Parameter sensing and fault processing method of electric energy measuring instrument verification system

Abstract

The invention relates to a parameter sensing and fault processing method for an electric energy measuring instrument verification system, in particular to a multi-parameter sensing and fault processing method for an automatic verification system, which is characterized by comprising the following steps of: collecting and analyzing signals, acquiring audio and video information of a verification line body by utilizing inspection equipment, and designing important inspection points and inspection periods. Fault information of a verification line body is collected in real time, classification and installation of sensors are carried out, and a sensor network is formed; accumulating a fault knowledge base and marking fault marks, and establishing an assembly line panoramic intelligent sensing model by using a digital twinning technology; carrying out fusion analysis on a multi-parameter sensing model taking digital twinning as a carrier, and carrying out principal component analysis on fault influence factors; performing data noise reduction and dimension reduction processing by using a principal component analysis algorithm; carrying out target detection by adopting an ensemble learning algorithm; and a time sequence event model of a self-attention mechanism is introduced to carry out fault early warning and defect elimination guidance, and the potential fault occurrence probability of the key equipment is predicted.

Description

Technical Field

[0001] The present invention belongs to the technical field of parameter sensing and fault handling methods of a calibration system, and in particular relates to a parameter sensing and fault handling method of a calibration system of an electric energy metering instrument. Background Art

[0002] In order to improve the operation and maintenance efficiency of the automated verification system, various provincial network companies and automated verification system manufacturers have conducted research on the intelligent operation and maintenance and fault handling strategies of the automated verification system. Most verification systems conduct simulation analysis of the pipeline operation based on their own pipeline electrical parameter information, and perform fault alarms by installing color lights on the online body to improve the efficiency of fault operation and maintenance. Regarding fault handling, this article introduces a fault online diagnosis method based on hierarchical symbolic directed graph (HSDG), models and hierarchizes the automated pipeline, and gives a hierarchical symbolic directed graph model of the automated verification pipeline; and combines the dynamic kernel partial least squares support vector regression model to calculate the symbols of each node and output the model sample. This method can determine whether the verification system has a fault to a certain extent, but it cannot help guide fault handling.

[0003] In foreign countries, the automatic testing line of single-phase smart energy meters mainly includes assembly area, energy meter function and insulation performance test area, metering characteristic test area, label printing area, and sealing and packing area. The automatic verification line of energy meters produced by foreign energy meter manufacturers is more suitable for the assembly and testing of energy meter manufacturers. Moreover, foreign manufacturers pay great attention to process control and the optimization of components. Each functional area has been automated, and the overall process still requires manual intervention [12-14]. Since the 1980s, foreign countries have developed various high-performance operation and maintenance monitoring systems, such as the 7200 series of B&N Company in the United States, the M700 monitoring system of Atlanta Company, the TC-PC periodic inspection system of IRD Company, and the 3542 mechanical condition monitoring system of B&K Company in Denmark. The characteristics of this type of system are: the hardware accounts for a large proportion, a large amount of information is collected, the processing speed is fast, the function is strong, and it has a simple fault diagnosis function. Many foreign scientific research institutions and colleges and universities have successively developed various equipment operation and maintenance monitoring and fault diagnosis systems. The characteristics of these systems are: based on microcomputers, centered on software development, with multiple functions, great flexibility, wide applicability, and low system prices. The main problems are: weak hardware functions, mainly patrol inspections, poor real-time performance; low system integration and software commercialization; low reliability of system sensors, secondary instruments, computer systems and software, and difficult to adapt to large-scale system integration applications and intelligent development requirements for data collection and analysis. Foreign research on automated verification systems is mostly aimed at certain individual devices, and in-depth research has not been conducted on the perception analysis and fault handling of electric energy metering instruments.

[0004] In the automated verification system, accurate monitoring and fault handling of equipment is crucial. At present, there are problems such as incomplete information collection, inaccurate fault analysis, and low processing efficiency in the monitoring and fault handling methods of the automated verification system. Summary of the invention

[0005] The main technical problem solved by the present invention is the key problem of multi-parameter intelligent perception and fault handling of the automatic verification system of electric energy metering instruments. First, the automatic verification pipeline modeling based on digital twin technology is carried out, and the multi-parameter perception model of the pipeline is established. Secondly, based on the machine vision integration algorithm, the target detection of typical faults is carried out. Finally, based on the perception model and fault detection, the fault elimination and guidance model is carried out, the fault handling strategy is established, and a pilot application is carried out. A multi-parameter sensing and fault handling method for an automated verification system comprises the following steps: Step 1: Collect and analyze the electrical parameter signals of the automated calibration system, including the normal and abnormal information of the appearance unit, labeling unit, coding unit, loading and unloading robots, the temperature and humidity information of the calibration unit, the fault information of the strong and weak current terminals, and the temperature and humidity information of the laboratory environment. Then, use the inspection equipment to collect the audio and video information of the calibration line, design important inspection points and inspection cycles according to the prone fault points, collect the fault information of the calibration line in real time, and classify and install sensors according to the line information that needs to be collected during the operation and maintenance of the assembly line, so as to form a sensor network for monitoring important information of the assembly line. Step 2: Based on the fault pictures collected on a daily basis, the fault knowledge base is accumulated and the fault marks are annotated to lay a foundation for modeling and analysis. The digital twin technology is used to establish a panoramic intelligent perception model of the assembly line based on the line body electrical parameter signals, audio and video information of the inspection device, and sensor information, thus realizing a 1:1 mapping of the automated verification system from the physical object to the digital model. Step 3: The multi-parameter perception model based on digital twins is used to perform fusion analysis on the electrical parameter signals, audio and video information, sensor data, etc. collected by the model. At the same time, based on the pipeline fault data, multi-modal information fusion analysis is performed. By performing principal component analysis on the factors affecting the fault, the relative collinearity between the factors is eliminated, and the high-dimensional variable space is reduced in dimensionality. Step 4: Data preprocessing: Use the principal component analysis algorithm to perform data noise reduction and dimension reduction on the electrical parameter signals, audio and video information, and sensor data collected at the fault information point, and construct a basic structure of fault data that is easy to analyze; Step 5: Target detection algorithm, using an integrated learning algorithm that combines the One-Stage algorithm and the Two-Stage algorithm in the deep learning algorithm to detect faults; Step 6: Research on pipeline fault handling strategies based on the self-attention mechanism timing model, implement model training and simulation monitoring of various fault data, and perform model adaptation and iterative optimization technology to improve the accuracy of model fault detection and elimination; Step 7: Input the multimodal information obtained on-site into the multi-parameter perception model to realize the model's visualization of the on-site equipment operation status information. By analyzing feasible data and comparing it with the equipment status database data, the possible causes of faults such as positioning errors between targets are predicted. Then, the temporal event model of the self-attention mechanism is introduced. The dependency between faults and the operation status of the assembly line is learned through the self-attention mechanism. Autonomous early warning and fault elimination guidance for different faults are performed. Based on the model status data, the probability of potential failures of key equipment is predicted, and a computational model for fault handling is implemented.

[0006] An automated verification system multi-parameter sensing and fault handling system, comprising: Information collection module, used to collect electrical parameter signals, audio and video information and sensor data of the automated verification system; Model building module, used to build pipeline panoramic intelligent perception model and self-attention mechanism timing model; Data processing module, used to perform data noise reduction, dimensionality reduction and target detection on the collected information; The fault handling module is used to provide fault warning and fault elimination guidance based on the model analysis results, and predict the probability of potential failure of key equipment.

[0007] The beneficial effects of the present invention are: comprehensive information collection: by collecting the electrical parameter signals, audio and video information and sensor data of the automated calibration system, all-round monitoring of each unit of the system is achieved, including the appearance unit, labeling unit, coding unit, loading and unloading robots, calibration unit, laboratory environment, etc., which can timely discover various potential problems and improve the reliability and stability of the system.

[0008] Intelligent perception model: Using digital twin technology to build a panoramic intelligent perception model of the assembly line, a 1:1 mapping from the physical object to the digital model is achieved, which can more intuitively understand the operating status of the system. At the same time, the multi-parameter perception model can integrate and analyze multiple types of information to improve the accuracy of fault diagnosis.

[0009] Efficient fault handling: The use of principal component analysis and BP neural network algorithms for data preprocessing can effectively reduce noise and dimension, and improve data quality. The integrated learning algorithm that combines the One-Stage algorithm with the Two-Stage algorithm can quickly and accurately detect faults. The pipeline fault handling strategy based on the self-attention mechanism timing model can achieve autonomous early warning and fault elimination guidance, and improve fault handling efficiency.

[0010] Predicting potential failures: By analyzing the comparison of multimodal information and equipment status database data, it is possible to predict possible causes of failures such as positioning errors between targets. At the same time, using deep learning tools to predict the probability of potential failures of key equipment provides strong support for preventive maintenance and reduces system failure risks and maintenance costs.

[0011] Strong applicability: The present invention is applicable to various automated verification systems and can be customized according to the characteristics of different systems, and has broad application prospects.

[0012] In summary, the present invention can improve the monitoring accuracy, fault diagnosis accuracy and processing efficiency of the automated verification system, reduce the risk of system failures and maintenance costs, and provide a strong guarantee for the stable operation of the automated verification system. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 :Roadmap for implementing the present invention; Figure 2 :Roadmap for implementing the present invention; Figure 3 : Network architecture diagram of the present invention. DETAILED DESCRIPTION

[0014] The idea of ​​digital twin was named "Information Mirroring Model" by Michael Grieves of the University of Michigan, and then evolved into the term "digital twin". Digital twin is also called digital twin and digital mapping. Digital twin is deeply developed on the basis of MBD. In the process of implementing model-based system engineering (MBSE), enterprises have generated a large number of physical and mathematical models, which laid the foundation for the development of digital twin. In 2012, NASA gave a conceptual description of digital twin: Digital twin refers to the integration of multidisciplinary and multi-scale simulation processes by making full use of physical models, sensors, operation history and other data. It is a mirror image of physical products in virtual space, reflecting the entire life cycle of the corresponding physical entity products. In order to facilitate the understanding of digital twin, Zhuang Cunbo et al. proposed the concept of digital twin, believing that digital twin is the process of digitally defining and modeling the composition, characteristics, functions and performance of physical entities using information technology. Digital twin refers to an information model that is completely equivalent to the physical entity in the computer virtual space, and the physical entity can be simulated, analyzed and optimized based on the digital twin. Digital twins are technology, processes, and methods, while digital twins are objects, models, and data.

[0015] A digital twin is a virtual model of a physical object. It spans the lifecycle of the object and uses real-time data sent from sensors on the object to simulate behavior and monitor operations. Digital twins can replicate many real-world items, from a single piece of equipment in a factory to complete installations, such as wind turbines, or even entire cities. Digital twin technology enables you to monitor the performance of your assets, identify potential failures, and make more informed decisions about maintenance and lifecycles.

[0016] Digital twins work by digitally replicating a physical asset in a virtual environment, including its functions, features, and behaviors. A real-time digital representation of an asset is created using smart sensors that collect data from the product. This representation can be used throughout the asset's lifecycle, from initial product testing to actual operation and decommissioning. Digital twins use a variety of technologies to provide a digital model of an asset. These include the following: The Internet of Things (IoT) refers to a network of connected devices and the technologies that facilitate communication between them and the cloud, as well as between the devices themselves. Billions of devices are now connected to the internet, thanks to the advent of cheap computer chips and high-bandwidth telecommunications. Digital twins rely on IoT sensor data to transfer information from real-world objects to digital-world objects. The data is entered into a software platform or dashboard where you can see the data update in real time. Artificial intelligence (AI) is a field of computer science that focuses on solving cognitive problems commonly associated with artificial intelligence, such as learning, problem solving, and pattern recognition. Machine learning (ML) is an AI technique that develops statistical models and algorithms that enable computer systems to perform tasks without explicit instructions, relying on patterns and reasoning. Digital twin technology uses machine learning algorithms to process large amounts of sensor data and identify patterns in the data.

[0017] Object detection algorithm based on deep learning Deep learning is a method in machine learning based on learning to represent data. Observations can be represented in a variety of ways, such as a vector of intensity values ​​for each pixel, or more abstractly represented as a series of edges, regions of a specific shape, etc. Using certain specific representation methods makes it easier to learn tasks from examples. The benefit of deep learning is that it replaces manual feature acquisition with unsupervised or semi-supervised feature learning and hierarchical feature extraction efficient algorithms.

[0018] Deep learning transforms the feature representation of samples in the original space into a new feature space by performing layer-by-layer feature transformation on the original signal, and automatically learns to obtain a hierarchical feature representation, which is more conducive to classification or feature visualization. The key is to establish a model to simulate the neural connection structure of the human brain. When processing signals such as images, sounds, and texts, the data features are described in layers through multiple transformation stages, thereby giving an explanation of the data.

[0019] Object detection is an important branch in the field of computer vision. Its main purpose is to analyze and process a given video or image, point out the category of each object in it, and draw a bounding box near the object to mark the location of the object. Traditional object detection algorithms are based on manually extracted features, and the accuracy of the algorithm is low and the generalization performance is weak. In recent years, deep learning technology has developed rapidly. The object detection algorithm based on deep learning uses convolutional neural networks to extract features, and has made breakthrough progress in detection accuracy. It has been widely used in video surveillance, smart logistics, medical image analysis, unmanned driving and other fields. Object detection algorithms are mainly divided into traditional detection algorithms and deep learning-based detection algorithms. Traditional detection methods are based on hand-crafted features and shallow trainable architectures. This method combines a large number of low-level image features and high-level semantic information from object detectors and scene classifiers to build a complex system with low performance.

[0020] Traditional target detection methods have achieved certain results, but they have also exposed their inherent drawbacks. First, the use of sliding windows for region selection will lead to high time complexity and window redundancy. Second, the diversity of appearance, uncertainty of illumination changes, and diversity of backgrounds lead to poor robustness and generalization of manual feature design methods. The complicated algorithm steps lead to slow detection efficiency and low accuracy. Traditional detection methods can no longer meet people's demand for high performance in target detection.

[0021] In recent years, deep learning has developed rapidly, and some tools that can learn semantic, high-level, and deep-level features have been introduced to solve problems in traditional architectures, improving the performance of models in terms of network architecture, training strategies, and optimization functions. In 2012, Hinton et al. first applied the deep convolutional neural network AlexNe to large-scale image classification and won the championship in the classification task of the large-scale visual recognition challenge of the target detection dataset ImageNet. With the significant research progress of deep convolutional neural networks, DCNN has gradually become the focus of people's attention and a new research direction for target detection. The mainstream deep learning target detection algorithms are mainly divided into two-stage detection algorithms and single-stage detection algorithms.

[0022] Two-stage detection algorithm The two-stage target detection method mainly selects candidate regions (Region Proposal) that may contain the detection target from the input image through algorithms such as Selective Search or Edge Boxes, and then classifies and regresses the candidate regions to obtain the detection results. Typical algorithms include the R-CNN series, R-FCN, Mask R-CNN, etc.

[0023] Single-stage detection algorithm The single-stage target detection algorithm adopts the idea of ​​regression analysis, also known as the target detection algorithm based on regression analysis. The single-stage target detector omits the candidate region generation stage and directly obtains the target classification and location information. The YOLO series and SSD series are typical representatives.

[0024] Object Detection Algorithm Based on Generative Adversarial Network In 2014, Goodfellow et al. proposed Generative Adversarial Networks (GANs), which is an unsupervised generative model that works based on the maximum likelihood principle and uses adversarial training. The idea of ​​adversarial learning is to train the detection network by generating occluded and deformed image samples through adversarial networks, which is one of the commonly used generative model methods for generating data distribution. GAN is not only an image generation tool, it also retrieves useful information from training data to perform target detection, segmentation and classification tasks in different fields. In 2017, Wang et al. introduced the idea of ​​adversarial networks and proposed the A-Fast-RCNN algorithm that uses adversarial networks to generate difficult positive samples. A-Fast-RCNN combines adversarial learning with Fast R-CNN, and generates difficult samples (Hard Example) through GANs to increase the number of targets with occlusion and posture changes. Different from the traditional method of directly generating sample images, this method uses some transformations on the feature map: (1) In the Adversarial Spatial Dropout Network (ASDN), a Mask layer is added to achieve partial occlusion of features, and the Mask is selected according to the loss; (2) In the Adversarial Spatial Transformer Network (ASTN), the feature correspondence is manipulated to achieve partial deformation of features. STN consists of three parts: a localization network, a grid generator, and a sampler. The localization network estimates deformation parameters (such as rotation, translation distance, and scaling factor), and the grid generator and sampler use these parameters to generate deformed feature maps. ASDN and ASTN provide two different variations. By combining these two deformations (ASDN output is used as the input of ASTN), the detector can be trained to be more robust. ASTN was compared with random jittering. Using AlexNet, the mAPs were 58.1% and 57.3%, respectively. Using VGG16, the mAPs were 69.9% and 68.6%, respectively. ASTN performed better than random jittering. In comparison with the OHEM (Online Hard Example Mining) method, the introduction of adversarial networks into target detection is indeed a pioneering approach. However, in terms of improvement, it is not much better than OHEM, and some occluded samples may lead to misclassification.

[0025] The target detection mode is implemented using a target detection algorithm based on deep learning. After inputting the original image with defects, the defect location and defect classification in the original image are directly output. The defect positioning is automatically completed by the model. The model is easy to use and the defect positioning is more accurate.

[0026] With the development of information technology, a huge amount of complex information is constantly coming to people, and information is everywhere. However, the information that humans can receive is limited. Researchers have found that the human visual system has a huge visual information processing ability under a limited field of view. In the early stage of processing visual data, the human visual system will quickly focus on the important areas in the scene. This selective perception mechanism greatly reduces the amount of data processed by the human visual system, so that humans can suppress unimportant stimuli when processing complex visual information, and allocate limited neural computing resources to key parts of the scene, providing easier to process and more relevant information for higher-level perceptual reasoning and more complex visual processing tasks (such as object recognition, scene classification, video understanding, etc.). Drawing on this feature of the human visual system, researchers have proposed the idea of ​​the attention mechanism. The importance of features is different for things, which is reflected in the convolutional network, that is, the importance of each feature map is different. The core idea of ​​the attention mechanism is to obtain the difference in the importance of each feature map through certain means, invest more computing resources of the neural network in more important tasks, and use the task results to reversely guide the weight update of the feature map, so as to complete the corresponding tasks efficiently and quickly.

[0027] The attention mechanism can focus on important information with high weights and ignore irrelevant information with low weights, and can also continuously adjust the weights so that important information can be selected in different situations. Therefore, it has higher scalability and robustness.

[0028] In addition, it can also exchange information with others by sharing important information (i.e. selected important information), thereby achieving the transmission of important information. Therefore, the attention mechanism has attracted the attention of many researchers, and some new research algorithms based on the attention mechanism are constantly being proposed, verified and applied.

[0029] There are three main reasons why the attention mechanism can develop rapidly in deep learning: This structure is the most advanced model for solving multiple tasks such as machine translation, question answering, sentiment analysis, part-of-speech tagging, dialogue systems, data monitoring, fault diagnosis, etc.

[0030] The significant advantage of the attention mechanism is that it focuses on relevant information and ignores irrelevant information, directly establishes the dependency between input and output without going through a loop, enhances the degree of parallelization, and greatly improves the running speed.

[0031] It overcomes some limitations of traditional neural networks, such as the performance degradation of the system as the input length increases, the inefficiency of the system's computation due to unreasonable input order, and the lack of feature extraction and enhancement in the system. However, the attention mechanism can well model sequence data with variable length, further enhancing its ability to capture long-range dependency information, reducing the depth of the hierarchy while effectively improving the accuracy.

[0032] In the past two years, attention models have been widely used in various types of deep learning tasks such as natural language processing, image recognition, and speech recognition, such as SANet and VSG-Net, and are one of the core technologies in deep learning that deserves the most attention and in-depth understanding. The attention mechanism in deep learning is essentially similar to the selective visual attention mechanism of humans, and its core goal is also to select the more critical information for the current task goal from a large amount of information.

[0033] As a lightweight module, the attention mechanism has the characteristics of plug-and-play. However, even though the number of its own parameters is not high, in some deep learning tasks, the attention module is often called repeatedly. When the attention module is called too many times, it will still cause a certain burden on the network as a whole. How to optimize the module structure, reduce the number of module parameters or reduce the number of module calls, and obtain more accurate attention information faster, will play a vital role in the promotion of the attention mechanism in other tasks in the future, and is also one of the important contents that need to be studied in the future.

[0034] The attention mechanism was first used in 2014 as part of the encoder-decoder framework in RNN (Recurrent Neural Network) to encode long input sentences, and was subsequently widely used in RNN. For example, in machine translation, an RNN encoder is usually used to read the context and obtain a context vector (the last hidden state of the RNN). An RNN decoder uses this hidden state as the starting state and generates each word of the target in turn. However, the disadvantage of this approach is that no matter how long the previous context is and how much information it contains, it will eventually be compressed into a vector of several hundred dimensions. This means that the larger the context, the more information will be lost in the final state vector. As the length of the input sentence increases, the final decoder translation result will deteriorate significantly. In fact, because the context is known at input, a model can fully utilize all the context information during the decoding process, not just the information of the last state. This is the basic idea of ​​the attention mechanism.

[0035] However, in CNN (Convolutional Neural Networks), there is no state, and the attention mechanism acts on the feature map, trying to obtain the available attention information in the feature map to achieve better task results.

[0036] When the idea of ​​attention mechanism is applied to convolutional neural network, it is necessary to focus on the key information in the feature map that is unique to the convolutional network. The mainstream method of the current attention mechanism is to deeply mine the potential attention information in the feature map. The most common method is to obtain the channel attention information between the channels of each feature map and the spatial attention information between the pixels inside the feature map through various means. The acquisition method also includes but is not limited to convolution operation, matrix operation to build correlation matrix, etc. The common purpose is to obtain the complete attention information in the feature map at a deeper level and more comprehensively. The core of the attention mechanism in the convolutional network is to deeply mine the information contained in the feature map. The channels for obtaining attention that have been discovered are relatively few, but the attention mechanism has been widely proven to have not only a small number of parameters and plug-and-play convenience for a large number of deep learning tasks, but also a relatively significant improvement in task results. This shows that in-depth research on the attention mechanism in the future is necessary and of great significance.

[0037] Embodiment 1, a method for sensing parameters and handling faults in a calibration system for electric energy metering equipment, 1. Invention name: Multi-parameter sensing and fault handling method and system for automated calibration system, Information collection phase: Information from different units in the automated calibration system is collected through various sensors, including normal and abnormal information of the appearance unit, labeling unit, coding unit, loading and unloading robots, etc., as well as temperature and humidity information of the calibration unit, fault information of the strong and weak current terminals, and temperature and humidity information of the laboratory environment.

[0038] Inspection equipment is used to collect audio and video information of the calibration line, and important inspection points and inspection cycles are designed according to fault-prone points to collect fault information in real time.

[0039] Model building phase: Based on the collected fault pictures, a fault knowledge base is accumulated and fault labels are annotated. The digital twin technology is used to establish a panoramic intelligent perception model of the assembly line to achieve a 1:1 mapping from the physical object to the digital model.

[0040] Based on the basic Encoder-Decoder framework, the study uses a timing model based on the self-attention mechanism to divide the input signals into three categories: electrical parameter signals of the pipeline model, audio and video signals collected by the inspection device, and multimodal signals fed back by sensors.

[0041] Data processing stage: The principal component analysis algorithm is used to perform data noise reduction and dimensionality reduction on the collected electrical parameter signals, audio and video information, and sensor data, and to construct a basic structure of fault data that is easy to analyze.

[0042] An integrated learning algorithm that combines the One-Stage algorithm and the Two-Stage algorithm is used for target detection. Positive and negative samples are marked first to ensure that the ratio of positive and negative samples in the sample library is within the range of 10:1, and then sample training, recognition and data classification or normalization are performed.

[0043] Troubleshooting phase: The multi-modal information obtained on-site is input into the multi-parameter perception model to visualize the equipment operation status information. By comparing with the equipment status database data, the possible cause of the failure is predicted.

[0044] The time series event model with self-attention mechanism is introduced to learn the dependency between faults and pipeline operation status, and provide autonomous early warning and fault elimination guidance. Deep learning tools are used to predict the probability of potential failure of key equipment based on model status data.

[0045] A parameter perception and fault handling method for an electric energy metering instrument calibration system. First, according to the existing construction plan of the automated calibration system, various types of electrical parameter signals are accessed and analyzed. By analyzing the common faults of the calibration system, sensors are selected and deployed to achieve multi-parameter perception of the assembly line. Then, a typical fault sample feature library is constructed. By analyzing various machine vision algorithms, algorithms are selected and integrated, and an assembly line fault target detection algorithm is designed. Finally, a fault elimination knowledge base and a guidance model are established based on the self-attention mechanism, so as to implement the fault handling strategy and conduct a pilot application.

[0046] The specific process is as follows: Step 1, collect and analyze the electrical parameter signals of the automatic verification system, including the normal and abnormal information of the appearance unit, labeling unit, coding unit, loading and unloading robots, the temperature and humidity information of the verification unit, the fault information of the strong and weak current terminals, and the temperature and humidity information of the laboratory environment. Then, use the inspection equipment to collect the audio and video information of the verification line, design important inspection points and inspection cycles according to the prone fault points, collect the fault information of the verification line in real time, and classify and install sensors according to the line information that needs to be collected during the operation and maintenance of the assembly line, so as to form a sensor network for monitoring important information of the assembly line. Step 2: Based on the fault pictures collected on a regular basis, the fault knowledge base is accumulated and the fault marks are annotated to lay a foundation for modeling and analysis. The digital twin technology is used to establish a panoramic intelligent perception model of the assembly line based on the line electrical parameter signals, audio and video information of the inspection device, and sensor information, thus realizing a 1:1 mapping of the automated verification system from the physical object to the digital model. Step 3: A multi-parameter perception model based on digital twins is used to perform fusion analysis on the electrical parameter signals, audio and video information, sensor data, etc. collected by the model. At the same time, a fusion analysis of multi-modal information is performed based on the pipeline fault data. Usually, a fault phenomenon contains multiple fault influencing factors, and there may be a high correlation between the factors. By performing principal component analysis on the fault influencing factors, the relative collinearity between the factors is eliminated, and the high-dimensional variable space is reduced in dimensionality. Step 4: Data preprocessing. Currently, principal component analysis and BP neural network algorithms are commonly used. Principal component analysis is a commonly used unsupervised learning method. This method uses orthogonal transformation to convert observation data represented by linearly correlated variables into data represented by a few linearly independent variables. Linearly independent variables are called principal components. In this study, the electrical parameter signals, audio and video information, and sensor data collected from the fault information point are subjected to data denoising and dimensionality reduction using the principal component analysis algorithm to construct a basic structure of fault data that is easy to analyze, thereby facilitating the application of the target detection algorithm. Step 5, target detection algorithm. This topic plans to use an integrated learning algorithm that integrates the One-Stage algorithm and the Two-Stage algorithm in the deep learning algorithm. In the One-Stage algorithm stage, the CNN algorithm is used to extract and classify the fault data. First, we will mark the positive and negative samples of the accumulated sample data. To ensure the accuracy of the algorithm, we ensure that the ratio of positive and negative samples in the sample library is within the range of 10:1. Then we train and identify the samples, and finally classify or normalize the data. In the Two-Stage algorithm stage, we first use various CNN networks to extract the features of the data, and use the extracted data as feature samples, and then perform a step of rough classification and rough positioning, and finally realize the detection of faults. By analyzing the accuracy and speed of the two types of algorithms, combined with the needs of pipeline fault identification, the algorithms are combined to form a specific integrated algorithm suitable for pipeline fault target detection. Step 6: Research on pipeline fault handling strategies based on the self-attention mechanism time series model, implement model training and simulation monitoring of various fault data, and perform model adaptation and iterative optimization technology to improve the accuracy of model fault detection and elimination, and conduct pilot applications; The electrical parameter signals, audio and video information, and sensor data collected during the operation of the automated verification system are subjected to data denoising and preprocessing, and are used as multi-phase inputs into the attention model. Based on the basic Encoder-Decoder framework, the timing model is studied. In this study, the input signals are mainly divided into three categories: electrical parameter signals of the assembly line model, audio and video signals collected by the inspection device, and multi-modal signals fed back by the sensor. Step 7: Input the multimodal information obtained on site into the multi-parameter perception model to realize the visualization of the model on site equipment operation status information. By analyzing the feasible data and comparing it with the equipment status database data, the possible causes of faults such as positioning errors between targets are predicted. Then, the temporal event model of the self-attention mechanism is introduced. The dependency relationship between the fault and the operation status of the assembly line is learned through the self-attention mechanism. Without any prior conditions and external guidance, the dependency path is calculated based on the inclusive relationship between the fault occurrence and the time expression, and more accurate autonomous warning and elimination guidance for different faults are performed. Finally, make full use of deep learning tools to predict the probability of potential faults of key equipment based on model status data, and realize a computational model for fault handling.

Claims

1. A method for parameter perception and fault handling of an electric energy metering instrument calibration system, characterized in that: The following steps are involved: Step 1: Collect and analyze the electrical parameter signals of the automated verification system, use inspection equipment to collect audio and video information of the verification line, design important inspection points and inspection cycles based on fault-prone points, collect fault information of the verification line in real time, classify and install sensors, and form a sensor network; Step 2: Accumulate the fault knowledge base and annotate the fault tags, and use digital twin technology to establish a panoramic intelligent perception model of the assembly line; Step 3: Use the multi-parameter perception model based on the digital twin to conduct fusion analysis and perform principal component analysis on the factors affecting the fault; Step 4: Use principal component analysis algorithm to perform data noise reduction and dimension reduction; Step 5: Use the integrated learning algorithm that combines the One-Stage algorithm and the Two-Stage algorithm to perform target detection; Step 6: Process pipeline faults based on the self-attention mechanism timing model; Step 7: Input the multimodal information into the multi-parameter perception model, introduce the temporal event model of the self-attention mechanism for fault warning and fault elimination guidance, and predict the probability of potential failure of key equipment.

2. The method for parameter sensing and fault handling of the electric energy metering instrument calibration system according to claim 1 is characterized in that: The information collected in step 1 includes normal and abnormal information of the appearance unit, labeling unit, coding unit, loading and unloading robots, temperature and humidity information of the verification unit, fault information of the strong and weak current terminals, and temperature and humidity information of the laboratory environment.

3. The method for parameter sensing and fault handling of the electric energy metering instrument calibration system according to claim 1 is characterized in that: In step 2, digital twin technology is used to establish a panoramic intelligent perception model of the assembly line based on the line electrical parameter signals, audio and video information of the inspection device, and sensor information.

4. The method for parameter sensing and fault handling of an electric energy metering instrument calibration system according to claim 1 is characterized in that: In step 5, the One-Stage algorithm uses the CNN algorithm to extract features and classify fault data, and the Two-Stage algorithm uses various CNN networks to extract features, roughly classify and roughly locate data.

5. The method for parameter sensing and fault handling of the electric energy metering instrument calibration system according to claim 1 is characterized in that: In step 6, the timing model is studied based on the basic Encoder-Decoder framework, and the input signals are divided into three categories: electrical parameter signals of the pipeline model, audio and video signals collected by the inspection device, and multi-modal signals fed back by sensors.

6. A parameter sensing and fault handling system for electric energy metering equipment calibration system, characterized in that: include: Information collection module, used to collect electrical parameter signals, audio and video information and sensor data of the automated verification system; Model building module, used to build pipeline panoramic intelligent perception model and self-attention mechanism timing model; Data processing module, used to perform data noise reduction, dimensionality reduction and target detection on the collected information; The fault handling module is used to provide fault warning and fault elimination guidance based on the model analysis results, and predict the probability of potential failure of key equipment.

7. The parameter sensing and fault handling system for electric energy metering instrument calibration system according to claim 6 is characterized in that: The information collection module includes sensors, inspection equipment, etc., which are used to collect normal and abnormal information, temperature and humidity information, fault information, and audio and video information of different units.

8. The parameter sensing and fault handling system for electric energy metering instrument calibration system according to claim 6 is characterized in that: The model building module uses digital twin technology and self-attention mechanism to build the model.

9. The parameter sensing and fault handling system for electric energy metering instrument calibration system according to claim 6 is characterized in that: The data processing module uses principal component analysis algorithm and integrated learning algorithm to process data.

10. The parameter sensing and fault handling system for electric energy metering equipment calibration system according to claim 6, characterized in that: The fault handling module introduces a time series event model with a self-attention mechanism to provide fault warning and fault elimination guidance, and uses deep learning tools to predict the probability of potential failures of key equipment.

Images