Intelligent electric meter online monitoring method based on multi-sensor fusion and deep learning

Abstract

The invention relates to the technical field of electric meter online monitoring, in particular to an intelligent electric meter online monitoring method based on multi-sensor fusion and deep learning. Comprising the following steps: acquiring electrical quantity, environment quantity and equipment state data of multiple sensors in real time, and supplementing missing values in a data acquisition process through a local weighted regression method; decomposing each collected data signal into a low-frequency part and a high-frequency part by using wavelet transform, then removing noise of the high-frequency part, and carrying out abnormal value detection on the denoised signal by using a standard deviation method; data collected by multiple sensors are processed in combination with multiple deep learning models, and feature extraction and fusion are carried out on the data in the processing process by adopting bidirectional LSTM, multi-scale cavity convolution, LSTM and Transform models. Through the multi-sensor fusion technology, the problem that the false alarm rate of real-time monitoring of the intelligent electric meter is high is solved, and high-precision intelligent online monitoring of the electric meter is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of online monitoring of electric meters, and in particular to an online monitoring method of an intelligent electric meter based on multi-sensor fusion and deep learning. Background Art

[0002] With the rapid development of smart grids, smart meters, as core equipment on the user side of the power system, undertake important functions such as electricity metering, data collection, and status monitoring. However, the monitoring capabilities of traditional smart meters mainly rely on single electricity data (such as current, voltage, and power), which makes it difficult to cope with the increasingly complex power consumption environment and diverse abnormal scenarios. For example, the means of electricity theft are becoming increasingly covert (such as neutral line shunting and magnetic field interference), and the equipment failure modes are complex and varied (such as overheating of terminal blocks and harmonic pollution). Traditional threshold alarms and statistical analysis methods can no longer meet the needs of high-precision and real-time monitoring. In addition, when smart meters operate in harsh environments such as high temperature and strong magnetic fields, their metering accuracy and reliability may be significantly affected, further increasing the difficulty of power grid management.

[0003] In the existing technology, anomaly detection of smart meters is mainly based on rule engines or shallow machine learning models (such as support vector machines and random forests). Although these methods can identify simple abnormal patterns to a certain extent, they are insufficient in the integration of high-dimensional, nonlinear multi-source data (such as electrical quantities, environmental quantities, and spatiotemporal information), resulting in high false alarm rates and low detection efficiency. At the same time, traditional methods usually rely on centralized cloud computing, with large data processing delays, making it difficult to meet the needs of real-time monitoring. Especially in the context of large-scale access to distributed energy (such as photovoltaics and energy storage), the bidirectional flow of electricity and intermittent power generation characteristics further increase the complexity of the power grid. Traditional meter monitoring methods face huge challenges in data collection, analysis, and decision-making response.

[0004] In recent years, deep learning technology has achieved remarkable results in image recognition, natural language processing and other fields, but its application in power systems is still in the exploratory stage. Existing research focuses on anomaly detection of a single data source (such as current waveform), lacking deep fusion and collaborative analysis of multi-sensor data. In addition, deep learning models usually have high computational complexity and are difficult to deploy directly on resource-constrained edge devices (such as smart meters), which limits their application in practical scenarios. Therefore, how to design an efficient and reliable online monitoring method for smart meters that combines multi-sensor fusion and deep learning has become a key issue that needs to be urgently addressed in the current smart grid field.

[0005] Existing intelligent meter monitoring technologies have obvious deficiencies in aspects such as multi-source data fusion, complex anomaly detection, and real-time guarantee, making it difficult to meet the requirements of the high-quality development of smart grids. The purpose of this invention is to construct a high-precision, low-latency, and strongly robust online monitoring method for intelligent meters through the organic combination of multi-sensor fusion and deep learning technologies, providing technical support for the safe and efficient operation of the power grid. Summary of the Invention

[0006] The purpose of this invention is to solve the problem of high false alarm rate in real-time monitoring of intelligent meters through multi-sensor fusion technology, and propose an online monitoring method for intelligent meters with high precision, low latency, and strong robustness.

[0007] The technical solution of this invention: An online monitoring method for intelligent meters based on multi-sensor fusion and deep learning, including:

[0008] Real-time collect electrical quantities, environmental quantities, and equipment status data of multiple sensors, and supplement missing values during the data collection process through the locally weighted regression method;

[0009] Use wavelet transform to decompose the collected data signals into low-frequency and high-frequency parts, then remove the noise in the high-frequency part, and detect outliers in the denoised signals through the standard deviation method;

[0010] Combine multiple deep learning models to process the collected data of multiple sensors, and use bidirectional LSTM, multi-scale dilated convolution, LSTM, and Transformer models to extract and fuse features from the data during the processing process;

[0011] Obtain the fused features, and combine classification and regression tasks and cloud-edge collaboration strategies to perform online monitoring of intelligent meters.

[0012] Optionally, in the cleaning strategy for missing values and outliers, assume that the data x at a certain moment t is missing, and use the non-missing data points {x 1 , x 2 , …, x k} near this moment for regression filling, and the weights are determined by the distance; the formula for locally weighted regression is as follows:

[0013]

[0014] Among them, w i is the weight of the data point x i , which is set according to the distance d i between the data point x t and the missing point x i . The commonly used weighting function is the Gaussian weight, which is expressed as:

[0015]

[0016] σ is the standard deviation of the Gaussian function.

[0017] Optionally, let the signal of sensor i be x i (t), and the low-frequency part obtained after wavelet transform is A i (t), and the high-frequency part is D i (t), then:

[0018] x i (t) = A i (t) + D i (t)

[0019] During the denoising process, the soft threshold method is used to perform threshold processing on the high-frequency part D i (t), and we get:

[0020]

[0021] where λ is the threshold, and the denoised signal is:

[0022]

[0023] Optionally, let the signal value at a certain moment deviate from the normal range of this sensor. The formula for the standard deviation method to detect outliers is:

[0024]

[0025] where μ i and σ i are the mean and standard deviation of the signal of sensor i respectively, and k is the set threshold.

[0026] Optionally, the feature extraction includes current data feature extraction, voltage data feature extraction, temperature data feature extraction, and vibration data feature extraction.

[0027] Optionally, the current data feature extraction and voltage data feature extraction use bidirectional LSTM to simultaneously capture the forward information and backward information in the current data and voltage data. The bidirectional LSTM consists of two LSTMs. One learns features from the front end of the sequence, and the other learns features from the back end of the sequence. The calculation process is as follows:

[0028]

[0029]

[0030]

[0031] where x tis the input of the current or voltage signal at time step t, and are the hidden states of the forward and backward LSTMs respectively, and the final output h t is the result of concatenating the features in both directions.

[0032] Optionally, the temperature data feature extraction uses multi-scale dilated convolutions to capture features at different time scales to adapt to the changes in temperature data. The calculation formula for multi-scale dilated convolutions is as follows:

[0033]

[0034] where y t is the output after the convolution operation, x t is the input of the temperature signal at time step t, w i is the weight of the convolution kernel, k is the length of the convolution kernel, and d is the dilation factor.

[0035] Optionally, for the multi-scale dilated convolution, it is processed through convolution kernels with multiple different dilation factors, and after the processing output, concatenation or weighted fusion is performed, expressed as:

[0036]

[0037] where d scale is the dilation factor set for different time steps.

[0038] Optionally, the vibration data feature extraction captures the temporal features in the signal through LSTM, and remembers and forgets the features in the vibration signal through the input gate, forget gate, and output gate, expressed as:

[0039] h t = LSTM(x t , h t-1 , c t-1 )

[0040] where x t is the input feature data collected by the sensor at time t, h t-1 is the hidden state at the previous time step, c t-1 is the cell state at the previous time step, and h t is the hidden state at the current time step.

[0041] Optionally, after extracting the data features of different sensors, they are concatenated to form a feature vector. The feature vector contains the temporal features and physical attributes of the current, voltage, temperature, and vibration sensor data. The concatenated feature vector is expressed as:

[0042] h concat = [h current , hvoltage , h temperature , h vibration

[0043] Among them, h current , h voltage , h temperature , h vibration are the feature vectors extracted from current, voltage, temperature, and vibration data respectively;

[0044] Then, the concatenated feature vectors are input into the Transformer model for further feature fusion. The Transformer model uses the self-attention mechanism to perform weighted fusion on each feature. The formula is as follows:

[0045] Q = XW q , K = XW k , V = XW v

[0046]

[0047] O = AV

[0048] The finally output O is the fused global feature representation, where Q, K, and V are the matrices of query, key, and value respectively, X is the input feature matrix, and W q , W k , W v are the weight matrices respectively.

[0049] Compared with the prior art, the present application includes at least one of the following beneficial technical effects:

[0050] 1. Multi-sensor data fusion improves the accuracy of fault detection and classification: The invention significantly improves the accuracy of device fault detection and classification by fusing multi-sensor data such as voltage, current, temperature, and vibration, and adopting deep learning models such as bidirectional LSTM (BiLSTM), LSTM, multi-scale dilated convolution, and Transformer. The multi-modal data fusion enhances the real-time monitoring ability of power grid devices and reduces the risks of missed fault reports and false alarms.

[0051] 2. Multi-modal feature fusion enhances the accuracy of remaining useful life prediction: The present invention improves the prediction accuracy of the remaining useful life of devices by fusing multi-modal sensor data. After the current, voltage, temperature, and vibration data are processed by deep learning, the Transformer model is used for weighted fusion, so as to accurately predict the device fault time, optimize the maintenance strategy, and reduce the downtime.

[0052] ​3. Optimizing task performance through multi-scale feature extraction: The present invention extracts features from temperature data through multi-scale dilated convolutions, and combines BiLSTM and LSTM to process other sensor data, improving the accuracy of fault detection and prediction of the remaining useful life of the device. Multi-scale feature extraction enhances the system's sensitivity to different levels of device features, improving the overall prediction performance and robustness.

[0053] 4. Collaborating edge computing and cloud computing to enhance the overall system performance: Through the collaborative work of edge computing and cloud computing in the present invention, edge devices quickly process data, and the cloud performs global optimization and training, improving the accuracy of fault detection and prediction. This architecture ensures low-latency response and efficient data analysis, giving full play to the computing advantages of the edge and the cloud and enhancing the overall performance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0055] The technical solutions of the present disclosure will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present disclosure. The components of the embodiments of the present disclosure described and shown herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the present disclosure claimed, but merely represents selected embodiments of the present disclosure. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present disclosure without creative efforts shall fall within the scope of protection of the present disclosure.

[0056] As Figure 1 shown, the online monitoring method for smart meters based on multi-sensor fusion and deep learning of the present invention mainly includes the following three major processes:

[0057] I. Multi-sensor data acquisition method

[0058] 1.1 Multi-sensor data acquisition

[0059] In the online monitoring of this smart meter, data acquisition relies on a variety of high-precision sensors, aiming to comprehensively monitor the operation status of the power grid and ensure the safety and stability of the equipment. This invention adopts a voltage sensor with an accuracy of ±0.2%, precisely acquiring the voltage signals of each node in the power grid, reflecting voltage fluctuations and abnormalities in real time, and ensuring the stability of the power grid voltage. The current sensor is used to monitor the current changes in the power grid, detecting potential fault risks such as load fluctuations, overloads, or short circuits. This invention is also equipped with a high-precision temperature sensor with a measuring range of -40°C to 85°C and a resolution of 0.1°C, specifically used to monitor the temperature changes of the meter and related power equipment in real time, effectively preventing equipment damage caused by overheating. In addition, this invention integrates a triaxial accelerometer with a bandwidth of 0 to 5 kHz as a vibration sensor, used to detect the vibration status of the meter equipment and its attached equipment, helping to identify mechanical failures or equipment aging problems in a timely manner, and improving the prediction and maintenance capabilities of the system. Through the collaborative work of these sensors, this invention can achieve efficient monitoring and fault warning of various operation indicators of the power grid.

[0060] 1.2 Multi-sensor data cleaning strategy

[0061] In the online monitoring of smart meters, the quality of data directly affects the accuracy of subsequent analysis and model prediction. To ensure data quality, this invention designs a cleaning strategy for missing values and outliers, ensuring that the final data can provide reliable monitoring results. Missing values are usually caused by reasons such as sensor failures, communication interruptions, or external factors. In the smart meter monitoring system, missing values will affect the continuity of data and the prediction ability of the model. This invention designs a locally weighted regression method to implement the supplementation of missing values. Assume that the data x t is missing at a certain moment. Use the non-missing data points {x 1 , x 2 , …, x k} near this moment for regression filling, and the weights are determined by the distance. The formula for locally weighted regression is as follows:

[0062]

[0063] where w i is the weight of the data point x i , which can be set according to the distance d i between the data point and the missing point. The commonly used weighting function is the Gaussian weight:

[0064]

[0065] Here, σ is the standard deviation of the Gaussian function, and d i is the distance from the data point x i to the missing point x t .

[0066] 1.3 Noise Removal Strategy Based on Wavelet Transform

[0067] Noise Removal Method

[0068] Wavelet transform can decompose a signal into a low-frequency part (i.e., the approximation signal) and a high-frequency part (i.e., the detail signal). In multi-sensor signals, the low-frequency part usually contains the main information of the signal, while the high-frequency part usually represents noise or irrelevant details. By processing the high-frequency part (such as denoising), the quality of the signal can be effectively improved. Assume that the signal of sensor i is x i (t), the low-frequency approximation part obtained after wavelet transform is A i (t), and the high-frequency detail part is D i (t), then:

[0069] x i (t) = A i (t) + D i (t)

[0070] During the denoising process, the soft threshold method is used to perform threshold processing on the high-frequency detail part D i (t):

[0071]

[0072] where λ is the threshold, usually selected by cross-validation. The denoised signal is:

[0073]

[0074] Outlier Detection Method

[0075] In multi-sensor data, outliers may be caused by device failures or external interferences. The signal after wavelet transform denoising is usually smoother, so outliers can be detected more effectively. Assume that in the denoised signal, the signal value at a certain moment significantly deviates from the normal range of this sensor. The outlier detection is performed by the standard deviation method:

[0076]

[0077] where μ i and σ i are the mean and standard deviation of the signal of sensor i respectively, and k is the set threshold (for example, k = 3). If this condition holds, then this signal point is considered an outlier, and the above-mentioned locally weighted regression method can be used for repair.

[0078] Through the above data acquisition method, high-quality multi-sensor data acquisition can be achieved.

[0079] II. Multi-Sensor Feature Fusion Network

[0080] In the present invention, by combining multiple deep learning models to process data from different sensors, efficient power grid status monitoring and fault prediction are achieved. The processing of current, voltage, temperature, and vibration data includes feature extraction and fusion, and the Bidirectional LSTM (BiLSTM), 1D Dilated Convolution, LSTM, and Transformer models are used. The following details the processing of each sensor's data and gives the corresponding formulas.

[0081] 2.1 Multi-Sensor Data Feature Extraction

[0082] 2.1.1 Current and Voltage Data Feature Extraction

[0083] Current and voltage data usually have strong temporal dependencies. The Bidirectional LSTM (BiLSTM) can capture both forward and backward information in the data, making it suitable for this type of signal processing. BiLSTM consists of two LSTMs, one learning features from the front end of the sequence and the other from the back end. The specific calculation process is as follows:

[0084]

[0085]

[0086]

[0087] where x t is the input of the current or voltage signal at time step t, and are the hidden states of the forward and backward LSTMs respectively, and the final output h t is the concatenation result of the features in both directions.

[0088] 2.1.2 Temperature Data Feature Extraction

[0089] Temperature data usually has dependencies over a long time span. Using multi-scale Dilated Convolution can effectively capture long-term dependencies and can capture multi-level features in the signal through convolutional kernels of different scales. By introducing a dilation factor, the convolution operation of Dilated Convolution at each time step expands the receptive field and increases the ability to capture long-term trends. Multi-scale Dilated Convolution can capture features at different time scales by setting different dilation rates and convolutional kernel sizes, thus better adapting to the changes in temperature data. The calculation formula of multi-scale Dilated Convolution is as follows:

[0090]

[0091] Among them, y t is the output after the convolution operation, x t is the input of the temperature signal at time step t, w i is the weight of the convolution kernel, k is the length of the convolution kernel, and d is the dilation factor.

[0092] For multi-scale dilated convolution, convolution kernels with multiple different dilation factors are used for processing, and finally their outputs are concatenated or weighted and fused. The specific formula is:

[0093]

[0094] Among them, d scale is the dilation factor set for different contexts.

[0095] 2.1.3 Vibration data feature extraction

[0096] Vibration data usually contains periodic variations and temporal characteristics. LSTM (Long Short-Term Memory Network) can effectively capture the temporal features in the signal. The LSTM network needs to remember and forget the important features in the vibration signal through the input gate, forget gate, and output gate, so as to better handle the long-term and short-term dependencies in the temporal data.

[0097] h t = LSTM(x t , h t-1 , c t-1 )

[0098] Among them, x t is the input feature data collected by the sensor at time t, h t-1 is the hidden state at the previous moment, c t-1 is the cell state at the previous moment, and h t is the hidden state at the current moment.

[0099] 2.2 Feature fusion module

[0100] After the feature extraction by each model, the present invention concatenates the data features from different sensors to form a comprehensive feature vector. This feature vector contains the temporal features of the current, voltage, temperature, and vibration sensor data and their important physical properties. The concatenated feature vector can be expressed as:

[0101] h concat = [h current , h voltage , h temperature , h vibration

[0102] Among them, h​current , h voltage , h temperature , h vibration They are feature vectors extracted from current, voltage, temperature, and vibration data respectively. Then, the concatenated feature vectors are input into the Transformer model for further feature fusion. Transformer uses the self-attention mechanism to perform weighted fusion on each feature, and the formula is as follows:

[0103] Q = XW q , K = XW k , V = XW v

[0104]

[0105] O = AV

[0106] The final output O is the fused global feature representation, where Q, K, and V are the matrices of query, key, and value respectively, X is the input feature matrix, and W q , W k , W v are weight matrices respectively, which are used for subsequent classification tasks such as fault detection and classification, status device classification, load prediction, etc., and regression tasks such as remaining life prediction and power consumption prediction.

[0107] III. Design of Edge-Cloud Collaborative Inference Framework

[0108] With the development of online monitoring technology for smart meters, the amount of data in the power system is increasing continuously. The traditional single-edge or cloud processing mode can no longer meet the requirements of real-time performance, accuracy, and computational efficiency.

[0109] The present invention designs an edge-cloud collaborative inference architecture to achieve efficient online monitoring of electric energy meters. The present invention will combine the advantages of edge computing and cloud computing, and through reasonable data transmission and processing strategies, achieve efficient data processing, real-time analysis, and intelligent decision-making.

[0110] 3.1 Collaborative Inference between Edge Computing Layer and Cloud Computing Layer

[0111] The core task of the edge computing layer is to perform a preliminary analysis of the real-time state of the electric energy meter through the fusion and processing of multi-sensor data. Lightweight deep learning models, such as LSTM, CNN, etc., can be deployed on edge devices to extract features and predict faults from sensor data. The main advantages of edge devices are low latency and fast response, which can immediately trigger an alarm or take response measures when an abnormality occurs in the device.

[0112] The cloud computing layer is responsible for more complex computing tasks, including the storage and processing of large-scale data, the training of deep learning models, and global optimization. By receiving the data reported by edge devices, the cloud can analyze the long-term trends of the power grid, the health status of devices, and perform tasks such as remaining useful life (RUL) prediction and equipment maintenance optimization. The cloud can access the power equipment data of the entire network and further improve the prediction accuracy through collaborative analysis.

[0113] 3.2 Data Flow and Inference Strategy

[0114] 3.2.1 Data Preprocessing and Preliminary Inference (Edge Side)

[0115] On edge devices, the multi-sensor data such as voltage, current, temperature, and vibration collected in real time is first subjected to preliminary cleaning, feature extraction, and processing. Different neural networks are used to process the data of each sensor respectively:

[0116] a) Current and voltage data are subjected to temporal feature extraction through BiLSTM.

[0117] b) Temperature data extracts multi-scale features through multi-scale dilated convolution.

[0118] c) Vibration data captures its temporal features through LSTM.

[0119] After preliminary processing, the data features are concatenated and transmitted to the cloud for further analysis. At this time, the edge device can perform preliminary classification according to the set thresholds or preset rules (for example, whether a fault occurs, whether it is overloaded, etc.) to reduce the burden of transmission to the cloud.

[0120] 3.2.2 Data Aggregation and Global Inference (Cloud)

[0121] In the cloud, the data uploaded by edge devices will be aggregated and stored. The cloud deep learning model uses the data of the entire network for further feature fusion and analysis. Here, the Transformer model weights and fuses the features of different sensor data and outputs high-level feature representations.

[0122] Based on these feature representations, the cloud can perform tasks such as remaining useful life prediction (RUL) and fault classification. The cloud can also improve the accuracy of the model through large-scale multi-sensor data fusion and optimization algorithms. Especially when dealing with large-scale data, the computing power of the cloud can provide strong support.

[0123] It should be noted that in the main substation of a certain city, the operating conditions of power equipment directly affect the power supply stability of the entire region. To ensure the safe and efficient operation of the power system, the substation adopts the intelligent electricity meter online monitoring system in this patent and integrates it with the edge-cloud collaborative inference architecture. The system uses high-precision voltage, current, temperature, and vibration sensors to collect and analyze power grid operation data in real time, and conducts intelligent analysis and fault prediction with the help of deep learning models, so as to achieve early warning of equipment and optimization of maintenance.

[0124] The intelligent electricity meter online monitoring system first preliminarily processes the collected current, voltage, temperature, and vibration data through edge computing devices, and uses BiLSTM, dilated convolution, and LSTM models to extract the temporal features of the signals respectively. After preliminary analysis, when the edge device detects abnormalities, such as current overload or too high temperature, it can quickly issue an alarm and send the relevant data to the cloud for further analysis and optimized decision-making. The cloud uses the Transformer model to perform weighted fusion on different sensor data, provides more accurate fault prediction, and predicts the remaining useful life (RUL) of the equipment, providing decision support for equipment maintenance for operation and maintenance personnel.

[0125] As the system operates, the cloud further optimizes the maintenance strategy according to the long-term health status of the equipment. Through the health assessment of all network equipment, the cloud can reasonably arrange maintenance tasks, avoid unnecessary maintenance, and thus reduce the operation and maintenance costs. Especially during long-term operation, through large-scale data fusion and optimization algorithms, the system not only improves the prediction accuracy, but also effectively reduces the incidence of sudden failures, improving the reliability and stability of the power grid.

[0126] This intelligent electricity meter online monitoring system provides strong guarantee for the equipment health management of the substation. Through the collaborative work of multi-sensor data, the intelligent analysis of deep learning models, and the efficient cooperation of the edge-cloud collaborative inference architecture, the substation can achieve early fault warning and precise maintenance of equipment. Through this intelligent management, the substation improves the safety and stability of power grid operation, successfully reduces the cost of over-maintenance, and extends the service life of equipment.

[0127] The above specific embodiments are only several alternative embodiments of the present invention. Based on the technical solution of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A smart meter online monitoring method based on multi-sensor fusion and deep learning, characterized in that: include: Collect electrical quantities, environmental quantities and equipment status data from multiple sensors in real time, and supplement missing values ​​in the data collection process through local weighted regression methods; The collected data signals are decomposed into low-frequency and high-frequency parts by wavelet transform, and the noise of the high-frequency part is removed. The outlier detection is performed on the denoised signal by standard deviation method. Combine multiple deep learning models to process the collected data from multiple sensors, and perform feature extraction and fusion on the data during the processing; The fused features are obtained, and the classification and regression tasks are combined with the cloud-edge collaborative strategy to perform online monitoring of smart meters.

2. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 1 is characterized in that: In the missing value and outlier cleaning strategy, suppose the data x at a certain time t missing, use the non-missing data points {x1,x2,…,x k }Regression filling, the weight is determined by the distance; the formula for local weighted regression is as follows: Among them, w i is the data point x i The weight of the data point x i With missing point x t The distance d i Assume that the commonly used weighting function is Gaussian weight expressed as: σ is the standard deviation of the Gaussian function.

3. The smart meter online monitoring method based on multi-sensor fusion and deep learning according to claim 1 is characterized in that: Let the signal of sensor i be x i (t), the low-frequency part obtained after wavelet transform is A i (t), the high frequency part is D i (t), then: x i (t)=A i (t)+D i (t) In the denoising process, the soft threshold method is used to i (t) is thresholded and we get: Among them, λ is the threshold, and the denoised signal is:

4. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 1 is characterized in that: The signal value at a certain time Deviating from the normal range of the sensor, the formula for detecting abnormal values ​​using the standard deviation method is: Among them, μ i and σ i are the mean and standard deviation of the sensor i signal, respectively, and k is the set threshold.

5. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 4 is characterized in that: The feature extraction includes current data feature extraction, voltage data feature extraction, temperature data feature extraction and vibration data feature extraction.

6. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 5 is characterized in that: The current data feature extraction and voltage data feature extraction use bidirectional LSTM to simultaneously capture the forward information and reverse information in the current data and voltage data. The bidirectional LSTM consists of two LSTMs, one learning features from the front end of the sequence and the other learning features from the back end of the sequence. The calculation process is as follows: Among them, x t is the input current or voltage signal at time step t, and They are the hidden states of the forward and reverse LSTMs, and the final output is h t It is the result of feature splicing in two directions.

7. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 5 is characterized in that: The temperature data feature extraction uses multi-scale dilated convolution to capture features at different time scales to adapt to changes in temperature data. The calculation formula of multi-scale dilated convolution is as follows: Among them, y t is the output after the convolution operation, x t is the input temperature signal at time step t, w i is the weight of the convolution kernel, k is the length of the convolution kernel, and d is the dilation factor.

8. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 7 is characterized in that: The multi-scale dilated convolution is processed by convolution kernels with multiple different dilation factors, and the processed outputs are concatenated or weighted fused, which is expressed as: Among them, d scale is the expansion factor for different swap settings.

9. The method for online monitoring of smart meters based on multi-sensor fusion and deep learning according to claim 5 is characterized in that: The vibration data feature extraction captures the time series features in the signal through LSTM, and memorizes and forgets the features in the vibration signal through the input gate, forget gate and output gate, which is expressed as: h t =LSTM(x t ,h t-1 ,c t-1 ) Among them, x t is the input feature data collected by the sensor at time t, h t-1 is the hidden state of the previous moment, c t-1 is the cell state at the previous moment, h t is the hidden state at the current moment.

10. The smart meter online monitoring method based on multi-sensor fusion and deep learning according to claim 8, characterized in that: The feature extracts the data features of different sensors and then splices them to form a feature vector. The feature vector contains the time series features and physical properties of the current, voltage, temperature and vibration sensor data. The spliced ​​feature vector is expressed as: h concat =[h current ,h vo l tage ,h temperature ,h vibration ] Among them, h current ,h voltage ,h temperature ,h vibration are the feature vectors extracted from current, voltage, temperature, and vibration data, respectively; The concatenated feature vector is then input into the Transformer model for further feature fusion. The Transformer model uses the self-attention mechanism to perform weighted fusion of each feature. The formula is as follows: Q=XW q ,K=XW k ,V=XW v O=AV The final output O is the fused global feature representation, where Q, K, V are the matrices of query, key, and value, respectively, X is the input feature matrix, and W q , W k , W v are the weight matrices respectively.

Images