A non-intrusive household appliance load identification method and system based on edge-cloud collaboration

Abstract

The present invention relates to the technical field of electrical appliance identification systems, and discloses a non-invasive household appliance load identification method and system based on edge-cloud collaboration. The system is an edge-cloud collaboration system, and its structure consists of a cloud server and edge intelligent devices. The edge intelligent devices are logically abstracted into a data acquisition layer, a data aggregation layer, an edge collaboration and reasoning layer. The data acquisition layer is responsible for capturing electrical characteristics from electrical appliances. These data from multiple electrical devices will be aggregated in the data aggregation layer. The edge collaboration and reasoning layer is responsible for interacting with the cloud and reasoning of the execution model. The entire electrical appliance identification system forms a system architecture by combining electrical category decomposition and identification network and socket hardware. The invented intelligent electrical identification socket integrates sampling, processing, analysis, and communication, and the recognition accuracy of the load identification algorithm is above 95%.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical appliance identification systems, and specifically to a non-invasive household appliance load identification method and system based on edge-cloud collaboration. Background Art

[0002] At present, with the improvement of living standards and the development of science and technology, the number of household electrical appliances owned by residents has increased rapidly. Providing safe and reliable power supply and ensuring the balance of power supply and demand have also become the basis and prerequisite for the stable and healthy development of the national economy.

[0003] Against this backdrop, we have developed an appliance identification system based on edge-cloud collaboration, leveraging IoT, signal processing, and artificial intelligence technologies. This system uses edge intelligent devices to collect current or voltage waveform data from appliances. After preprocessing, this data is fed into a model for inference, which occurs at the edge. Once inference is complete, the results are uploaded to the cloud. When a new appliance is connected, the collected data is sent to the cloud for model training. After training, the new model weights are downloaded to the edge to update the entire model. The cloud stores the inference results uniformly and, upon request, calls the edge database to display the appliance status to the user.

[0004] The entire appliance identification system consists of a cloud server and edge intelligent devices. The edge intelligent devices are logically divided into a data acquisition layer, a data aggregation layer, and an edge collaboration and inference layer. They consist of an analog front end, a microcontroller, a Wi-Fi module, and an OLED display. When an appliance is connected to the analog front end, the collected current data is processed and stored in a database. The database is divided into a training database and an inference database. Data before model training is stored in the training database, as is data from newly connected appliances. All other data is stored in the inference database. During system operation, users can view the operating status of their home appliances through a webpage.

[0005] Existing load identification algorithms have not yet been well integrated with hardware, and the models are relatively large, making it impossible to form a system-level solution. Current hardware devices cannot meet the needs of smart home systems for electrical device management and cannot automatically identify electrical appliances, resulting in less information being displayed and users being unable to obtain feedback on the real-time status of electrical appliances. Based on the above points, this appliance identification system, with self-developed hardware devices and software algorithms at its core, utilizes an edge-cloud collaborative model to effectively combine the two, providing an automated solution for smart home systems that can be integrated into smart home systems to serve users. Summary of the Invention

[0006] (1) Technical problems solved

[0007] In response to the shortcomings of the existing technology, the present invention provides a non-invasive household appliance load identification method and system based on edge-cloud collaboration.

[0008] (2) Technical solution

[0009] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a non-invasive household appliance load identification method and system based on edge-cloud collaboration, wherein the system is an edge-cloud collaborative system, and its structure consists of a cloud server and edge intelligent devices. The edge intelligent devices are logically abstracted into a data acquisition layer, a data aggregation layer, an edge collaboration and reasoning layer. The data acquisition layer is responsible for capturing electrical characteristics from electrical appliances. These data from multiple electrical devices will be aggregated in the data aggregation layer. The edge collaboration and reasoning layer is responsible for interacting with the cloud and executing model reasoning.

[0010] Preferably, the data acquisition layer is designed to capture the electrical characteristics of the equipment at high / low frequencies. The system adopts a sampling frequency of 6.4KHz to capture 128 voltage and current data in a cyclic manner, so as to reflect both the overall trend and the instantaneous changes. At this sampling frequency, the electrical characteristics of different electrical appliances are collected separately by the data acquisition layer in the system.

[0011] Preferably, the data aggregation layer is that the data from the acquisition layer will be aggregated in the data aggregation layer of the edge intelligent device, and the electrical data will be temporarily stored in the training database. These data will be regularly extracted by the cloud server for further analysis. At the same time, the inference database will be used to update the weight parameters of the inference model. The training database is more inclined to use a relational database, while the inference database is a non-relational database and uses cache to store data.

[0012] Preferably, the edge coordination and reasoning layer establishes a transmission link with the cloud server, the model is deployed on the server side, and the data used for reasoning and training comes from the database of the data aggregation layer. The edge intelligent device cooperates with the cloud server in the following stages.

[0013] Preferably, the method is as follows: Phase 1: Initial training model, the cloud server trains the first neural network TD0 from the initial data set N0 stored in the training database, and generates a weight file W0 after the training is completed.

[0014] Preferably, the second stage: obtaining model weights. After the training of the neural network N0 is completed, the server will send the weight file W0 to the data coordination and reasoning layer of the edge smart device. Therefore, the model deployed in the edge smart device will obtain the initial weights for performing model reasoning and return the reasoning results to the cloud. Finally, the cloud server will unify the reasoning results and display the real-time status of the appliance to the user.

[0015] Preferably, the third stage: updating data. During the reasoning process performed by the edge intelligent device, when a new device category is added to the system, the device data will be automatically extracted and added to the training database, and the TD1 data will be sent to the cloud server to prompt the training of the model. After the cloud server trains the new neural network N1 based on the latest data set, it will send the updated weight file W1 to the edge intelligent device. In this stage, an asynchronous mechanism is adopted to perform the model update process on the server and perform the edge model reasoning at the edge. The data stage update operations are continuously performed, and the model is dynamically updated to adapt to changes in the real environment.

[0016] Preferably, the trained model is designed to extract internal features and classify them into different categories. The instantaneous voltage and current data of the equipment are used as the input of the model. The classification of electrical data can be achieved through a time series classification model. Wavelet decomposition is similar to multi-level discrete wavelet decomposition (MDWD), which is a well-known method for capturing time series features in the time domain and frequency domain. On this basis, a multi-level wavelet decomposition network (mWDN) is proposed in the field of time series analysis to extract sequence feature information. Combined with mWDN, an improved multi-level wavelet decomposition network-residual feature vector (mWDN-RFV) model is proposed. The model is divided into an mWDN structure and an RFV structure.

[0017] The preferred edge intelligent device for appliance identification systems is designed to retrieve and analyze electrical data, which is also used to evaluate the system's performance in real-world environments. The edge intelligent device consists of an analog front end, an MCU, a Wi-Fi module, and an OLED display. Electrical characteristics, including transient current and voltage information, are captured through data acquisition circuitry, including a ZMPT107 transformer and an RN8302B three-phase energy measurement chip. The chip includes seven 24-bit ADC channels, three phase current sampling channels, one zero-line current sampling channel, and three voltage sampling channels. The buffer has an 8kHz refresh rate. The controller module is based on an STM32F103 microprocessor, which controls the other modules. The wireless communication module includes an ESP-07S, which provides a 2.4GHz 802.11b / g / n radio and a 32-bit MCU for data transmission.

[0018] (3) Beneficial effects

[0019] Compared with the existing technology, the present invention provides a non-intrusive household appliance load identification method and system based on edge-cloud collaboration, which has the following beneficial effects:

[0020] 1. This non-invasive household appliance load identification method and system based on edge-cloud collaboration, the entire appliance identification system forms a system architecture by combining electrical category decomposition and identification network and socket hardware. The invented intelligent electrical identification socket integrates sampling, processing, analysis, and communication, and the recognition accuracy of the load identification algorithm is above 95%.

[0021] 2. This non-invasive household appliance load identification method and system based on edge-cloud collaboration uses 6.4 kHz sampling of analog current signals for data acquisition. At this frequency, the waveform data exhibits significant periodicity and variation. Furthermore, a buffer is set up to receive data from high-refresh-rate registers, ensuring data continuity and combining the advantages of both high- and low-frequency sampling.

[0022] 3. This non-invasive household appliance load identification method and system based on edge-cloud collaboration proposes an edge-cloud collaboration model for data processing, analysis, and communication, establishing training and inference databases to provide data support for different functions. This edge-cloud collaboration model combines the advantages of both the edge and cloud. Edge-side reasoning addresses user-sensitive appliance identification delays, while offloading model training tasks, which are less noticeable to users, to the cloud. The entire model operates asynchronously, with the edge and cloud each completing different tasks. The system communicates between the edge and cloud only when necessary, reducing pressure on the cloud.

[0023] 4. This non-intrusive household appliance load identification method and system, based on edge-cloud collaboration, combines current data with time series data, innovatively introducing models from the time series field. This improved model forms the load identification model mWDN-RFV (multilevel Wavelet Decomposition Network-Residual Feature Vector). This model achieves an average class error of 1.4% on the public UCR time series dataset and an average classification error of 2.7% on a self-built appliance dataset.

[0024] 5. This non-intrusive household appliance load identification method and system, based on edge-cloud collaboration, can assist smart home systems in better managing electrical devices, eliminating the need for users to manually enter device information and making appliances more automated. This can be used in home life scenarios to provide users with real-time appliance status. It can also be used in places like schools to prevent the connection of dangerous appliances. When a dangerous appliance is identified, the system can automatically cut off power to ensure safety.

[0025] 6. This non-invasive household appliance load identification method and system based on edge-cloud collaboration has an original logical architecture and hardware structure. There is currently no device with integrated collection and computing capabilities that has load identification capabilities. Compared with the existing load identification algorithm, the load identification model mWDN-RFV introduces a multi-level wavelet decomposition network, which makes the data show new features in both time and frequency domains. It defines a triplet loss function based on Euclidean distance and optimizes the model to the best performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a structural diagram of the edge-cloud collaborative system of the present invention;

[0027] Figure 2 Schematic diagram of the learning process of the edge coordination and reasoning layer of the present invention;

[0028] Figure 3 Schematic diagram of the mWDN-RFV model structure of the present invention;

[0029] Figure 4 Schematic diagram of the ψ(·) function structure in the mWDN-RFV model of the present invention;

[0030] Figure 5 Schematic diagram of the mWDN-RFV model training flow chart of the present invention;

[0031] Figure 6 This is a schematic diagram of the comparison of the classification performance of the UCR dataset of the present invention;

[0032] Figure 7 This is a schematic diagram of the experimental results of the electrical appliance dataset of the present invention;

[0033] Figure 8 This is a schematic diagram of the confusion matrix of the test set of the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] See also Figure 1-8A method and system for non-invasive household appliance load identification based on edge-cloud collaboration, wherein the non-invasive household appliance load identification includes training a multi-level wavelet decomposition network-residual feature vector model based on the voltage and current data of the household appliance load known in the database on the cloud, and obtaining a high-precision appliance category classification model through the triplet loss function. The real-time voltage and current raw data of various electrical loads at the entrance of the target user's home are collected at the edge smart device, and the appliance category is classified and the working status of the appliance is recorded in real time through the inference model deployed at the edge. The feature information of unrecognized appliances will be uploaded to the cloud for model update, and the model weights will be sent to the edge end regularly to improve the recognition accuracy.

[0036] The proposed multi-level wavelet decomposition network-residual eigenvector model captures time series features in both the time and frequency domains through the wavelet decomposition network, achieving highly accurate time series data analysis. Using a triplet loss function, the distinction between different electrical categories is enhanced, further increasing the classification accuracy of the model. Testing has demonstrated classification accuracy exceeding 95%.

[0037] This method utilizes cloud-edge collaboration for deployment. It leverages the computing power of large cloud servers to train a large amount of appliance category and feature data stored in a database. The trained weights are then periodically transmitted to the edge for model inference. Edge devices can then use lower-frequency chips for real-time inference to obtain the final classification results for user appliances.

[0038] Edge-cloud collaboration can also be used to manually label appliances with low confidence levels when classified at the edge. This can be uploaded to the cloud to update model weights, thereby continuously improving the model's coverage of household appliances.

[0039] The system utilizes an edge data acquisition control board and a cloud server. The board samples analog current signals at a 6.4 kHz frequency and uses an energy metering chip to acquire real-time current, voltage, and power data. This data is filtered, decomposed, and inferred using a 1.43 GHz quad-core Cortex-A57 CPU. The board also features network communication capabilities and can send recorded appliance classification results and status to the cloud for display.

[0040] 1. Edge-Cloud Collaborative System Structure

[0041] The overall structure of the electrical appliance identification system is as follows Figure 1As shown in Figure 2, the edge-cloud collaborative system architecture consists of cloud servers and edge intelligent devices. The edge intelligent devices are logically abstracted into a data acquisition layer, a data aggregation layer, and an edge collaboration and inference layer. The data acquisition layer is responsible for capturing electrical characteristics from electrical appliances. This data from multiple electrical devices is aggregated in the data aggregation layer. The edge collaboration and inference layer is responsible for interacting with the cloud and executing model inference.

[0042] 1) Data Collection Layer

[0043] The data acquisition layer is designed to capture the electrical characteristics of devices at both high and low frequencies. The system uses a 6.4 kHz sampling frequency to capture 128 voltage and current data points in a cyclical manner, reflecting both overall trends and instantaneous changes. At this sampling frequency, the electrical characteristics of different appliances are collected separately by the data acquisition layer within the system.

[0044] 2) Data aggregation layer

[0045] Data from the acquisition layer is aggregated at the data aggregation layer of the edge intelligent device. Electrical data is temporarily stored in the training database. This data is periodically retrieved by the cloud server for further analysis. Simultaneously, the inference database is used to update the weight parameters of the inference model. The training database is typically a relational database, while the inference database is non-relational and uses a cache to store data.

[0046] 3) Edge coordination and reasoning layer

[0047] The edge coordination and reasoning layer establishes a transmission link with the cloud server, such as Figure 2 As shown in Figure 2, the model is deployed on the server side, and the data used for inference and training comes from the database at the data aggregation layer. Edge intelligent devices collaborate with cloud servers in the following stages.

[0048] Phase 1: Initial training model.

[0049] The cloud server trains the first neural network from the initial dataset stored in the training database and generates a weight file after training is complete.

[0050] Phase 2: Obtaining model weights.

[0051] Once neural network training is complete, the server sends the weight file to the data coordination and inference layer of the edge smart device. The model deployed on the edge smart device then obtains the initial weights for performing model inference and returns the inference results to the cloud. Ultimately, the cloud server consolidates the inference results and displays the real-time status of the appliance to the user.

[0052] Phase 3: Update data.

[0053] During the inference process performed by edge intelligent devices, when a new device category is added to the system, device data is automatically extracted and added to the training database. This data is then sent to the cloud server to prompt model training. After training a new neural network based on the latest dataset, the cloud server sends the updated weight file to the edge intelligent device. During this phase, an asynchronous mechanism is used, with the model update process occurring on the server and the edge model inference performed at the edge. Data updates are continuously performed, and the model is dynamically updated to adapt to changes in the real environment.

[0054] 2. mWDN-RFV Model Design

[0055] The trained model is designed to extract internal features and classify them into different categories. The instantaneous voltage and current data of the equipment are used as the input of the model. In fact, the classification of these electrical data can be achieved through the time series classification model. Wavelet decomposition is similar to multi-level discrete wavelet decomposition (MDWD), which is a well-known method for capturing time series features in the time domain and frequency domain. On this basis, the multi-level wavelet decomposition network (mWDN) was proposed in the field of time series analysis to extract sequence feature information. Combined with mWDN, we proposed an improved multi-level wavelet decomposition network-residual feature vector (mWDN-RFV) model, which is divided into mWDN structure and RFV structure. The overall structure of the model is as follows Figure 3 shown.

[0056] 1) Application of MDWD algorithm

[0057] In the MDWD algorithm, each level of wavelet function includes a low-pass filter and a high-pass filter. The input time series data is recorded as s={s1,...,s t ,...,s T}, decomposed by wavelet transform to obtain a low-frequency subsequence and a high-frequency subsequence, expressed as s l (i) and s h (i), i represents the i-th level decomposition. Low-pass filter l={l1,...,l k ,...,l K} and high-pass filter h={h1,...,h k ,...,h K}, K<<T is used for the (i+1)th level of MDWD. The following formula is used to further convolve the obtained low-frequency subsequence data

[0058]

[0059] Here represents the nth element of the i-th level low-frequency subseries, s l (0) is the input sequence. The low-frequency and high-frequency subsequences s in the i-th levell (i) and s h (i) is composed of the intermediate variable sequence v l (i) and v h (i) is generated by 1 / 2 downsampling. Subsequence set S(i) = {s h (1),s h (2),...,s h (i),s l (i)} is called the i-level decomposition result of s.

[0060] 2) mWDN model structure

[0061] When MDWD is combined with neural network, it becomes mWDN, and its structure is as follows: Figure 3 As shown. The mWDN model uses the following formula to decompose the time series hierarchically

[0062]

[0063] in Is an activation function. The low-frequency and high-frequency sub-level sequences s in the i-th level l (i) and s h (i) is the intermediate variable v l (i) and v h (i) After inputting the average pooling layer, the average pooling layer is

[0064]

[0065] Set the filter coefficients to the weight matrix W l and W h This is how the convolution operation is implemented.

[0066] 3) RFV network

[0067] The RFV network uses the residual flow method to extract feature vectors, and the triplet loss function is used to train the network to obtain the optimal feature vector. Then the support vector machine (SVM) is trained as a classifier to classify the obtained feature vector to determine the category of the appliance.

[0068] RFV network connection in the mWDN generated by the i-th level sub-sequence s h (i) and s l (i) Then, using the feedforward neural network

[0069] u(i)=ψ(s h (i),s l (i),θ ψ )

[0070] To connect, where θ ψRepresents a trainable parameter. The structure of ψ(·) is as follows Figure 4 As shown in , it consists of 5 residual blocks and 1 fully connected layer. Each residual block consists of three groups of networks, each of which includes a one-dimensional convolution layer, a batch normalization layer, and a linear rectification function layer. The convolution operation in each residual block is completed by three one-dimensional convolution kernels of sizes {9, 5, 3}, and the number of each convolution kernel in the 5 residual blocks is {4, 16, 16, 32, 16}. The result of the residual block is input into the fully connected layer so that the feature vector can learn global information, and finally the output result u(i) is obtained. The result after each level of subsequence input ψ(·) is connected using the residual learning method, which is specifically expressed as the following formula

[0071] v(i)=v(i-1)+u(i)

[0072] Where v(i) represents the feature vector obtained after residual learning of the i-th level subsequence, and the last level v(i) represents the final feature vector.

[0073] The feature vectors of the same category should be as close as possible, while the feature vectors of different categories need to be kept far apart. Therefore, the Euclidean distance is used to calculate the difference between feature vectors, which is expressed as the following formula

[0074]

[0075] Where R represents the total number of elements contained in the eigenvector, s ai Represents sequence s a The i-th element in s bi Represents sequence s b The i-th element in , f(·) represents the feature extractor, whose output is the final feature vector.

[0076] The Euclidean distance provides the basis for the loss function, but it is also necessary to filter out data that is not meaningful for training. Therefore, a triple is defined to select data suitable for training. The triple is constructed as follows: a sample is randomly selected from the training data set, called Anchor (denoted as s a ), and then select a sample that belongs to the same category as the Anchor, called Positive (denoted as s p ), and a sample belonging to a different category from the Anchor, called Negative (denoted as s n ). A triplet contains the data of 3 samples.

[0077] After determining the triplets, the triplet loss function can be implemented as follows

[0078]

[0079] Where N represents the number of triplets, and α is a bias value that prevents the result of the triple loss function from approaching 0. In order to make the loss function effective for training, the selection of triplets is particularly important. The method for selecting triplets is as follows

[0080]

[0081] The entire dataset contains many triplets. Assuming the dataset contains N time series in total, and each category has K time series, the number of triplets is on the order of N × N × K. If a triplet is selected for each data point during training, training efficiency will be severely impacted. Therefore, selecting a triplet when training a batch of data balances training efficiency and triplet optimization. Specifically, the training strategy involves extracting a triplet based on the feature vector during training for each batch of data. Once all triplets in the dataset are obtained, the loss function for these triplets is calculated and the filter coefficients of the mWDN are updated. This entire process is repeated until convergence is achieved or a specified number of training iterations are completed.

[0082] Support Vector Machines (SVMs) are a method for partitioning the entire feature space according to categories. By training and adjusting the triplet loss function, subsequent SVM classification can be considered as an approximately linearly separable problem. The final classifier can be obtained by training the trained feature vectors in the SVM.

[0083] The pseudo code of the entire training process of mWDN-RFV is shown in Algorithm 1.

[0084] 3. Edge Intelligent Devices

[0085] IoT-based edge smart devices are designed to retrieve and analyze electrical data, which is also used to evaluate system performance in real-world environments. The hardware of the edge smart device consists of an analog front end, an MCU, a Wi-Fi module, and an OLED display. Electrical characteristics, including transient current and voltage information, are captured through data acquisition circuitry. It includes a ZMPT107 transformer and an RN8302B three-phase energy measurement chip. The chip features seven 24-bit ADC channels, three phase current sampling channels, one zero-line current sampling channel, and three voltage sampling channels. The buffer has an 8kHz refresh rate. The controller module is based on the STM32F103 microprocessor, which controls the other modules. The wireless communication module includes the ESP-07S, which provides a 2.4GHz 802.11b / g / n radio and a 32-bit MCU for data transmission. 802.11b / g / n is a series of wireless local area network (WLAN) technology standards developed by the Institute of Electrical and Electronics Engineers (IEEE).

[0086] The algorithm model was tested on both public and private datasets, as follows:

[0087] 1) UCR public dataset test

[0088] UCR is a public time series dataset, each of which is labeled with a specific sample category. It is widely used in the field of time series mining and contains 44 data sets. We selected 13 relatively representative datasets for our experiments.

[0089] Five classification models are compared, including (1) multi-layer perception (MLP) (2) fully convolutional network (FCN) (3) residual network (ResNet) (4) multi-level wavelet decomposition network-residual classification flow model (mWDN-RCF) (5) mWDN-RFV, and the feature vector dimension of mWDN-RFV is set to 128. In these models, MLP has 3 layers of fully connected layers, each with 500 neurons, and ends with a softmax layer. FCN consists of three convolutional layers and a softmax layer. ResNet uses three layers of residual blocks and a softmax layer. In the mWDN-RCF model, the decomposition results of each level of mWDN are input into the classifier for classification, and the residual learning method is used to connect all classifiers to obtain the final result.

[0090] For each UCR dataset, each classification model was run 10 times and the average classification error rate of the test set was returned for evaluation. The results are shown in Table 1. The winning time represents the number of times the model outperformed other models in these 13 datasets. In order to compare the overall performance of the models on these 13 datasets, we used the mean per-class error (MPCE) as the performance metric for each model. k Indicates the number of categories of the kth dataset, e k Represents the error rate of the model, and the MPCE of the model is defined as

[0091]

[0092] When MPCE is used for evaluation, the influence of the number of categories is ignored, and the smaller the MPCE value, the better the overall performance. The evaluation results show that the model based on the multi-level wavelet decomposition network outperforms the first three benchmark models, indicating that the use of multi-level wavelet decomposition as the basis for time series decomposition problems significantly improves the classification accuracy. Secondly, compared with mWDN-RCF, mWDN-RFV has a lower average classification error rate and a smaller MPCE value. This result also proves that using the RFV network structure has better performance than directly using the RCF network structure. The reason is that as the number of categories or the number of imbalanced categories increases, the accuracy of mWDN-RFV will drop significantly. Figure 6 shown.

[0093] 2) Private appliance dataset testing

[0094] We created a private appliance dataset for further analysis. We collected data from seven types of appliances using IoT edge smart devices, with 240 sets of data for each. Each set of data consisted of a sequence of 1,000 points. Figure 8 Several data examples are shown. The seven categories include light bulbs, soldering irons, razors, projectors, kettles, empty loads, and computers. The training and validation sets are split in a 7:3 ratio, and a test set of 60 data sets for each category is created.

[0095] After testing the aforementioned appliance dataset, we calculated the average classification error rate. The average classification error rate of the models is shown in Table 2. It can be seen that mWDN-RFV outperforms other models in terms of accuracy, further demonstrating the effectiveness of the mWDN-RFV model. Figure 8 The classification effect of the trained model on the test set is shown in the form of a confusion matrix. The horizontal axis represents the predicted category of the sample, and the vertical axis represents the true category of the sample. It can be seen that except for a few misclassified data, the classification results of other categories are correct. Figure 8 shown.

[0096] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A non-intrusive household appliance load identification system based on edge-cloud collaboration, characterized by: The system is an edge-cloud collaborative system, consisting of cloud servers and edge intelligent devices. The edge intelligent devices are logically abstracted into a data acquisition layer, a data aggregation layer, and an edge collaboration and reasoning layer. The data acquisition layer is responsible for capturing electrical characteristics from electrical appliances. These data from multiple electrical devices are aggregated in the data aggregation layer. The edge collaboration and reasoning layer is responsible for interacting with the cloud and executing model reasoning. Edge intelligent devices collaborate with cloud servers in the following stages; The method is as follows: Phase 1: Initial training model, the cloud server trains the first neural network TD0 from the initial data set N0 stored in the training database, and generates the weight file W0 after the training is completed; Phase 2: Obtaining model weights. After the neural network N0 is trained, the server sends the weight file W0 to the data coordination and inference layer of the edge smart device. The model deployed in the edge smart device obtains the initial weights for performing model inference and returns the inference results to the cloud. Ultimately, the cloud server unifies the inference results and displays the real-time status of the appliance to the user. Phase 3: Update data. During the inference process performed by the edge intelligent device, when a new device category is added to the system, the device data will be automatically extracted and added to the training database, and the TD1 data will be sent to the cloud server to prompt the model training. After the cloud trains the new neural network N1 based on the latest data set, it will send the updated weight file W1 to the edge intelligent device. In this phase, an asynchronous mechanism is adopted to perform the model update process on the server and perform edge model inference on the edge. The data stage update operations are continuously performed, and the model is dynamically updated to adapt to changes in the real environment.

2. The non-intrusive household appliance load identification system based on edge-cloud collaboration according to claim 1 is characterized by: The data acquisition layer is designed to capture the electrical characteristics of the equipment at high / low frequencies. The system uses a sampling frequency of 6.4KHz to capture 128 voltage and current data in a cyclic manner, so as to reflect both overall trends and instantaneous changes. At this sampling frequency, the electrical characteristics of different electrical appliances are collected separately by the data acquisition layer in the system.

3. The non-intrusive household appliance load identification system based on edge-cloud collaboration according to claim 1 is characterized by: The data aggregation layer aggregates the data from the acquisition layer on the edge smart device. The electrical data will be temporarily stored in the training database. These data will be regularly extracted by the cloud server for further analysis. At the same time, the inference database will be used to update the weight parameters of the inference model. The training database tends to use a relational database, while the inference database is a non-relational database and uses cache to store data.

4. The non-intrusive household appliance load identification system based on edge-cloud collaboration according to claim 1 is characterized by: The edge coordination and reasoning layer establishes a transmission link with the cloud server, the model is deployed on the server side, and the data used for reasoning and training comes from the database of the data aggregation layer.

5. The non-intrusive household appliance load identification system based on edge-cloud collaboration according to claim 1 is characterized by: The trained model is designed to extract internal features and classify them into different categories. The instantaneous voltage and current data of the equipment are used as input to the model. The electrical data is classified through a time series classification model, an improved multi-level wavelet decomposition network-residual feature vector (mWDN-RFV) model, which is divided into an mWDN structure and an RFV structure. The model is based on the MDWD algorithm. MDWD is a multi-level discrete wavelet decomposition algorithm that captures time series characteristics in the time domain and frequency domain. In the MDWD algorithm, each level of the wavelet function includes a low-pass filter and a high-pass filter. The input time series data is denoted as s = {s1,...,s t ,...,s T }, decomposed by wavelet transform to obtain a low-frequency subsequence and a high-frequency subsequence, expressed as s l (i) and s h (i), i represents the i-th level decomposition, and the low-pass filter l={l1,...,l k ,...,l K } and high-pass filter h={h1,...,h k ,...,h K }, K<<T is used for the (i+1)th level of MDWD, and the formula is used to perform convolution operation on the low-frequency subsequence data to prepare for feature extraction of the subsequent model; By combining MDWD with neural networks, we can produce the mWDN model, which has been proposed in the field of time series analysis to extract sequence feature information. The mWDN model uses the following formula to decompose the time series hierarchically: in is an activation function, the low-frequency and high-frequency sub-level sequences s in the i-th level l (i) and s h (i) is the intermediate variable, v l (i) and v h (i) After inputting the average pooling layer, the average pooling layer is Set the filter coefficients to the weight matrix W l and W h In this way, convolution operation is realized to extract sequence feature information; The RFV network uses the residual flow method to extract feature vectors, and uses the triplet loss function to train the network to obtain the best feature vectors. Then the support vector machine (SVM) will be trained as a classifier to classify the obtained feature vectors, thereby determining the category of the electrical appliance and connecting it to the i-th level subsequence s generated by the mWDN. h (i) and s l (i) After that, use the forward neural network u(i) = ψ(s h (i),s l (i),θ ψ ) are connected, where θ ψ Represents trainable parameters; Its structure consists of 5 residual blocks and 1 fully connected layer. Each residual block consists of three groups of networks, each of which includes a one-dimensional convolutional layer, a batch normalization layer, and a linear rectification function layer. The network is trained using a triplet loss function to obtain the optimal feature vector. The triplet consists of anchor, positive, and negative samples. The difference between the feature vectors is calculated using Euclidean distance to optimize the model. During training, a triplet is selected from each batch of data, the loss function is calculated, and the filter coefficients of the mWDN are updated until convergence is achieved or a specified number of training iterations are completed.

6. The non-intrusive household appliance load identification system based on edge-cloud collaboration according to claim 1 is characterized by: The edge intelligent device for the appliance identification system is designed to retrieve and analyze electrical data. This data is also used to evaluate the performance of the system in a real-world environment. The edge intelligent device consists of an analog front end, an MCU part, a Wi-Fi module, and an OLED display. The electrical characteristics, including transient current and voltage information, are captured through the data acquisition circuit, which includes a ZMPT107 transformer and a three-phase power measurement chip RN8302B. The chip includes 7 24-bit ADC channels, 3 phase line current sampling channels, 1 Zero Line current sampling channel, and 3 voltage sampling channels. The refresh rate of the buffer is 8kHz. The controller module is based on the STM32F103 microprocessor, which controls other modules. The wireless communication module includes ESP-07S, which provides a 2.4GHz radio and a 32-bit MCU for data transmission.

Images