Multi-task non-intrusive load detection method and system based on feature fusion

By using time-frequency feature fusion and a lightweight network structure, the problems of single load data features and limited edge computing resources in non-intrusive load detection are solved, and efficient multi-device load identification and power grid management are achieved.

CN118981676AActive Publication Date: 2024-11-19NANJING TECH UNIV

Patent Information

Application Number
CN202410993882.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-19
Estimated Expiration
2044-07-23

AI Technical Summary

Technical Problem

In non-invasive load detection, the single feature of load data leads to low recognition accuracy, and the limited edge computing resources result in a large amount of computation, making it difficult to meet the recognition needs of multiple devices operating.

Method used

A data processing method that integrates time-frequency features is adopted, combined with adaptive multi-window sliding bilateral CUSUM event detection, data dimensionality reduction, and a lightweight multi-task recognition network, including a Light Block module. Through feature fusion and a lightweight network structure, the computational load is reduced and the recognition accuracy is improved.

Benefits of technology

It improves the accuracy of load identification and the response speed of edge computing devices, reduces the demand for computing resources, and enables efficient multi-device load identification and power grid management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981676A_ABST
    Figure CN118981676A_ABST
Patent Text Reader

Abstract

The invention provides a multi-task non-intrusive load detection method and system based on feature fusion. The system is mainly composed of an edge computing end and a central control end. The method comprises the following steps: firstly, carrying out event detection on original data, collecting steady-state current and voltage data of equipment, carrying out downsampling on data needing to be processed, and calculating a p-norm distance for the current data in edge computing equipment to form current distance matrix information; and then, the amplitude spectrum after the current short-time Fourier transform and the amplitude spectrum after the voltage Fourier transform are respectively used as three-channel input of a matrix. And then inputting the feature image into a built lightweight multi-task network model to carry out load identification. According to the method, the problem of insufficient single charge characteristic is solved by fusing the time domain and frequency domain characteristics, the problem of computing power cost of computing equipment is solved by providing a lightweight multi-task network, and the detection accuracy is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of non-intrusive load detection, and in particular to a multi-task non-intrusive load detection method and system based on feature fusion. Background Art

[0002] With the progress of society and the continuous development of science and technology, sustainable energy utilization and environmental protection have become the focus of global attention. By accurately monitoring and diagnosing the energy consumption of each electrical appliance in the power system, users can gain a deep understanding of energy consumption patterns, identify equipment with high energy consumption and low efficiency, and take corresponding measures to adjust them, thereby achieving effective utilization and conservation of energy. The emergence of non-intrusive load detection technology provides new support for achieving energy conservation and emission reduction. On the other hand, with the continuous growth of electricity demand, the safe operation of the power grid and the stability of power supply have also become urgent issues to be solved. Traditional power demand management methods have been unable to meet the growing power demand. Therefore, it is particularly important to improve the power demand side management capabilities of the power grid. Non-intrusive load detection technology provides new ideas and means for power grid management. By accurately monitoring and analyzing users' electricity consumption behaviors and demand patterns, it can help power companies understand users' electricity needs more accurately, reasonably allocate power resources, and optimize power grid load distribution, thereby improving the safety and stability of the power grid.

[0003] Non-intrusive load identification is an important goal in the field of non-intrusive load monitoring. Due to the wide variety of electrical appliances in the power system and their complex working conditions, single feature information cannot meet the needs of accurate identification of various electrical appliances. How to improve the accuracy and reliability of identification during the operation of single and multiple devices has become a problem that needs to be solved. However, non-intrusive load identification usually relies on neural networks to learn a large amount of load data for training and reasoning, which requires a lot of computing resources and computing power support. This brings some challenges to edge deployment. It is necessary to explore how to reduce the amount of data calculation and improve the system response speed in non-intrusive load detection systems to adapt to the computing resource limitations of edge devices. Summary of the invention

[0004] Technical problem: In view of the problems described in the background above, the present invention proposes a data processing method that integrates time-frequency features to solve the problem of single load data features. In view of the problem of limited resources in edge computing, a lightweight multi-task recognition network is proposed. This solves the problem that the load identification accuracy is not high and the computing power requirement is too high when multiple devices are running in edge deployment.

[0005] Technical solution:

[0006] A multi-task non-intrusive load detection method and system based on feature fusion includes the following steps:

[0007] Step S100, collecting steady-state current and voltage data of a single device and multiple devices when running through a data acquisition module;

[0008] Step S200, in the data processing module, firstly, the distance matrix of the calculated current data, the short-time Fourier transform matrix of the current and voltage data, and the features are integrated into a three-dimensional matrix, and secondly, the load data information of each device is calculated;

[0009] Step S300: In the network identification module, the characteristic data is identified to obtain the specific device name of the data;

[0010] Step S400, uploading the acquired device names and power usage of each device to the central control terminal through the data transmission module;

[0011] The step S100 includes the first aspect of the present invention and provides a steady-state data acquisition method.

[0012] Step S101, for the original current data, an adaptive multi-window sliding bilateral CUSUM is used to perform event detection. Specifically, the following steps are included:

[0013] 1. Define sliding windows: Define two sliding windows: short window N s and long window N l ;

[0014] 2. Calculate the mean and standard deviation: At each time point t, calculate the mean and standard deviation of the current signal in the short window and the long window;

[0015] 3. Dynamically adjust the offset and threshold: Dynamically adjust the offset k of CUSUM detection based on the statistics in the sliding window t and threshold h t ;

[0016] 4. Bilateral CUSUM detection: Apply the bilateral CUSUM algorithm to detect the initial change point in the current signal;

[0017] 5. Introduce detection window: further verify the initially detected change points to reduce false positives;

[0018] Step S102, extract the steady-state current and voltage data when a single device and multiple devices are running, and use the voltage zero-crossing sampling method to extract the steady-state current and voltage data. By calculating and screening the zero-crossing points that meet the required voltage, the current data between the two zero-crossing points when a single device and multiple devices are running are then collected as the steady-state periodic current signal.

[0019] The step S200 includes the second aspect of the present invention, and provides a method for processing data dimension reduction and time-frequency domain feature fusion:

[0020] Step S201: for the collected high-frequency current steady-state signal, X = {x1, x2, ..., x n}, the segmented aggregation approximation method is used to reduce the data dimension. Furthermore, in order to maintain the uniformity and integrity of the data as much as possible, the dimension reduction step can be divided into two cases:

[0021] First, when the input sequence can divide the length m that needs to be reduced, the current sequence X will be divided into equal parts subsequences, calculate the average value of each subsequence, and form all the calculated average values ​​into the reduced dimension data X p ;

[0022] Secondly, when the length of the input sequence cannot be divided into the length m of the dimension reduction, the input sequence needs to be expanded. The expanded sequence is X′=[x1,x2,…,x n , x1, x2, ..., x n , ...], repeat until the sequence length can divide the length of the subsequence m. According to the first step, the expanded sequence is divided into subsequences, and the average value of each subsequence is calculated. All the calculated average values ​​are combined into the reduced dimension data X p ;

[0023] Step S202, constructing a current p-norm distance matrix, first repeating the dimension-reduced data m times along the column direction to form a new m×m matrix X repeat , and transpose it with the matrix The result of the difference operation is recorded as diff. The calculation formula of the p-norm distance matrix D1 of the current is as follows:

[0024]

[0025] in, The i-th row of the matrix diff represents the power operation result of the difference between the i-th data point and all other data points;

[0026] Step S203, construct the matrix after the short-time Fourier transform of the current, and transform the collected original signal X={x1, x2, ..., x n}, without data dimensionality reduction, directly through the short-time Fourier transform function, and then return an m×m current short-time Fourier amplitude spectrum matrix D2;

[0027] The short-time Fourier amplitude spectrum matrix D3 transformation method of voltage is the same as that of current, so I will not elaborate on it;

[0028] Furthermore, the matrices D1, D2, and D3 are fused into a three-dimensional matrix D using a method similar to the three-channel processing of the image R, G, and B, and used as the input of the network recognition module;

[0029] Step S204, for the collected voltage and current signals, real-time calculation of load data information within the collection period, such as the use time of each device in the central control system, the power of the device, the use time, the use frequency and other information;

[0030] The step S300 includes the third aspect of the present invention, providing a lightweight multi-task recognition network for the field of non-intrusive load monitoring:

[0031] Furthermore, the present invention proposes a new lightweight network module Light Block;

[0032] Furthermore, each Light Block contains two Light modules;

[0033] Specifically, the layered convolution in the Light module replaces the 2D convolution. The first layer uses depthwise separable convolution to extract features, and the second layer uses layered convolution to expand the dimension. The feature map is added to the original feature result of the depthwise separable convolution and then output. The input of the first module and the output of the second module form a residual structure to form a complete Light Block.

[0034] Furthermore, the structure of the inference network is as follows: the data is input into a Conv2D convolution module, then enters four Light Block modules in sequence, then enters the average pooling layer, and then passes through a Conv2D convolution module, and then branch one passes through a Linear layer to output recognition result one, and branch two passes through a Linear layer to output recognition result two. Recognition result one outputs a binary classification result of a single device or multiple devices, and recognition result two outputs the specific name of the device;

[0035] In a fourth aspect of the present invention, a multi-task non-intrusive load detection system based on feature fusion is proposed, comprising:

[0036] Edge computing end and central control end, where the edge computing end includes:

[0037] Data acquisition module, used for collecting load voltage and current data;

[0038] A data storage module, used to temporarily save the collected data to the edge computing end;

[0039] Data preprocessing module, used to process the collected load data, including preprocessing such as data dimension reduction, and calculation and processing of equipment data information;

[0040] The network identification module is used to perform deep learning network inference on the data and identify the device;

[0041] Communication module, used for communication between power-consuming devices and edge computing terminals, and between edge computing terminals and central control terminals;

[0042] The central control end is responsible for unified management and further processing of data of all edge devices, as well as sending generated strategies and solutions to edge devices to guide their operations;

[0043] The edge computing system will be installed in the secondary distribution box in the factory production environment, so that the power consumption of the power boxes in each area of ​​the project can be displayed one by one on the large screen of the central control end;

[0044] Compared with the above technical solutions, the present invention has the following advantages:

[0045] 1. Provide an adaptive multi-window sliding bilateral CUSUM event detection method for non-intrusive load detection. The short window is used to capture rapid changes in the short term, while the long window is used to capture slow changes in the long term. By combining the two windows, the sensitivity and accuracy of change point detection can be improved. By dynamically adjusting the detection offset and threshold, it can sensitively capture the sudden change in the current signal. After the event is initially detected, the change point is further confirmed by creating a new detection window and calculating whether the mean value significantly deviates from the historical mean, which significantly improves the sensitivity and accuracy of event detection.

[0046] 2. Provide a data preprocessing method for multi-feature fusion in the time and frequency domains in the field of non-intrusive load detection. Use the p-norm distance matrix to extract the current time domain features, and use the short-time Fourier transform to extract the current and voltage frequency domain features. Without increasing the size of the data dimension, it largely solves the problem of low accuracy of using a single feature for network identification.

[0047] 3. Provide a lightweight multi-task recognition network for the field of non-intrusive load detection. The main part uses Light Block, which makes extensive use of layer-by-layer convolution, point convolution and residual structure, greatly reducing the parameter size of the network model and the amount of floating-point operations for calculation and reasoning. In addition, the multi-task recognition network can better identify the accuracy of multi-device runtime recognition;

[0048] 4. Provide a multi-task non-intrusive load detection system based on feature fusion, combining the central control system and the edge computing end. The edge computing end is close to the data source and can quickly process data and respond, reducing the delay of data transmission. The central control system is responsible for the collection and comprehensive analysis of global data, and can provide load detection, remote supervision and optimization strategies from a global perspective, improving the overall efficiency and power safety of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flowchart of a multi-task non-intrusive load detection method and system based on feature fusion provided in an embodiment of the present invention;

[0050] Figure 2 A structural diagram of a multi-task non-intrusive load detection system based on feature fusion provided in an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of a Light Block module in a lightweight multi-task non-intrusive load network according to an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of a Light module in a lightweight multi-task non-intrusive load network in an embodiment of the present invention;

[0053] Figure 5 Schematic diagram of a deeply separable DWConv module in a lightweight multi-task non-intrusive load network in an embodiment of the present invention;

[0054] Figure 6 A schematic diagram of a lightweight multi-task non-intrusive load network structure in an embodiment of the present invention; DETAILED DESCRIPTION

[0055] The technical solutions in the examples of the present invention are described in more detail below in conjunction with the drawings in the embodiments of the present invention. It is to be understood that the specific embodiments described herein are only used to explain the present invention and cannot be used to limit the scope of protection of the present invention. Based on one or more embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should belong to the scope of protection of the present invention.

[0056] like Figure 1 As shown in the flow chart, the embodiment of the present invention provides a multi-task non-intrusive load detection method and system based on feature fusion, including the following steps:

[0057] Step S100, collecting steady-state current and voltage data of a single device and multiple devices when running through a data acquisition module;

[0058] Step S200, in the data processing module, firstly, the distance matrix of the calculated current data, the short-time Fourier transform matrix of the current and voltage data, and the features are integrated into a three-dimensional matrix, and secondly, the load data information of each device is calculated;

[0059] Step S300: In the network identification module, the characteristic data is identified to obtain the specific device name of the data;

[0060] Step S400, uploading the acquired device name and the power usage of the device to the central control terminal through the data transmission module;

[0061] In the step S100, data collection is performed by using a high sampling rate device, where the collector is connected in series to the secondary electrical box of the factory to collect current and voltage data of the device;

[0062] Specifically, the step S100 includes the first aspect of the present invention, which provides a steady-state data acquisition method;

[0063] Step S101, for the original current data, an adaptive multi-window sliding bilateral CUSUM is used to perform event detection. Specifically, the following steps are included:

[0064] Step 1: Define two sliding windows: short window N s and long window N l Usually the short window length can be set to 5%-10% of the cycle length, and the long window length is generally much larger than the short window, and can usually be set to 50%-100% of the cycle length. The short window is responsible for capturing short-term changes in the signal, and the long window is responsible for capturing long-term changes in the signal.

[0065] Step 2: Calculate the mean and standard deviation. The short window calculates the most recent N s The mean and standard deviation of samples are used to detect rapid changes that occur in a short period of time. At each time point t, the long window calculates the most recent N l The mean and standard deviation of the samples are used to detect slow changes over a longer period of time.

[0066] Short window mean The calculation method is:

[0067]

[0068] Long window mean The calculation method is:

[0069]

[0070] Short window standard deviation The calculation method is:

[0071]

[0072] Long window standard deviation The calculation method is:

[0073]

[0074] Step 3: Dynamically adjust the offset and threshold. At each time point t, dynamically adjust the offset and threshold of CUSUM detection based on the calculated standard deviation:

[0075] Offset k t :

[0076]

[0077] Among them, β is a proportionality coefficient.

[0078] Threshold h t :

[0079]

[0080] Among them, γ is a proportionality coefficient.

[0081] Step 4: bilateral CUSUM detection. For each time point t, the bilateral CUSUM algorithm is used to perform preliminary change point detection:

[0082] First, calculate the increment d t :

[0083]

[0084] Next, calculate the forward cumulative sum statistic

[0085]

[0086] Inverse CUSUM statistic

[0087]

[0088] Finally, perform preliminary change point detection: when the positive cumulative sum Exceeding the positive threshold h t When a rising change is detected, the reverse accumulation and Below the reverse threshold -h t , a falling change is detected.

[0089] Step 4: Introduce the detection window. For each initially detected change point t, introduce the detection window W d Conduct further verification.

[0090] First, select the size W d The detection window is further analyzed within the time period after the change point t is initially detected.

[0091] Then, the mean and standard deviation within the detection window are calculated as follows:

[0092] The mean μ in the detection window d :

[0093]

[0094] Standard deviation within the detection window σ d :

[0095]

[0096] If the mean within the detection window deviates significantly from the historical mean, a change point is confirmed.

[0097]

[0098] Among them, α is a significance coefficient.

[0099] In the above steps, β, γ, and α are all constant coefficients, which control the sensitivity of the event detection algorithm and should be adjusted appropriately for different data;

[0100] Finally, save the current data after the change point and save the voltage data at the same time.

[0101] Step S102, extracting periodic steady-state current data, using a voltage zero-crossing sampling method to extract the steady-state current data;

[0102] Specifically, the first step is to calculate the sign of the voltage signal. If the voltage is positive, it is recorded as +1, if it is negative, it is -1, and if it is zero, it is 0. The formula is:

[0103] The sign difference of adjacent points is then calculated to determine the zero crossing point. When the voltage signal passes through zero from positive to negative or from negative to positive, the sign difference is not zero and is recorded as the zero crossing point. The formula is as follows:

[0104] ZERO_CROSS={i|sign(v i+1 )-sign(v i )≠0}

[0105] Check whether the first data point after each zero crossing is a positive value. If it is a positive value, keep the zero crossing point. After processing all the data, count the number of all zero crossings. If the number of zero crossings is an odd number, remove the last zero crossing point to ensure that the zero crossings appear in pairs so that complete current data can be extracted. Check whether the data points after the last zero crossing point contain a complete current cycle. If there are not enough data points, remove the last pair of zero crossings to ensure the integrity of the extracted current data. Collect the data between the two zero crossings when a single device is running, and the data between the two zero crossings when multiple devices are running, as steady-state periodic current signals;

[0106] The step S200 mainly processes the collected data, extracts the time-frequency domain characteristics of the data, and calculates the load data information required by the central control system;

[0107] Specifically, the step S200 includes the second aspect of the present invention, and provides a method for processing data dimension reduction and time-frequency domain feature fusion:

[0108] Step S201: for the collected high-frequency current steady-state signal, X = {x1, x2, ..., x n}, use the segmented aggregation approximation method to reduce data dimension;

[0109] Furthermore, the dimension reduction step can be divided into two cases:

[0110] First, when the input sequence can divide the length m that needs to be reduced, the current sequence X will be divided into equal parts subsequences, each subsequence X i The length of is m. The average value of each subsequence is X i The calculation is as follows:

[0111]

[0112] Among them, x (i-1)m+j Represents the jth element in the i-th subsequence. is the average value of the ith subsequence;

[0113] Calculate the average value of each subsequence, and form all the calculated average values ​​into the reduced dimension data X p ;

[0114] Secondly, when the length of the input sequence cannot be divided into the length m of the dimension reduction, the input sequence needs to be expanded. The expanded sequence is X′=[x1,x2,...,x n , x1, x2, ..., x n, ...], repeat until the sequence length can divide the length of the subsequence m. According to the first step, the expanded sequence is divided into subsequences, and the average value of each subsequence is X i The calculation is as follows:

[0115]

[0116] Calculate the average value of each subsequence, and form all the calculated average values ​​into the reduced dimension data X p ;

[0117] Step S202, constructing a current p-norm distance matrix, first repeating the dimension-reduced data m times along the column direction to form a new m×m matrix X repeat , then calculate the difference between the new matrix and its transposed matrix, which can be expressed as

[0118]

[0119] where it is combined with the transposed matrix The result of the difference operation is recorded as diff. Then, we perform a power operation on the difference according to the given norm order p, which is 1 / 2 in this implementation. Finally, we accumulate the power operation results of each difference and raise the result to the power of p to obtain the final distance matrix D. This operation can be expressed as:

[0120]

[0121] in, The i-th row of the matrix diff represents the power operation result of the difference between the i-th data point and all other data points;

[0122] Step S203, construct the matrix after the short-time Fourier transform of the current, and transform the collected original signal X={x1, x2, ..., x n}, without data dimensionality reduction, directly through the short-time Fourier transform function:

[0123]

[0124] Where x(n) is the input signal, w(nm) is the window function around time m, w is the frequency, n is the time index, and then an m×m current short-time Fourier amplitude spectrum matrix D2 is returned;

[0125] The short-time Fourier amplitude spectrum matrix D3 transformation method of voltage is the same as that of current, so I will not elaborate on it;

[0126] Furthermore, the matrices D1, D2, and D3 are fused into a three-dimensional matrix D using a method similar to the three-channel processing method of the image R, G, and B, and used as the input of the network recognition module.

[0127] Step S204, for the collected voltage and current signals, before data processing, they are low-pass filtered with a cutoff frequency of 50 Hz, and the device information within the collection period is calculated in real time through the current and voltage data information, such as the usage time of each device in the central control system, the device power, the usage time, the usage frequency and other information.

[0128] The step S300 includes the third aspect of the present invention, providing a lightweight multi-task recognition network for the field of non-intrusive load monitoring:

[0129] Furthermore, the present invention proposes a new lightweight network module Light Block, such as Figure 3 As shown;

[0130] Furthermore, each Light Block contains two Light modules, such as Figure 4 As shown;

[0131] Specifically, in the Light module, the data input goes through a depth-wise separable convolution with the following structure: Figure 5 As shown in the figure, the main operation is to first enter the deep convolution to convolve each channel of the data to reduce the number of parameters of the convolution network. After the deep convolution is completed, it is followed by the batch normalization layer (BN) and the ReLU activation function. The next step is point-by-point convolution, i.e., 1×1 convolution, to merge the channels of the deep convolution and increase the nonlinear ability of the network to process more feature information. The feature results of the deep separable convolution are linearly transformed, and then a batch normalization layer (BN) is added to generate a new feature map. The feature map is added to the original feature results of the deep separable convolution, and the final output result is consistent with the output result of the Conv2d convolution layer in terms of the number of channels. This module is a Light module. The BN layers in the middle of the two Light modules are connected, and the input of the first module and the output of the second module form a residual structure to form a complete Light Block.

[0132] Furthermore, the network model is Figure 6As shown in the figure, its specific structure is as follows: assuming that the input data size is 50*50*3, the first step enters a convolution layer with a 3x3 convolution kernel and a stride of 1, and the output is 50*50*16. Input it into the first LightBlock and the output is 50*50*24, input it into the second Light Block and the output is 25*25*40, input it into the third Light Block and the output is 13*13*80, input it into the fourth Light Block and the output is 13*13*112, and pass it through an average pooling layer, the output is 1*1*112. Input it into a 1x1 convolution kernel and a stride of 1 convolution layer, the output is 1*1*256, then branch 1 passes through a Linear layer and outputs recognition result 1, and branch 2 passes through a Linear layer and outputs recognition result 2. Among them, recognition result 1 outputs the binary classification result of single device or multiple devices, and recognition result 2 outputs the specific name of the device. Only when both recognition result 1 and recognition result 2 output the correct result, the recognized device name is output, ensuring the accuracy of classification for multiple devices. In terms of network parameters, the training parameters of the same structure of the Conv2d convolutional neural network are 234703, and the calculated FLOPs are 23.793M, while the network training parameters of the present invention are 65503, and the calculated FLOPs are 13.419M. The network parameters are almost 1 / 4 of the Conv2d convolutional neural network, and the floating-point calculation amount is almost 1 / 2 of the ordinary convolutional neural network, which greatly reduces the number of parameters and floating-point calculations, and has a huge advantage in edge computing devices.

[0133] In step S400, the load monitoring and power consumption of each device can be displayed on the central control screen for intelligent management through the central control system.

[0134] In order to explain the interaction between software and hardware in more detail, the fourth aspect of the present invention is introduced here, which proposes a multi-task non-intrusive load detection system based on feature fusion, and its structure diagram is shown in FIG. Figure 2 As shown, including:

[0135] Edge computing end and central control end, where the edge computing end includes:

[0136] Data acquisition module, used for collecting load voltage and current data;

[0137] Specifically, the data acquisition module extracts steady-state current data after the load data event is detected; in the embodiment of the present invention, the data acquisition module uses a 1.7k sampling rate for data acquisition;

[0138] A data storage module, used to temporarily save the collected data to the edge computing end;

[0139] Specifically, the storage of data in the embodiments of the present invention mainly provides classified retrieval and management functions for the data to facilitate subsequent data processing and analysis, and provides a temporary cache function before the data is uploaded to the central control system to ensure the integrity and continuity of the data.

[0140] Data preprocessing module, used to process the collected load data, including preprocessing such as data dimension reduction, and calculation and processing of equipment data information;

[0141] Specifically, the data preprocessing module is mainly divided into two parts. The first part is data reception and processing. In the embodiment of the present invention, by receiving the current and voltage data transmitted by the sensor, the steady-state current and voltage signals are extracted from the received data, and the steady-state current and voltage signals are downsampled to reduce the amount of data and reduce the calculation burden. The p-norm distance matrix of the downsampled current signal is calculated, and the steady-state current and voltage signals are short-time Fourier transformed to generate a frequency domain feature matrix. Finally, the above-processed data is sorted into the input format required by the network model. The second part is feature extraction and analysis. In this embodiment, the extracted current and voltage signals are low-pass filtered, and a cutoff frequency of 50Hz is used to remove high-frequency noise. By analyzing the filtered current and voltage signals, the cumulative use time of each device is calculated, and the power consumption of each device is calculated based on the current and voltage signals. Through statistical analysis of the operating status of the device, the use frequency and use cycle of each device are calculated.

[0142] The network identification module is used to perform deep learning network inference on the data and identify the device;

[0143] Specifically, the optimal model obtained through training is deployed to the edge device. The edge device receives the real-time transmitted current and voltage data, processes and recognizes the real-time data using the deployed model, and outputs the recognition results.

[0144] Communication module, used for communication between power-consuming devices and edge computing terminals, and between edge computing terminals and central control terminals;

[0145] Specifically, the communication module is mainly responsible for two parts. One is the communication between the electrical equipment and the edge computing end. The current, voltage and other relevant data are collected from the electrical equipment in real time. The load data is required to achieve reliable data transmission. Data transmission can be achieved by using, including but not limited to, Modbus RTU or Modbus TCP. The second is the communication between the edge computing end and the central control end. The aggregated data processed by the edge computing device needs to be transmitted to the central control system for centralized monitoring and analysis. At the same time, the control commands received from the central control system need to be passed to the electrical equipment to achieve remote monitoring and intelligent power consumption. The communication transmission method between the edge computing end and the central control end is through wireless transmission. The wireless transmission protocol includes, but is not limited to, one or several combined communication methods such as WiFi, Bluetooth, 4G / 5G, LoRa, UWB, etc.;

[0146] The edge computing system will be installed in the secondary distribution box in the factory production environment.

[0147] The central control end is responsible for unified management and further processing of the data of all edge devices, further analysis, processing and visualization of the aggregated data, real-time monitoring of the status of edge devices, detection of anomalies and triggering alarm mechanisms, and sending generated strategies and plans to edge devices to guide their operations.

Claims

1. A multi-task non-intrusive load detection method based on feature fusion, characterized in that: include: Step 1: Use a collector to collect steady-state current and voltage signals when a single device or multiple devices are running; Step 2: Down-sampling and low-frequency filtering of the collected voltage and current signals; Step 3: Preprocess the current and voltage signals to adapt them to the network input; Step 4: Analyze the filtered current and voltage signals and calculate relevant statistical results of the equipment operation status; Step 5: Send the preprocessed power current signal to the saved optimal lightweight multi-task network model for real-time identification to obtain the identification result of the device.

2. The method for collecting steady-state periodic voltage and current signals according to claim 1, characterized in that: The adaptive multi-window sliding bilateral CUSUM is used for event detection to collect steady-state current data. The method includes: using two sliding windows, short window N s and long window N l The short window is responsible for detecting rapid changes that occur in a short time, and the long window is responsible for detecting slow changes that occur in a long time. The parameters are dynamically adjusted according to the mean and standard deviation of the current signal in the short window and the long window to control the offset and threshold of the CUSUM detection, so that it can better adapt to different types of data.

3. According to claim 2, in the method for performing event detection using the adaptive multi-window sliding bilateral CUSUM, after the event is initially detected, in order to reduce misjudgment, a detection window W is added in the time period after the change point t. d , calculate the mean μ within the detection window d and standard deviation σ d If the mean within the detection window deviates significantly from the historical mean, a change point is confirmed.

4. The method for collecting steady-state current and voltage signals when a single device or multiple devices are running according to claim 1, characterized in that: The voltage zero-crossing method is used to collect current and voltage data. First, ensure that the collected data is a complete data cycle and the first data point is a positive value. The zero-crossing point that meets the required voltage is calculated and screened, and then the current data between the two zero-crossing points when single-device and multi-device are operated is collected as a steady-state periodic current signal.

5. The voltage signal and current signal downsampling method according to claim 1, characterized in that: The segmented aggregation approximation method is used to maintain the uniformity and integrity of the data as much as possible. The dimensionality reduction step is divided into two cases. First, when the input sequence length is divisible by the dimensionality reduction length, the current sequence is divided into several subsequences in equal proportions, the average value of each subsequence is calculated, and all the average values ​​are used to form the data after dimensionality reduction. When the input sequence length is not divisible by the dimensionality reduction length, the repeated data is expanded to be divisible by the input sequence, and then processed according to the above method, the average value of each subsequence is calculated, and the data after dimensionality reduction is formed.

6. The current and voltage preprocessing method according to claim 1, characterized in that: The following steps are included. The first step is to construct the current p-norm distance matrix, repeat the reduced-dimensional data m times along the column direction to form a new m×m matrix X repeat , and transpose it with the matrix The result of the difference operation is recorded as diff. The calculation formula of the p-norm distance matrix D1 of the current is as follows: Secondly, the matrices of current and voltage after short-time Fourier transform are constructed and recorded as D2 and D3 respectively. The matrices D1, D2, and D3 are fused into a three-dimensional matrix D using a three-channel processing method similar to that of R, G, and B in an image, which serves as the input of the network recognition module.

7. The lightweight multi-task network according to claim 1, characterized in that: A new lightweight network module Light Block is proposed, in which each Light Block contains two Light modules; in the Light module, layered convolution replaces 2D convolution. The first layer uses depthwise separable convolution to extract features, and the second layer uses layer convolution to expand the dimension. The feature map is added to the original feature result of the depthwise separable convolution and then output. The input of the first module and the output of the second module form a residual structure to form a complete Light Block.

8. According to claim 1, the model structure of the lightweight multi-task network is as follows: first pass through a Conv2D convolution module, then pass through four Light Blocks in sequence, then pass through an average pooling layer, then pass through another Conv2D convolution module, then branch one passes through a Linear layer, outputs recognition result one, branch two passes through a Linear layer, and outputs recognition result two.

9. A multi-task non-intrusive load detection method and system based on feature fusion, characterized in that: include: The edge computing end and the central control end, the edge computing end includes: data acquisition module, data storage module, network identification module, data preprocessing module, communication module. The central control end is responsible for unified management and further processing of the data of all edge devices, real-time monitoring of the status of edge devices, and sending the generated strategies and solutions to edge devices to guide their operation.

Citation Information

Patent Citations

  • Non-invasive load identification method based on MLCDTL

    CN110555369A

  • Load event detection method and system and storage medium

    CN111191908A

  • Transient event detection method based on composite sliding window and bilateral correction

    CN114298223A

  • CUSUM-Bi-LSTM-based edge side online non-intrusive load identification method

    CN115186258A

  • Non-intrusive load identification method, device, equipment and medium

    CN115758188A

Cited By

  • Event detection and feature extraction method of CUSUM algorithm based on bidirectional detection window

    CN120508815A

  • An event detection and feature extraction method based on a bidirectional detection window CUSUM algorithm

    CN120508815B