, public buildings and large commercial complexes. In such systems, key energy-using equipment such as electric boilers, heat pumps, refrigeration units, energy storage equipment, etc. exhibit complex and variable operating characteristics under different operating conditions, and their energy efficiency, operating state and safety have important influence on the stable operation and energy utilization efficiency of the entire system. Therefore, realizing the state recognition and abnormal energy-using behavior monitoring of key energy-using equipment has become a key supporting technology for improving the intelligent level of multi-energy system operation.

[0003] However, the traditional energy-using equipment monitoring methods in the prior art have obvious drawbacks. Because the types of equipment in the multi-energy system are diverse, the operating mechanisms and variation laws of different equipment are significantly different, making it difficult for traditional recognition methods that rely on specific sensor data collection and analysis to fully adapt. Furthermore, as the number of monitoring equipment increases, the cost of building the operation control system also increases. Further, based on real-world needs, different analysis models need to be built for different equipment, and the significant increase in monitoring data will also seriously affect the system's response speed. Furthermore, because a large number of monitoring equipment need to be added for each energy-using equipment, the probability of false positives will also increase significantly, thereby affecting the overall stable operation efficiency of the system.

[0004] In view of the above, there is a need to provide a new energy-using equipment monitoring method that can monitor the energy-using state of equipment based on existing hardware, through data monitoring combined with intelligent analysis and recognition, and separate abnormal energy-using conditions, and is suitable for complex and diverse equipment scenarios, thereby realizing cost-reducing and efficiency-increasing intelligent monitoring. SUMMARY

[0005] The present application provides a key equipment state recognition and abnormal behavior detection method combining data enhancement and similarity learning mechanism to provide intelligent support for safe operation and energy efficiency improvement of multi-energy systems.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows:

[0007] The present application provides a device energy state monitoring method based on electrical signals of a multi-energy complementary energy supply system, comprising the following steps:

[0008] Obtaining the time series electrical signals of each energy-using equipment in the energy supply system;

[0009] Preprocessing the time-series electrical signal to obtain a preprocessed electrical signal;

[0010] Extracting multi-dimensional features from the preprocessed electrical signal, and constructing a feature vector of the energy-using device based on the multi-dimensional features; and

[0011] Applying a perturbation enhancement operation to the original samples of the known energy-using states to generate an abnormal behavior proxy sample set;

[0012] Constructing a device energy-using state monitoring model based on a two-stage model training mechanism; the device energy-using state monitoring model is an Encoder-Classifier architecture model, which is provided with a shared feature extraction network and a linear classifier;

[0013] Real-time acquisition of electrical signals of each energy-using device in the energy supply system, and output of real-time time-series electrical signals based on a fixed window;

[0014] Preprocessing the real-time time-series electrical signals, and inputting the feature extraction results into the device energy-using state monitoring model to obtain a monitoring result.

[0015] Optionally, in the two-stage model training, the following steps are included:

[0016] The first stage is based on the original sample set for closed-set supervised classification training;

[0017] The second stage is to construct a training pair by combining the original sample set and the abnormal behavior proxy sample set, and to introduce a similarity supervision loss based on the training pair for joint modeling to obtain the device energy-using state monitoring model.

[0018] Optionally, in the first stage, the optimization target is a classification loss function.

[0019] Optionally, in the first stage, the loss weight coefficient is set to 0.

[0020] After the first stage training is completed, the loss weight coefficient is adjusted to 1 for the second stage training.

[0021] Optionally, the multi-dimensional features include time domain features, frequency domain features, and statistical features.

[0022] Optionally, the perturbation enhancement operation includes at least one of amplitude perturbation, time scale stretching and compression, Gaussian noise perturbation, frequency domain coefficient perturbation, or feature channel random masking.

[0023] Optionally, the following steps are further included:

[0024] The monitoring result includes normal behavior and abnormal behavior.

[0025] outputting a device state category when the monitoring result is normal behavior;

[0026] triggering an alarm mechanism and outputting an abnormal behavior probability and a corresponding abnormal behavior label when the monitoring result is abnormal behavior.

[0027] Optionally, the obtaining of the monitoring result further comprises the following steps:

[0028] calculating a category confidence sum of the monitoring sample on a known category, calculating an abnormal confidence score based on the category confidence sum, and comparing the abnormal confidence score with a preset threshold;

[0029] outputting a category of the monitoring sample as abnormal behavior when the abnormal confidence score is higher than the preset threshold;

[0030] determining a category of the monitoring sample as normal behavior when the abnormal confidence score is not higher than the preset threshold.

[0031] Optionally, the time-series electrical signal at least comprises one of current, voltage, active power or reactive power.

[0032] Optionally, the preprocessing of the time-series electrical signal at least comprises one of denoising, outlier rejection or normalization.

[0033] Compared with the prior art, the present application has the following beneficial effects:

[0034] The present application can realize state monitoring of each energy-using device through electrical signal monitoring and analysis, and provides a novel state monitoring method. Further, the present application combines time-series characteristics of energy-using electrical signals and device operation rules, has the ability to automatically generate abnormal behavior proxy samples after disturbance enhancement, does not need to rely on abnormal labels, and is suitable for state recognition and abnormal monitoring methods in complex and diverse device scenarios. Furthermore, the present application combines data enhancement and similarity learning mechanisms for two-stage model training, so that the model can identify device states and abnormal behaviors, and provides intelligent support for safe operation and energy efficiency improvement of multi-energy systems. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0036] Figure 1 Method flowchart of the embodiments of the present application;

[0037] Figure 2The voltage, current and active power raw data diagram of the load event section of the embodiment 1 of the present application is shown in the figure;

[0038] Figure 3 The voltage, current and reactive current curve diagram of the energy saving lamp device of the embodiment 1 of the present application is shown in the figure;

[0039] Figure 4 The identification score statistics diagram of each device type of the embodiment 1 of the present application is shown in the figure. DETAILED DESCRIPTION

[0040] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0041] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in the subsequent drawings.

[0042] It is worth mentioning that, the methods used in the present application are conventional methods unless otherwise specified; the raw materials and devices used are conventional commercially available products unless otherwise specified, and their sources are not specifically limited.

[0043] It should be further mentioned that, in order to facilitate understanding, the method steps in the specific embodiments of the present application are described in a certain order, but those of ordinary skill in the art can change the order of the steps according to actual needs, therefore, it should not be regarded as a limitation; it is further mentioned that, in the description of the following specific embodiments, the upper and lower indexes of each parameter should be understood as the distinguishing marks of similar indexes unless otherwise specified, representing the parameters of the index-related or corresponding devices, and should not be understood as specific models or special marks.

[0044] Please refer to Figure 1 The present application provides a device energy state monitoring method based on a multi-energy complementary energy supply system electric signal, which comprises the following steps:

[0045] Firstly, the time sequence electric signal of each energy-using device in the energy supply system is obtained; optionally, the present application collects the current, voltage, active power, reactive power and other electric signals of each energy-using device during operation through a smart meter and / or a sensor, and combines the sampling time to mark the electric signal data, thereby obtaining the time sequence electric signal.

[0046] Timing signals can fully capture the operating characteristics of equipment under different conditions such as startup, stable operation, and load switching, and are the core data foundation for subsequent status analysis and anomaly identification. To ensure the signal coverage of the details of equipment operation, the sampling frequency is set to 1Hz or higher to ensure that no characteristic information of key operating states is missed.

[0047] The acquired time-series electrical signals are preprocessed. Optionally, signal preprocessing includes at least one of denoising, outlier removal, or normalization. This application specifically uses a combination of the above three methods for preprocessing. Denoising is used to eliminate interference from environmental electromagnetic noise and sensor thermal noise during signal transmission, specifically including wavelet transform denoising or moving average denoising. Outlier removal is used to remove isolated outlier data points caused by instantaneous sensor failures or sudden line interference, achieved through the 3σ criterion or box plot method. Normalization maps electrical signal indicators of different magnitudes, such as current in amperes and power in kilowatts, to the same data range of [0,1] or [-1,1], avoiding excessive bias of the model towards large-scale features due to numerical differences, which would affect the training effect.

[0048] After the above preprocessing, a stable preprocessed electrical signal that can be directly used for feature extraction can be obtained.

[0049] Multidimensional features are extracted based on preprocessed electrical signals. These multidimensional features include time-domain features, frequency-domain features, and statistical features. Time-domain features reflect the signal's variation over time, such as peak value, valley value, mean, variance, and waveform factor. Frequency-domain features are extracted after converting the time-domain signal to the frequency domain using Fourier transform or wavelet transform, reflecting the signal's energy distribution characteristics in different frequency bands, such as dominant frequency amplitude and band energy proportion. Statistical features are a quantitative description of the overall statistical regularity of the signal, such as skewness (reflecting the degree of asymmetry in signal distribution), kurtosis (reflecting the steepness of signal distribution), and kurtosis. The extracted multidimensional features are combined according to device type and signal category to construct feature vectors for each energy-consuming device. The construction process can be represented as follows:

[0050] ;

[0051] In the formula, For the first i Feature vectors of energy-consuming equipment in Taiwan For the equipment k Statistical representations of time-series characteristics, such as the mean of current signals and the amplitude of the dominant frequency of voltage signals; n This represents the total number of feature dimensions.

[0052] The feature vectors of all power-consuming devices are collected to form the original sample set for model training.

[0053] To solve the problem that the abnormal energy consumption behavior in the actual scene is sudden and non-repetitive, leading to the scarcity of abnormal samples and the difficulty of the model in learning abnormal characteristics, the original samples of multiple known energy consumption states are subjected to disturbance enhancement operation to generate an abnormal behavior proxy sample set, i.e. pseudo-unseen class samples representing the "unknown class" behavior.

[0054] Further, the disturbance enhancement operation of the present application at least includes one of amplitude disturbance, time scale stretching and compression, Gaussian noise disturbance, frequency domain coefficient disturbance, and feature channel random masking.

[0055] The amplitude disturbance randomly multiplies the signal amplitude of the original sample by an amplitude factor to simulate the instantaneous fluctuation of the device voltage or current;

[0056] The time scale stretching and compression stretches or compresses the time axis of the signal by a ratio of 0.9-1.1 to simulate the instantaneous change of the device running speed, such as the slight fluctuation of the motor speed.

[0057] The Gaussian noise disturbance adds a noise term to the original signal , such as a Gaussian distribution with a mean of 0 and a variance of 0.01-0.05, to simulate electromagnetic interference in actual operation.

[0058] The frequency domain coefficient disturbance randomly fine-tunes the frequency domain coefficients of the signal after Fourier transform (the adjustment amplitude does not exceed 10% of the original coefficient) to simulate the frequency characteristic change caused by unstable device working conditions.

[0059] The feature channel random masking randomly sets 10%-20% of the feature dimensions in the feature vector to zero to simulate the scene of temporary missing of local sensor data.

[0060] All disturbance operations ensure that the generated samples have significant differences from the original class characteristics, but conform to the physical operation rules of the device, such as the current amplitude not appearing negative and the power not exceeding the rated range of the device.

[0061] Further, the present application takes amplitude disturbance and Gaussian disturbance as examples to form abnormal behavior proxy samples which can be represented as:

[0062] ;

[0063] In the formula, is the disturbance enhancement operation.

[0064] The abnormal behavior agent sample set is collected to form an abnormal behavior agent sample set required for training. Abnormal energy use behaviors usually have the characteristics of suddenness, non-repeatability and unpredictability, and there is a lack of sufficient labeled abnormal data, which leads to insufficient generalization ability of conventional supervised learning methods in identifying unknown behaviors; the above operation of the application can solve the problem of scarcity of abnormal samples, that is, it is difficult to construct an effective supervised model.

[0065] The device energy state monitoring model is constructed based on a two-stage model training mechanism. The model adopts an Encoder-Classifier architecture, includes a shared feature extraction network Encoder() and a linear classifier Classifier(), and the output dimension of the model is , wherein is the number of known device states, is the number of pre-allocated abnormal behavior agent sample categories, that is, the number of pseudo-unseen classes, which is usually set to 3-5 classes according to the device type and the number of historical abnormal types.

[0066] (1) The first stage;

[0067] The first stage training is based on closed set supervised classification training of the original sample set. The core goal is to let the model learn the feature distribution of the known device state, and the optimization goal is the cross-entropy loss function, and the formula is:

[0068]

[0069] In the formula, is the cross-entropy loss, N is the number of original samples in the first stage training, is the true label of the i-th original sample, which adopts one-hot encoding, such as i [1, 0, …, 0], the dimension is consistent with the number of known classes , the prediction probability distribution of the model for the i-th original sample is output by the linear classifier Classifier(), and each dimension corresponds to the prediction probability of a known class, that is, i .

[0070] In order to ensure that the first stage model only focuses on known class classification, the weight coefficient λ2 of the subsequent similarity supervision loss is set to 0, and only the cross-entropy loss is optimized. Through this stage of training, the model can initially have accurate classification ability for known device states, and the training is usually ended when the cross-entropy loss tends to be stable on the validation set, such as when the loss decreases by less than 0.001 for 5 consecutive epochs.

[0071] ​​​​After the first stage training is completed, the loss weight coefficient λ2 is adjusted to 1, and the second stage training is entered.

[0072] (2) the second stage;

[0073] The core of the second stage is to let the model learn the feature difference between the known class samples and the abnormal behavior agent samples, and to construct a training pair with similarity by combining the original sample set and the abnormal behavior agent sample set , and introduce a similarity supervision loss for joint modeling.

[0074] Among them, the joint training set is constructed as ;

[0075] Further, the label similarity of any two samples , ( ) is defined, and the formula is:

[0076] ;

[0077] In the formula, is the label similarity of sample i and sample j . For example, when both samples belong to "heat pump normal refrigeration" (known class) or both belong to "voltage amplitude disturbance generated agent sample" (abnormal behavior agent class), , otherwise it is 0.

[0078] Define the feature encoding of the model to the sample and the similarity supervision loss: sample , The feature encoding output by the shared feature extraction network Encoder() is: , ; The cosine similarity of the two is , which reflects the similarity of the two samples in the feature space. The similarity supervision loss adopts binary similarity supervision loss, and its formula is:

[0079] ;

[0080] In the formula, is the similarity supervision loss, M 2 is the total number of training pairs constructed in the second stage. The final joint loss function is:

[0081] ;

[0082] In the formula, L is the total loss; during training, the first stage is set , After initial convergence, the second stage is switched to .

[0083] Through stage training, the model can learn the feature boundary between known classes and abnormal behavior agent samples while retaining the ability to classify known classes, laying the foundation for subsequent abnormal behavior detection.

[0084] Because the actual scene is open, the monitoring method will have an "open class problem". During the operation in the industrial field, there are often device states or abnormal behaviors that have not been seen in the model training stage, such as soft failure, misoperation, electrical interference, etc., which belong to the category of open set recognition (OSR) problems, and traditional classification models are prone to misjudgment of such unknown classes as known classes. Further, a series of monitoring technologies including traditional classifiers, abnormal detection models, clustering recognition methods, such as device abnormal detection methods based on support vector machine (SVM), isolation forest (Isolation Forest), autoencoder (AutoEncoder) and other models, generally have weak identification ability for unknown classes, poor model migration, strong dependence on specific devices, and other problems, which are difficult to meet the intelligent monitoring needs of multi-competent systems across devices and working conditions. However, the present application uses the above-mentioned two-stage training method, with the help of open set recognition technology, to detect unknown classes based on discriminative boundary expansion, contrastive learning, and generative adversarial samples, which show a certain robustness in an open environment.

[0085] In the actual monitoring scene, real-time acquisition of electrical signals of each energy-using device in the energy supply system is performed, and real-time time series electrical signals are output based on a fixed time window. The window length is set to 30-60 seconds to ensure complete coverage of the features of a minimum running period of the device; the sliding step is set to 10 seconds to balance real-time and data redundancy, avoiding monitoring delay. The real-time time series electrical signals are subjected to the above-mentioned preprocessing steps (denoising, outlier rejection, normalization), and the corresponding time domain features, frequency domain features and statistical features are extracted. After the extracted features are arranged according to the feature dimensions and order of the original sample set, they are input into the trained device energy consumption state monitoring model to obtain the prediction probability distribution output by the model.

[0086] The determination of the monitoring result is realized by calculating the sum of the class confidence and monitoring the sample abnormal confidence score. First, in the calculation of the sum of the class confidence, for any sample The classification probability is calculated as:

[0087] ;

[0088] ;

[0089] In the formula, cThe category for known device status.

[0090] Next, calculate the anomaly confidence score. The formula is:

[0091] ;

[0092] The anomaly confidence score reflects the probability that a sample exhibits abnormal behavior. The anomaly confidence score is then compared with a preset threshold. τ Comparison; where the threshold τ By selecting the validation set and AUROC curve, the threshold can be determined. τ Optimization is performed, specifically by comparing different τ The abnormal sample identification rate and normal sample accuracy on the validation set are selected when the AUROC value is maximized. τ The value is generally between 0.6 and 0.8.

[0093] when (Abnormal energy consumption behavior) indicates that the sample The corresponding energy-consuming equipment exhibits abnormal behavior, and the category of the abnormal behavior is output; when When the monitored sample is determined to be in normal behavior, the category of normal behavior is output, and the known category with the highest predicted probability is selected as the equipment status category.

[0094] For monitoring results that are determined to be normal behavior, the status category of the equipment is directly output to provide data support for energy efficiency management of the power supply system, such as load scheduling and energy consumption optimization.

[0095] For monitoring results identified as abnormal behavior, an alarm mechanism is immediately triggered, such as system pop-up notifications, audible and visual alarms, and remote push notifications to maintenance personnel. At the same time, the probability of abnormal behavior, i.e., the abnormal confidence score, and the corresponding abnormal behavior label are output, i.e., the abnormal behavior proxy sample category most similar to the monitored sample features. Additionally, the feature space similarity between the abnormal behavior and the most recently known class is also output, i.e., the cosine similarity between the feature code of the abnormal sample and the feature code of the most recently known class sample, to provide a reference for subsequent manual interpretation or expert analysis of the cause of the anomaly.

[0096] In the model deployment process, to adapt to the edge computing environment in the industrial field, the shared feature extraction network Encoder() can be quantitatively deployed to an embedded edge device (such as an edge gateway or an industrial controller) with a quantization precision of 8 bits or 16 bits. On the premise that the model precision loss is less than 3%, the device computing power and storage occupancy are reduced, and the real-time response speed is improved. To avoid false alarms caused by single data fluctuations, the alarm logic is set to trigger an alarm when the monitoring results of three consecutive sliding windows are all determined to be abnormal behavior. At the same time, a model self-adaptive updating mechanism is set. When the number of abnormal behavior proxy samples, i.e., Open samples, reaches a preset threshold, such as 50 samples per abnormal proxy class, these samples are added to the training set for semi-supervised backfill training to update the model parameters and further improve the model's ability to identify new types of abnormal behavior.

[0097] Embodiment 1;

[0098] To verify the effectiveness of the method of the embodiment, the public load data set PLAID is used as the verification basis. The data set contains voltage and current waveform data of multiple common household and commercial appliances under multiple states, which is suitable for constructing an experimental scene of typical load behavior of multi-energy devices.

[0099] The verification task selects eight common devices in the data set, including air conditioners (AC), energy-saving lamps (CFL), refrigerators (Fridge), hair dryers (Hairdryer), heaters (Heater), incandescent lamps (ILB), computers (Laptop), and vacuum cleaners (Vacuum), which are constructed as known class samples. In addition, three types of devices not involved in training (unknown class, Unknown) are randomly selected, such as fans, microwaves, and washing machines, as representatives of simulated "abnormal energy use behavior". These unseen classes are simulated as unknown devices or abnormal energy use behaviors that may occur in the test phase.

[0100] During the experiment, first, the original sampling signal is preprocessed, including normalization, filtering, sliding window segmentation, and extraction of time domain and frequency domain features to form a unified sample vector. Then, the known class training set and pseudo-unseen class enhanced samples are constructed according to the method proposed in the application, and two-stage training is performed: the first stage identifies the standard known device state, and the second stage introduces pseudo-unseen samples for open set recognition ability enhancement learning.

[0101] In the extraction of time domain and frequency domain features to form a unified sample vector, taking the "off-on" load event of the energy-saving lamp sample as an example: when extracting the corresponding features, the user-side total original voltage, current, and active power signals of the load event section shown in FIG. 2 are first collected, and then preprocessed. Further, the load event in the figure corresponds to the change of the electrical signal , voltage signal recovery , current signal on , current signal off ; voltage of the energy saving lamp according to the event of this load change , current , reactive current The curves are shown in FIG. 1, with units of (V), (A), and (A), respectively. Figure 3

[0102] In the verification phase, the PLAID dataset is divided into a training set and a test set. The training set only contains 8 known types of devices, and the test set contains both known types and unseen types. When identifying the samples in the test set using the method of the present application, most of the unknown types can be correctly identified as "open types", i.e., abnormal energy use behaviors, thereby significantly reducing the risk of misclassification of unknown behaviors as known types.

[0103] The results of the embodiments show that the method has an accuracy of more than 96% in the known device state identification task, and has good identification ability in detecting open class samples, with an identification rate of more than 94%, as shown in FIG. 3. Compared with traditional Softmax classifier, One-Class SVM and Isolation Forest method, the present application significantly improves the detection ability of unknown behaviors, and has lower misjudgment rate, which shows that it has better robustness and generalization ability in real multi-energy supply systems. Figure 4

[0104] Further analysis of the prediction results shows that the "abnormal confidence" score of abnormal energy use behaviors in the method is significantly higher than that of normal class samples, and the model can effectively construct a discrimination boundary and make abnormal alarm judgments. Overall, the monitoring method proposed in the present application has excellent verification effect on the standard load dataset PLAID, and has good migratability and engineering practicability.

[0105] Finally, it should be noted that the above content is only used to illustrate the technical solutions of the present application, and is not a limitation on the protection scope of the present application. Simple modifications or equivalent replacements of the technical solutions of the present application made by those skilled in the art do not deviate from the essence and scope of the technical solutions of the present application.​​

Claims

1. A method for monitoring the energy consumption state of a device based on the electrical signal of a multi-energy complementary energy supply system, comprising the following steps: obtaining the time series electrical signal of each energy-consuming device in the energy supply system; preprocessing the time series electrical signal to obtain a preprocessed electrical signal; extracting multi-dimensional features from the preprocessed electrical signal and constructing a feature vector of the energy-consuming device based on the multi-dimensional features; collecting the feature vectors of each energy-consuming device to obtain an original sample set; applying a perturbation enhancement operation to the original sample set of known energy-consuming states to generate an abnormal behavior proxy sample set; constructing a device energy consumption state monitoring model based on a two-stage model training mechanism; the device energy consumption state monitoring model is an Encoder-Classifier architecture model, which has a shared feature extraction network and a linear classifier; collecting the electrical signal of each energy-consuming device in the energy supply system in real time, and outputting real-time time series electrical signal based on a fixed window; preprocessing the real-time time series electrical signal and inputting it into the device energy consumption state monitoring model after extracting features to obtain a monitoring result; in the two-stage model training, the following steps are included: the first stage training is based on the original sample set for closed-set supervised classification training, and the optimization target is the cross-entropy loss function, the formula is: the multi-dimensional features include time domain features, frequency domain features and statistical features; the perturbation enhancement operation at least includes one of amplitude perturbation, time scale stretching and compression, Gaussian noise perturbation, frequency domain coefficient perturbation or feature channel random masking; further comprising the following steps: the monitoring result includes normal behavior and abnormal behavior; when the monitoring result is normal behavior, output the device state category; when the monitoring result is abnormal behavior, trigger the alarm mechanism and output the abnormal behavior probability and the corresponding abnormal behavior label; the monitoring result further includes the following steps: calculate the class confidence sum of the monitoring sample on the known category, calculate the abnormal confidence score based on the class confidence sum, and compare the abnormal confidence score with a preset threshold; when the abnormal confidence score is higher than the preset threshold, output the monitoring sample as an abnormal behavior category; when the abnormal confidence score is not higher than the preset threshold, determine the monitoring sample as a normal behavior category; the time series electrical signal at least includes one of current, voltage, active power or reactive power; the preprocessing of the time series electrical signal at least includes one of denoising, outlier rejection or normalization. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ； In the formula, For cross-entropy loss, N This represents the number of original samples for the first stage of training. For the first i The true labels of the original samples are encoded using one-hot encoding. For the model to the first i The predicted probability distribution of the original samples; The second stage is to let the model learn the feature difference between the known class samples and the abnormal behavior agent samples, and to construct a training pair with similarity by combining the original sample set and the abnormal behavior agent sample set And a similarity supervision loss is introduced for joint modeling. wherein the joint training set is constructed as ; The label similarity of any two samples , ( ) is defined as follows: ； In the formula, is a sample i has a label similar to that of the sample j has a label similar to that of the sample The definition model encodes the characteristics of the sample and the similarity supervision loss: the sample , The feature encoding output by the shared feature extraction network Encoder() is respectively: , The cosine similarity of the two is The similarity supervision loss adopts binary similarity supervision loss, and its formula is: ； wherein is the similarity supervision loss, M 2 is the total number of training pairs constructed for the second stage; the final joint loss function is: ； In the formula, L Total loss; first phase set during training , Second phase switched after initial convergence .

2. The method for monitoring the state of energy use of a device based on the electrical signal of a multi-energy complementary energy supply system according to claim 1, characterized in that: ​ 3. The method of claim 1, wherein: ​ 4. The method of claim 1, wherein: ​ ​ ​ ​ 5. The method of claim 4, wherein: ​ ​ ​ ​ 6. The method of claim 1, wherein: ​ 7. The method of claim 1 or 6, wherein: ​

Images