CN-121792885-B - Transformer substation wireless meter reading method, system, electronic device and storage medium

CN121792885BCN 121792885 BCN121792885 BCN 121792885BCN-121792885-B

Abstract

The invention discloses a method, a system, an electronic device and a storage medium for reading a wireless meter of a transformer substation, belongs to the technical field of electronic communication, and solves the technical problem that the data acquisition precision of the existing wireless meter is insufficient. The method comprises the steps of collecting multi-mode original data through a plurality of wireless collecting nodes deployed in a Transformer substation, preprocessing the multi-mode original data to obtain preprocessed multi-mode data sets of all the nodes, extracting semantic features of the preprocessed multi-mode data sets to generate unified multi-mode feature representation of each node, processing the unified multi-mode feature representation through a lightweight transducer model to obtain meter reading predicted values and corresponding reliability assessment results of the nodes, carrying out weighted fusion on the meter reading predicted values of all the nodes based on the reliability assessment results to obtain global meter reading, and optimizing parameters of the lightweight transducer model based on errors between the global meter reading and reference values. The reliability and fault tolerance of data reading are realized.

Inventors

ZHANG YONG
CHANG JUN
MA SHAOQIANG
SHI XIAOMIN
ZHONG YI
SHEN JIAWEI
WU MING
ZHANG YING
GUO LEI

Assignees

国网上海市电力公司

Dates

Publication Date: 20260508
Application Date: 20260306

Claims (10)

1. The method for reading the wireless meter of the transformer substation is characterized by comprising the following steps of: S1, acquiring multi-mode original data by using a plurality of wireless acquisition nodes deployed in a transformer substation and preprocessing the multi-mode original data to obtain a preprocessed multi-mode data set of each node; S2, carrying out semantic feature extraction on the preprocessed multi-modal dataset to generate a unified multi-modal feature representation of each node; S3, processing the unified multi-mode feature representation by using a lightweight transducer model to obtain a meter reading predicted value and a corresponding reliability evaluation result of the node; The step S3 includes: performing input projection on the unified multi-mode feature representation of each node to the dimension of the lightweight transducer model, and adding and fusing the position codes of the unified multi-mode features of each node to obtain a first feature vector; The first feature vector is subjected to addition fusion with the first feature vector after passing through a transducer layer, and a second feature vector is obtained; carrying out batch normalization on the second feature vectors, carrying out nonlinear conversion through a feedforward network, and then carrying out addition fusion on the second feature vectors to obtain third feature vectors; carrying out batch normalization and average pooling on the third feature vector to obtain a fourth feature vector; inputting the fourth feature vector into a dual-task output layer to obtain a meter reading predicted value and a reliability evaluation result for representing the reliability of the meter reading predicted value; The output layer of the lightweight transducer model comprises a meter reading prediction branch and a reading reliability judging branch which are connected in parallel; based on the fourth feature vector, the meter reading prediction branch obtains the meter reading predicted value as follows: Wherein, the Is a node A meter reading predictive value for (1); Learning a weight vector for regression; Is a fourth feature vector; learning bias terms for regression; Based on the fourth feature vector, the reading reliability judging branch outputs the reliability evaluation result representing the reliability of the predicted value as follows: Wherein, the Predicted value for meter reading Reliable probability; activating a function for Sigmoid; a learnable weight vector for classification; Learning bias items for classification; And obtaining a probability distribution vector of a reliability evaluation result by adopting multi-stage reliability, wherein the probability distribution vector is as follows: Wherein, the For the probability distribution vector of the reliability evaluation result, Is the reliability category number; S4, weighting and fusing meter reading predicted values of all nodes based on the reliability evaluation result to obtain global meter reading; The step S4 includes: the method comprises the steps that a master control node is used for receiving ternary group information comprising meter reading predicted values, reliability assessment results and time stamps, which are periodically reported by all acquisition nodes, and fusion weights are distributed to the meter reading predicted values of all the acquisition nodes based on the reliability assessment results, and weighted average is carried out, so that global meter reading is obtained; The main control node calculates the variance of the global meter readings of all reporting nodes at the current moment, and if the variance exceeds a first preset threshold or the mean value of the reliability evaluation result is lower than a second preset threshold, an abnormal alarm and data re-acquisition process is triggered; constructing a global loss function with an error term between the global meter reading and a reference value and a consistency constraint term between the meter reading predicted values of all the acquisition nodes as cores, minimizing the global loss function by using a gradient descent method, and updating global parameters of a lightweight transducer model local to a main control node; Updating the parameters of the local lightweight transducer model of each edge computing device associated with each wireless acquisition node by using the updated global parameters of the local lightweight transducer model of the main control node; when the master control node detects that the global meter reading fluctuation is larger than a preset fluctuation threshold value or the reliability evaluation result is smaller than a preset reliability threshold value, automatically adjusting and updating parameters of a lightweight transducer model of the local master control node; The global loss function is as follows: Wherein, the Reading for a global meter; Is a reference value; Is a consistency constraint coefficient; for consistency constraint terms, for Time, meter reading predicted value predicted by all edge computing devices Is a variance of (2); automatically adjusting and updating parameters of a lightweight transducer model local to a master node, wherein the parameters are as follows: Wherein, the 、 The global learning rates after the t+1st iteration and the t th iteration are respectively; 、 Feedback fusion factors after t+1st iteration and t iteration respectively; For adjusting the coefficients; Is an average reliability index; a second preset threshold value; Wherein parameters of the lightweight transducer model are optimized based on an error between the global meter reading and a reference value.
2. The method for reading a wireless meter of a transformer substation according to claim 1, wherein each wireless acquisition node is integrated with at least a multi-sensor integrated terminal device of an image acquisition unit, an acoustic acquisition unit and a voltage acquisition unit; The image acquisition unit acquires image data of a noisy meter; The acoustic acquisition unit acquires equipment audio data; the voltage acquisition unit acquires equipment voltage data; the noisy meter image data, the device audio data and the voltage data form the multi-modal raw data.
3. The method for reading a wireless meter of a transformer substation according to claim 2, wherein preprocessing the multi-mode raw data comprises: estimating and removing noise from the noisy meter image by minimizing a loss function comprising a data fidelity term and a spatial smoothness regularization term, thereby obtaining a denoised meter image; performing Fourier transform on the equipment audio data to convert the equipment audio data into a frequency domain, and filtering the frequency domain to remove noise to obtain denoised audio data; And carrying out moving average filtering on the voltage data to smooth random interference, and obtaining filtered voltage data.
4. A method for reading a wireless meter of a transformer substation according to claim 3, wherein the step S2 comprises: extracting visual semantic features of the denoised meter image by using a convolutional neural network; The method comprises the steps of denoising audio data, carrying out Mel frequency cepstrum coefficient transformation on the denoised audio data to obtain corresponding frequency spectrum characteristics, and encoding the frequency spectrum characteristics by using a cyclic neural network to obtain acoustic modal context characteristics; Performing nonlinear dimension reduction and feature extraction on the filtered voltage data by using a self-encoder to obtain semantic features of the voltage data; After the visual semantic features, the acoustic mode context features and the semantic features of the voltage data are respectively normalized, the normalized features are respectively projected to a shared semantic space of the same dimension through a learnable parameter matrix, and embedded vectors of the visual semantic features, the acoustic mode context features and the semantic features of the voltage data are respectively obtained through linear transformation; And calculating the attention weights of the embedded vectors of the visual semantic features, the acoustic modal context feature representations and the semantic features of the voltage data by using the learnable query vectors, and carrying out weighted summation on the embedded vectors of the visual semantic features, the acoustic modal context feature representations and the semantic features of the voltage data state based on the corresponding attention weights to generate a unified multi-modal feature representation of each node.
5. The substation wireless meter reading method of claim 1, wherein when the global loss function is in continuous And when the optimization period is not reduced, the main control node rolls back the parameters of the lightweight transducer model local to the main control node to a latest stable model parameter set.
6. The method for reading a wireless meter of a transformer substation according to claim 5, wherein, The value is 5.
7. The method for reading a wireless meter of a transformer substation according to any one of claims 1 to 6, wherein the reference value is obtained by measuring the target meter of the transformer substation in a standard laboratory environment using a high-precision standard meter.
8. The transformer substation wireless meter reading system is characterized by comprising a multi-mode data acquisition and preprocessing module M1, a semantic feature extraction and fusion module M2, a lightweight model reasoning and reliability evaluation module M3 and a global fusion decision and model optimization module M4; the multi-mode data acquisition and preprocessing module M1 is used for acquiring multi-mode original data by using a plurality of wireless acquisition nodes deployed in a transformer substation and preprocessing the multi-mode original data to obtain a preprocessed multi-mode data set of each node; the semantic feature extraction and fusion module M2 is used for extracting semantic features of the preprocessed multi-modal dataset and generating unified multi-modal feature representation of each node; The lightweight model reasoning and reliability evaluation module M3 is used for processing the unified multi-mode characteristic representation by using a lightweight transducer model to obtain a meter reading predicted value and a corresponding reliability evaluation result of the node; the lightweight model reasoning and reliability assessment module M3 comprises: performing input projection on the unified multi-mode feature representation of each node to the dimension of the lightweight transducer model, and adding and fusing the position codes of the unified multi-mode features of each node to obtain a first feature vector; The first feature vector is subjected to addition fusion with the first feature vector after passing through a transducer layer, and a second feature vector is obtained; carrying out batch normalization on the second feature vectors, carrying out nonlinear conversion through a feedforward network, and then carrying out addition fusion on the second feature vectors to obtain third feature vectors; carrying out batch normalization and average pooling on the third feature vector to obtain a fourth feature vector; inputting the fourth feature vector into a dual-task output layer to obtain a meter reading predicted value and a reliability evaluation result for representing the reliability of the meter reading predicted value; The output layer of the lightweight transducer model comprises a meter reading prediction branch and a reading reliability judging branch which are connected in parallel; based on the fourth feature vector, the meter reading prediction branch obtains the meter reading predicted value as follows: Wherein, the Is a node A meter reading predictive value for (1); Learning a weight vector for regression; Is a fourth feature vector; learning bias terms for regression; Based on the fourth feature vector, the reading reliability judging branch outputs the reliability evaluation result representing the reliability of the predicted value as follows: Wherein, the Predicted value for meter reading Reliable probability; activating a function for Sigmoid; a learnable weight vector for classification; Learning bias items for classification; And obtaining a probability distribution vector of a reliability evaluation result by adopting multi-stage reliability, wherein the probability distribution vector is as follows: Wherein, the For the probability distribution vector of the reliability evaluation result, Is the reliability category number; The global fusion decision and model optimization module M4 is used for carrying out weighted fusion on the meter reading predicted values of all the nodes based on the reliability evaluation result to obtain global meter reading; the global fusion decision and model optimization module M4 includes: the method comprises the steps that a master control node is used for receiving ternary group information comprising meter reading predicted values, reliability assessment results and time stamps, which are periodically reported by all acquisition nodes, and fusion weights are distributed to the meter reading predicted values of all the acquisition nodes based on the reliability assessment results, and weighted average is carried out, so that global meter reading is obtained; The main control node calculates the variance of the global meter readings of all reporting nodes at the current moment, and if the variance exceeds a first preset threshold or the mean value of the reliability evaluation result is lower than a second preset threshold, an abnormal alarm and data re-acquisition process is triggered; constructing a global loss function with an error term between the global meter reading and a reference value and a consistency constraint term between the meter reading predicted values of all the acquisition nodes as cores, minimizing the global loss function by using a gradient descent method, and updating global parameters of a lightweight transducer model local to a main control node; Updating the parameters of the local lightweight transducer model of each edge computing device associated with each wireless acquisition node by using the updated global parameters of the local lightweight transducer model of the main control node; when the master control node detects that the global meter reading fluctuation is larger than a preset fluctuation threshold value or the reliability evaluation result is smaller than a preset reliability threshold value, automatically adjusting and updating parameters of a lightweight transducer model of the local master control node; The global loss function is as follows: Wherein, the Reading for a global meter; Is a reference value; Is a consistency constraint coefficient; for consistency constraint terms, for Time, meter reading predicted value predicted by all edge computing devices Is a variance of (2); automatically adjusting and updating parameters of a lightweight transducer model local to a master node, wherein the parameters are as follows: Wherein, the 、 The global learning rates after the t+1st iteration and the t th iteration are respectively; 、 Feedback fusion factors after t+1st iteration and t iteration respectively; For adjusting the coefficients; Is an average reliability index; a second preset threshold value; Wherein parameters of the lightweight transducer model are optimized based on an error between the global meter reading and a reference value.
9. An electronic device for reading a wireless meter of a transformer substation, the electronic device comprising: A memory for storing a computer program; A processor for executing the computer program to implement the substation wireless meter reading method according to any one of claims 1 to 7.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the steps of the substation wireless meter reading method according to any of the claims 1-7.

Description

Transformer substation wireless meter reading method, system, electronic device and storage medium Technical Field The invention belongs to the technical field of electronic communication, and particularly relates to a method, a system, an electronic device and a storage medium for reading a wireless meter of a transformer substation. Background Under the development background of the electric power Internet of things and the intelligent substation, the wireless meter in the substation is increasingly used for metering data, and the data types show multi-mode characteristics such as images, voices, vibration signals, environmental parameters and the like. These heterogeneous data are susceptible to signal attenuation, interference and noise when transmitted in a wireless channel, resulting in reduced reading accuracy and reliability. Meanwhile, the traditional centralized data acquisition and processing mode has the following problems: The data acquisition mode is single, and the characteristics of data of different modes are difficult to consider; the stability of the communication link is insufficient, and data packet loss and delay are easy to occur in a high-density node environment; the data fusion and semantic understanding capability are limited, and the joint intelligent analysis of the multi-mode data is difficult to realize; the system lacks an adaptive optimization mechanism and cannot dynamically adjust the reading strategy according to environmental changes. With the development of 6G (6 Generation, sixth generation mobile communication technology) communication and multi-mode intelligent sensing technology, distributed semantic understanding and fusion decision become key directions for improving the reading reliability of a wireless meter. How to construct a wireless meter data reading method which can realize high-precision multi-mode fusion in a complex environment and has self-adaptive optimization capability becomes the key point and the difficulty of the current research. Disclosure of Invention In view of the above analysis, the embodiment of the invention aims to provide a method, a system, an electronic device and a storage medium for reading wireless meters of a transformer substation, so as to solve the technical problem of insufficient data acquisition precision of the wireless meters in the existing method, and realize efficient, reliable and intelligent reading of multi-source heterogeneous wireless meter data of the transformer substation. The invention provides a method for reading a wireless meter of a transformer substation, which comprises the following steps: S1, acquiring multi-mode original data by using a plurality of wireless acquisition nodes deployed in a transformer substation and preprocessing the multi-mode original data to obtain a preprocessed multi-mode data set of each node; S2, carrying out semantic feature extraction on the preprocessed multi-modal dataset to generate a unified multi-modal feature representation of each node; S3, processing the unified multi-mode feature representation by using a lightweight transducer model to obtain a meter reading predicted value and a corresponding reliability evaluation result of the node; S4, weighting and fusing meter reading predicted values of all nodes based on the reliability evaluation result to obtain global meter reading; Wherein parameters of the lightweight transducer model are optimized based on an error between the global meter reading and a reference value. Further, each wireless acquisition node is at least integrated with a multi-sensor integrated terminal device of an image acquisition unit, an acoustic acquisition unit and a voltage acquisition unit; The image acquisition unit acquires image data of a noisy meter; The acoustic sensing unit collects equipment audio data; the voltage acquisition unit acquires equipment voltage data; the noisy meter image data, the device audio data and the voltage data form the multi-modal raw data. Further, preprocessing the multi-mode raw data, including: estimating and removing noise from the noisy meter image by minimizing a loss function comprising a data fidelity term and a spatial smoothness regularization term, thereby obtaining a denoised meter image; performing Fourier transform on the equipment audio data to convert the equipment audio data into a frequency domain, and filtering the frequency domain to remove noise to obtain denoised audio data; And carrying out moving average filtering on the voltage data to smooth random interference, and obtaining filtered voltage data. Further, the step S2 includes: extracting visual semantic features of the denoised meter image by using a convolutional neural network; The method comprises the steps of denoising audio data, carrying out Mel frequency cepstrum coefficient transformation on the denoised audio data to obtain corresponding frequency spectrum characteristics, and encoding the frequency spectrum characteristics b