Charger baby load identification technology

Abstract

The invention provides a charge pal load identification technology, which specifically comprises a data acquisition step, a model construction step, a model training step, a data enhancement step, an event extraction step, a first judgment step, a second judgment step, a third judgment step and a long time scale identification step, and is characterized in that an fft + MLP network structure is used to replace an LSTM in a TimeGAN model; and a TimeGAN-fft model is obtained. Through the trained TimeGAN-fft model, data enhancement processing is carried out on the high-frequency current and voltage data, it is guaranteed that small-power load events in the high-frequency current and voltage data can be extracted, the accuracy of small-power load event extraction is improved, a small-power load event sliding window is analyzed in combination with short-time high-frequency features and long-time low-frequency features, and the accuracy of small-power load event extraction is improved. And when the detected sliding window meets the preset judgment standard, the low-power load event is judged to be the charge pal event, so that the identification accuracy of the low-power load event is improved, and the charging potential safety hazard of the charge pal is eliminated in the daily power utilization scene of the user.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electricity utilization, in particular to a power bank load identification technology. BACKGROUND

[0002] With the wide popularity of electronic devices, power banks as a convenient mobile power source have been widely used in daily life.

[0003] In personnel-intensive scenarios such as campus dormitories, with the illegal use of inferior power banks, power banks have become an important hidden danger for electrical fire incidents.

[0004] The existing load identification technology has the defects of insufficient high-frequency current and voltage feature extraction capability, poor small-power electricity event extraction capability, relying only on short-term electricity features for load analysis, and lack of multi-scale feature analysis capability. Therefore, it is of great significance to identify power banks from the dual dimensions of "short-term high-frequency features and long-term power trends". SUMMARY

[0005] The present application provides a power bank load identification technology to solve the problem of poor small-power electricity event extraction capability and inaccurate load event identification of the existing load identification technology.

[0006] The present application provides a power bank load identification technology, which specifically includes a data collection step, a model construction step, a model training step, a data enhancement step, an event extraction step, a first judgment step, a second judgment step, and a third judgment step.

[0007] The data collection step is to collect real-time high-frequency current and voltage data of small-power charging appliances by using high-precision current and voltage sensors; the model construction step is to replace the LSTM in the TimeGAN model with the network structure of fft+MLP to obtain the TimeGAN-fft model; the model training step is to obtain the reconstruction loss function L R , the unsupervised loss function L U , and the supervised loss function L S based on the collected high-frequency current and voltage data and the TimeGAN-fft model, and then train the TimeGAN-fft model by using the reconstruction loss function L R , the unsupervised loss function L S , and the supervised loss function L U , and the supervised loss function L SThe TimeGAN-fft model is trained; the data enhancement step is to input the collected high-frequency current voltage data into the trained TimeGAN-fft model, and the high-frequency current voltage data is subjected to data enhancement processing to obtain enhanced data; the event extraction step is to extract a small power charging event and a detection window corresponding to the small power charging event from the enhanced data according to a dynamic window triggering mechanism and a multi-scale event perception method; the first judgment step is to filter the detection window, extract pure high-frequency current voltage in the filtered detection window, extract a feature vector of the pure high-frequency current, and use a random forest model to preliminarily judge the detection window based on the obtained feature vector; if the judgment result is true, the detection window is determined as a preliminary judgment window, and the next step is executed; the second judgment step is to use a frequency domain conversion FITS model as a classifier, input the pure high-frequency current voltage data of the preliminary judgment window into the FITS model for secondary judgment, and if the judgment result is true, the preliminary judgment window is determined as a power bank sub-event; the third judgment step is used to define a time T when the power bank sub-event is first identified, record a first cycle at the time T as an anchor point, and continue to slide the window for the small power charging event based on the time T as a starting point to obtain a subsequent detection window of the small power charging event, execute the subsequent detection window and the second judgment step on the subsequent detection window, and when the subsequent detection window is determined as false, the anchor point is reacquired; when the subsequent detection window is determined as true, the third judgment step is repeatedly executed until a preset number of subsequent detection windows are determined as power bank sub-events, and the small power charging event is determined as a power bank event.

[0008] Further, the TimeGAN-fft model includes an embedding layer, a reconstruction layer, a generator, and a classifier, and the model training step specifically includes a random noise calculation step, a hidden state acquisition step, a time threshold data restoration step, a frequency domain component classification step, a first loss function calculation step, a second loss function calculation step, and a third loss function calculation step.

[0009] The random noise calculation step is that the high-frequency current voltage data obtains a noise sequence Z through a generator, and the calculation formula is

[0010]

[0011]

[0012] wherein V={v1,v2,v3,...,v l}, I={i1,i2,i3,...,il},ZV represents the noise sequence of high-frequency voltage data, Z represents the noise sequence of high-frequency current data, l represents the length of voltage current data, the length is 1, and z represents the noise value, V={v1,v2,v3,...,v l},I={i1,i2,i3,...,i l} can obtain the original time sequence data pair X=[V,I].

[0013] The hidden state acquisition step is that the high-frequency current voltage data and the noise sequence acquire the hidden state of the high-frequency current voltage data and the hidden state of the noise sequence through the embedding layer, and the hidden state of the high-frequency current voltage data is defined as the first hidden state and the noise sequence is the second hidden state.

[0014] The time domain data restoration step is that the first hidden state restores the original time sequence data pair to obtain a reconstruction sequence X’=[V’,I’] through a reconstruction layer.

[0015] The frequency domain component classification step is that the first hidden state and the second hidden state obtain the first hidden state discrimination result and the second hidden state discrimination result through a classifier, and the first hidden state discrimination result is defined as the first discrimination result and the second hidden state discrimination result is the second discrimination result.

[0016] The first loss function calculation step is used to calculate the reconstruction loss function L R , and the formula is

[0017] Among them, E s,x1:l~p represents the simulation of the generation process of the high-frequency current voltage data, an attribute s is sampled from the static feature distribution, a time sequence data x_{1:T} is generated according to the attribute s, and the statistical average value of the target function is finally calculated, t represents the time step, and x t represents the real observation value of the input data, represents the reconstruction value of the input data.

[0018] The second loss function calculation step is used to calculate the unsupervised loss function L U , and the formula is

[0019] Among them, y t represents the output of the classifier to the high-frequency current voltage data; represents the output of the classifier to the enhanced data.

[0020] The third loss function calculation step is used to calculate the supervised loss function L S , and the formula is

[0021] , wherein h t represents a hidden state, represents a random noise vector.

[0022] Further, the event extraction step specifically includes a threshold setting step, a window division step, and a detection target acquisition step.

[0023] The threshold setting step is used to acquire the background power P bg of the to-be-detected window, calculate the standard deviation of the background power in a preset time period , and the power standard deviation P std between the current time and a preset time, when P std > , the dynamic power change rate threshold P anchor = r*P bg , r is a proportionality coefficient, the power change rate is , and P start represents the power at the starting point of the dynamic window.

[0024] The window division step is when P rate > P anchor , the dynamic window is started, and the power change rate P rate in the dynamic window is continuously detected, when P rate < P anchor or the window length of the dynamic window , T max represents a maximum duration threshold, and the dynamic window is defined as a small power charging event.

[0025] The detection target acquisition step is used to preset a window length threshold, and the dynamic window corresponding to the small power charging event is screened, when the window length of the dynamic window is within the window length threshold range, the dynamic window is defined as a to-be-detected window.

[0026] Further, the first judgment step specifically includes a window screening step, a first feature extraction step, a feature screening step, and a model judgment step.

[0027] The window screening step is used to calculate the difference P3-P0 between the power P0 at the opening of the to-be-detected window and the power P3 at 3s after the window is opened, when the difference P3-P0 is within a preset charging power threshold range, the screened to-be-detected window is obtained, the screened to-be-detected window is defined as a screened window, and then the next step is executed.

[0028] The first feature extraction step is based on the first cycle at the screening window time t and the second cycle at the time t+3s, the first cycle is subtracted from the second cycle to extract the pure high-frequency current voltage data of the charging class event in the screening window, and the entropy, stationarity and frequency of the pure high-frequency current voltage are extracted.

[0029] The feature screening step is to perform vertical feature vector clustering on the extracted entropy, stationarity and frequency from the intra-class and inter-class directions respectively to obtain similar feature vectors.

[0030] The model judgment step is to calculate the first feature, the second feature and the third feature of the screening window, take the first feature, the second feature and the third feature as model input, and use the trained high-recall version of the random forest model to preliminarily judge the small power charging class event corresponding to the screening window from the perspective of similar feature vectors.

[0031] Further, the model judgment step specifically includes a first feature calculation step, a second feature calculation step and a third feature calculation step.

[0032] The first feature calculation step is used to calculate the probability of pattern repetition in the screening window, and is used to measure the randomness and regularity of the screening window, and the formula is ; ; ; ; ; wherein X(i) represents an m-dimensional vector reconstructed for an input sequence, is a first pattern, i=1,2,...,N-m+1, j=1,2,...,N-m+1, u(i) represents the input after current voltage combination in the screening window, d[X(i),X(j)] represents the maximum value of the distance difference between the vector X(i) and the vector X(j), represents another m-dimensional vector reconstructed for an input sequence, is a second pattern, k=0,1,...,m-1, represents the proportion of the second pattern that satisfies d[X(i),X(j)]≤r, r represents a threshold radius for similarity determination, represents the logarithmic probability of the pattern under the m-dimensional vector, represents the logarithmic probability of the pattern under the m+1-dimensional vector, ApEn(m,r,N) represents the randomness and regularity of the difference value measurement sequence, i.e. the first feature.

[0033] The second feature calculation step is used to calculate the first-order difference term of the filtering window and determine whether the filtering window is stationary. The formula is as follows:

[0034] Among them, y t This represents the raw observations of the filter window at time t. The first-order difference term is represented by α, which is the second characteristic term, and the lag term y is represented by α. t-1 The coefficient, β i Represents the difference lag term The coefficient, k, represents the maximum lag of the difference term. This represents the white noise error term.

[0035] The third feature calculation step is used to calculate the power spectral density of the filtering window, quantifying the energy distribution of the signal at different frequencies within the filtering window, and its formula is as follows:

[0036] in, x represents the power spectral density. k (n) represents the nth sampling point of the kth subsequence. Represents the window function. The normalization coefficients of the window function are represented by K, where K represents the total number of segments. denoted by angular frequency, and N represents the length of each subsequence.

[0037] Furthermore, the second judgment step specifically includes a frequency domain conversion step, a high-frequency interference cancellation step, a second feature extraction step, and a feature transformation step.

[0038] The frequency domain transformation step converts the pure high-frequency current and voltage data into frequency domain information based on Fourier transform, and extracts amplitude and phase information; the high-frequency interference elimination step eliminates high-frequency interference in the background of the frequency domain information through a low-pass filter; the second feature extraction step adjusts the frequency domain information components with a linear layer with complex values ​​to extract features in the frequency domain and obtain frequency domain features; the feature transformation step converts the frequency domain features into time domain features through inverse Fourier transform, inputs the time domain features into the TimeGAN-fft model, and uses a classifier to determine whether the preliminary judgment window is a power bank event.

[0039] Further, the third determining step further comprises a long-time scale identification step. Based on the obtained pure power curve of the power bank event, the pure power curve of the power bank event is matched with a standard curve in a power bank curve database. If the matching is successful and the power bank event meets a preset condition, it is determined that the electrical appliance corresponding to the power bank event is a power bank.

[0040] Further, the long-time scale identification step specifically comprises a power time sequence obtaining step, a power curve extracting step and a DWT distance calculating step.

[0041] The power time sequence obtaining step is used to obtain active power, reactive power and distortion power of 10 Hz during 5 min before the power bank event is started and 3 h after the power bank event is started, to generate a power time sequence. The power curve extracting step is based on the obtained power time sequence, uses a fixed window length non-overlapping sliding window to segment the power time sequence, extracts time domain features of the power time sequence, and obtains a pure power change curve of the power bank event. The DWT distance calculating step is used to calculate DWT distance between the pure power change curve and each standard curve in the power bank curve database. If the DWT distance is within a preset DWT distance threshold range, it is determined that the matching is successful.

[0042] The application provides a power bank load identification technology. An LSTM in a TimeGAN model is replaced by using an fft+MLP network structure to obtain a TimeGAN-fft model. The high-frequency current and voltage data are processed by using the trained TimeGAN-fft model for data enhancement, so that small power load events in the high-frequency current and voltage data can be extracted, the accuracy of small power load event extraction is improved, short-time high-frequency features and long-time low-frequency features are combined to analyze a small power load event sliding window, and when the detected sliding window meets a preset judgment standard, the small power load event is determined to be a power bank event, the accuracy of small power load event identification is improved, and the charging safety hazard of the power bank is eliminated in a user daily power consumption scene. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0044] Figure 1 is a flowchart of the power bank load identification technology described in the embodiments of the application. Figure 2is a flowchart of the model training step described in the embodiments of the present application. Figure 3 is a structural diagram of the TimeGAN-fft model described in the embodiments of the present application. Figure 4 is a flowchart of the event extraction step described in the embodiments of the present application. Figure 5 is a flowchart of the first judgment step described in the embodiments of the present application. Figure 6 is a flowchart of the model judgment step described in the embodiments of the present application. Figure 7 is a flowchart of the second judgment step described in the embodiments of the present application. Figure 8 is a flowchart of the long time scale identification step described in the embodiments of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. Please refer to Figure 1 The present application provides a power bank load identification technology, which specifically includes the following steps: step S1) data acquisition step, step S2) model construction step, step S3) model training step, step S4) data enhancement step, step S5) event extraction step, step S6) first judgment step, step S7) second judgment step, step S8) third judgment step, and step S9) long time scale identification step.

[0046] Step S1) data acquisition step, high-precision current and voltage sensors are used to collect real-time high-frequency current and voltage data of small-power charging appliances.

[0047] In this embodiment, the current and voltage data of small-power appliances such as power banks, mobile phones, routers, and computers are collected in real time at a sampling frequency of 6400 Hz. This high-frequency sampling can capture the subtle changes in current and voltage during the use of charging appliances, providing a data basis for subsequent analysis.

[0048] Step S2) model construction step, the LSTM in the TimeGAN model is replaced by using the network structure of fft+MLP to obtain the TimeGAN-fft model, which enhances the processing capability of TimeGAN for high-frequency data and significantly reduces the event and spatial complexity of the network.

[0049] Step S3) model training step, based on the collected high-frequency current voltage data and TimeGAN-fft model, the reconstruction loss function L is obtained R Unsupervised loss function L U And supervised loss function L S Through L R +L S And L U +L S The TimeGAN-fft model is trained, based on L R +L S And L U +L S交替 The TimeGAN-fft model realizes the purpose of data enhancement.

[0050] Please refer to Figure 2 And Figure 3 The TimeGAN-fft model includes embedding layer, reconstruction layer, generator and classifier, embedding layer and reconstruction layer are used for encoding and reconstruction of real appliance data, generator and classifier are used for generating data conforming to real distribution, step S3) model training step specifically includes step S31) random noise calculation step, step S32) hidden state acquisition step, step S33) time threshold data restoration step, step S34) frequency domain component classification step, step S35) first loss function calculation step, step S36) second loss function calculation step and step S37) third loss function calculation step.

[0051] Step S31) random noise calculation step, the high-frequency current voltage data obtains noise sequence Z through generator, its calculation formula is

[0052]

[0053]

[0054] Among them, V={v1,v2,v3,...,v l}, I={i1,i2,i3,...,i l}, ZV represents the noise sequence of high-frequency voltage data, Z I represents the noise sequence of high-frequency current data, l represents the length of voltage current data, the size of the length is 1, z represents noise value, V={v1,v2,v3,...,v l}, I={i1,i2,i3,...,i l} can obtain original time series data pair X=[V,I].

[0055] Step S32) Hidden state acquisition step, the high-frequency current voltage data and the noise sequence acquire the hidden state of the high-frequency current voltage data and the hidden state of the noise sequence through the embedding layer, define the hidden state of the high-frequency current voltage data as a first hidden state, and the noise sequence as a second hidden state.

[0056] Step S33) Time threshold data restoration step, the first hidden state is restored to a reconstructed sequence X'=[V', I'] through the reconstruction layer on the original time sequence data pair.

[0057] Step S34) Frequency domain component classification step, the first hidden state and the second hidden state obtain the first hidden state discrimination result and the second hidden state discrimination result through the classifier, define the first hidden state discrimination result as a first discrimination result, and the second hidden state discrimination result as a second discrimination result, and distinguish the true and false of the input sequence as accurately as possible through the first discrimination result and the second discrimination result.

[0058] In this embodiment, for the high-frequency current voltage data, the hidden state is extracted and restored from the frequency domain by combining the original TimeGAN and the fast Fourier transform; in the embedding layer and the generator, first, the time sequence information of the input high-frequency current voltage data is converted using fft, then the hidden state is obtained by using MLP to extract the frequency domain component of the high-frequency current voltage data, in the reconstruction layer, ifft and MLP are used to restore the time domain data of the high-frequency current voltage data, and in the classifier, mlp is directly used to classify the frequency domain component obtained by fft.

[0059] Step S35) First loss function calculation step, calculate the reconstruction loss function L R , the formula is

[0060] Wherein, E s,x1:l~p represents the simulation of the generation process of the high-frequency current voltage data, samples an attribute s from the static feature distribution, generates a time sequence data x_{1:T} according to the attribute s, and finally calculates the statistical average value of the objective function, the purpose is to optimize the model to approximate the real data distribution, t represents the time step, x t represents the real observation value of the input data, represents the reconstruction value of the input data.

[0061] Step S36) Second loss function calculation step, calculate the unsupervised loss function L U , the formula is

[0062] Wherein, yt This represents the classifier's output on the high-frequency current and voltage data; This represents the output of the classifier for the enhanced data.

[0063] Step S37) Third loss function calculation step: Calculate the supervised loss function L. S Its formula is

[0064] Among them, h t Indicates a hidden state. This represents a random noise vector.

[0065] Step S4) Data augmentation step: Input the collected high-frequency current and voltage data into the trained TimeGAN-fft model, and perform data augmentation processing on the high-frequency current and voltage data to obtain enhanced data.

[0066] Step S5) Event extraction step: Extract low-power charging events from the enhanced data according to the dynamic window triggering mechanism and multi-scale event perception method to obtain the low-power charging events and the detection window corresponding to the low-power charging events.

[0067] like Figure 4 As shown, step S5) event extraction step specifically includes step S51) threshold setting step, step S52) window division step and step S53) target acquisition step.

[0068] Step S51) Threshold setting step: Obtain the background power P of the window to be detected. bg Calculate the standard deviation of background power over a preset time period. The preset time period is set according to the desired value. In this embodiment, the preset time period is 10 seconds, and the power standard deviation P between the current time and the preset time is also considered. std In this embodiment, the current time is defined as t, and the preset time is t+3s. When P std > At that time, the dynamic power change rate threshold P anchor =r*P bg r is the proportionality coefficient, and the power change rate is... P start This indicates the power at the starting point of the dynamic window.

[0069] Step S52) Window partitioning step, when P rate >P anchor At that time, the dynamic window is activated, and the power change rate P within the dynamic window is monitored. rate Continuous monitoring is performed when a dynamic window exists in P. rate <Panchor Or the window length of the dynamic window At that time, T max This represents the maximum duration threshold, defining the dynamic window as a low-power charging event.

[0070] Step S53) Target acquisition step: A preset window length threshold is defined as the window length threshold. len For a period of 5 to 30 seconds, dynamic windows corresponding to low-power charging events are filtered. When the window length of the dynamic window is within the window length threshold range, the dynamic window is defined as the window to be detected.

[0071] Step S6) First judgment step: Filter the window to be detected, extract the pure high-frequency current voltage in the filtered window to be detected, extract the feature vector of the pure high-frequency current, and use a random forest model to make a preliminary judgment on the window to be detected based on the obtained feature vector. If the judgment result is true, the window to be detected is determined to be a preliminary judgment window, and the next step is executed.

[0072] Furthermore, step S6) the first judgment step specifically includes step S61) window filtering step, step S62) first feature extraction step, step S63) feature filtering step and step S64) model judgment step.

[0073] Step S61) Window filtering step: Calculate the difference P3-P0 between the power P0 at the opening of the window to be detected and the power P3 3s after the window opens. When the difference P3-P0 is within a preset charging power threshold range, the charging power threshold is a set charging power range that can be set according to user needs. The filtered window to be detected is obtained, the filtered window to be detected is defined as the filtering window, and the next step is executed.

[0074] Step S62) First feature extraction step: Based on the first cycle at time t and the second cycle at time t+3s in the filtering window, extract the pure high-frequency current and voltage data of charging events within the filtering window by subtracting the first cycle from the second cycle, and extract the entropy, stability and frequency of the pure high-frequency current and voltage.

[0075] In this embodiment, the first cycle is the aligned complete cycle W obtained at time t of the filtering window. p The second cycle is the 10 aligned complete cycles W at time t+3s of the filtering window. b .

[0076] Step S63) Feature filtering step: By performing vertical feature vector clustering on the extracted entropy, stationarity and frequency from both intra-category and inter-category directions, similar feature vectors are obtained. The filtered entropy, stationarity and frequency no longer participate in the model's judgment, thereby improving the accuracy of the model's judgment.

[0077] Step S64) Model judgment step: By calculating the first feature, the second feature and the third feature of the filtering window, the first feature, the second feature and the third feature are used as model input, and a trained high-recall version of the random forest model is used to make a preliminary judgment on the low-power charging events corresponding to the filtering window from the perspective of similar feature vectors.

[0078] like Figure 6 As shown, step S64) model judgment step specifically includes step S65) first feature calculation step, step S66) second feature calculation step and step S67) third feature calculation step.

[0079] Step S65) First feature calculation step: Calculate the probability of pattern repetition in the filtering window to measure the randomness and regularity of the filtering window. The formula is as follows: ; ; ; ; ; Where X(i) represents the m-dimensional vector reconstructed from the input sequence, representing the first mode, i=1,2,...,N-m+1, j=1,2,...,N-m+1, u(i) represents the input after combining current and voltage in the filtering window, and d[X(i),X(j)] represents the maximum value of the distance difference between vector X(i) and vector X(j). Let m be another m-dimensional vector reconstructed from the input sequence, representing the second pattern, k=0,1,...,m-1. Let d[X(i),X(j)]≤r represent the proportion of the second pattern satisfying d[X(i),X(j)]≤r, where r represents the threshold radius for similarity determination. Represents the pattern log probability in an m-dimensional vector. ApEn(m,r,N) represents the pattern log probability under an m+1 dimensional vector. ApEn(m,r,N) represents the difference that measures the randomness and regularity of the sequence, i.e., the first feature. The first feature is actually an approximate entropy. The higher the value, the stronger the randomness of the sequence (such as noise), and the lower the value, the stronger the regularity (such as periodic signals).

[0080] Step S66) Second feature calculation step: Calculate the first-order difference term of the filtering window, determine whether the filtering window is stationary, add a lag difference term to the calculation formula to eliminate the influence of autocorrelation, effectively improving the processing capability of high-order autocorrelation data. The formula is as follows:

[0081] Among them, y t This represents the raw observations of the filter window at time t. The first-order difference term is represented by α, which is the second characteristic term, and the lag term y is represented by α. t-1 The coefficient, β i Represents the difference lag term The coefficient, k, represents the maximum lag of the difference term. The second feature essentially tests whether the sequence has statistical properties that change over time (such as trends or abrupt changes). It is actually a hypothesis testing method. If α is significant, the time series of the screening window is stationary; if α is not significant, the time series of the screening window is not stationary.

[0082] Step S67) The third feature calculation step calculates the power spectral density of the filtering window, quantifies the energy distribution of the signal at different frequencies within the filtering window, effectively limits spectral leakage, and is suitable for processing short data or non-stationary signals. The formula is as follows:

[0083] in, x represents the power spectral density. k (n) represents the nth sampling point of the kth subsequence. Represents the window function. The normalization coefficients of the window function are represented by K, where K represents the total number of segments. denoted by angular frequency, and N represents the length of each subsequence.

[0084] Step S7) Second judgment step: Using the FITS model with frequency domain transformation as a classifier, input the pure high-frequency current and voltage data of the preliminary judgment window into the FITS model for secondary judgment. If the judgment result is true, the preliminary judgment window is determined to be a power bank event.

[0085] like Figure 7 As shown, step S7) the second judgment step specifically includes step S71) frequency domain conversion step, step S72) high frequency interference cancellation step, step S73) second feature extraction step and step S74) feature conversion step.

[0086] Step S71) Frequency domain conversion step: Based on Fourier transform, the pure high-frequency current and voltage data is converted into frequency domain information, and amplitude and phase information are extracted.

[0087] Step S72) High-frequency interference cancellation step: High-frequency interference in the frequency domain information background is eliminated by using a low-pass filter.

[0088] Step S73) The second feature extraction step involves adjusting the frequency domain information components using a linear layer with complex values ​​to extract features in the frequency domain, thereby obtaining frequency domain features.

[0089] Step S74) Feature transformation step: Convert the frequency domain features into time domain features through inverse Fourier transform, input the time domain features into the TimeGAN-fft model, and use a classifier to determine whether the preliminary judgment window is a power bank event.

[0090] Step S8) The third judgment step defines the time T when the power bank sub-event is first identified. The first cycle obtained at time T is recorded as the anchor point. Starting from time T, the sliding window is continued for the low-power charging event to obtain the subsequent detection window of the low-power charging event. The subsequent detection window and the second judgment step are executed for the subsequent detection window. When the subsequent detection window is determined to be false, the anchor point is re-acquired. When the subsequent detection window is determined to be true, the third judgment step is repeated until a preset number of subsequent detection windows are determined to be power bank sub-events. The preset number is not less than 9 subsequent detection windows. The low-power charging event is then determined to be a power bank event.

[0091] In this embodiment, after a power bank event is detected, if there is no requirement for the immediacy of the identification result, a longer-term determination can be enabled to further enhance the accuracy of the identification result.

[0092] Step S9) Long-term scale identification step: Based on the obtained power bank event pure power curve, the power bank pure power curve is matched with the standard curve in the power bank curve database. If the match is successful and the power bank event meets a preset condition, then the electrical appliance corresponding to the power bank event is determined to be a power bank. In this embodiment, the preset condition is that the charging time of the power bank event lasts for more than 1 hour and the power bank pure power curve shows a three-segment feature of "rise-plateau-fall".

[0093] like Figure 8 As shown, step S9) long-term scale identification step specifically includes step S91) power time series acquisition step, step S92) power curve extraction step and step S93) DWT distance calculation step.

[0094] Step S91) Power timing sequence acquisition step: acquire the active power, reactive power, and distortion power at 10Hz during the 5 minutes before and 3 hours after the power bank event is turned on, and generate a power timing sequence.

[0095] Step S92) Power curve extraction step: Based on the obtained power time series, the power time series is segmented using a non-overlapping sliding window with a fixed window length. The time-domain features of the power time series are extracted to obtain the pure power change curve of the power bank event. Windows containing large active power jumps are removed to eliminate background appliance interference. The power rise and fall slopes of each window are statistically analyzed. Based on the slope distribution of power changes in all windows, the actual charging time, power change curve, and other information of the appliance are analyzed. Based on the slope information, the power change curve of the appliance is redrawn to eliminate background appliance interference.

[0096] Step S93) DWT distance calculation step: Calculate the DWT distance between the pure power change curve and each standard curve in the power bank curve database. When the DWT distance is within a preset DWT distance threshold range, the match is determined to be successful. The specific calculation process of DWT distance is easy for those skilled in the art to understand, and will not be described in detail here.

[0097] The advantage of this application lies in providing a power bank load identification technology. It uses an FFT+MLP network structure to replace the LSTM in the TimeGAN model to obtain a TimeGAN-FFT model. Through the trained TimeGAN-FFT model, data augmentation processing is performed on the high-frequency current and voltage data to ensure the extraction of low-power load events from the high-frequency current and voltage data, thus improving the accuracy of low-power load event extraction. By combining short-time high-frequency features and long-time low-frequency features to analyze the sliding window of low-power load events, when the detected sliding window meets the preset judgment criteria, the low-power load event is determined to be a power bank event, improving the accuracy of low-power load event identification and eliminating charging safety hazards of power banks in users' daily electricity usage scenarios.

[0098] The above provides a detailed description of a power bank load identification technology. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A power bank load identification technology, characterized in that, Specifically, the steps include the following: The data acquisition step involves using high-precision current and voltage sensors to collect high-frequency current and voltage data of low-power charging appliances in real time. The model construction steps involve replacing the LSTM in the TimeGAN model with an FFT+MLP network structure to obtain the TimeGAN-FFT model. The model training steps, based on the collected high-frequency current and voltage data and the TimeGAN-fft model, yield the reconstruction loss function L. R Unsupervised loss function L U and the supervised loss function L S Through L R +L S and L U +L S The TimeGAN-fft model is trained. The data augmentation step involves inputting the collected high-frequency current and voltage data into the trained TimeGAN-fft model to perform data augmentation processing on the high-frequency current and voltage data, thereby obtaining augmented data. The event extraction step involves extracting low-power charging events from the enhanced data based on a dynamic window triggering mechanism and a multi-scale event perception method, thereby obtaining the low-power charging events and the corresponding detection windows. The first judgment step involves filtering the window to be detected, extracting the pure high-frequency current and voltage within the filtered window, extracting the feature vector of the pure high-frequency current, and using a random forest model to make a preliminary judgment on the window to be detected based on the obtained feature vector. If the judgment result is true, the window to be detected is determined to be a preliminary judgment window, and the next step is executed. The second judgment step involves using the FITS model with frequency domain conversion as a classifier, inputting the clean high-frequency current and voltage data of the preliminary judgment window into the FITS model for secondary judgment, and if the judgment result is true, determining that the preliminary judgment window is a power bank event. as well as The third judgment step defines the time T at which the power bank sub-event is first identified. The first cycle obtained at time T is recorded as the anchor point. Starting from time T, the sliding window is continued for the low-power charging event to obtain the subsequent detection window of the low-power charging event. The first judgment step and the second judgment step are performed on the subsequent detection window. When the subsequent detection window is determined to be false, the anchor point is re-acquired. When the subsequent detection window is determined to be true, the third judgment step is repeated until a preset number of subsequent detection windows are determined to be power bank sub-events, and the low-power charging event is determined to be a power bank event.

2. The power bank load identification technology as described in claim 1, characterized in that, The TimeGAN-fft model includes an embedding layer, a reconstruction layer, a generator, and a classifier. The model training steps specifically include the following steps: In the random noise calculation step, the high-frequency current and voltage data are used to obtain a noise sequence Z through a generator, and its calculation formula is as follows: Where V = {v1, v2, v3, ..., v l }, I={i1,i2,i3,...,i l }, ZV represents the noise sequence of high-frequency voltage data, Z This represents a noise sequence of high-frequency current data, where l represents the length of the voltage and current data (which is 1), and z represents the noise value, derived from V = {v1, v2, v3, ..., v...}. l }, I={i1,i2,i3,...,i l } This allows us to obtain the original time series data pair X=[V,I]; In the hidden state acquisition step, the high-frequency current and voltage data and the noise sequence are obtained through the embedding layer to acquire the hidden state of the high-frequency current and voltage data and the hidden state of the noise sequence. The hidden state of the high-frequency current and voltage data is defined as the first hidden state H, and the noise sequence is defined as the second hidden state. ; In the time-domain data restoration step, the first hidden state is restored by the reconstruction layer to obtain the reconstructed sequence X'=[V',I'] from the original time-series data pair; In the frequency domain component classification step, the first hidden state and the second hidden state are processed by a classifier to obtain the discrimination result of the first hidden state and the discrimination result of the second hidden state. The discrimination result of the first hidden state is defined as the first discrimination result Y, and the discrimination result of the second hidden state is defined as the second discrimination result. ; The first loss function calculation step calculates the reconstruction loss function L. R Its formula is Among them, E s,x1:l~p This represents a simulation of the high-frequency current and voltage data generation process. An attribute 's' is sampled from a static feature distribution, and a time-series data segment x_{1:T} is generated based on attribute 's'. Finally, the statistical average of the objective function is calculated, where 't' represents the time step and x_{1:T} represents the time step. t Represents the actual observed values ​​of the input data. Reconstructed values ​​representing the input data; The second loss function calculation step involves calculating the unsupervised loss function L. U Its formula is Among them, y t This represents the classifier's output on the high-frequency current and voltage data; The output of the classifier for the augmented data; and The third step is to calculate the supervised loss function L. S Its formula is Among them, h t Indicates a hidden state. This represents a random noise vector.

3. The power bank load identification technology as described in claim 1, characterized in that, The event extraction step specifically includes the following steps: The threshold setting step involves obtaining the background power P of the window to be detected. bg Calculate the standard deviation of background power over a preset time period. And the power standard deviation P between the current time and the preset time. std When P std > At that time, the dynamic power change rate threshold P anchor =r*P bg r is the proportionality coefficient, and the power change rate is... P start Indicates the power at the starting point of the dynamic window; Window partitioning steps, when P rate >P anchor At that time, the dynamic window is activated, and the power change rate P within the dynamic window is monitored. rate Continuous monitoring is performed when a dynamic window exists in P. rate <P anchor Or the window length of the dynamic window At that time, T max This represents the maximum duration threshold, defining the dynamic window as a low-power charging event. as well as The target acquisition step involves setting a window length threshold and filtering dynamic windows corresponding to low-power charging events. When the window length of the dynamic window is within the window length threshold range, the dynamic window is defined as the window to be detected.

4. The power bank load identification technology as described in claim 1, characterized in that, The first determination step specifically includes the following steps: In the window filtering step, the difference P3-P0 between the power P0 at the opening of the window to be detected and the power P3 3 seconds after the window opens is calculated. When the difference P3-P0 is within a preset charging power threshold range, the filtered window to be detected is obtained. The filtered window to be detected is defined as the filtering window, and the next step is executed. The first feature extraction step involves extracting the pure high-frequency current and voltage data of charging events within the filtering window by subtracting the first and second cycles at time t+3s based on the first cycle and the second cycle in the filtering window, and extracting the entropy, stability, and frequency of the pure high-frequency current and voltage. The feature selection step involves vertically clustering the extracted entropy, stationarity, and frequency feature vectors from both intra- and inter-category perspectives to obtain similar feature vectors. as well as The model judgment step involves calculating the first feature, second feature, and third feature of the filtering window, using these features as model inputs, and then using a trained high-recall version of the random forest model to make a preliminary judgment on the low-power charging events corresponding to the filtering window from the perspective of similar feature vectors.

5. The power bank load identification technology as described in claim 4, characterized in that, The model judgment step specifically includes the following steps: The first feature calculation step involves calculating the probability of pattern repetition within the filtering window, used to measure the randomness and regularity of the filtering window. The formula is as follows: ； ； ； ； ； Where X(i) represents the m-dimensional vector reconstructed from the input sequence, representing the first mode, i=1,2,...,N-m+1, j=1,2,...,N-m+1, u(i) represents the input after combining current and voltage in the filtering window, and d[X(i),X(j)] represents the maximum value of the distance difference between vector X(i) and vector X(j). Let m be another m-dimensional vector reconstructed from the input sequence, representing the second pattern, k=0,1,...,m-1. Let d[X(i),X(j)]≤r represent the proportion of the second pattern satisfying d[X(i),X(j)]≤r, where r represents the threshold radius for similarity determination. Represents the pattern log probability in an m-dimensional vector. The pattern log probability is represented by the m+1 dimensional vector, and ApEn(m,r,N) represents the difference that measures the randomness and regularity of the sequence, i.e., the first feature; The second feature calculation step involves calculating the first-order difference term of the filtering window and determining whether the filtering window is stationary. The formula is as follows: Among them, y t This represents the raw observations of the filter window at time t. The first-order difference term is represented by α, which is the second characteristic term, and the lag term y is represented by α. t-1 The coefficient, β i Represents the difference lag term The coefficient, k, represents the maximum lag of the difference term. Represents the white noise error term; and The third feature calculation step involves calculating the power spectral density of the filtering window, quantifying the energy distribution of the signal at different frequencies within the filtering window, using the following formula: in, x represents the power spectral density. k (n) represents the nth sampling point of the kth subsequence. Represents the window function. The normalization coefficients of the window function are represented by K, where K represents the total number of segments. denoted by angular frequency, and N represents the length of each subsequence.

6. The power bank load identification technology as described in claim 1, characterized in that, The second judgment step specifically includes the following steps: The frequency domain conversion step converts the pure high-frequency current and voltage data into frequency domain information based on Fourier transform, and extracts amplitude and phase information; The high-frequency interference cancellation step involves using a low-pass filter to eliminate high-frequency interference in the frequency domain information background. The second feature extraction step involves adjusting the frequency domain information components using a linear layer with complex values ​​to extract features in the frequency domain, thereby obtaining frequency domain features. as well as The feature transformation step involves converting the frequency domain features into time domain features through inverse Fourier transform, inputting the time domain features into the TimeGAN-fft model, and using a classifier to determine whether the preliminary judgment window is a power bank event.

7. The power bank load identification technology as described in claim 1, characterized in that, The third determination step is followed by the following steps: In the long-term scale identification step, based on obtaining the pure power curve of the power bank event, the pure power curve of the power bank is matched with the standard curve in the power bank curve database. If the match is successful and the power bank event meets a preset condition, then the electrical appliance corresponding to the power bank event is determined to be a power bank.

8. The power bank load identification technology as described in claim 7, characterized in that, The long-term scale recognition step specifically includes the following steps: The power timing sequence acquisition step involves acquiring the active power, reactive power, and distortion power at 10Hz during the 5 minutes before and 3 hours after the power bank event is activated, and generating a power timing sequence. The power curve extraction step involves segmenting the power time series based on the obtained power time series using a non-overlapping sliding window with a fixed window length, extracting the time-domain features of the power time series, and obtaining the pure power change curve of the power bank event. as well as The DWT distance calculation step involves calculating the DWT distance between the pure power change curve and each standard curve in the power bank curve database. When the DWT distance is within a preset DWT distance threshold range, a successful match is determined.

Images