A non-intrusive device state recognition method and device based on federated learning

Abstract

The application relates to a non-intrusive device state recognition method and device based on federal learning, and the method comprises the following steps: arranging a non-intrusive device data acquisition system in a plurality of local user nodes, high-frequency sampling of total current data of a trunk line in the local user nodes, and low-frequency sampling of power data of each parallel branch in the local user nodes; constructing a historical data set; constructing a deep learning neural network, arranging a feature extraction subnetwork in the local user nodes, and arranging a device prediction subnetwork in the cloud; training the deep learning neural network through the historical data set, and performing data transmission between the local user nodes and the cloud through an encryption algorithm during the training process; and using a trained recognition model to recognize the device state. Compared with the prior art, the application solves the data reliability problem, solves the data transmission bottleneck problem existing in high-frequency sampling, and solves the user privacy security problem in the training and prediction processes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of non-intrusive device identification, and in particular to a non-intrusive device state identification method and device based on federated learning. BACKGROUND

[0002] In a complex power system such as a building with a large number of electrical equipment, identifying and determining the state of the equipment is the focus of researchers. Non-intrusive device identification is a method that only uses a set of sensors to measure aggregate data to obtain the running state of various devices in the internal system. This method does not need to involve many sets of sensors, has the characteristics of convenient deployment, low price and easy popularization, and is widely used.

[0003] In the non-intrusive device identification method, from the perspective of signal acquisition, the first type is to analyze 50Hz alternating current, and the sampling frequency is from several Hz to several Hz. Generally speaking, such low-frequency data is used to analyze the state switching between steady states (such as on and off), and the data itself is mainly average power, power factor, etc. The second type is to analyze 50Hz alternating current and its harmonics, and the sampling frequency is from 1kHz to several tens of kHz. Most household electrical equipment is nonlinear, such as televisions and LED lights, so important harmonic distortion information is contained in the current waveform. In order to capture such detailed information, a higher sampling rate is necessary. The third type is to analyze high-frequency transient signals, and the sampling frequency is from several tens of kHz to several tens of MHz. Most electronic devices use switch mode power supply (SMPS), and its frequency band is in this interval. From the perspective of data processing, most of the existing research work focuses on supervised machine learning techniques, including traditional Markov models and deep learning models.

[0004] In data acquisition, data acquisition can be high-frequency transient signals or lower-frequency steady-state signals. Undoubtedly, high-frequency transient signals contain more information and can provide more accurate identification. At the same time, low-frequency steady-state signals can be obtained by sampling, and there is no doubt that it is the first choice for data acquisition. However, high-frequency data sampling will result in large flow of data transmission, which will replace the computing power bottleneck and become a greater bottleneck, significantly limiting the carrying capacity of the server.

[0005] In data processing, deep learning models are obviously more optimal. Intrusive device identification systems require a large amount of training data, and these training data come from a large number of users, and it is necessary to aggregate the data of different users to form a training set before neural network training. In the data aggregation link, how to protect the privacy and security of each user has become a challenge that must be faced. Potential attackers may steal the original training data, identify the device usage, and further infer the user's state information, and then achieve the purpose of stealing the user's privacy.

[0006] In summary, large flow data transmission and user privacy security are problems to be solved in the non-intrusive device state identification method. SUMMARY

[0007] The purpose of the present application is to overcome the defects of the prior art and provide a non-intrusive device state identification method and device based on federated learning.

[0008] The purpose of the present application can be achieved by the following technical solutions:

[0009] A non-intrusive device state identification method based on federated learning, comprising the following steps:

[0010] S1, arranging a non-intrusive device data acquisition system in a plurality of local user nodes, high-frequency sampling the total current data of the main line in the local user nodes, and low-frequency sampling the power data of each parallel branch in the local user nodes, the local user nodes being home, laboratory, office and the like environment;

[0011] S2, constructing a historical data set according to the data of each local user node;

[0012] S3, constructing a deep learning neural network, including a feature extraction subnetwork and a device prediction subnetwork, arranging the feature extraction subnetwork in the local user nodes, and arranging the device prediction subnetwork in the cloud;

[0013] S4, training the deep learning neural network through the historical data set, the local user nodes and the cloud transmitting data through an encryption algorithm during the training process, and obtaining a trained identification model;

[0014] S5, using the trained identification model to identify the device state and monitor the running state of each device.

[0015] Further, the non-intrusive device data acquisition system in step S1 non-intrusively acquires the total current data of the main line in the local user nodes at a frequency of 180 kHz, and acquires the power data of each parallel branch in the local user nodes at a frequency of 1 Hz. Each device in the local user nodes is connected to the parallel branch.

[0016] Further, step S2 comprises the following steps:

[0017] S21, acquiring the total current data and power data continuously acquired by the non-intrusive device data acquisition system;

[0018] S22, setting the minimum starting power P min_start , the maximum stopping power P max_end , the shortest duration t min and the longest interruption time t maxThe power data is processed, and a streaming algorithm is used to determine the working state of each device according to the power data, including starting and stopping.

[0019] S23, introducing a sliding window, applying a streaming algorithm to converge the working state of each device for the data between the current time t and the previous time t-1, retaining valid data according to the pre-set judgment rule, and adding a label to the device according to the working state to construct a historical data set in each local user node.

[0020] Further, the pre-set judgment rule is: if the working state of all devices does not switch from starting to stopping or from stopping to starting within the same window, the window data is considered legal; if all devices are in the stopping state within the same window, the window is considered invalid.

[0021] Further, the deep learning neural network is used to realize multi-label classification, the feature extraction sub-network is a residual network with one-dimensional convolution units as the basic module, the device prediction sub-network is a residual network of multi-layer perceptron, the loss function used for multi-label classification is Asymmetric Loss, the feature extraction sub-network and the device prediction sub-network are optimized by model compression, pruning and other methods, and related parameters are adjusted to reduce the amount of calculation while ensuring the accuracy, so that it can be deployed on edge devices.

[0022] Further, the training process of the deep learning neural network is as follows:

[0023] S41, for each local user node, obtaining a historical data set and preprocessing, sending the historical data set into the feature extraction sub-network to obtain a calculation result;

[0024] S42, the calculation result in each local user node is encrypted by an encryption algorithm and uploaded to the cloud;

[0025] S43, the cloud summarizes the calculation results of all local user nodes, and trains the device prediction sub-network in the cloud, and returns the gradient to each local user node;

[0026] S44, the local user node updates the feature extraction sub-network according to the received gradient.

[0027] Further, the encryption algorithm is a differential encryption algorithm.

[0028] Further, the cloud returns the gradient to each local user node, the structure of the feature extraction sub-network of each local user node is independent of each other, and the update of the feature extraction sub-network of each local user node is independent of each other.

[0029] A non-intrusive device state recognition device based on federated learning, comprising:

[0030] A non-intrusive device data acquisition system is arranged in each local user node to sample the total current data of the main line in the local user node at high frequency and sample the power data of each parallel branch in the local user node at low frequency.

[0031] A local data set construction module is arranged in each local user node to construct a historical data set according to the data of each local user node.

[0032] A local computing unit is arranged in each local user node and is in communication connection with the local data set construction module and stores a feature extraction sub-network to perform calculation and update of the feature extraction sub-network in the training process.

[0033] A cloud computing unit is arranged in the cloud and stores a device prediction sub-network to perform calculation and update of the device prediction sub-network in the training process, and the local computing unit and the cloud computing unit perform data transmission through an encryption algorithm in the training process.

[0034] A detection module uses the trained recognition model to recognize the device state.

[0035] Further, a high-frequency digital-to-analog conversion chip is used to non-intrusively acquire the total current data of the main line in the local user node at a frequency of 180 kHz, and a smart socket with a power measurement function is used to acquire the power data of each parallel branch in the local user node at a frequency of 1 Hz, and each device in the local user node is connected to the parallel branch.

[0036] Further, the total current data is transmitted to the local data set construction module in a wired transmission manner, and the smart socket has a wireless communication function, and the power data is transmitted to the local data set construction module in a wireless transmission manner.

[0037] Compared with the prior art, the present application has the following beneficial effects:

[0038] (1) High-frequency sampling of total current and data processing through a deep learning neural network, high-frequency sampling can retain more information, and at the same time, the power data on the parallel branch of each device is collected, the legality and effectiveness are verified based on the power data, and a label is added, solving the data reliability problem.

[0039] (2) The deep learning neural network is divided into a local feature extraction sub-network and a cloud device prediction sub-network, the high-frequency data collected locally is directly calculated locally, and the calculation results and gradients are transmitted between the cloud and the local, the data transmission amount is small, and the data transmission bottleneck problem existing in high-frequency sampling is solved.

[0040] (3) The features extracted by the local feature extraction subnetwork are sent to the cloud through an encryption algorithm, and the cloud returns the gradient for updating. On the one hand, the encryption algorithm ensures the security of data transmission, and on the other hand, the calculation result and the gradient cannot be traced back to the original state data of the user, solving the problem of user privacy security in the training and prediction process. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 A flowchart of a device state recognition method;

[0042] Figure 2 An architectural diagram of a device state recognition system;

[0043] Reference signs: 1, non-intrusive device data acquisition system, 2, local data set construction module, 3, local computing unit, 4, cloud computing unit. DETAILED DESCRIPTION

[0044] The present application will be described in detail below in conjunction with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and gives a detailed implementation manner and specific operation process, but the protection scope of the present application is not limited to the following embodiments.

[0045] In the drawings, components of the same structure are denoted by the same reference numerals, and components having similar structures or functions are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present application does not limit the size and thickness of each component. In order to make the drawing clearer, the components are appropriately exaggerated in some places in the drawing.

[0046] Example 1:

[0047] A non-intrusive device state recognition method based on federated learning, as shown in Figure 1 includes the following steps:

[0048] S1, arranging a non-intrusive device data acquisition system in a plurality of local user nodes, high-frequency sampling the total current data of the main line in the local user node, and low-frequency sampling the power data of each parallel branch in the local user node, the local user node being a home, a laboratory, an office, etc.

[0049] The non-intrusive device data acquisition system non-intrusively acquires the total current data of the main line in the local user node at a frequency of 180 kHz, and acquires the power data of each parallel branch in the local user node at a frequency of 1 Hz. Each device in the local user node is connected to the parallel branch.

[0050] Specifically, in the present embodiment, the following is true:

[0051] Step S11: high-frequency digital-analog conversion chip, design sampling circuit, and non-invasive collection of total current data I at a frequency of 180 kHz 总 The total current is the sum of the currents of each parallel branch, and theoretically n represents the total number of devices, where I i is the current of each parallel branch, each device is connected to the parallel branch, and there is no high-frequency filter device between each branch;

[0052] Step S12: use multiple smart sockets with wifi communication and power measurement functions to collect low-frequency power data P of each parallel branch at a frequency of 1 Hz i ;

[0053] Step S13: deploy a micro host as a local computing center to obtain high-frequency total current data through usb wired transmission and low-frequency power data through local area network wifi wireless transmission. The software buffer layer writes data quickly into memory and continuously writes into the hard disk of the micro host, ensuring reliable data transmission.

[0054] S2, construct historical data sets according to the data of each local user node;

[0055] S21, obtain total current data and power data continuously collected by the non-invasive device data acquisition system, such as power data and total current data of devices with high-frequency characteristics in actual use within a week, half a month, etc. after the deployment of the non-invasive device data acquisition system is completed;

[0056] S22, according to expert experience, set the minimum startup power P min_start , maximum stop power P max_end , minimum duration t min , and maximum interruption time t max of each device, process the power data, and use a streaming algorithm to determine the working state of each device, including startup and shutdown; when the real-time power of the device is greater than P min_start and the duration is longer than t min , it can be considered that the device has started; when the real-time power of the device is less than P max_end and the duration is longer than t max , it can be considered that the device has been turned off;

[0057] S23, introduce a sliding window, for the data between the current time t and the previous time t-1s, apply the stream algorithm to converge the working status of each device, according to the pre-set judgment rule, retain the legal and effective data, and according to the working status, add a label to the device, and construct a historical data set in each local user node. For example: if the working status of all devices in the same window does not change from start to stop or from stop to start, the window data is considered legal; if all devices are in the off state in the same window, the window is considered invalid. For legal and effective data, record the label and save it to construct a historical data set. The data sample is the total current data and the corresponding device status, and the rest of the data is discarded. Power data is used to determine the legality and effectiveness of the data and the determination of the label. In this way, according to the power, the label can be automatically added and the data can be screened, greatly reducing the manual work.

[0058] S3, construct a deep learning neural network, including a feature extraction subnetwork and a device prediction subnetwork, the feature extraction subnetwork is arranged in the local user node, and the device prediction subnetwork is arranged in the cloud;

[0059] A hybrid model based on a neural network, namely a deep learning neural network, is constructed. The deep learning neural network is used to realize multi-label classification of input data. The hybrid model is modified to be divided into a local calculation part and a cloud calculation part. In this embodiment, the feature extraction subnetwork is a residual network with a one-dimensional convolution unit as a basic module, which is arranged in the local micro host (edge device). The device prediction subnetwork is a residual network of a multi-layer perceptron, which is arranged in the cloud server. The loss function used for multi-label classification is Asymmetric Loss.

[0060] The high-frequency total current data collected at 180kHz is a high-dimensional data. If dimension reduction processing is performed, the original intention of using high-frequency data to retain more information is violated. If a conventional neural network is used, a large number of input neurons need to be designed, resulting in a complex network structure. Convolutional networks are generally used for image processing. Taking a 240x240 picture as an example, the input is actually a 240x240 pixel matrix, which can process high-dimensional data. Therefore, the feature extraction subnetwork of the present application is a residual network with a one-dimensional convolution unit as a basic module. The high-frequency data is extracted by convolution operation, which reduces the complexity of the feature extraction subnetwork, and the calculation amount of the local edge device can meet the requirements.

[0061] In order to cope with the output of multi-label classification, the device prediction subnetwork is a residual network of a multi-layer perceptron, and the loss function used for multi-label classification is Asymmetric Loss.

[0062] Finally, the feature extraction sub-network and the device prediction sub-network are optimized by model compression, pruning and other methods, and the related parameters are adjusted to reduce the calculation amount while ensuring the accuracy, so that they can be deployed on local edge devices.

[0063] The specific structure of the feature extraction sub-network and the device prediction sub-network is not limited in this embodiment, and related practitioners can design as needed, which can apply a variety of high-performance neural network hybrid modeling to ensure high accuracy at low computational cost.

[0064] S4, training a deep learning neural network through a historical data set, during the training process, the local user node and the cloud transmit data through an encryption algorithm, and obtain a trained recognition model;

[0065] The training process is as follows:

[0066] S41, for each local user node, a historical data set is obtained and preprocessed, including denoising, normalization and other preprocessing operations, and the historical data set is input into the feature extraction sub-network to obtain a calculation result;

[0067] S42, the calculation result in each local user node is encrypted through an encryption algorithm and uploaded to the cloud, such as a differential encryption algorithm;

[0068] S43, the cloud is equipped with a device prediction sub-network, which integrates the calculation results of all local user nodes, integrates the calculation results, and trains the device prediction sub-network in the cloud, and returns the gradient to each local user node;

[0069] S44, the local user node updates the feature extraction sub-network according to the received gradient.

[0070] Considering that the number and working mode of devices in different local user nodes are different, the structure of the feature extraction sub-network of each local user node is independent of each other, and different structures can be used, so the cloud returns to each local user node is the gradient, which does not affect the update of different feature extraction sub-networks, and the final recognition model can also adapt to each local user node.

[0071] The update of each local user node feature extraction sub-network is independent, and the update according to the gradient is different due to different network structures. Considering the peak and valley of device use, network state and other factors of different local user nodes, the gradient can be returned at different time periods, and each feature extraction sub-network is updated asynchronously.

[0072] S5, using the trained recognition model to identify the device state and monitor the running state of each device.

[0073] After the training is completed, the device prediction subnetwork can also be arranged locally, that is, the local edge device is equipped with an identification model. Real-time total current data is obtained locally, and the real-time total current data is sent to the identification model, so that the state of each device can be monitored in real time, and analysis can be performed, such as the operating state of each device in a target time period. The device prediction subnetwork can also be arranged in the cloud, and the local edge device is equipped with a feature extraction subnetwork. Real-time total current data is sent to the feature extraction subnetwork, and the features extracted by the feature extraction subnetwork are sent to the cloud device prediction subnetwork to obtain a device state identification result.

[0074] If a new device is newly introduced in the local user node, steps S1-S4 can be repeated to update the feature extraction subnetwork and the device prediction subnetwork to obtain a new identification model.

[0075] Embodiment 2:

[0076] A non-intrusive device state identification device based on federated learning, as shown in Figure 2 , comprises:

[0077] A non-intrusive device data acquisition system 1 is arranged in each local user node, and is used for high-frequency sampling of total current data of a main circuit in the local user node and low-frequency sampling of power data of each parallel branch circuit in the local user node.

[0078] A local data set construction module 2 is arranged in each local user node, and is used for constructing a historical data set according to data of each local user node collected by the non-intrusive device data acquisition system 1.

[0079] A local computing unit 3 is arranged in each local user node, is in communication connection with the local data set construction module 2, stores a feature extraction subnetwork, and is used for calculation and update of the feature extraction subnetwork in a training process.

[0080] A cloud computing unit 4 is arranged in the cloud, stores a device prediction subnetwork, and is used for calculation and update of the device prediction subnetwork in a training process. In the training process, data transmission is performed between the local computing unit 3 and the cloud computing unit 4 through an encryption algorithm.

[0081] A detection module uses a trained identification model to identify a device state.

[0082] As shown in Figure 2As shown, in the process of building the non-intrusive device data acquisition system 1, a high-frequency digital-to-analog conversion chip is used to design a sampling circuit to non-intrusively collect the total current data of the main line in the local user node at a frequency of 180 kHz, and a smart socket with power measurement function is used to collect the power data of each parallel branch in the local user node at a frequency of 1 Hz, and each device in the local user node is connected to the parallel branch.

[0083] The local data set construction module 2 and the local computing unit 3 can be integrated on a micro host, i.e. an edge computing device, the total current data is transmitted to the local data set construction module 2 in the form of USB wired transmission, and the smart socket has a wireless communication function, and the power data is transmitted to the local data set construction module 3 in the form of WIFI wireless transmission.

[0084] The high-frequency digital-to-analog conversion chip and the sampling circuit are mature technologies for high-frequency sampling, and the micro host and the smart socket are existing designs, and ordinary computers and smart sockets can be used to achieve the same, so the cost is low and the reliability is high.

[0085] The above describes the preferred embodiments of the application in detail. It should be understood that those skilled in the art can make many modifications and changes without creative labor according to the concept of the application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment on the basis of the prior art according to the concept of the application shall be within the protection scope determined by the claims.

Claims

1. A non-intrusive device state recognition method based on federated learning, characterized in that, The method comprises the following steps: S1, arranging a non-intrusive device data acquisition system in a plurality of local user nodes to perform high-frequency sampling on total current data of a main line in the local user nodes and low-frequency sampling on power data of each parallel branch in the local user nodes, wherein the power data of each parallel branch in the local user nodes is collected at a frequency of 1 Hz, each device in the local user nodes is connected into a parallel branch, and the frequency of the low-frequency sampling is 1 Hz; S2, constructing a historical data set according to the data of each local user node respectively; S3, constructing a deep learning neural network, including a feature extraction subnetwork and a device prediction subnetwork, arranging the feature extraction subnetwork in the local user nodes, and arranging the device prediction subnetwork in the cloud; S4, training the deep learning neural network through the historical data set, performing data transmission between the local user nodes and the cloud through an encryption algorithm during the training process, and obtaining a trained identification model; S5, identifying the device state by using the trained identification model; Step S2 comprises the following steps: S21, obtaining total current data and power data continuously collected by the non-intrusive device data acquisition system; S22, set the minimum start power of each device , maximum stop power , minimum duration and maximum interruption time , process the power data, and determine the working state of each device according to the power data using a streaming algorithm; S23, introducing a sliding window, applying a streaming algorithm to converge the working states of each device for the data between a current time t and a previous time t-1s, reserving legal and effective data according to a pre-set judgment rule, adding a label to the device according to the working state, and constructing a historical data set in each local user node, wherein the working state of the device includes starting and stopping, and the pre-set judgment rule is that if the working state of all devices does not switch from starting to stopping or from stopping to starting, the window data is considered legal, and the data sample is the total current data and the corresponding device state.

2. The non-intrusive device state recognition method based on federated learning according to claim 1, characterized in that, The non-intrusive device data acquisition system in step S1 non-intrusively collects the total current data of the main line in the local user nodes at a frequency of 180 kHz.

3. The non-intrusive device state recognition method based on federated learning according to claim 1, characterized in that, The deep learning neural network is used to realize multi-label classification, the feature extraction subnetwork is a residual network with one-dimensional convolution units as basic modules, the device prediction subnetwork is a residual network of a multi-layer perceptron, and the loss function used for multi-label classification is Asymmetric Loss.

4. The non-intrusive device state recognition method based on federated learning according to claim 1, characterized in that, The training process of the deep learning neural network is as follows: S41, for each local user node, obtaining a historical data set and performing preprocessing, and sending the historical data set into the feature extraction subnetwork to obtain a calculation result; S42, uploading the calculation result in each local user node to the cloud after encryption by an encryption algorithm; S43, the cloud summarizes the calculation results of all local user nodes, trains the device prediction subnetwork in the cloud, and returns the gradient to each local user node; S44, the local user node updates the feature extraction subnetwork according to the received gradient.

5. The non-intrusive device state recognition method based on federated learning according to claim 4, characterized in that, The encryption algorithm is a differential encryption algorithm.

6. The non-intrusive device state recognition method based on federated learning according to claim 4, characterized in that, The cloud returns the gradient to each local user node, the structure of the feature extraction subnetwork of each local user node is independent of each other, and the update of the feature extraction subnetwork of each local user node is independent of each other.

7. A non-intrusive device state recognition apparatus based on federated learning, characterized in that, The non-intrusive device state recognition method based on federated learning according to any one of claims 1-6, comprising: a non-intrusive device data acquisition system arranged in each local user node, configured to perform high-frequency sampling on the total current data of the main line in the local user node and low-frequency sampling on the power data of each parallel branch in the local user node, wherein the power data of each parallel branch in the local user node is collected at a frequency of 1 Hz, each device in the local user node is connected to the parallel branch, and the frequency of the low-frequency sampling is 1 Hz; a local data set construction module arranged in each local user node, configured to construct a historical data set according to the data of each local user node; a local computing unit arranged in each local user node and communicatively connected to the local data set construction module, and storing a feature extraction sub-network, configured to perform calculation and update of the feature extraction sub-network during the training process; a cloud computing unit arranged in the cloud and storing a device prediction sub-network, configured to perform calculation and update of the device prediction sub-network during the training process, and the local computing unit and the cloud computing unit perform data transmission through an encryption algorithm during the training process; a detection module configured to recognize the device state using the trained recognition model.

8. The non-intrusive device state recognition apparatus based on federated learning of claim 7, wherein, A high-frequency digital-to-analog conversion chip is used to non-intrusively collect the total current data of the main line in the local user node at a frequency of 180 kHz, and a smart socket with power measurement function is used.

9. The non-intrusive device state recognition apparatus based on federated learning of claim 8, wherein, The total current data is transmitted to the local data set construction module in a wired transmission manner, and the smart socket has a wireless communication function, and the power data is transmitted to the local data set construction module in a wireless transmission manner.

Images