A non-intrusive load state monitoring system and method based on multi-domain state fusion features

Abstract

The present application discloses a non-intrusive load state monitoring system and method with multi-domain fusion features. By collecting a number of static data and a number of dynamic data, performing a short-time Fourier transform on the dynamic data, and performing domain and form transformation on the static data, a spectrum multi-dimensional matrix and a multi-dimensional identification matrix are obtained, and the data are input into the static and dynamic feature extractors respectively to obtain potential feature vectors and dynamic feature vectors. The potential feature vectors and dynamic feature vectors are merged, convolved and weighted fused to obtain a fused feature vector matrix. The fused feature vector matrix is ​​input into the student network to obtain output devices and load events, which are finally sent to the cloud computing center. It can be seen that when analyzing load events, static data is introduced and feature extraction is performed on the static data to analyze static variable factors. Taking into account the influence of external static variable factors on potential load events, the accuracy of load event monitoring is guaranteed.

Description

Technical Field

[0001] The present application relates to the technical field of power grid load state assessment, and more specifically, to a non-invasive load state monitoring system and method with multi-domain state fusion features. Background Art

[0002] As living standards improve, the demand for electricity continues to rise. With the establishment of smart cities and smart grids, the importance of electricity data will only grow. In a smart grid environment, companies producing electricity can use acquired electricity-related data to analyze relevant power events and conduct daily planning from power generation to distribution. Therefore, efficiently acquiring electricity-related data and promptly identifying power load events can significantly reduce costs and improve the efficiency of power generation and distribution.

[0003] At present, in the research of load monitoring, most of them only consider the dynamic data characteristics, but fail to consider the impact of external static variable factors on potential load events, resulting in low accuracy of load event monitoring. Summary of the Invention

[0004] In view of the above problems, the present application is proposed to provide a non-intrusive load state monitoring system and method with multi-domain state fusion features to improve the accuracy of load event monitoring.

[0005] In order to achieve the above objectives, the following specific plans are proposed:

[0006] A non-intrusive load state monitoring system with multi-domain state fusion features, comprising a data acquisition unit, a data conversion unit, a feature extraction unit, a feature fusion unit and a network calculation unit;

[0007] The data acquisition unit is used to collect a plurality of static data and a plurality of dynamic data;

[0008] The data conversion unit is used to perform short-time Fourier transform on each dynamic data to obtain a frequency spectrum multidimensional matrix of the dynamic data, and to perform domain and form transformation on each static data to obtain a multidimensional identification matrix of each static data;

[0009] The feature extraction unit is configured to input the multidimensional recognition matrix into a pre-trained static feature extractor to obtain a plurality of potential feature vectors of the static data output by the static feature extractor, and input the spectral multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model;

[0010] The feature fusion unit is used to merge, convolve and weightedly fuse each potential feature vector and each dynamic feature vector to obtain a fused feature vector matrix;

[0011] The network computing unit is used to input the fused feature vector matrix into a pre-trained student network, obtain the device information of several load devices output by the student network and the load events occurring in each load device, and send the device information of each load device and the load events occurring in each load device to the cloud computing center through the communication unit.

[0012] Optionally, the data acquisition unit includes a static data acquisition unit and a dynamic data acquisition unit;

[0013] The static data collection unit is used to collect a number of static data;

[0014] The dynamic data acquisition unit is used to acquire a number of dynamic data.

[0015] Optionally, the static data collection unit collects a number of static data, including:

[0016] The static data acquisition unit acquires a plurality of static data collected by a plurality of static data meters connected to the input port of the power grid.

[0017] Optionally, the dynamic data acquisition unit acquires a number of dynamic data, including:

[0018] The dynamic data unit acquires a plurality of dynamic data collected by a plurality of dynamic data meters connected to the input port of the power grid.

[0019] Optionally, the data conversion unit performs short-time Fourier transform on each dynamic data to obtain a multidimensional matrix of the spectrum of the dynamic data, including:

[0020] The data conversion unit obtains a dynamic data sample set of each dynamic data;

[0021] The data conversion unit determines the mean square error sequence of each dynamic data sample set by calculating the mean square error of the dynamic data sample set;

[0022] The data conversion unit determines the frequency spectrum of each dynamic data using the following formula:

[0023]

[0024] Among them, x' rms (n) is a mean square error reciprocal sequence obtained by performing short-time Fourier transform on the mean square error sequence of the dynamic data sample set of the dynamic data, n is a sequence index, w(·) is a window function, τ is the center position of the window function, and ω is the phase of the window function;

[0025] The data conversion unit calculates the cumulative mean square error of each dynamic data using the following formula:

[0026]

[0027] The data conversion unit determines the frequency spectrum multi-dimensional matrix of each dynamic data using the following formula:

[0028] S(ω,τ)=S n (ω,τ)·sgn(S N )

[0029] where sgn(·) is the sign function.

[0030] Optionally, the data transformation unit performs domain and form transformation on each static data to obtain a multi-dimensional identification matrix of each static data, including:

[0031] The data transformation unit determines the static variable probability density function of each static data using the following formula:

[0032]

[0033] Among them, x is the variable of the static data, is the static variable distribution probability that the static data obeys the standard normal distribution, σ is the standard deviation of the static data, b is the preset static constant of the static data, and N is the total number of samples collected within a fixed time.

[0034] The data conversion unit multiplies the static variable probability density function of each static data by the static variable of the static data to obtain a time series vector of the static data;

[0035] The data conversion unit uses time as an index for matrix splicing, performs matrix splicing on the time series vectors of each static data, and obtains a multi-dimensional identification matrix of each static data.

[0036] Optionally, the dynamic feature extraction model is obtained by training historical dynamic data in a pre-built convolutional network, and the convolutional network is composed of multiple convolutional auto-encoding networks, each of which has the same structure and parallel positions, and is composed of three convolutional layers, three pooling layers and three fully connected layers.

[0037] Optionally, the feature extraction unit inputs the multidimensional recognition matrix into a pre-trained static feature extractor to obtain several potential feature vectors of the static data output by the static feature extractor, including:

[0038] The feature extraction unit determines a normal variable vector and a temporal variable vector in the multidimensional identification matrix;

[0039] The feature extraction unit converts each normal variable vector into a normal latent feature vector using the following formula:

[0040] h c =e c (c)

[0041] Among them, e c (c) A function that extracts the latent low-dimensional features of the normal variable c for the pre-trained static feature extractor;

[0042] The feature extraction unit converts each temporal variable vector into a temporal latent feature vector using the following formula:

[0043] h t =e t (h c ,h t-1 ,x t )

[0044] Among them, e t (h c ,h t-1 ,x t ) is the static feature extractor extracting the temporal variable x t function of the potential low-dimensional features of .

[0045] Optionally, the feature fusion unit merges, convolves, and weighted-fuses each potential feature vector and each dynamic feature vector to obtain a fused feature vector matrix, including:

[0046] The feature fusion unit merges each potential feature vector and each dynamic feature vector through a preset convolutional attention module to obtain a comprehensive feature matrix;

[0047] The feature fusion unit performs average pooling on the comprehensive feature matrix to obtain a first pooling vector;

[0048] The feature fusion unit performs maximum pooling on the comprehensive feature matrix to obtain a second pooling vector;

[0049] The feature fusion unit performs two layers of one-dimensional convolution on the first pooling vector to obtain a first convolution result;

[0050] The feature fusion unit performs two layers of one-dimensional convolution on the second pooling vector to obtain a second convolution result;

[0051] The feature fusion unit activates the sum of the first convolution result and the second convolution result through a sigmoid function to obtain a first weight vector corresponding to the first pooling vector and a second weight vector corresponding to the second pooling vector;

[0052] The feature fusion unit multiplies the first weight vector by the first pooling vector to obtain a first weighted feature vector;

[0053] The feature fusion unit multiplies the second weight vector by the second pooling vector to obtain a second weighted feature vector;

[0054] The feature fusion unit concatenates the first weighted feature vector and the second weighted feature vector to obtain a fused feature vector matrix.

[0055] A non-intrusive load state monitoring method based on multi-domain state fusion features is applied to a non-intrusive load state monitoring system. The method comprises:

[0056] Collect some static data and some dynamic data;

[0057] Perform short-time Fourier transform on each dynamic data to obtain the frequency spectrum multidimensional matrix of the dynamic data, and perform domain and form transformation on each static data to obtain the multidimensional identification matrix of each static data;

[0058] Inputting the multidimensional recognition matrix into a pre-trained static feature extractor to obtain a plurality of potential feature vectors of the static data output by the static feature extractor, and inputting the spectrum multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model;

[0059] Merge, convolve and weightedly fuse each latent eigenvector and each dynamic eigenvector to obtain a fused eigenvector matrix;

[0060] The fused eigenvector matrix is ​​input into a pre-trained student network to obtain the device information of several load devices and the load events occurring in each load device output by the student network. The device information of each load device and the load events occurring in each load device are sent to the cloud computing center through the communication unit in the non-invasive load status monitoring.

[0061] Optionally, the training process of the student network includes:

[0062] Under the constraints of the teacher network loss function, the teacher network is trained through training samples in historical data. The teacher network loss function is:

[0063]

[0064] Where A is the sum of the number of static variables in the historical data and the number of dynamic variables in the historical data, N is the number of training samples, y is the actual value of the training sample, and y i ' is the output value of the teacher network for the training sample during the training process, and σ is the sigmoid function;

[0065] Inputting the test samples in the historical data into the trained teacher network to obtain the soft labels output by the trained teacher network;

[0066] Covering the test sample with the soft label to obtain a soft label test sample;

[0067] The soft label test samples and the preset real label test samples are used as training samples. Under the constraint of the total loss function, the student network is trained. The total loss function is:

[0068]

[0069] Among them, y′ is the soft label, y ～ is the output value of the student network for the true label test sample during training, L s (y,y′) is the first loss function when the true label test sample is used as the training sample, L st (y ～ ,y′) is the second loss function when the soft label test sample is used as a training sample, ω(τ) is the weight ramp-up function, and τ is the preset training value.

[0070] By means of the above technical solution, the present application collects a number of static data and a number of dynamic data through the data acquisition unit, the data conversion unit performs short-time Fourier transform on each dynamic data to obtain a spectral multidimensional matrix of the dynamic data, performs domain and form transformation on each static data to obtain a multidimensional identification matrix of each static data, the feature extraction unit inputs the multidimensional identification matrix into a pre-trained static feature extractor to obtain a number of potential feature vectors of the static data output by the static feature extractor, inputs the spectral multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model, the feature fusion unit merges, convolves and weightedly fuses each potential feature vector and each dynamic feature vector to obtain a fused feature vector matrix, the network computing unit inputs the fused feature vector matrix into a pre-trained student network to obtain device information of several load devices output by the student network and load events occurring on each load device, and sends the device information of each load device and the load events occurring on each load device to the cloud computing center through the communication unit. It can be seen that the non-intrusive load state monitoring with multi-domain fusion features introduces static data when analyzing load events, and extracts features from the static data to analyze static variable factors. It takes into account the impact of external static variable factors on potential load events, thereby ensuring the accuracy of load event monitoring.

[0071] Furthermore, compared with analysis based only on frequency domain features or only on time domain features, the accuracy of device identification is improved by transforming the domain and form of static data. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present application. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0073] Figure 1 This is an architecture diagram of a non-intrusive load status monitoring system with multi-domain state fusion features provided in an embodiment of the present application;

[0074] Figure 2 This is an architecture diagram of another non-intrusive load status monitoring system with multi-domain state fusion features provided in an embodiment of the present application;

[0075] Figure 3 A schematic diagram of a feature data processing method for determining a fused feature vector matrix provided in an embodiment of the present application;

[0076] Figure 4 A schematic diagram of a flow chart of a non-intrusive load status monitoring system implementing a multi-domain state fusion feature for non-intrusive load status monitoring provided in an embodiment of the present application;

[0077] Figure 5 A schematic diagram of a process for training a student network provided in an embodiment of the present application;

[0078] Figure 6 A schematic diagram of the structure of a teacher network trained using knowledge distillation technology provided in an embodiment of the present application;

[0079] Figure 7 A schematic diagram of the structure of a student network trained using knowledge distillation technology provided in an embodiment of the present application. DETAILED DESCRIPTION

[0080] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0081] Figure 1 An optional architecture of the non-intrusive load state monitoring system with multi-domain state fusion features provided in the embodiment of the present application, such as Figure 1As shown, the architecture may include:

[0082] Data acquisition unit 10, data transformation unit 20, feature extraction unit 30, feature fusion unit 40 and network calculation unit 50.

[0083] The data collection unit 10 can be used to collect some static data and some dynamic data.

[0084] Specifically, static data can represent data on variables that have external influences on the power system and loads, with a change rate less than a preset threshold rate, such as light intensity data, temperature data, wind speed data, weather data, etc. Dynamic data can represent data on internal equipment operations that have a change rate greater than the preset threshold rate and can intuitively reflect the operating status and changes of the equipment, such as current, voltage, and power.

[0085] The data conversion unit 20 can be used to perform short-time Fourier transform on each dynamic data to obtain a frequency spectrum multidimensional matrix of the dynamic data, and perform domain and form transformation on each static data to obtain a multidimensional identification matrix of each static data.

[0086] Specifically, the data conversion unit 20 may include a dynamic data conversion unit and a static data conversion unit.

[0087] Among them, the dynamic data conversion unit can be used to perform short-time Fourier transform on each dynamic data to obtain a multidimensional spectrum matrix of the dynamic data, and the static data conversion unit can be used to perform domain and form transformation on each static data to obtain a multidimensional identification matrix of each static data.

[0088] The feature extraction unit 30 can be used to input the multidimensional recognition matrix into a pre-trained static feature extractor to obtain several potential feature vectors of the static data output by the static feature extractor, and input the spectral multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain the dynamic feature vector of the dynamic data output by the dynamic feature extraction model.

[0089] The feature fusion unit 40 may be used to merge, convolve, and weightedly fuse each potential feature vector and each dynamic feature vector to obtain a fused feature vector matrix.

[0090] The network computing unit 50 can be used to input the fused feature vector matrix into a pre-trained student network, obtain the device information of several load devices output by the student network and the load events occurring in each load device, and send the device information of each load device and the load events occurring in each load device to the cloud computing center through the communication unit.

[0091] Specifically, the student network may be a student network trained by a knowledge distillation method.

[0092] The non-intrusive load state monitoring system with multi-domain fusion features provided in this embodiment collects a plurality of static data and a plurality of dynamic data through the data acquisition unit. The data conversion unit performs a short-time Fourier transform on each dynamic data to obtain a spectral multidimensional matrix of the dynamic data. The domain and form conversion is performed on each static data to obtain a multidimensional identification matrix of each static data. The feature extraction unit inputs the multidimensional identification matrix into a pre-trained static feature extractor to obtain a plurality of latent feature vectors of the static data output by the static feature extractor. The spectral multidimensional matrix of each dynamic data is input into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model. The feature fusion unit merges, convolves, and weightedly fuses the latent feature vectors and the dynamic feature vectors to obtain a fused feature vector matrix. The network computing unit inputs the fused feature vector matrix into a pre-trained student network to obtain device information of a plurality of load devices and load events occurring on each load device output by the student network. The device information of each load device and the load events occurring on each load device are sent to a cloud computing center via a communication unit. It can be seen that the non-intrusive load status monitoring system with multi-domain fusion features introduces static data when analyzing load events, and extracts features from the static data to analyze static variable factors. It takes into account the impact of external static variable factors on potential load events, thereby ensuring the accuracy of load event monitoring.

[0093] Furthermore, compared with analysis based only on frequency domain features or only on time domain features, the accuracy of device identification is improved by transforming the domain and form of static data.

[0094] In some embodiments of the present application, the data acquisition unit 10 mentioned in the above embodiment is introduced. The data acquisition unit 10 may include a static data acquisition unit and a dynamic data acquisition unit.

[0095] The static data collection unit may be used to collect a number of static data.

[0096] Specifically, the process of the static data collection unit collecting a number of static data may include:

[0097] The static data acquisition unit acquires a plurality of static data collected by a plurality of static data meters connected to the input port of the power grid.

[0098] Specifically, the static data acquisition unit can be connected to multiple static data meters, and these static data meters may not be placed in the non-intrusive load status monitoring system. The static data acquisition unit sends an instruction to each static data meter to request static data. After each static data meter responds to the instruction, the static data acquisition unit can receive the response instruction sent by each static data meter and convert it into the actual measured value of the variable according to the corresponding communication protocol of each static data meter.

[0099] At the same time, the dynamic data acquisition unit can be used to acquire a number of dynamic data.

[0100] Specifically, the process of the dynamic data acquisition unit acquiring a plurality of dynamic data may include:

[0101] The dynamic data unit acquires a plurality of dynamic data collected by a plurality of dynamic data meters connected to the input port of the power grid.

[0102] Specifically, the dynamic data acquisition unit can be connected to multiple dynamic data instruments, and these dynamic data instruments may not be placed in the non-intrusive load state monitoring system. The dynamic data acquisition unit sends an instruction to each dynamic data instrument to request dynamic data. After each dynamic data instrument responds to the instruction, the dynamic data acquisition unit can receive a response instruction sent by each dynamic data instrument and convert it into an actual measured value of the variable according to the corresponding communication protocol of each dynamic data instrument.

[0103] Based on this, Figure 1 Based on the architecture shown, Figure 2 The architecture of another non-intrusive load status monitoring system with multi-domain state fusion features provided by an embodiment of the present application is shown.

[0104] The non-invasive load status monitoring with multi-domain fusion characteristics provided in this embodiment, in which the static data acquisition unit collects static data by connecting to multiple external static data instruments, and the dynamic data acquisition unit collects dynamic data by connecting to multiple external dynamic data instruments, can reduce the load and maintenance difficulty of non-invasive load status monitoring.

[0105] In some embodiments of the present application, the process of the data conversion unit 20 performing a short-time Fourier transform on each dynamic data to obtain a multidimensional matrix of the spectrum of the dynamic data mentioned in the above embodiment is introduced. This process may include:

[0106] S1. The data conversion unit 20 obtains a dynamic data sample set of each dynamic data.

[0107] Specifically, the data transformation unit 20 may apply a sliding window method to each dynamic data to obtain a dynamic data sample set for each dynamic data.

[0108] S2. The data transformation unit 20 determines a mean square error sequence of each dynamic data sample set by calculating the mean square error of the dynamic data sample set.

[0109] S3. The data conversion unit 20 determines the frequency spectrum of each dynamic data using the following formula:

[0110]

[0111] Among them, x' rms (n) is the mean square error reciprocal sequence sequence obtained by performing short-time Fourier transform on the mean square error sequence of the dynamic data sample set of the dynamic data, n is the sequence index, w(·) is the window function, τ is the center position of the window function, and ω is the phase of the window function.

[0112] S4. The data conversion unit 20 calculates the cumulative mean square error of each dynamic data using the following formula:

[0113]

[0114] S5. The data conversion unit 20 determines the frequency spectrum multidimensional matrix of each dynamic data using the following formula:

[0115] S(ω,τ)=S n (ω,τ)·sgn(S N )

[0116] where sgn(·) is the sign function.

[0117] It is understandable that the frequency spectra of dynamic data of completely opposite events such as turning on and off may be the same. Therefore, the situation where the frequency spectra of opposite events are the same can be avoided by calculating the cumulative mean square error and combining it with the sign function.

[0118] The non-intrusive load state monitoring with multi-domain state fusion features provided in this embodiment processes dynamic data through the data conversion unit 20 to obtain a spectrum multidimensional matrix of each dynamic data, so that the feature extraction unit 30 extracts dynamic feature information from the spectrum multidimensional matrix.

[0119] In some embodiments of the present application, the process of the data transformation unit 20 performing domain and form transformation on each static data to obtain a multidimensional identification matrix of each static data mentioned in the above embodiment is introduced. This process may include:

[0120] S1, the data conversion unit 20 uses the following formula to determine the static variable probability density function of each static data:

[0121]

[0122] Among them, x is the variable of the static data, is the static variable distribution probability that the static data obeys the standard normal distribution, σ is the standard deviation of the static data, b is the preset static constant of the static data, and N is the total number of samples collected within a fixed time for the static data.

[0123] Specifically, a region R is taken within the value range of the static variable of the static data. When the width of the region is less than a preset width threshold, the probability p of the data falling into the region R can be expressed by the following formula:

[0124] p=f(x)h

[0125] because When the region width is l,

[0126] Furthermore, the standard normal distribution kernel function can be used to approximate x i The static variable distribution probability of a point:

[0127]

[0128] Then the number of samples k distributed in region R in the total samples can be estimated by the following formula:

[0129]

[0130] So we can get the static variable probability density function of each static data as:

[0131]

[0132] S2. The data conversion unit 20 multiplies the static variable probability density function of each static data by the static variable of the static data to obtain a time series vector of the static data.

[0133] S3. The data transformation unit 20 uses time as an index for matrix concatenation to concatenate the time series vectors of each static data into a matrix to obtain a multi-dimensional identification matrix of each static data.

[0134] The non-intrusive load state monitoring with multi-domain state fusion features provided in this embodiment processes dynamic data through the data conversion unit 20 to obtain a spectrum multidimensional matrix of each dynamic data, so that the feature extraction unit 30 extracts dynamic feature information from the spectrum multidimensional matrix.

[0135] In some embodiments of the present application, the dynamic feature extraction model mentioned in the above embodiments is introduced. Specifically, the dynamic feature extraction model is obtained by training historical dynamic data in a pre-built convolutional network.

[0136] The convolutional network can be composed of multiple convolutional auto-encoding networks, each of which has the same structure and is positioned in parallel, and is composed of three convolutional layers, three pooling layers and three fully connected layers.

[0137] By inputting a two-dimensional matrix of size N*N into the trained dynamic feature extraction model, the dynamic feature extraction model can output a feature vector of length N.

[0138] Based on this, when the size of the spectrum multidimensional matrix of each dynamic data is N*N, the spectrum multidimensional matrix of each dynamic data is input into the dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data with a length of N.

[0139] In some embodiments of the present application, the process mentioned in the above embodiment, in which the feature extraction unit 30 inputs the multidimensional recognition matrix into a pre-trained static feature extractor to obtain several potential feature vectors of the static data output by the static feature extractor, is introduced. This process may include:

[0140] S1. The feature extraction unit determines a normal variable vector and a temporal variable vector in the multidimensional identification matrix.

[0141] Specifically, the normal variable vector is a variable vector that does not exist in a time series sequence, and the temporal variable vector is a variable vector that exists in a time series sequence. Static variables can be represented by a vector, such as (x1,…,x n ,c1,…,c m ). X is the vector space of temporal variables, C is the vector space of normal variables, and (X, C) is a random vector belonging to the vector space X, C.

[0142] It is understandable that for a static variable vector, it must obey a certain target distribution p(C 1:m ,X 1:n ), a distribution can be trained through the generator learning in PD-GAN to approximate the target distribution.

[0143] S2. The feature extraction unit converts each normal variable vector into a normal potential feature vector using the following formula:

[0144] h c =e c (c)

[0145] Among them, e c (c) Function for extracting latent low-dimensional features of the normal variable c by the pre-trained static feature extractor.

[0146] S3. The feature extraction unit converts each temporal variable vector into a temporal potential feature vector using the following formula:

[0147] h t =e t (h c ,h t-1 ,x t )

[0148] Among them, e t (h c ,h t-1 ,x t ) is the static feature extractor extracting the temporal variable x t function of the potential low-dimensional features of .

[0149] It is understandable that during the training process of the static feature extractor, the potential low-dimensional feature quantities are inversely mapped to verify and correct the extraction results, thereby ensuring that the extracted features are the true features of the static data. Therefore, the loss function during the training process of the static feature extractor is:

[0150]

[0151] in, is the normal variable obtained by inverse mapping, is the temporal variable obtained by inverse mapping.

[0152] Furthermore, the result of the inverse mapping is input to the discriminator, so the discriminator's discriminant process can be expressed as follows:

[0153]

[0154] Among them, d c (·) is the normal discriminant function when the discriminator discriminates the normal potential low-dimensional feature quantity, y c is the discrimination result of the normal discriminant function, d x (·) is the temporal discriminant function when the discriminator discriminates the temporal potential low-dimensional feature quantity, y t is the discrimination result of the temporal discriminant function.

[0155] Then, the generator and the discriminator will use the following two loss functions during the game:

[0156]

[0157]

[0158] in, To train and learn normal variable features, and They are the results of the discriminator distinguishing the normal low-dimensional feature vector and the temporal low-dimensional feature vector when a random vector is input to the generator, For training and learning temporal variable features, g x (·) is the function of the generator to generate temporal features, z t is the input random temporal data vector.

[0159] In some embodiments of the present application, the process of the feature fusion unit 40 merging, convolving, and weighted fusion of the potential feature vectors and the dynamic feature vectors mentioned in the above embodiment to obtain a fused feature vector matrix is ​​introduced. This process may include:

[0160] S1. The feature fusion unit 40 merges each potential feature vector and each dynamic feature vector through a preset convolutional attention module to obtain a comprehensive feature matrix.

[0161] Specifically, each latent feature vector and each dynamic feature vector can be merged through the concat layer in the convolutional attention block.

[0162] S2. The feature fusion unit 40 performs average pooling on the comprehensive feature matrix to obtain a first pooling vector.

[0163] S3. The feature fusion unit 40 performs maximum pooling on the comprehensive feature matrix to obtain a second pooling vector.

[0164] S4. The feature fusion unit 40 performs two layers of one-dimensional convolution on the first pooling vector to obtain a first convolution result.

[0165] S5. The feature fusion unit 40 performs two layers of one-dimensional convolution on the second pooling vector to obtain a second convolution result.

[0166] S6. The feature fusion unit 40 activates the sum of the first convolution result and the second convolution result through a sigmoid function to obtain a first weight vector corresponding to the first pooling vector and a second weight vector corresponding to the second pooling vector.

[0167] S7. The feature fusion unit 40 multiplies the first weight vector by the first pooling vector to obtain a first weighted feature vector.

[0168] S8. The feature fusion unit 40 multiplies the second weight vector by the second pooling vector to obtain a second weighted feature vector.

[0169] S9. The feature fusion unit 40 concatenates the first weighted feature vector and the second weighted feature vector to obtain a fused feature vector matrix.

[0170] The process of obtaining the fused feature vector matrix by the method implemented by S1-S9 in this embodiment can be referred to Figure 3 shown.

[0171] based on Figure 1 The architecture shown, Figure 4 A flow chart of a non-intrusive load state monitoring method for realizing multi-domain state fusion features provided by an embodiment of the present application is shown. Figure 4 , the process may include:

[0172] Step S110: Collect some static data and some dynamic data.

[0173] Step S120 , performing short-time Fourier transform on each dynamic data to obtain a frequency spectrum multidimensional matrix of the dynamic data, and performing domain and form transform on each static data to obtain a multidimensional identification matrix of each static data.

[0174] Step S130: input the multidimensional recognition matrix into a pre-trained static feature extractor to obtain a number of potential feature vectors of the static data output by the static feature extractor, and input the spectral multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain the dynamic feature vector of the dynamic data output by the dynamic feature extraction model.

[0175] Step S140: Merge, convolve, and weighted fuse each latent eigenvector and each dynamic eigenvector to obtain a fused eigenvector matrix.

[0176] Step S150: input the fused eigenvector matrix into a pre-trained student network to obtain the device information of several load devices and the load events occurring in each load device output by the student network, and send the device information of each load device and the load events occurring in each load device to the cloud computing center through the communication unit in the non-invasive load status monitoring system.

[0177] The present embodiment provides a non-intrusive load state monitoring method using multi-domain fusion features. The data acquisition unit collects a plurality of static data and a plurality of dynamic data. The data conversion unit performs a short-time Fourier transform on each dynamic data to obtain a spectral multidimensional matrix of the dynamic data. The domain and form conversion is performed on each static data to obtain a multidimensional identification matrix of each static data. The feature extraction unit inputs the multidimensional identification matrix into a pre-trained static feature extractor to obtain a plurality of latent feature vectors of the static data output by the static feature extractor. The spectral multidimensional matrix of each dynamic data is input into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model. The feature fusion unit merges, convolves, and weightedly fuses the latent feature vectors and the dynamic feature vectors to obtain a fused feature vector matrix. The network computing unit inputs the fused feature vector matrix into a pre-trained student network to obtain device information of a plurality of load devices and load events occurring on each load device output by the student network. The device information of each load device and the load events occurring on each load device are sent to a cloud computing center via a communication unit. It can be seen that the non-intrusive load status monitoring system introduces static data when analyzing load events and extracts features from the static data to analyze static variable factors. It takes into account the impact of external static variable factors on potential load events and ensures the accuracy of load event monitoring.

[0178] Furthermore, compared with analysis based only on frequency domain features or only on time domain features, the accuracy of device identification is improved by transforming the domain and form of static data.

[0179] In some embodiments of the present application, the training process of the student network mentioned in the above embodiment is introduced, combined with Figure 5 , the process may include:

[0180] Step S210: Under the constraints of the teacher network loss function, the teacher network is trained using training samples in the historical data.

[0181] Specifically, the teacher network loss function is:

[0182]

[0183] Where A is the sum of the number of static variables in the historical data and the number of dynamic variables in the historical data, N is the number of training samples, y is the actual value of the training sample, and y i ' is the output value of the teacher network for the training sample during the training process, and σ is the sigmoid function.

[0184] like Figure 6As shown, the teacher network can use knowledge distillation technology during training.

[0185] Step S220: Input the test samples in the historical data into the trained teacher network to obtain the soft labels output by the trained teacher network.

[0186] Step S230: Cover the test sample with the soft label to obtain a soft label test sample.

[0187] Step S240: Using the soft label test sample and the preset real label test sample as training samples, under the constraint of the total loss function, train a student network.

[0188] Specifically, the total loss function is:

[0189] L=L s (y,y′)+ω(τ)*L st (y ～ ,y′)

[0190] Among them, y′ is the soft tag, y ～ is the output value of the student network for the true label test sample during training, L s (y,y′) is the first loss function when the true label test sample is used as the training sample, L st (y ～ ,y′) is the second loss function when the soft label test sample is used as a training sample, ω(τ) is the weight ramp-up function, and τ is the preset training value.

[0191] like Figure 7 As shown, the student network can use knowledge distillation technology during the training process.

[0192] The non-intrusive load state monitoring method with multi-domain state fusion features provided in this embodiment combines the pre-trained teacher network in the process of training the student network, and the training process of the student network and the teacher network uses the knowledge distillation method to perform network compression, which reduces the model size while further reducing the amount of calculation and the computing pressure of the supporting equipment, making it possible to run on edge embedded devices.

[0193] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0194] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referenced to each other.

[0195] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-intrusive load state monitoring system with multi-domain state fusion features, characterized in that: It includes a data acquisition unit, a data conversion unit, a feature extraction unit, a feature fusion unit and a network calculation unit; The data acquisition unit is used to collect a plurality of static data and a plurality of dynamic data; The data conversion unit is used to perform short-time Fourier transform on each dynamic data to obtain a frequency spectrum multidimensional matrix of the dynamic data, and to perform domain and form transformation on each static data to obtain a multidimensional identification matrix of each static data; The feature extraction unit is configured to input the multidimensional recognition matrix into a pre-trained static feature extractor to obtain a plurality of potential feature vectors of the static data output by the static feature extractor, and input the spectral multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model; The feature fusion unit is used to merge, convolve and weightedly fuse each potential feature vector and each dynamic feature vector to obtain a fused feature vector matrix; The network computing unit is configured to input the fused eigenvector matrix into a pre-trained student network, obtain device information of a plurality of load devices and a load event occurring on each load device output by the student network, and transmit the device information of each load device and the load event occurring on each load device to a cloud computing center via a communication unit; The data conversion unit performs short-time Fourier transform on each dynamic data to obtain a frequency spectrum multidimensional matrix of the dynamic data, including: The data conversion unit obtains a dynamic data sample set of each dynamic data; The data conversion unit determines the mean square error sequence of each dynamic data sample set by calculating the mean square error of the dynamic data sample set; The data conversion unit determines the frequency spectrum of each dynamic data using the following formula: ； in, is the mean square error reciprocal sequence obtained by performing short-time Fourier transform on the mean square error sequence of the dynamic data sample set, n is the sequence index, is the window function, is the center position of the window function, is the phase of the window function; The data conversion unit calculates the cumulative mean square error of each dynamic data using the following formula: ； The data conversion unit determines the frequency spectrum multi-dimensional matrix of each dynamic data using the following formula: ； in, is a symbolic function; The data transformation unit performs domain and form transformation on each static data to obtain a multi-dimensional identification matrix of each static data, including: The data transformation unit determines the static variable probability density function of each static data using the following formula: ； Among them, x is the variable of the static data, is the static variable distribution probability that the static data obeys the standard normal distribution, , is the standard deviation of the static data, is the preset static constant of the static data, and N is the total number of samples of the static data collected within a fixed time; The data conversion unit multiplies the static variable probability density function of each static data by the static variable of the static data to obtain a time series vector of the static data; The data conversion unit uses time as an index for matrix splicing, performs matrix splicing on the time series vectors of each static data, and obtains a multi-dimensional identification matrix of each static data.

2. The system according to claim 1, wherein: The data acquisition unit includes a static data acquisition unit and a dynamic data acquisition unit; The static data collection unit is used to collect a number of static data; The dynamic data acquisition unit is used to acquire a number of dynamic data.

3. The system according to claim 2, characterized in that The static data collection unit collects a number of static data, including: The static data acquisition unit acquires a plurality of static data collected by a plurality of static data meters connected to the input port of the power grid.

4. The system according to claim 2, wherein: The dynamic data acquisition unit acquires a number of dynamic data, including: The dynamic data unit acquires a plurality of dynamic data collected by a plurality of dynamic data meters connected to the input port of the power grid.

5. The system according to claim 1, wherein: The dynamic feature extraction model is obtained by training historical dynamic data in a pre-built convolutional network. The convolutional network is composed of multiple convolutional autoencoder networks. Each convolutional autoencoder network has the same structure and is positioned in parallel. They are composed of three convolutional layers, three pooling layers and three fully connected layers.

6. The system according to claim 1, wherein: The feature extraction unit inputs the multidimensional recognition matrix into a pre-trained static feature extractor to obtain several potential feature vectors of the static data output by the static feature extractor, including: The feature extraction unit determines a normal variable vector and a temporal variable vector in the multidimensional identification matrix; The feature extraction unit converts each normal variable vector into a normal latent feature vector using the following formula: ； in, A function that extracts latent low-dimensional features of the normal variable c for the pre-trained static feature extractor; The feature extraction unit converts each temporal variable vector into a temporal latent feature vector using the following formula: ； in, Extract temporal variables for the static feature extractor function of the potential low-dimensional features of .

7. The system according to claim 1, wherein: The feature fusion unit merges, convolves and weights each potential feature vector and each dynamic feature vector to obtain a fused feature vector matrix, including: The feature fusion unit merges each potential feature vector and each dynamic feature vector through a preset convolutional attention module to obtain a comprehensive feature matrix; The feature fusion unit performs average pooling on the comprehensive feature matrix to obtain a first pooling vector; The feature fusion unit performs maximum pooling on the comprehensive feature matrix to obtain a second pooling vector; The feature fusion unit performs two layers of one-dimensional convolution on the first pooling vector to obtain a first convolution result; The feature fusion unit performs two layers of one-dimensional convolution on the second pooling vector to obtain a second convolution result; The feature fusion unit activates the sum of the first convolution result and the second convolution result through a sigmoid function to obtain a first weight vector corresponding to the first pooling vector and a second weight vector corresponding to the second pooling vector; The feature fusion unit multiplies the first weight vector by the first pooling vector to obtain a first weighted feature vector; The feature fusion unit multiplies the second weight vector by the second pooling vector to obtain a second weighted feature vector; The feature fusion unit concatenates the first weighted feature vector and the second weighted feature vector to obtain a fused feature vector matrix.

8. A non-intrusive load state monitoring method based on multi-domain state fusion features, characterized in that: The non-intrusive load state monitoring system using the multi-domain state fusion feature according to claim 1 comprises: Collect some static data and some dynamic data; Perform short-time Fourier transform on each dynamic data to obtain the frequency spectrum multidimensional matrix of the dynamic data, and perform domain and form transformation on each static data to obtain the multidimensional identification matrix of each static data; Inputting the multidimensional recognition matrix into a pre-trained static feature extractor to obtain a plurality of potential feature vectors of the static data output by the static feature extractor, and inputting the spectrum multidimensional matrix of each dynamic data into a pre-trained dynamic feature extraction model to obtain a dynamic feature vector of the dynamic data output by the dynamic feature extraction model; Merge, convolve and weightedly fuse each latent eigenvector and each dynamic eigenvector to obtain a fused eigenvector matrix; The fused eigenvector matrix is ​​input into a pre-trained student network to obtain the device information of several load devices and the load events occurring in each load device output by the student network. The device information of each load device and the load events occurring in each load device are sent to the cloud computing center through the communication unit in the non-invasive load status monitoring system.

9. The method according to claim 8, characterized in that The training process of the student network includes: Under the constraints of the teacher network loss function, the teacher network is trained through training samples in historical data. The teacher network loss function is: ； Wherein, A is the sum of the number of static variables in the historical data and the number of dynamic variables in the historical data, B is the number of the training samples, and y is the actual value of the training sample. is the output value of the teacher network for the training sample during the training process, is the sigmoid function; Inputting the test samples in the historical data into the trained teacher network to obtain the soft labels output by the trained teacher network; Covering the test sample with the soft label to obtain a soft label test sample; The soft label test samples and the preset real label test samples are used as training samples. Under the constraint of the total loss function, the student network is trained. The total loss function is: ； in, is the soft tag, is the output value of the student network for the true label test sample during training, is the first loss function when the true label test sample is used as the training sample, is the second loss function when the soft label test sample is used as the training sample, is the weight ramp-up function, The default training value.

Images