CN114511007B - Non-invasive electrical fingerprint identification method based on multi-scale feature perception - Google Patents
Non-invasive electrical fingerprint identification method based on multi-scale feature perception Download PDFInfo
- Publication number
- CN114511007B CN114511007B CN202210049795.XA CN202210049795A CN114511007B CN 114511007 B CN114511007 B CN 114511007B CN 202210049795 A CN202210049795 A CN 202210049795A CN 114511007 B CN114511007 B CN 114511007B
- Authority
- CN
- China
- Prior art keywords
- layer
- data
- network
- input
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 230000008447 perception Effects 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 title claims abstract description 19
- 238000012549 training Methods 0.000 claims abstract description 29
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 14
- 238000012544 monitoring process Methods 0.000 claims abstract description 13
- 238000007781 pre-processing Methods 0.000 claims abstract description 12
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 10
- 238000012545 processing Methods 0.000 claims abstract description 10
- 238000012360 testing method Methods 0.000 claims abstract description 10
- 230000000694 effects Effects 0.000 claims abstract description 9
- 238000012795 verification Methods 0.000 claims abstract description 9
- 238000006116 polymerization reaction Methods 0.000 claims abstract description 5
- 238000011049 filling Methods 0.000 claims description 26
- 230000006870 function Effects 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 8
- 210000002364 input neuron Anatomy 0.000 claims description 6
- 210000004205 output neuron Anatomy 0.000 claims description 6
- 238000001228 spectrum Methods 0.000 claims description 4
- 238000000137 annealing Methods 0.000 claims description 3
- 238000000605 extraction Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 2
- 230000001932 seasonal effect Effects 0.000 claims description 2
- 239000000126 substance Substances 0.000 claims description 2
- 238000009825 accumulation Methods 0.000 abstract description 2
- 238000001514 detection method Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 230000002776 aggregation Effects 0.000 description 3
- 238000004220 aggregation Methods 0.000 description 3
- XEEYBQQBJWHFJM-UHFFFAOYSA-N Iron Chemical compound [Fe] XEEYBQQBJWHFJM-UHFFFAOYSA-N 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 238000005065 mining Methods 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000012800 visualization Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 229910052742 iron Inorganic materials 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Economics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Strategic Management (AREA)
- Computational Linguistics (AREA)
- Human Resources & Organizations (AREA)
- Biophysics (AREA)
- Marketing (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Primary Health Care (AREA)
- Evolutionary Biology (AREA)
- Water Supply & Treatment (AREA)
- Public Health (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the technical field of electrical fingerprint identification, and particularly relates to a non-invasive electrical fingerprint identification method based on multi-scale feature perception. The invention comprises the following steps: acquiring electrical data of a target electrical appliance, carrying out block processing according to a specific length, marking, and dividing a training set, a verification set and a test set; preprocessing the acquired data; constructing a multi-scale perception convolutional neural network, and determining the input and output formats of the network; training the network, selecting proper super parameters to obtain the best training effect; accessing the trained network model into an electrical fingerprint identification task, and performing real-time electrical fingerprint monitoring by taking polymerization current and voltage as input; the invention detects the working state of the electric appliance of the corresponding circuit by constructing and training the multi-scale feature perception convolutional neural network. The invention realizes end-to-end integrated work, reduces error accumulation of cooperative work of all parts after task decomposition, and achieves better electrical fingerprint identification effect.
Description
Technical Field
The invention belongs to the technical field of electrical fingerprint identification, and particularly relates to a non-invasive electrical fingerprint identification method.
Background
Everyone has a fingerprint and is considered a unique "identification card", as is the case with electrical fingerprints, each appliance having its own "representation". The electric fingerprint identifies the working state of certain equipment by various characteristics of voltage and current of the electric appliance during power utilization and combining a deep learning method, and even warns dangerous power supply (illegal charging of a battery car, illegal use of electric appliances in a dormitory and the like), so that the power utilization safety is ensured. The convolution neural network based on multi-scale perception provided by the invention performs convolution operation under multi-scale based on original current and voltage data so as to obtain multi-scale and multi-dimensional time sequence data characteristics, and then performs 'portrait' on the data characteristics by utilizing the high-dimensional characteristic extraction capability of the convolution neural network, so that the characteristic information belonging to specific equipment in a data segment is identified, and the corresponding electric appliance identity is identified. When the method is applied to real-time circuit monitoring, the detailed working state of equipment in the circuit can be obtained, and the detailed information monitoring of the circuit power consumption or the warning and reminding of dangerous power utilization are realized.
Disclosure of Invention
The invention aims to provide a non-invasive electrical fingerprint identification method with high accuracy for identifying the working state of electrical equipment.
The non-invasive electrical fingerprint identification method provided by the invention is based on a multi-scale feature perception neural network model technology, and identifies the working state of an electrical appliance in the current circuit by acquiring the current and voltage data of the circuit in real time through a non-invasive sensor so as to achieve the effect of identifying the working state of the electrical appliance with high accuracy.
The neural network model provided by the invention improves the existing neural network for Non-Intrusive Load Monitoring (NILM) tasks, and has stronger distinguishing and identifying capability for special characteristics belonging to different electrical appliances and different working states by extracting characteristics of different scales of the acquired raw data after preprocessing. The working state of the electrical equipment in the circuit can be identified with higher accuracy, and the circuit is monitored and dangerous power utilization is warned.
The invention provides a non-invasive electrical fingerprint identification method based on multi-scale feature perception, which comprises the following specific steps:
step 1: acquiring electrical data of a target electrical appliance, carrying out block processing according to a specific length, marking, and dividing a training set, a verification set and a test set;
and 2, step: preprocessing the data obtained in the previous step;
and 3, step 3: constructing a multi-scale perception convolutional neural network, and determining the input and output formats of the network;
and 4, step 4: training the network, selecting proper super parameters to obtain the best training effect;
and 5: and (3) accessing the trained network model into an electrical fingerprint identification task, and performing real-time electrical fingerprint monitoring by taking the aggregation current and voltage as input.
The following is a further detailed description of the various steps:
step 1: and (3) carrying out electrical data acquisition on the target electrical appliance, carrying out block processing according to a specific length, marking, and dividing a training set, a verification set and a test set.
The existing public data sets in the non-intrusive load monitoring field are not many, the problems of lack of unified standardization, less maximum concurrent load and the like exist, and therefore the data set is automatically acquired by taking the own requirements as targets according to the network characteristics of the invention. The acquisition mode is as follows: firstly, defining specific quantity and types of electric appliances to be identified, fixing sampling frequency by taking the electric appliances as targets, and firstly, respectively acquiring current and voltage data of single equipment in each working state; and then all the devices are arranged and combined, and data acquisition is carried out on the multiple devices which are not started simultaneously but are superposed in working states. Then, the acquired data are segmented: dividing the acquired continuous time sequence into a plurality of data segments in a sliding window mode, wherein the window size is fixed (W), the step is fixed (S), and the number of the obtained segmented data blocks is as follows under the assumption that the length of the time sequence data is L: (L-W)/S. For the divided data, labeling is performed according to the electric appliance state contained in the divided data, for example: no load, no load + blower start + blower first gear, etc. And then dividing the data set according to a specific proportion to obtain a training set, a verification set and a test set.
And 2, step: preprocessing the data obtained in the last step
The content of the data set obtained in the step one only comprises original continuous current and voltage values and a label corresponding to the data block, and in order to carry out deeper mining on data characteristics, multi-angle and multi-level preprocessing is carried out on data. The method specifically comprises the following steps: calculating active power, performing spectrum decomposition (fourier transform) and time series decomposition (STL decomposition) seasonal trend decomposition, and the like. And training and predicting by using the preprocessed results and the original data as input data of the network.
The data preprocessing specifically comprises the following steps:
(1) Calculating corresponding active power according to the real-time current and the voltage;
(2) Taking a section of data as a unit, and carrying out Fourier transform on the current and voltage data;
(3) And taking a section of data as a unit, and performing frequency domain information extraction and time series decomposition.
And step 3: constructing a multi-scale perception convolutional neural network and determining the input and output formats of the network
The network model constructed in the invention comprises three parts in total, wherein the first part is a multi-scale feature perception network module and is used for realizing multi-scale feature perception; the second part is a global feature perception network module used for realizing global feature perception; and the third part is a multi-label classification network and is used for finishing classification tasks, specifically, mapping the extracted features of the previous part to finally obtain the prediction of the identity of the electric appliance.
The specific structure and input/output format of each part of the network are as follows:
(I) multiscale feature aware network module
The module has a two-layer structure:
a first layer: the method comprises the following steps of (1) forming 3 one-dimensional convolution blocks, wherein convolution kernels of the blocks are different in size and are respectively 1, 5 and 9, and filling lengths of the corresponding convolution kernels are different and are respectively 0, 2 and 4; the convolution step length of each block is 1, the number of input channels is 6, and the number of output channels is 8;
a second layer: and (3) two-dimensional convolution, wherein the size of a convolution kernel is 3*3, the step size is 1, the filling is 1*1, the number of input channels is 24, the number of output channels is 32, and then a ReLU layer is connected.
(II) Global feature aware network Module
The module has nine layers of networks:
a first layer: two-dimensional convolution, the size of a convolution kernel is 3*3, the step length is 1, the filling is 1*1, the number of input channels is 32, the number of output channels is 64, and then a ReLU layer is connected;
a second layer: two-dimensional convolution, the size of a convolution kernel is 3*3, the step length is 1, the filling is 1*1, the number of input channels is 64, the number of output channels is 64, and then a ReLU layer is connected;
and a third layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 64, and then a ReLU layer is connected;
a fourth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 64, and then a ReLU layer is connected;
and a fifth layer: two-dimensional convolution, the convolution kernel size is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 128, then the global maximum pooling layer is connected, and then the ReLU layer is connected;
a sixth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 128, the number of output channels is 64, and then a ReLU layer is connected;
a seventh layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 32, and then a ReLU layer is connected;
an eighth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 32, the number of output channels is 16, and then a ReLU layer is connected;
a ninth layer: and (3) two-dimensional convolution, wherein the size of a convolution kernel is 5*5, the step size is 1, the filling is 2*2, the number of input channels is 16, the number of output channels is 4, and then a ReLU layer is connected.
(III) Multi-Label Classification network
The network maps the extracted characteristics of the previous part to finally obtain the prediction of the identity of the electric appliance; the network comprises two layers:
a first layer: fully connecting layers, wherein the number of input neuron nodes is 900 (4 × 15), the number of output neuron nodes is 256, and then a ReLU activation function is connected;
a second layer: and in the full connection layer, the number of input neuron nodes is 256, the number of output neuron nodes is the length of an array of onehot results, and then a Sigmoid activation function is connected.
And 4, step 4: training network, selecting proper super parameter to obtain best training effect
In the training of the network, a random gradient descent algorithm SGD is selected as an optimizer of the network, a cosine annealing algorithm is used for dynamically adjusting the learning rate of the network, and a cross entropy loss function is selected for carrying out loss calculation and back propagation on a prediction result of the network;
wherein the content of the first and second substances,is the prediction result of the model, y is the true label corresponding to the sample, and Cross EntrophyLoss is the cross entropy loss calculation function of the self-contained in the nn.
And 5: and (3) accessing the trained network model into an electrical fingerprint identification task, and performing real-time electrical fingerprint monitoring by taking the aggregation current and voltage as input.
Example is carried out by taking the identification of the working state of an electric appliance of a building as a target: non-invasive current and voltage sensing equipment is installed on a general power supply line of the building, the sensor acquires current and voltage equipment in a circuit at a specific frequency, acquired data are transmitted to a data cache module in real time according to a time sequence, the module maintains latest time sequence data within two seconds in real time, and the data are transmitted to a data processing module of the model at a specific interval. After receiving data with a specific length, a data processing module of the model performs non-repeated division on the data according to the size of a window consistent with a training set, then performs preprocessing on the divided data, finally sends the data into a prediction module of the model for prediction, sends the obtained prediction result to a visualization module in real time, converts the predicted equipment state label into a corresponding equipment state and displays the equipment state, and therefore real-time monitoring on the state of a working electric appliance is achieved.
The invention detects the working state of the electric appliance of the corresponding circuit by constructing and training a multi-scale feature perception convolutional neural network. The invention realizes the integrated work of end to end (from the original data of the sensor to the corresponding electric appliance identity), reduces the error accumulation of the cooperative work of all parts after the task decomposition, and achieves better electric fingerprint identification effect.
Drawings
Fig. 1 is a flowchart of a non-intrusive load monitoring electrical fingerprint identification method based on multi-scale feature sensing according to the present invention.
Fig. 2 is a diagram of a network model architecture of the present invention.
Detailed Description
The technical scheme of the invention is explained in detail by combining the drawings and the embodiment.
Examples
The invention provides a method for extracting electrical fingerprint characteristics by using current and voltage polymerization data obtained in a non-invasive load monitoring mode as input, which comprises the following specific steps:
step 1: acquiring electrical data of a target electrical appliance, carrying out block processing according to a specific length, marking, and dividing a training set, a verification set and a test set;
and 2, step: preprocessing the data obtained in the last step in a specific mode;
and 3, step 3: constructing a multi-scale perception convolutional neural network, and determining the input and output formats of the network;
and 4, step 4: training the network, selecting proper hyper parameters to obtain the best training effect;
and 5: and (3) accessing the trained network model into an electrical fingerprint identification task, and performing real-time electrical fingerprint monitoring by taking the aggregation current and voltage as input.
The following is a further detailed description of the various steps:
step 1: and (3) carrying out electrical data acquisition on the target electrical appliance, carrying out block processing according to a specific length, marking, and dividing a training set, a verification set and a test set.
Firstly, defining specific quantity and types of electric appliances to be identified, specifically comprising: the electric hair drier, the electric iron, the dust collector, the small electric cooker, the warm air blower and the blower are used as targets, firstly, current and voltage data of single equipment in each working state are respectively collected, the collection frequency is 1KHz, the equipment to be collected is started from a closed state, then works in a stable state, and is then closed, and the whole process is about 20s to 40 s; and then all the devices are arranged and combined, and data acquisition with different starting of the devices and overlapping of working states is carried out, wherein the time duration is different from 40s to 140 s. Then, the acquired data are segmented: the collected continuous time sequence is divided into a plurality of data segments in a sliding window mode, the window size is fixed to 900, and the pace is fixed to 3. For the divided data, labeling is performed according to the electric appliance state contained in the divided data, for example: no load ([ 1, 0, 0, 0, 0, 0, 0 ]), no load + blower activation ([ 1, 1, 0, 0, 0, 0, 0, 0, 1, 0, 0 ]), no load + blower activation + blower first gear, etc ([ 1, 1, 0, 0, 0, 0, 0, 1, 1, 0 ]). And then dividing the data set according to the ratio of 6: 1: 3 to obtain a training set, a verification set and a test set.
Step 2: and preprocessing the data obtained in the last step in a specific mode.
The content of the data set obtained in the step one only comprises original continuous current and voltage values and a label corresponding to the data block, and in order to carry out deeper mining on data characteristics, multi-angle and multi-level preprocessing is carried out on data. The method specifically comprises the following steps: calculating corresponding active power according to the data segments to obtain an active power data sequence; respectively carrying out Fourier transform on the current data and the voltage data to obtain corresponding frequency spectrum information; and respectively carrying out STL decomposition on the current data and the voltage data to obtain corresponding season, trend and residual error information. And training and predicting by taking the preprocessed results and the original data as input data of the network, wherein the final format of the data is as follows: (8, 900).
And step 3: and constructing a multi-scale perception convolutional neural network, and determining the input and output formats of the network.
The network model constructed in the invention totally comprises three parts, wherein the first part is a one-dimensional convolution part for realizing multi-scale feature perception, the second part is a two-dimensional convolution neural network for realizing global feature perception, and the third part is a fully-connected network for completing classification tasks. The input length 900 and the number of channels of the network are 8, and the specific implementation and input and output formats of each part of the network are as follows:
(I) multiscale feature aware network module
The module has a two-layer structure:
a first layer: the method comprises the following steps that (1) each block consists of 3 one-dimensional convolution blocks, convolution kernels among the blocks are different in size and are respectively 1, 5 and 9, and filling lengths of the corresponding convolution kernels are different and are respectively 0, 2 and 4; the convolution step of each block is 1, the number of input channels is 6, the number of output channels is 8 (if the layer input is (8, 900), the output of each block is (8, 900), the outputs of the three blocks are spliced to obtain (24, 900), and then the output is deformed into a feature matrix with the size of 30 × 30 of 24 channels, and finally the output is (24, 30, 30));
a second layer: two-dimensional convolution, convolution kernel size 3*3, step size 1, padding 1*1, input lane number 24, output lane number 32, followed by the ReLU layer (final output (32, 30, 30)).
(II) Global feature aware network Module
The module has nine layers of networks:
a first layer: two-dimensional convolution, the convolution kernel size is 3*3, the step length is 1, the padding is 1*1, the number of input channels is 32, the number of output channels is 64, and the subsequent ReLU layer (final output (64, 30, 30));
a second layer: two-dimensional convolution, the convolution kernel size is 3*3, the step size is 1, the padding is 1*1, the number of input channels is 64, the number of output channels is 64, and the subsequent ReLU layer (final output (64, 30, 30));
and a third layer: two-dimensional convolution, the convolution kernel size is 5*5, the step size is 1, the padding is 2*2, the number of input channels is 64, the number of output channels is 64, and the subsequent ReLU layer (final output (64, 30, 30));
a fourth layer: two-dimensional convolution, the convolution kernel size is 5*5, the step size is 1, the padding is 2*2, the number of input channels is 64, the number of output channels is 64, and the subsequent ReLU layer (final output (64, 30, 30));
a fifth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 128, then the global maximum pooling layer is connected, and then the ReLU layer is connected (final output is (128, 15, 15));
a sixth layer: two-dimensional convolution, the convolution kernel size is 5*5, the step size is 1, the padding is 2*2, the number of input channels is 128, the number of output channels is 64, and the subsequent ReLU layer (final output (64, 15, 15));
a seventh layer: two-dimensional convolution, the convolution kernel size is 5*5, the step size is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 32, and then a ReLU layer (final output (32, 15, 15));
an eighth layer: two-dimensional convolution, the convolution kernel size is 5*5, the step size is 1, the filling is 2*2, the number of input channels is 32, the number of output channels is 16, and then a ReLU layer (final output (16, 15, 15));
ninth layer: and (3) two-dimensional convolution, wherein the size of a convolution kernel is 5*5, the step size is 1, the padding is 2*2, the number of input channels is 16, the number of output channels is 4, and then a ReLU layer is connected (finally output (4, 15, 15)).
(III) Multi-Label Classification network
The network maps the extracted characteristics of the previous part to finally obtain the prediction of the identity of the electric appliance; the network comprises two layers:
a first layer: a fully-connected layer, wherein the number of input neuron nodes is 900 (4 × 15), the number of output neuron nodes is 256, and then a ReLU activation function is connected;
a second layer: and in the full connection layer, the number of input neuron nodes is 256, the number of output neuron nodes is the length of an array of onehot results, and then a Sigmoid activation function is connected.
And 4, step 4: training network, selecting proper super parameter to obtain best training effect
In the training of the network, a random gradient descent algorithm SGD is selected as an optimizer of the network, the learning rate of the network is dynamically adjusted by a cosine annealing algorithm, and a cross entropy loss function is selected to perform loss calculation and back propagation on a prediction result of the network;
wherein, the first and the second end of the pipe are connected with each other,the cross entropy loss calculation function is a self-contained cross entropy loss calculation function in an nn.functional module of the pytorech; through experiments, the model provided by the invention gradually converges when 150 EPOCHs, the initial learning rate is set to be 0.1, and the batch size is 64.
And 5: and connecting the trained network model into an electrical fingerprint identification task, and performing real-time electrical fingerprint monitoring by taking the polymerization current and the polymerization voltage as input.
In the invention, the power supply of a room is taken as a detection target: non-invasive current and voltage sensing equipment is installed on a main power supply line of the house, the sensors acquire current and voltage data in the circuit at the frequency of 1KHz, the acquired data are transmitted to a data cache module in real time according to the time sequence, the module maintains the latest time sequence data within two seconds in real time, and 1100 time sequence data (100 repeated data) are transmitted to a data processing module of the model at the interval of 1 s. After receiving data with a specific length, a data processing module of the model divides the data according to the window size 900, the step length is 11, then the divided data is preprocessed and finally sent to a prediction module of the model for prediction, voting is carried out on a plurality of obtained predictions, a final prediction result is selected, and the result is converted into a corresponding equipment state through a visualization module.
In the invention, the accuracy of the detection result of the equipment state in the plurality of data in the detection time period is used as an index for evaluation. The detection accuracy of the model for each selected device in the validation set, the test set and the actual experiment is given in table 1.
TABLE 1 detection accuracy (%) of each apparatus
Claims (1)
1. A non-invasive electrical fingerprint identification method based on multi-scale feature perception is characterized by comprising the following specific steps:
step 1: acquiring electrical data of a target electrical appliance, carrying out block processing according to a certain length, marking, and dividing a training set, a verification set and a test set;
and 2, step: pre-processing the collected electrical data, comprising:
calculating active power, and performing frequency spectrum decomposition and time sequence decomposition; using the preprocessed results and the original data as input data of the network together for training and predicting;
and step 3: constructing a multi-scale perception convolutional neural network and determining the input and output formats of the network
The constructed multi-scale perception convolutional neural network model comprises three parts: the multi-scale feature perception network module is used for realizing multi-scale feature perception; the global feature perception network module is used for realizing global feature perception; the multi-label classification network is used for completing classification tasks, and specifically maps the extracted features of the previous part to finally obtain the prediction of the identity of the electric appliance;
and 4, step 4: training network, selecting proper super parameter to obtain best training effect
Selecting a random gradient descent algorithm SGD as an optimizer of the network, dynamically adjusting the learning rate of the network by using a cosine annealing algorithm, and performing loss calculation and back propagation on a prediction result of the network by using a cross entropy loss function;
wherein the content of the first and second substances,the cross entropy loss calculation function is a self-contained cross entropy loss calculation function in an nn.functional module of the pytorech;
and 5: accessing the trained network model into an electrical fingerprint identification task, and performing real-time electrical fingerprint monitoring by taking polymerization current and voltage as input;
the process of the step 1 is as follows:
data are collected in the following mode: firstly, defining specific quantity and types of electric appliances to be identified, taking the electric appliances as targets, fixing the sampling frequency to be 1KHz, and firstly, respectively collecting current and voltage data of single equipment in each working state; then all the devices are arranged and combined, and data acquisition is carried out, wherein the multiple devices are not started at the same time, but the working states are overlapped;
then, the acquired data are segmented: dividing the collected continuous time sequence into a plurality of data segments in a sliding window mode, wherein the window size W is fixed, the step S is fixed, and the length of the time sequence data is L, so that the number of the obtained segmented data blocks is as follows: (L-W)/S; marking the divided data according to the electric appliance state contained in the divided data;
then dividing the data set according to a specific proportion to obtain a training set, a verification set and a test set;
the data preprocessing in the step 2 comprises the following steps: calculating active power, and performing frequency spectrum decomposition, namely Fourier transform and time sequence decomposition, namely STL seasonal trend decomposition; the method specifically comprises the following steps:
(1) Calculating corresponding active power according to the real-time current and voltage;
(2) Taking a section of data as a unit, and carrying out Fourier transform on the current and voltage data;
(3) Taking a section of data as a unit, and performing frequency domain information extraction and time series decomposition;
the specific structure of the multi-scale perception convolutional neural network and the input and output format of the network in the step 3 are as follows:
the multi-scale feature perception network module has a two-layer structure:
a first layer: the system is composed of 3 one-dimensional convolution blocks, convolution kernels among the blocks are respectively 1, 5 and 9, and filling lengths corresponding to the convolution kernels are respectively 0, 2 and 4; the convolution step length of each block is 1, the number of input channels is 6, and the number of output channels is 8;
a second layer: two-dimensional convolution, the size of a convolution kernel is 3*3, the step length is 1, the filling is 1*1, the number of input channels is 24, the number of output channels is 32, and then a ReLU layer is connected;
(II) the global feature perception network module comprises nine layers of networks:
a first layer: two-dimensional convolution, the size of a convolution kernel is 3*3, the step length is 1, the filling is 1*1, the number of input channels is 32, the number of output channels is 64, and then a ReLU layer is connected;
a second layer: two-dimensional convolution, the size of a convolution kernel is 3*3, the step length is 1, the filling is 1*1, the number of input channels is 64, the number of output channels is 64, and then a ReLU layer is connected;
and a third layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 64, and then a ReLU layer is connected;
a fourth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 64, and then a ReLU layer is connected;
a fifth layer: two-dimensional convolution, the convolution kernel size is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 128, then the global maximum pooling layer is connected, and then the ReLU layer is connected;
a sixth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 128, the number of output channels is 64, and then a ReLU layer is connected;
a seventh layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 64, the number of output channels is 32, and then a ReLU layer is connected;
an eighth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 32, the number of output channels is 16, and then a ReLU layer is connected;
ninth layer: two-dimensional convolution, the size of a convolution kernel is 5*5, the step length is 1, the filling is 2*2, the number of input channels is 16, the number of output channels is 4, and then a ReLU layer is connected;
(III) the multi-label classification network comprises two layers:
a first layer: the full connection layer, the number of input neuron nodes is 900, the number of output neuron nodes is 256, and then a ReLU activation function is connected;
a second layer: and in the full connection layer, the number of input neuron nodes is 256, the number of output neuron nodes is the length of an array of onehot results, and then a Sigmoid activation function is connected.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210049795.XA CN114511007B (en) | 2022-01-17 | 2022-01-17 | Non-invasive electrical fingerprint identification method based on multi-scale feature perception |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210049795.XA CN114511007B (en) | 2022-01-17 | 2022-01-17 | Non-invasive electrical fingerprint identification method based on multi-scale feature perception |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114511007A CN114511007A (en) | 2022-05-17 |
CN114511007B true CN114511007B (en) | 2022-12-09 |
Family
ID=81550560
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210049795.XA Active CN114511007B (en) | 2022-01-17 | 2022-01-17 | Non-invasive electrical fingerprint identification method based on multi-scale feature perception |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114511007B (en) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2020103901A4 (en) * | 2020-12-04 | 2021-02-11 | Chongqing Normal University | Image Semantic Segmentation Method Based on Deep Full Convolutional Network and Conditional Random Field |
CN112435142A (en) * | 2020-12-16 | 2021-03-02 | 北京航空航天大学 | Power load identification method and load power utilization facility knowledge base construction method thereof |
CN113807225A (en) * | 2021-09-07 | 2021-12-17 | 中国海洋大学 | Load identification method based on feature fusion |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU1330501A (en) * | 1999-10-07 | 2001-05-10 | Veridicom, Inc. | Spoof detection for biometric sensing systems |
US9104189B2 (en) * | 2009-07-01 | 2015-08-11 | Mario E. Berges Gonzalez | Methods and apparatuses for monitoring energy consumption and related operations |
CN106097353B (en) * | 2016-06-15 | 2018-06-22 | 北京市商汤科技开发有限公司 | Method for segmenting objects and device, computing device based on the fusion of multi-level regional area |
CN106199347B (en) * | 2016-06-23 | 2019-02-22 | 福州大学 | Fault arc detection method and detection device based on interference fingerprint identification |
CN109376753B (en) * | 2018-08-31 | 2022-06-28 | 南京理工大学 | Probability calculation method for three-dimensional spatial spectrum space dimension pixel generic |
CN109815339B (en) * | 2019-01-02 | 2022-02-08 | 平安科技(深圳)有限公司 | Knowledge extraction method and device based on TextCNN, computer equipment and storage medium |
CN111582007A (en) * | 2019-02-19 | 2020-08-25 | 富士通株式会社 | Object identification method, device and network |
CN113033633B (en) * | 2021-03-12 | 2022-12-09 | 贵州电网有限责任公司 | Equipment type identification method combining power fingerprint knowledge and neural network |
-
2022
- 2022-01-17 CN CN202210049795.XA patent/CN114511007B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
AU2020103901A4 (en) * | 2020-12-04 | 2021-02-11 | Chongqing Normal University | Image Semantic Segmentation Method Based on Deep Full Convolutional Network and Conditional Random Field |
CN112435142A (en) * | 2020-12-16 | 2021-03-02 | 北京航空航天大学 | Power load identification method and load power utilization facility knowledge base construction method thereof |
CN113807225A (en) * | 2021-09-07 | 2021-12-17 | 中国海洋大学 | Load identification method based on feature fusion |
Non-Patent Citations (2)
Title |
---|
"基于数据增强和深度学习的非侵入式负荷分解方法";刘睿迪;《中国优秀高级论文全文数据库工程科技Ⅱ辑》;20210815(第08期);1-64页 * |
"基于机器学习理论的智能电网数据分析及算法研究";杨延东;《中国博士学位论文全文数据库工程科技Ⅱ辑》;20210115(第01期);1-14、30-81页 * |
Also Published As
Publication number | Publication date |
---|---|
CN114511007A (en) | 2022-05-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111612650B (en) | DTW distance-based power consumer grouping method and system | |
CN108681747B (en) | Rotary machine fault diagnosis and state monitoring system and method based on deep learning | |
CN107229602B (en) | Method for identifying electricity consumption behavior of intelligent building microgrid | |
CN109685314B (en) | Non-intrusive load decomposition method and system based on long-term and short-term memory network | |
CN110188920A (en) | A kind of lithium battery method for predicting residual useful life | |
CN111368904B (en) | Electrical equipment identification method based on electric power fingerprint | |
CN115170000B (en) | Remote monitoring method and system based on electric energy meter communication module | |
CN111950596A (en) | Training method for neural network and related equipment | |
CN112396087B (en) | Method and device for analyzing power consumption data of solitary old people based on intelligent ammeter | |
CN104346503A (en) | Human face image based emotional health monitoring method and mobile phone | |
CN103268519A (en) | Electric power system short-term load forecast method and device based on improved Lyapunov exponent | |
CN112308124B (en) | Intelligent electricity larceny prevention method for electricity consumption information acquisition system | |
CN111242276B (en) | One-dimensional convolutional neural network construction method for load current signal identification | |
CN114201989A (en) | Alternating current motor bearing fault diagnosis method adopting convolutional neural network and bidirectional long-time and short-time memory network | |
CN116502175A (en) | Method, device and storage medium for diagnosing fault of graph neural network | |
CN112396098A (en) | Non-embedded apartment electrical appliance load identification and analysis method, system and application | |
CN114693624A (en) | Image detection method, device and equipment and readable storage medium | |
CN114970841A (en) | Training method of battery state prediction model and related device | |
CN113554361A (en) | Comprehensive energy system data processing and calculating method and processing system | |
CN114511007B (en) | Non-invasive electrical fingerprint identification method based on multi-scale feature perception | |
CN116523681A (en) | Load decomposition method and device for electric automobile, electronic equipment and storage medium | |
CN117055726A (en) | Micro-motion control method for brain-computer interaction | |
CN113780457B (en) | Abnormality detection method, device, equipment and medium for traditional Chinese medicine resource consumption | |
CN115169405A (en) | Hotel guest room equipment fault diagnosis method and system based on support vector machine | |
CN110033082B (en) | Method for identifying deep learning model in AI (Artificial intelligence) equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |