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

Abstract

This invention provides a multi-task non-intrusive load detection method and system based on feature fusion. The system mainly consists of two parts: an edge computing terminal and a central control terminal. The method includes the following steps: First, event detection is performed on the raw data, collecting steady-state current and voltage data from the device. The data to be processed is downsampled, and the p-norm distance of the current data is calculated in the edge computing device to form a current distance matrix. Then, this matrix, along with the amplitude spectrum after short-time Fourier transform of the current and the amplitude spectrum after Fourier transform of the voltage, are used as the three-channel inputs of the matrix. Next, the feature image is input to a lightweight multi-task network model for load identification. This invention solves the problem of insufficient single charge features by fusing time-domain and frequency-domain features, and addresses the computational cost problem of computing devices by proposing a lightweight multi-task network, effectively improving the detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of non-invasive load detection technology, and in particular to a multi-task non-invasive load detection method and system based on feature fusion. Background Technology

[0002] With societal progress and continuous technological development, sustainable energy utilization and environmental protection have become global concerns. Precise monitoring and diagnosis of the energy consumption of various electrical appliances in the power system can help users gain a deeper understanding of energy consumption patterns, identify high-energy-consuming and inefficient equipment, and take corresponding measures for adjustment, thereby achieving efficient energy utilization and conservation. The emergence of non-intrusive load sensing 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 become urgent issues to be addressed. Traditional electricity demand management methods are no longer sufficient to meet the ever-increasing electricity demand; therefore, improving the power grid's demand-side management capabilities is particularly important. Non-intrusive load sensing technology provides new ideas and methods for power grid management. By accurately monitoring and analyzing users' electricity consumption behavior and demand patterns, it can help power companies more accurately understand users' electricity needs, rationally allocate power resources, optimize power grid load distribution, and thus improve the safety and stability of the power grid.

[0003] Non-intrusive load identification is a key objective in the field of non-intrusive load monitoring. Due to the diverse types and complex operating states of electrical appliances in power systems, single feature information cannot accurately identify all of them. Improving the accuracy and reliability of identification during single-device and multi-device operation is a problem that needs to be solved. However, non-intrusive load identification typically relies on neural networks to learn from large amounts of load data for training and inference, requiring significant computing resources and power. This presents challenges for edge deployment, necessitating exploration of ways to reduce data computation and improve system response speed in non-intrusive load detection systems, while adapting to the computing resource limitations of edge devices. Summary of the Invention

[0004] Technical Problems: This invention addresses the problems described in the background above, specifically the issue of single-feature load data, by proposing a data processing method that fuses time and frequency features. It also addresses the resource constraints in edge computing by proposing a lightweight multi-task identification network. This solves the problems of low load identification accuracy and excessive computing power requirements when multiple devices are running in edge deployments.

[0005] Technical solution:

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

[0007] Step S100: Collect steady-state current and voltage data of single device and multiple devices during operation through the data acquisition module;

[0008] In step S200, in the data processing module, the first step is to fuse the distance matrix of the calculated current data, the short-time Fourier transform matrix of the current and voltage data, and the features into a three-dimensional matrix; the second step is to calculate the load data information of each device.

[0009] In step S300, the network identification module identifies the feature data and obtains the specific device name of the data;

[0010] Step S400: The obtained device names and power consumption information of each device are uploaded to the central control terminal through the data transmission module.

[0011] Step S100 includes the first aspect of the present invention, which provides a steady-state data acquisition method.

[0012] Step S101: For the raw current data, event detection is performed using an adaptive multi-window sliding bilateral CUSUM. Specifically, this includes the following steps:

[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 within the short and long windows;

[0015] 3. Dynamically adjust offset and threshold: Based on the statistics within the sliding window, dynamically adjust the offset k detected by CUSUM. t and threshold h t ;

[0016] 4. Bilateral CUSUM Detection: The bilateral CUSUM algorithm is applied to detect the initial change points in the current signal;

[0017] 5. Introduce a detection window: Further verify the initially detected changes to reduce false alarms;

[0018] Step S102: Extract steady-state current and voltage data during single-device and multi-device operation. Use voltage zero-crossing sampling method to extract steady-state current and voltage data. Calculate and filter zero-crossing points that meet the voltage requirements. Then, collect the current data between two zero-crossing points during single-device and multi-device operation as steady-state periodic current signal.

[0019] Step S200 includes the second aspect of the present invention, providing a data dimensionality reduction and time-frequency domain feature fusion processing method:

[0020] Step S201, for the acquired high-frequency current steady-state signal, X = {x1, x2, ..., x...} n The data dimensionality reduction is performed using a piecewise aggregation approximation method. Furthermore, to maintain data uniformity and integrity as much as possible, the dimensionality reduction steps can be divided into two cases:

[0021] Firstly, when the input sequence is divisible by the length m to be reduced in dimension, the current sequence X will be equally divided into: Given a set of subsequences, calculate the average of each subsequence, and combine all the calculated averages to form the dimensionality-reduced data X. p ;

[0022] Secondly, when the length of the input sequence is not a divisible by the length m to be reduced in dimension, 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 is divisible by the length of the subsequence m. Following the method in step one, divide the expanded sequence into subsequences and calculate the average of each subsequence. Combine all the calculated averages to form the dimensionality-reduced data X. p ;

[0023] Step S202: Construct the current p-norm distance matrix. First, repeat the dimensionality-reduced data m times along the column direction to form a new m×m matrix X. repeat Combine it with the transpose matrix The result of the interpolation operation is denoted as diff. The formula for calculating 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 result of exponentiation 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 use the acquired original signal X = {x1, x2, ..., x...} n Without data dimensionality reduction, it directly uses the short-time Fourier transform function to return an m×m current short-time Fourier amplitude spectrum matrix D2.

[0027] The method for the short-time Fourier amplitude spectrum matrix D3 transformation of voltage is the same as that for current, so I will not go into details.

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

[0029] Step S204: For the collected voltage and current signals, calculate the load data information within the collection period in real time, such as the usage time, power consumption, usage time, and usage frequency of each device in the central control system.

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

[0031] Furthermore, this invention proposes a novel lightweight network module, Light Block;

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

[0033] Specifically, in the Light module, layered convolutions replace 2D convolutions. The first layer uses depthwise separable convolutions to extract features, and the second layer uses layered convolutions to expand the dimensions. This feature map is then added to the original feature result of the depthwise separable convolutions before being output. The input of the first module and the output of the second module form a residual structure, which constitutes a complete Light Block.

[0034] Furthermore, the inference network structure is as follows: data is input into a Conv2D convolutional module, then sequentially into four Light Block modules, followed by an average pooling layer, then another Conv2D convolutional module. Branch one passes through a Linear layer, outputting recognition result one, and branch two passes through a Linear layer, outputting recognition result two. Recognition result one outputs the binary classification result indicating whether it's a single-device or multi-device classification, while recognition result two outputs the specific name of the device.

[0035] A fourth aspect of the present invention proposes a multi-task non-invasive load detection system based on feature fusion, comprising:

[0036] Edge computing endpoints and central control endpoints, wherein the edge computing endpoints include:

[0037] The data acquisition module is used to acquire load voltage and current data.

[0038] The data storage module is used to temporarily save the collected data to the edge computing terminal;

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

[0040] The network recognition module is used to perform inference on the data using deep learning networks and to identify the device.

[0041] The communication module is used for communication between the electrical equipment and the edge computing terminal, as well as communication between the edge computing terminal and the central control terminal.

[0042] The central control unit is responsible for the unified management and further processing of data from all edge devices, as well as distributing the generated strategies and schemes to the edge devices to guide their operation.

[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 electrical boxes in each area of ​​the project can be displayed on the large screen of the central control terminal.

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

[0045] 1. An adaptive multi-window sliding bilateral CUSUM event detection method is provided for the field of non-intrusive load sensing. A short window is used to capture rapid changes in the short term, while a 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, abrupt changes in the current signal can be sensitively captured. After initial event detection, a new detection window is created, and the mean is calculated to see if it significantly deviates from the historical mean, further confirming the change point and significantly improving the sensitivity and accuracy of event detection.

[0046] 2. This paper provides a data preprocessing method for time-frequency domain multi-feature fusion in the field of non-invasive load detection. It uses the p-norm distance matrix to extract the current time-domain features and 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 when using a single feature for network recognition.

[0047] 3. Provides a lightweight multi-task recognition network for non-intrusive load detection. The main part uses Light Block and extensively uses layered convolution, point convolution and residual structure, which greatly reduces the parameter size of the network model and the floating-point operation of computation and inference. In addition, the multi-task recognition network can better recognize the accuracy of multiple devices running at the same time.

[0048] 4. A feature-fusion-based, multi-task, non-intrusive load detection system is provided, combining a central control system and an edge computing terminal. The edge computing terminal, located close to the data source, can quickly process data and respond, reducing data transmission latency. The central control system is responsible for global data collection and comprehensive analysis, providing load detection from a global perspective, remote monitoring, and optimization strategies, thus improving overall system efficiency and power safety. Attached Figure Description

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

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

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

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

[0053] Figure 5 This is a 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 This is a schematic diagram of a lightweight, multi-task, non-intrusive load network structure in an embodiment of the present invention; Detailed Implementation

[0055] The technical solutions of the present invention will be described in more detail below with reference to the accompanying drawings of the embodiments of the present invention. It is understood that the specific embodiments described herein are merely illustrative of the invention and should not be construed as limiting the scope of protection of the invention. All other embodiments obtained by those skilled in the art based on one or more embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0056] like Figure 1 As shown in the flowchart, this invention provides a multi-task non-invasive load detection method and system based on feature fusion, including the following steps:

[0057] Step S100: Collect steady-state current and voltage data of single device and multiple devices during operation through the data acquisition module;

[0058] In step S200, in the data processing module, the first step is to fuse the distance matrix of the calculated current data, the short-time Fourier transform matrix of the current and voltage data, and the features into a three-dimensional matrix; the second step is to calculate the load data information of each device.

[0059] In step S300, the network identification module identifies the feature data and obtains the specific device name of the data;

[0060] Step S400: Upload the obtained device name and the device's power consumption to the central control terminal through the data transmission module;

[0061] In step S100, the data acquisition uses a high sampling rate device. The data acquisition device is connected in series in the factory's secondary electrical box to collect the current and voltage data of the device.

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

[0063] Step S101: For the raw current data, event detection is performed using an adaptive multi-window sliding bilateral CUSUM. Specifically, this includes the following steps:

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

[0065] Step 2: Calculate the mean and standard deviation. For each time point t, calculate the most recent N values ​​within a short window. s The mean and standard deviation of each sample are used to detect rapid changes occurring over a short period of time. A long window is used to calculate the most recent N values ​​at each time point t. l The mean and standard deviation of a sample are used to detect slow changes that occur over a longer period of time.

[0066] Short window mean The calculation method is as follows:

[0067]

[0068] Long window mean The calculation method is as follows:

[0069]

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

[0071]

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

[0073]

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

[0075] Offset k t :

[0076]

[0077] Where β is a proportionality coefficient.

[0078] Threshold h t :

[0079]

[0080] Here, γ is a proportionality coefficient.

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

[0082] First, calculate the increment d. t :

[0083]

[0084] Next, calculate the positive cumulative sum statistic.

[0085]

[0086] Inverse cumulative sum statistic

[0087]

[0088] Finally, preliminary change point detection is performed: when the positive cumulative sum... Exceeding the positive threshold h t When an upward change is detected, the reverse cumulative sum is... Below the reverse threshold -h t At that time, a downward change was detected.

[0089] Step 4: Introduce a detection window. For each initially detected change point t, introduce a detection window W. d Further verification is needed.

[0090] First, select a size of W. d The detection window is set to perform further analysis within the time period after the initial detection of the change point t.

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

[0092] Mean μ within the detection window d :

[0093]

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

[0095]

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

[0097]

[0098] Here, α is a significance coefficient.

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

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

[0101] Step S102: Extract periodic steady-state current data by using voltage zero-crossing sampling method.

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

[0103] The zero-crossing point is then determined by calculating the sign difference between adjacent points. When the voltage signal crosses zero from positive to negative or from negative to positive, the sign difference is not zero, and this is recorded as a zero-crossing point. The formula is as follows:

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

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

[0106] Step S200 mainly involves processing the collected data, extracting the time-frequency domain characteristics of the data, and calculating the load data information required by the central control system.

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

[0108] Step S201, for the acquired high-frequency current steady-state signal, X = {x1, x2, ..., x...} n The data dimensionality reduction is achieved using a segmented aggregation approximation method.

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

[0110] Firstly, when the input sequence is divisible by the length m to be reduced in dimension, the current sequence X will be equally divided into: There are subsequences, each subsequence X i The length is m. The average value X of each subsequence is... i The calculation is as follows:

[0111]

[0112] Where, x (i-1)m+j It represents the j-th element in the i-th subsequence. It is the average value of the i-th subsequence;

[0113] Calculate the mean of each subsequence, and combine all the calculated means to form the dimensionality-reduced data X. p ;

[0114] Secondly, when the length of the input sequence is not a divisible by the length m to be reduced in dimension, the input sequence needs to be expanded. The expanded sequence is X′=[x1,x2,...,x n x1, x2, ..., x nRepeat this process until the sequence length is divisible by the length of the subsequence m. Following the method in step one, divide the expanded sequence into subsequences, with each subsequence having an average value X. i The calculation is as follows:

[0115]

[0116] Calculate the mean of each subsequence, and combine all the calculated means to form the dimensionality-reduced data X. p ;

[0117] Step S202: Construct the current p-norm distance matrix. First, repeat the dimensionality-reduced data m times along the column direction to form a new m×m matrix X. repeat Next, we calculate the difference between the new matrix and its transpose, which can be expressed as:

[0118]

[0119] Here it is combined with the transpose matrix The result of the interpolation operation is denoted as `diff`. Then, based on the given norm order `p` (in this implementation, `p` is set to 1 / 2), we perform exponentiation on the differences. Finally, we sum the exponentiation results of each difference and take the p-th root of the result to obtain the final distance matrix `D`. This operation can be represented as:

[0120]

[0121] in, The i-th row of the matrix diff represents the result of exponentiation 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 use the acquired original signal X = {x1, x2, ..., x...} n Without data dimensionality reduction, directly using 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, and n is the time index, then an m×m short-time Fourier amplitude spectrum matrix D2 of the current is returned;

[0125] The method for the short-time Fourier amplitude spectrum matrix D3 transformation of voltage is the same as that for current, so I will not go into details.

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

[0127] Step S204: Before data processing, the collected voltage and current signals are low-pass filtered with a cutoff frequency of 50 Hz. The device information within the acquisition period is calculated in real time using the current and voltage data, such as the usage time, power consumption, usage time, and usage frequency of each device in the central control system.

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

[0129] Furthermore, this invention proposes a novel lightweight network module, the 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 is processed through a depthwise separable convolution, with the structure as follows: Figure 5 As shown, its main operations are as follows: First, the input is fed into a depthwise convolution layer to convolve each channel of the data, reducing the number of parameters in the convolutional network. After the depthwise convolution, a batch normalization (BN) layer and a ReLU activation function follow. The next step is a pointwise convolution, i.e., a 1×1 convolution, which merges the channels of the depthwise convolution, increasing the network's non-linearity to process more feature information. A linear transformation operation is performed on the feature result of the depthwise separable convolution, followed by a batch normalization (BN) layer to generate a new feature map. This feature map is added to the original feature result of the depthwise separable convolution, and the final output result is consistent with the output result of the Conv2d convolutional layer in terms of the number of channels. This module is a Light module. Connecting the BN layer between two Light modules, the input of the first module and the output of the second module form a residual structure, which constitutes a complete Light Block.

[0132] Furthermore, network models such as Figure 6As shown, the specific structure is as follows: Assuming the input data size is 50*50*3, the first step involves a convolutional layer with a 3x3 kernel and a stride of 1, outputting 50*50*16. This is then fed into the first LightBlock, outputting 50*50*24; into the second LightBlock, outputting 25*25*40; into the third LightBlock, outputting 13*13*80; and into the fourth LightBlock, outputting 13*13*112. This is then passed through an average pooling layer, outputting 1*1*112. Finally, it is fed into a convolutional layer with a 1x1 kernel and a stride of 1, outputting 1*1*256. Branch 1 then passes through a Linear layer, outputting recognition result one, and branch 2 passes through a Linear layer, outputting recognition result two. Recognition result one outputs the binary classification result (single-device or multi-device), while recognition result two outputs the specific name of the device. Only when both recognition result one and recognition result two output the correct result will the device name be output, ensuring the accuracy of classification when multiple devices are involved. Regarding network parameters, the Conv2d convolutional neural network with the same structure has 234703 training parameters and 23.793M FLOPs, while the network of this invention has 65503 training parameters and 13.419M FLOPs. The network parameters are almost 1 / 4 of those of the Conv2d convolutional neural network, and the floating-point operation volume is almost 1 / 2 of that of a regular convolutional neural network, significantly reducing the number of parameters and floating-point operations, giving it 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 through the central control system for intelligent management.

[0134] To illustrate the interaction between the software and hardware in more detail, the fourth aspect of this invention is introduced here, proposing a multi-task non-intrusive load detection system based on feature fusion, the structure of which is shown in the figure below. Figure 2 As shown, it includes:

[0135] Edge computing endpoints and central control endpoints, wherein the edge computing endpoints include:

[0136] The data acquisition module is used to acquire load voltage and current data.

[0137] Specifically, the data acquisition module extracts steady-state current data after detecting load data events; in the embodiments of the present invention, the data acquisition device uses a sampling rate of 1.7k for data acquisition.

[0138] The data storage module is used to temporarily save the collected data to the edge computing terminal;

[0139] Specifically, in this embodiment of the invention, data storage mainly provides classification, retrieval, and management functions to facilitate subsequent data processing and analysis, and provides a temporary caching function before the data is uploaded to the central control system to ensure data integrity and continuity.

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

[0141] Specifically, the data preprocessing module is mainly divided into two parts. First, data reception and processing: In this embodiment, by receiving current and voltage data transmitted from sensors, steady-state current and voltage signals are extracted from the received data. The steady-state current and voltage signals are downsampled to reduce the amount of data and computational burden. The p-norm distance matrix of the downsampled current signal is calculated. A short-time Fourier transform is performed on the steady-state current and voltage signals to generate a frequency domain feature matrix. Finally, the processed data is organized into the input format required by the network model. Second, feature extraction and analysis: In this embodiment, the extracted current and voltage signals are low-pass filtered using a 50Hz cutoff frequency to remove high-frequency noise. By analyzing the filtered current and voltage signals, the cumulative usage time of each device is calculated. The power consumption of each device is calculated based on the current and voltage signals. Through statistical analysis of the device operating status, the usage frequency and usage cycle of each device are calculated.

[0142] The network recognition module is used to perform inference on the data using deep learning networks and to identify the device.

[0143] Specifically, the optimal model obtained from training is deployed to edge devices. The edge devices receive real-time transmitted current and voltage data, use the deployed model to process and identify the real-time data, and output the identification results.

[0144] The communication module is used for communication between the electrical equipment and the edge computing terminal, as well as communication between the edge computing terminal and the central control terminal.

[0145] Specifically, the communication module is mainly responsible for two parts. First, it facilitates communication between the power-consuming equipment and the edge computing terminal. This involves real-time acquisition of current, voltage, and other relevant data from the equipment. For load data, reliable data transmission is required, and methods including, but not limited to, Modbus RTU or Modbus TCP can be used. Second, it facilitates communication between the edge computing terminal and the central control terminal. This involves transmitting the aggregated data processed by the edge computing device to the central control system for centralized monitoring and analysis. Simultaneously, it receives control commands from the central control system and transmits them to the power-consuming equipment, enabling remote monitoring and intelligent power consumption. The communication between the edge computing terminal and the central control terminal is conducted wirelessly, using protocols including, but not limited to, WiFi, Bluetooth, 4G / 5G, LoRa, UWB, and one or more combinations thereof.

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

[0147] The central control unit is responsible for the unified management and further processing of data from 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 of alarm mechanisms, and distribution of generated strategies and solutions to edge devices to guide their operation.

Claims

1. A multi-task non-intrusive load detection method based on feature fusion, characterized in that, include: Step 1: Use a data acquisition device to collect steady-state current and voltage signals during single-device and multi-device operation. Event detection is performed using an adaptive multi-window sliding bilateral CUSUM. Short and long windows are used to detect rapid changes occurring in a short time and slow changes occurring in a long time, respectively. The offset and threshold of CUSUM detection are dynamically adjusted based on the mean and standard deviation of the current signal within the short and long windows. The steady-state periodic current and voltage signals are extracted using the voltage zero-crossing method. Step 2: Perform downsampling and low-frequency filtering on the acquired voltage and current signals; Step 3: Preprocess the current and voltage signals to construct the current p-norm distance matrix, the current short-time Fourier amplitude spectrum matrix, and the voltage short-time Fourier amplitude spectrum matrix, and fuse the three into a three-dimensional matrix as the input of the lightweight multi-task network model. The current p-norm distance matrix D1 is obtained by repeatedly forming a matrix along the column direction of the dimension-reduced current data, subtracting it from its transpose matrix, and calculating according to the p-norm. The three-dimensional matrix D is formed by fusing D1, D2, and D3 as three channels. Step 4: Analyze the filtered current and voltage signals, and calculate the relevant statistical results of the equipment operating status; Step 5: Input the preprocessed 3D matrix into the saved optimal lightweight multi-task network model in real time for real-time identification, and obtain the identification result of the device.

2. In the multi-task non-intrusive load detection method based on feature fusion as described in claim 1, after an event is initially detected, to reduce misjudgments, a detection window W is added for 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, then the point of change is confirmed. 3.The method of claim 1, wherein, The current and voltage data are collected by using the voltage zero-crossing point method. First, it is ensured that the collected data is a complete data cycle and the first data point is positive. The zero-crossing points that meet the voltage requirements are calculated and selected. Then, the current data between the two zero-crossing points is collected when the single device and multiple devices are running, as the steady-state periodic current signal. 4.The method of claim 1, wherein, A segmented aggregation approximation method is adopted 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 length of the input sequence is divisible by the dimensionality reduction length, the current sequence is divided into several subsequences of equal length, the average value of each subsequence is calculated, and all the average values ​​are combined to form the dimensionality reduction data. Second, when the length of the input sequence is not divisible by the dimensionality reduction length, the repeated data is extended to make the input sequence divisible, and then processed in the above way to calculate the average value of each subsequence and form the dimensionality reduction data. 5.The method of claim 1, wherein, A novel lightweight network module, Light Block, is proposed, in which each Light Block contains two Light modules. In the Light module, hierarchical convolution replaces 2D convolution. The first layer uses depthwise separable convolution to extract features, and the second layer uses hierarchical 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, which constitutes a complete Light Block. 6.The method of claim 1, wherein, First, the image passes through a Conv2D convolutional module, then through four Light Blocks, followed by an average pooling layer, and then another Conv2D convolutional module. Branch 1 passes through a Linear layer and outputs recognition result 1, while branch 2 passes through a Linear layer and outputs recognition result 2.

7. A multi-task non-intrusive load detection system based on feature fusion, characterized in that, The system is used to implement the feature fusion-based multi-task non-intrusive load detection method according to any one of claims 1-6. The system includes an edge computing terminal and a central control terminal. The edge computing terminal includes a data acquisition module, a data storage module, a data preprocessing module, a network identification module, and a communication module. The data acquisition module is used to acquire steady-state current and voltage signals during single-device and multi-device operation. The data storage module is used to temporarily store the acquired data to the edge computing terminal. The data preprocessing module is used to perform downsampling, low-frequency filtering, p-norm distance matrix construction, short-time Fourier transform, and feature fusion processing on the acquired load data, and calculate the relevant statistical results of the device operating status. The network identification module is used to input the preprocessed three-dimensional matrix into the lightweight multi-task network model to obtain the device identification result; the communication module is used to realize the communication between the power-consuming device and the edge computing terminal, as well as the communication between the edge computing terminal and the central control terminal; the central control terminal is used to uniformly manage and further process the data of all edge computing terminals, monitor the status of edge devices in real time, and distribute the generated strategies and schemes to the edge computing terminals.

Images