A Method and System for Quality Inspection of Electricity Meter Assembly Based on Production Environment Control

Abstract

This invention provides a method and system for inspecting the assembly quality of electricity meters based on production environment control, belonging to the field of electricity meter inspection technology. The method includes: real-time acquisition of data during the assembly process of the electricity meter to obtain dynamic environmental parameters and assembly status data of key processes; environmental noise compensation to generate environmentally adaptive assembly features; construction of an assembly quality twin model based on a deep residual network, performing forward calculations to obtain the predicted assembly quality level of the electricity meter; root cause analysis when the level falls below a preset threshold to identify key environmental disturbance factors; generation of closed-loop control commands for local environmental regulation, and construction of an optimization strategy for electricity meter assembly quality. This invention solves the technical problem that existing technologies often neglect dynamic changes in the production environment, resulting in the failure to timely compensate or correct the impact of environmental factors on electricity meter assembly quality, thus leading to insufficient accuracy and stability of quality inspection results.

Description

Technical Field

[0001] This invention relates to the field of electricity meter testing technology, and specifically to an assembly quality testing method and system for electricity meters based on production environment control. Background Technology

[0002] As a crucial power metering device, electricity meters play a vital role in the power system. With the development of smart grids, the assembly quality of electricity meters directly affects their measurement accuracy and equipment lifespan. Therefore, the assembly quality inspection of electricity meters is particularly important, especially during the production process, where fluctuations in environmental factors such as temperature, humidity, and static electricity can have a significant impact on assembly quality. However, existing technologies often overlook these dynamic changes in the production environment, resulting in the failure to compensate or correct for the impact of environmental factors on assembly quality in a timely manner. This leads to insufficient accuracy and stability of the quality inspection results. For example, in traditional testing methods, if temperature changes are not compensated for, it may lead to inaccurate sensor data during the assembly process, thereby affecting the final assembly quality assessment. Summary of the Invention

[0003] This application provides a method and system for testing the assembly quality of electricity meters based on production environment control. It aims to solve the technical problem that existing technologies often ignore dynamic changes in the production environment, resulting in the failure to compensate or correct the impact of environmental factors on the assembly quality of electricity meters in a timely manner, which in turn leads to insufficient accuracy and stability of the quality test results.

[0004] The first aspect disclosed in this application provides a method for detecting the assembly quality of electricity meters based on production environment control. The method includes: real-time acquisition of dynamic environmental parameters and assembly status data of key processes in the electricity meter assembly process via a multi-source sensor network; environmental noise compensation of the assembly status data based on the dynamic environmental parameters to generate environmentally adaptive assembly features; construction of an assembly quality twin model based on a deep residual network, synchronizing the environmentally adaptive assembly features to the assembly quality twin model for forward calculation to obtain a predicted assembly quality level of the electricity meter; when the predicted assembly quality level is lower than a preset threshold, root cause analysis based on the temporal correlation between the dynamic environmental parameters and the assembly status data to determine key environmental disturbance factors; generation of closed-loop control commands based on the key environmental disturbance factors, and local environmental regulation of the electricity meter through the closed-loop control commands to construct an optimization strategy for electricity meter assembly quality.

[0005] The second aspect of this application discloses a power meter assembly quality inspection system based on production environment control. The system is used in the aforementioned power meter assembly quality inspection method based on production environment control. The system includes: a real-time acquisition module for real-time acquisition of the power meter assembly process via a multi-source sensor network to obtain dynamic environmental parameters and assembly status data of key power meter processes; an environmental noise compensation module for performing environmental noise compensation on the assembly status data based on the dynamic environmental parameters to generate environmentally adaptive assembly features; a forward calculation module for constructing an assembly quality twin model based on a deep residual network, synchronizing the environmentally adaptive assembly features to the assembly quality twin model for forward calculation, and obtaining a predicted assembly quality level of the power meter; a root cause analysis module for performing root cause analysis based on the temporal correlation between the dynamic environmental parameters and the assembly status data when the predicted assembly quality level is lower than a preset threshold, to determine key environmental disturbance factors; and a local environmental control module for generating closed-loop control commands based on the key environmental disturbance factors, performing local environmental control on the power meter through the closed-loop control commands, and constructing a power meter assembly quality optimization strategy.

[0006] One or more technical solutions provided in this application have at least the following beneficial effects:

[0007] By using a multi-source sensor network to collect dynamic environmental parameters and assembly status data in real time during the assembly process of electricity meters, accurate real-time data support is provided for subsequent data analysis and quality prediction. Analysis of dynamic environmental parameters allows for the identification and compensation of environmental noise, generating environmentally adaptive assembly features. These features reflect the adaptability of the assembly process to environmental changes, reducing the impact of environmental variations and improving data stability and accuracy. A twin model of assembly quality is constructed using a deep residual network, and the generated environmentally adaptive assembly features are input into this model for forward computation to obtain the predicted assembly quality level of the electricity meter. The deep residual network can handle complex nonlinear relationships and automatically learn... The study extracts the relationship between assembly quality and environmental factors, making the prediction process more accurate and providing stable quality assessments under different production conditions. When the predicted assembly quality level is lower than a preset threshold, root cause analysis is triggered. Based on the temporal correlation between dynamic environmental parameters and assembly status data, key environmental disturbance factors causing quality problems are identified, which helps reduce failure rates and improve production efficiency. Based on the identified key environmental disturbance factors, closed-loop control commands are generated, and these commands are used to regulate the local environment during the assembly process, thereby optimizing assembly quality. This dynamic optimization and adaptive regulation can effectively reduce the negative impact of environmental fluctuations on assembly quality, further improving the assembly quality of the energy meter.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 A schematic flowchart of an energy meter assembly quality inspection method based on production environment control provided in this application embodiment.

[0010] Figure 2 A schematic diagram of the structure of an energy meter assembly quality inspection system based on production environment control provided in this application embodiment.

[0011] Figure labeling: Real-time acquisition module 10, environmental noise compensation module 20, forward calculation module 30, root cause tracing module 40, local environmental control module 50. Detailed Implementation

[0012] This application provides a method and system for testing the assembly quality of electricity meters based on production environment control. It solves the technical problem that the prior art often ignores the dynamic changes in the production environment, resulting in the failure to compensate or correct the impact of environmental factors on the assembly quality of electricity meters in a timely manner, which in turn leads to insufficient accuracy and stability of the quality test results.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0014] Example 1, as Figure 1 As shown in the embodiment of this application, a method for inspecting the assembly quality of electricity meters based on production environment control is provided. The method includes: The assembly process of the electricity meter is collected in real time through a multi-source sensor network to obtain dynamic environmental parameters and assembly status data of key processes of the electricity meter.

[0015] Multi-source sensor networks include various types of sensors, such as temperature and humidity sensors, electric field strength sensors, image sensors, and torque sensors. These sensors collect real-time data on different environmental parameters and assembly status during the assembly of the electricity meter. Through real-time acquisition by the multi-source sensor network, two types of data are obtained: dynamic environmental parameters, such as ambient temperature, humidity, and electric field strength, reflecting changes in environmental conditions during assembly; and assembly status data for key processes in the electricity meter assembly, such as installation torque, image quality, and assembly status, providing specific data about the electricity meter assembly process.

[0016] Environmental noise compensation is performed on the assembly state data based on the dynamic environmental parameters to generate environmentally adaptive assembly features.

[0017] In actual production environments, environmental noise, such as temperature fluctuations, humidity changes, and electric field interference, can adversely affect data during the assembly process. To improve the accuracy and reliability of the data, it is necessary to compensate for this noise. First, time-series analysis is performed based on dynamic environmental parameters to identify the impact of environmental noise on the data. Then, signal processing techniques, such as low-pass filtering, noise suppression algorithms, or generative adversarial networks (GANs), are used to eliminate the interference of environmental noise on the assembly status data. After noise removal, environmentally adaptive assembly features are generated based on the compensated data. These features reflect the changes in the assembly process under different environmental conditions, ensuring that the prediction of assembly quality remains accurate even in different production environments.

[0018] An assembly quality twin model is constructed based on a deep residual network. The environmental adaptive assembly features are synchronized to the assembly quality twin model for forward calculation to obtain the predicted assembly quality level of the electricity meter.

[0019] Deep residual networks (DRNNs) are neural networks with skip connections that effectively mitigate the vanishing gradient problem during deep network training. Utilizing residual learning, they can better capture complex nonlinear relationships, thus modeling the factors influencing the assembly quality of electricity meters. A twin model is a technique for digitally representing physical systems to simulate behavior during assembly. In this model, a deep residual network serves as the core network architecture, associating the input environmentally adaptive assembly features with assembly quality. These environmentally adaptive assembly features are synchronized as input data to the assembly quality twin model. Based on this input data, the model performs forward computation to simulate the assembly quality of the electricity meter and determine whether it meets requirements. For example, the predicted results can be categorized into several levels, such as acceptable, poor, and unacceptable. If the predicted result is below a pre-set threshold, it indicates a problem with the electricity meter's assembly quality.

[0020] When the predicted assembly quality level is lower than a preset threshold, the root cause is traced based on the temporal correlation between the dynamic environmental parameters and the assembly status data to determine the key environmental disturbance factors.

[0021] If the predicted assembly quality level is lower than the preset quality standard threshold, such as a quality level of unqualified or poor, the root cause tracing process is triggered. Specifically, the temporal correlation between assembly status data and dynamic environmental parameters is analyzed to identify the correlation between environmental changes and changes in assembly quality. In this way, it is possible to pinpoint the adverse impact of environmental disturbances on assembly quality at a specific point in time or under specific environmental conditions. Through correlation analysis, key environmental factors affecting assembly quality are identified; for example, excessive temperature fluctuations can affect the assembly accuracy of electricity meters, and excessive humidity can damage electrical components.

[0022] Based on the key environmental disturbance factors, a closed-loop control command is generated, and the electricity meter is locally regulated through the closed-loop control command to construct an optimization strategy for the assembly quality of the electricity meter.

[0023] Based on the impact of key environmental disturbance factors, closed-loop control commands are generated. The purpose of these commands is to adjust the production environment to restore it to a state optimal for assembly quality. According to these commands, local factors in the production environment are controlled. For example, if temperature fluctuations are a key disturbance factor, the air conditioning temperature control system in the assembly workshop is adjusted to ensure that temperature fluctuations are within a suitable range; if humidity is too high, dehumidification equipment is activated. The closed-loop control commands adjust the production environment in real time through the automated control system to ensure effective optimization of the electricity meter assembly quality. Through continuous optimization of the control strategy and production environment adjustments, an electricity meter assembly quality optimization strategy is ultimately formed. This strategy can adjust the production environment in real time based on environmental changes and assembly quality monitoring results to ensure high-quality assembly of the electricity meters.

[0024] Furthermore, based on the dynamic environmental parameters, environmental noise compensation is performed on the assembly state data to generate environmentally adaptive assembly features. The method includes: Based on the dynamic environmental parameters, time-series analysis is performed to obtain time-series environmental parameter vectors. Fluctuation features of the production environment are extracted to construct an environmental fluctuation feature map. A generative adversarial network (GAN) is constructed, comprising a generator and a discriminator. High-resolution image data of the production environment is collected. The generator performs environmental noise removal on the high-resolution image data according to the environmental fluctuation feature map, generating a denoised image set. The discriminator performs real-world discrimination on the denoised image set, generating image discrimination results. Adversarial training is performed based on the denoised image set and the image discrimination results to obtain an image feature set. Real-time torque data is introduced, and the environmental fluctuation feature map is fused with the real-time torque data to construct an environment-torque mapping relationship for temperature influence analysis, determining the torque drift influence coefficient. The real-time torque data is nonlinearly corrected according to the torque drift influence coefficient to generate torque compensation features. The image feature set and the torque compensation features are multimodal feature concatenated to generate the environmental adaptive assembly features.

[0025] Time-series analysis of dynamic environmental parameters involves monitoring changes in these parameters within the production environment to obtain datasets that reflect environmental volatility and indicate their potential impact on the electricity meter assembly process. Features are extracted from the time-series data to reveal patterns of environmental fluctuations, including: temperature fluctuations (rate of change and amplitude), humidity fluctuations (rate of change and amplitude), and electric field strength fluctuations. These extracted environmental fluctuation features are then fused to construct an environmental fluctuation feature map. This map encompasses the fluctuation patterns of different environmental factors and provides detailed environmental background information for subsequent noise compensation, model training, and quality optimization.

[0026] Generative Adversarial Networks (GANs) are deep learning frameworks consisting of two neural networks: a generator and a discriminator. The generator generates images based on the input environment features, attempting to produce images as realistic as possible to deceive the discriminator and prevent it from distinguishing between real and generated images. The discriminator distinguishes between real and generated input images by learning from comparing the generator's output with real images, continuously improving its judgment ability.

[0027] High-resolution image data is acquired from a production environment. This data contains noise caused by environmental fluctuations, which affects image quality. The generator performs noise removal on this high-resolution image data based on features from an environmental fluctuation feature map. The generator learns how to reduce noise in images based on environmental features, generating clearer and more realistic images. The generator ultimately outputs a set of denoised images that have been corrected for environmental noise, resulting in clearer and more realistic images.

[0028] The discriminator determines whether the input set of denoised images is real or a fake image generated by the generator. After the judgment, the output image discrimination result is a binary classification, that is, real or fake, which indicates the authenticity of the corresponding image.

[0029] In Generative Adversarial Networks (GANs), the generator and discriminator compete to optimize their respective capabilities. Specifically, they continuously update their parameters through backpropagation. The generator learns from a set of denoised images and improves its generated images based on feedback from the discriminator (image discrimination results). The discriminator provides feedback based on the realism of each generated image, and the generator continuously improves the quality of its generated images based on this feedback. Through adversarial training, the generator will be able to generate high-quality, low-noise images, and the discriminator will be able to more accurately distinguish between real and generated images. The final result is a set of image features that can effectively describe the environmentally relevant features during the assembly of electricity meters, providing more accurate predictions and analysis of assembly quality.

[0030] In the assembly process of electricity meters, torque represents the rotational force applied to components. Real-time torque data is collected to clarify torque changes during assembly. Environmental fluctuation characteristic maps are fused with real-time torque data. Through data fusion, the impact of environmental changes on torque during assembly is identified. In other words, does the torque data change accordingly when temperature changes, and are there torque drifts caused by environmental factors? Through fusion analysis, a mapping relationship between the environment and torque is constructed. This mapping relationship describes the trend and pattern of torque changes under different environmental conditions. Based on this mapping relationship, the influence of temperature on torque is analyzed, particularly the degree of influence of temperature changes on torque data. Based on the analysis results, a torque drift influence coefficient is determined, i.e., the degree of torque data deviation caused by temperature changes. This coefficient will serve as the basis for subsequent torque data correction.

[0031] Since the influence of environmental factors on torque is not linear, a nonlinear correction method is adopted in the correction process. This nonlinear correction aims to accurately reflect the impact of the environment on torque data, especially the distortion effect of temperature changes. Based on the torque drift influence coefficient, the real-time torque data is adjusted to eliminate torque data drift caused by temperature fluctuations, generating a torque compensation feature. This feature reflects the adjustment result of the torque data, ensuring it more accurately reflects the actual assembly force during the electricity meter assembly process.

[0032] Image feature sets and torque compensation features are concatenated. Multimodal feature concatenation combines features from different types of data to more comprehensively describe the state during the energy meter assembly process. Image features and torque features reflect different assembly information, such as visual quality and assembly force, respectively. Combining them enhances the model's expressiveness and predictive ability for assembly quality. This ultimately constitutes an environment-adaptive assembly feature, which can adapt to different environmental conditions and provide more accurate input for subsequent assembly quality prediction and optimization.

[0033] Furthermore, based on the dynamic environmental parameters, time-series analysis is performed to obtain a time-series environmental parameter vector for extracting fluctuation characteristics of the production environment and constructing an environmental fluctuation feature map. The method includes: Based on the time-series environmental parameter vector, environmental temperature time-series analysis is performed to determine the environmental temperature time series. Temperature change is calculated based on the environmental temperature time series to construct temperature fluctuation characteristics, which include temperature change rate characteristics and temperature fluctuation amplitude characteristics. Based on the time-series environmental parameter vector, environmental humidity time-series analysis is performed to determine the environmental humidity time series. Humidity change is calculated based on the environmental humidity time series to construct humidity fluctuation characteristics, which include humidity change rate characteristics and humidity fluctuation amplitude characteristics. Based on the time-series environmental parameter vector, electrostatic electric field strength time-series analysis is performed to determine the electrostatic electric field strength time series. Intensity change is calculated based on the electrostatic electric field strength time series to construct electric field fluctuation characteristics, which include electric field strength change rate and electric field strength peak value. The temperature change rate characteristics, temperature fluctuation amplitude characteristics, humidity change rate characteristics, humidity fluctuation amplitude characteristics, electric field strength change rate, and electric field strength peak value are fused to construct the environmental fluctuation feature map.

[0034] A time-series environmental parameter vector refers to an environmental dataset that changes over time, including data on changes in ambient temperature. Temperature data at each time point forms a time series. By performing time-series analysis on this data, we can obtain the trend of temperature changes and clarify the fluctuation patterns at different time points, such as the trend of temperature increase or decrease, and the fluctuation period. We construct temperature fluctuation features: the temperature change rate feature represents the speed of temperature change, obtained by calculating the difference in temperature between adjacent time points and dividing by the time interval; this feature reflects the rate of temperature change. The temperature fluctuation amplitude feature represents the magnitude of temperature fluctuations, i.e., the maximum amount of temperature change over a period of time; this feature indicates the range and intensity of temperature changes.

[0035] Similar to temperature analysis, time-series analysis of humidity indicates the trend of humidity changes over time, whether there are fluctuations, and whether humidity changes periodically. Similar to the rate of change characteristic of temperature, the rate of change of humidity reflects the speed of humidity change, that is, the amount of change in humidity per unit time; the humidity fluctuation amplitude characteristic reflects the amplitude of humidity fluctuations, that is, the maximum amount of change in humidity within a certain period of time.

[0036] Electrostatic field strength time-series data is obtained by real-time monitoring of electric field strength changes through sensors. Time-series analysis clarifies the fluctuations in electric field strength over time. These fluctuations can affect the assembly process of electricity meters, particularly concerning the installation of electronic components and interactions between materials. By analyzing the time-series data, the changes in electric field strength can be calculated, such as the rate of change (the amount of change per unit time) and the amplitude of the change. Electric field fluctuation characteristics are constructed: the rate of change represents the speed at which the electric field strength changes; this is obtained by dividing the difference in electric field strength between adjacent time points by the time interval. The peak value of the electric field strength represents the maximum value of the electric field strength within a certain time range. These extreme values ​​can affect certain electronic components during assembly, especially when the electric field strength is too high, potentially causing electrostatic interference.

[0037] By integrating the above features, an environmental fluctuation feature map is obtained. This map not only provides the fluctuation of individual environmental factors such as temperature, humidity and electric field, but also integrates the comprehensive influence of these environmental factors, providing multi-dimensional data support for subsequent assembly quality optimization.

[0038] Furthermore, the high-resolution image data is subjected to environmental noise removal by a generator based on environmental fluctuation feature maps to generate a denoised image set. The method includes: A generator is constructed based on an encoder-decoder structure. The generator includes multiple convolutional layers, a feature embedding layer, a bottleneck layer, and multiple deconvolutional layers. The high-resolution image data is synchronized to the encoder and downsampled through the multiple convolutional layers to extract multi-scale image features. The environmental fluctuation feature map is fully mapped to the environmental fluctuation feature map through the feature embedding layer to obtain an environmental modulation vector. The environmental modulation vector is fused with the multi-scale image features through the bottleneck layer to obtain environmental adaptive image features. The environmental adaptive image features are synchronized to the decoder and upsampled through the multiple deconvolutional layers to generate the denoised image set.

[0039] The encoder-decoder architecture is a neural network architecture suitable for image generation tasks. Its basic construction includes: an encoder, which downsamples the high-dimensional input data through a series of convolutional layers to extract low-dimensional feature representations of the image; the encoder's goal is to capture the key features of the input data; and a decoder, which reconstructs the low-dimensional features into a high-dimensional output through deconvolutional layers, such as restoring high-resolution details of the image.

[0040] In the generator, convolutional layers are used to spatially downsample the input data, progressively extracting feature information from the image. Through convolution operations, the network can learn the local patterns and global structure of the image. The role of feature embedding layers is to map different types of input features to a shared feature space, which helps to combine environmental data with image data to generate environment-adaptive images. The bottleneck layer is a key layer in the network that compresses the input features into a smaller spatial representation and encodes information here. The output of the bottleneck layer contains the compressed features of the image, which are used to generate the final image. Deconvolutional layers are used to restore the spatial resolution of the image. They map low-dimensional features back to high-dimensional space through inverse convolution operations, generating a clear image.

[0041] High-resolution image data is input into the encoder part of the generator. The encoder's goal is to convert the image from high resolution to low resolution and extract key features in the process. The image is downsampled through multiple convolutional layers. Downsampling means progressively reducing the spatial size of the image, and each convolutional layer extracts different features from the image, such as edges, textures, and colors. With each convolutional layer, the image's detail and resolution decrease, but the network captures more abstract, high-level features. Through multiple convolutional operations, the network can extract multi-scale image features from details to the global picture. These multi-scale image features allow the generator to better capture the global structure and local details of the image, providing a foundation for subsequent image restoration and noise reduction.

[0042] The main function of the feature embedding layer is to transform different types of data into a unified representation. By mapping the environmental fluctuation feature map to a common feature space, this layer allows for a better understanding and handling of the impact of the environment on the assembly process, enabling subsequent neural networks to process and learn these features. In this process, the environmental fluctuation feature map is mapped into an environmental modulation vector, which contains key information about environmental changes and can influence the generation and processing of image data.

[0043] By processing and fusing information from environmental modulation vectors and multi-scale image features through a bottleneck layer, the bottleneck layer not only integrates these features but also performs information compression and efficient representation, ensuring the effective combination of environmental data and image features. Through the bottleneck layer, environmentally adaptive image features are formed. This feature set integrates environmental change information and image content, enabling the generated image to better adapt to actual environmental conditions. This allows the generated image to more accurately reflect the actual situation during the electricity meter assembly process.

[0044] The decoder takes environment-adaptive image features as input, gradually restoring the spatial resolution of the image and converting it into a high-quality image. The deconvolution layer, also known as the upsampling layer, is used to restore the image's spatial resolution. Through deconvolution operations, image features are progressively restored from low resolution to high resolution. The role of the deconvolution layer is to generate clearer, higher-resolution images while minimizing noise. Finally, the decoder outputs the images restored through deconvolution operations, generating a denoised image set. These images are modulated by the environment during generation, enabling them to adapt to the fluctuating characteristics of the current environment and reflect the actual quality status during the assembly of the electricity meter.

[0045] Furthermore, the method for constructing an assembly quality twin model based on deep residual networks includes: An initial deep residual network is constructed, and a historical assembly dataset is introduced to iteratively train the initial deep residual network to generate training results. The initial deep residual network is updated according to the training results to obtain a deep residual network. Backpropagation is performed based on the deep residual network to obtain an initial assembly quality twin model. The initial assembly quality twin model is validated and evaluated to generate evaluation indicators. When the evaluation indicators reach a preset indicator threshold, the assembly quality twin model is generated.

[0046] Deep residual networks are a deep learning architecture used to address the vanishing gradient and information decay problems in deep neural networks. By using residual connections—that is, directly skipping connections from the input to deeper layers of the network—deep residual networks can better train deep neural networks and typically achieve better performance. The initial deep residual network is constructed by using multiple residual modules, each of which can learn different features of the input data.

[0047] To enable the initial deep residual network to effectively predict the assembly quality of electricity meters, a set of historical assembly datasets is provided as training data. These datasets include various features from the historical assembly process, such as image data, torque data, and environmental data, along with corresponding quality labels. This data is used to iteratively train the initial deep residual network, enabling it to recognize and learn the relationship between assembly quality and various features. In each training iteration, the network parameters, such as weights and biases, are updated based on the backpropagation results. Through continuous training and optimization, the network gradually improves its ability to predict assembly quality. After training, the network structure and parameters are optimized, resulting in a final deep residual network used for assembly quality prediction.

[0048] Backpropagation is a neural network training method that calculates errors and propagates them back to the network to update its parameters. Specifically, it first calculates the error between the network's output and the actual assembly quality, and then uses this error information to adjust the network's parameters using gradient descent or other optimization algorithms. This process is repeated multiple times until the network can predict assembly quality as accurately as possible. During network training, the deep residual network learns the mapping relationship of assembly quality through backpropagation. After the network training is complete and backpropagation optimization is performed, the resulting model is the initial assembly quality twin model. A twin model is a digital simulation model used to simulate quality changes during the actual assembly process. This model can predict the assembly quality level when given assembly data.

[0049] The initial assembly quality twin model was validated to examine its performance on real-world data. Validation and evaluation included using a test dataset, distinct from the training dataset, to test the model's generalization ability. The model's accuracy was assessed by comparing its predictions with actual results. During evaluation, evaluation metrics were generated based on the differences between the model's output and the true labels. These metrics included accuracy, mean squared error, precision, and recall. If the model's evaluation metrics reached predetermined thresholds, it was considered sufficiently accurate for practical application. At this point, the assembly quality twin model could predict the assembly quality of the energy meter in real-time during subsequent assembly processes, helping to optimize the assembly process and improve product quality.

[0050] Furthermore, the method for synchronizing the environmental adaptive assembly features to the assembly quality twin model before forward computation includes: The feature dimension information of the environment-adaptive assembly features is extracted, including the feature vector length and the number of feature channels. The preset expected input dimension of the assembly quality twin model input layer is read, including the expected vector length and the expected number of channels. The feature dimension information is compared with the expected input dimension: when the feature vector length and the number of feature channels both match the expected number, a format verification pass signal is generated; when the feature vector length does not match the expected vector length, but the number of feature channels matches the expected number, a first feature alignment instruction is generated, and feature interpolation is performed on the feature vector length using the first feature alignment instruction to update the feature vector length until it matches the expected vector length; when the feature vector length matches the expected vector length, but the number of feature channels does not match the expected number, a second feature alignment instruction is generated, and feature dimensionality reduction is performed on the number of feature channels using the second feature alignment instruction to update the number of feature channels until it matches the expected number. When the feature vector length does not conform to the expected vector length, and the number of feature channels does not conform to the expected number of channels, a third feature alignment instruction is generated. The feature vector length is then interpolated using the third feature alignment instruction, and the feature vector length is updated until it conforms to the expected vector length. The feature channel number is then reduced using the third feature alignment instruction, and the number of feature channels is updated until it conforms to the expected number of channels.

[0051] The feature vector length represents the length of each feature vector, which is the number of elements in each input sample in the data. The feature vector length reflects the spatial dimension of the input data. The number of feature channels represents the number of different types of feature channels in the data. Each channel can be regarded as an independent feature set. For example, for image data, the number of channels is the number of color channels, such as RGB images having 3 channels.

[0052] The input layer of an assembly quality twin model has specific dimensional requirements; data can only be processed correctly when the dimensions of the input data match these requirements. The expected vector length is the expected length of the model's input feature vectors, determined by the model's architecture during design, reflecting the number of features per sample. The expected number of channels is the number of input channels the model expects, related to the data type or the number of input sources; for example, if the input data includes images and temperature data, there will be multiple channels.

[0053] The feature dimension information of the extracted environment adaptive assembly features is compared with the expected input dimension, including comparing the feature vector length and the number of feature channels.

[0054] When the feature dimension of the input data matches the expected input dimension, a format verification pass signal is generated, indicating that the input data is ready and can be sent to the twin model for prediction.

[0055] If the number of feature channels already meets the model requirements, but only the feature vector length is mismatched, then only the vector length needs to be adjusted. In this case, a first feature alignment instruction is generated, which instructs the adjustment of the feature vector length. An interpolation method, such as linear interpolation or other interpolation techniques, is called to adjust the length of the vector so that it matches the desired vector length. Through the interpolation method, elements are added or removed from the original feature vector, and the length of the feature vector is updated until it matches the desired vector length.

[0056] If the number of feature channels in the input data does not match the model's requirements, dimensionality reduction is needed to reduce the number of channels. Based on this mismatch, a second feature alignment instruction is generated, directing the adjustment of the number of channels through feature dimensionality reduction. Feature dimensionality reduction uses techniques such as principal component analysis and pooling layers to reduce the number of channels in the input data until it meets the desired number of channels. The purpose of dimensionality reduction is to retain the most important features while reducing redundant information, making the data dimensions fit the model input.

[0057] When the feature vector length and number of feature channels in the input data do not meet the model's expectations, both dimensions need to be adjusted simultaneously. This is where a third feature alignment instruction is generated. First, the feature vector length is interpolated to the desired length. Then, the number of channels is reduced to the desired number using feature dimensionality reduction methods. By performing interpolation and dimensionality reduction operations, and simultaneously updating the feature vector length and number of feature channels, it ensures that the input data meets the model's input dimensionality requirements.

[0058] Furthermore, the method of synchronizing the environmental adaptive assembly features to the assembly quality twin model for forward calculation to obtain the predicted assembly quality level of the energy meter includes: The environmentally adaptive assembly features are passed from the input layer of the assembly quality twin model to the first convolutional layer. The environmentally adaptive assembly features are then activated by convolution in the first convolutional layer to generate a shallow feature map. The shallow feature map is then passed to multiple cascaded residual modules for feature extraction to generate residual features. The environmentally adaptive assembly features and the residual features are fused through a skip connection structure to generate a high-level semantic feature map. Global dimensionality reduction is performed based on the high-level semantic feature map to generate a pooled feature vector, which is then linearly transformed to generate a mapped feature vector. Classification calculations are performed according to the mapped feature vector to set the predicted assembly quality level of the energy meter.

[0059] The environmental adaptive assembly features describe key factors in the electricity meter assembly process, such as environmental changes and assembly status. These features are passed as input data to the assembly quality twin model for further processing. These features are passed from the input layer to the first convolutional layer, a basic building block in convolutional neural networks used to extract local features. Here, the input features are filtered through convolution operations to extract low-level local features such as edges and textures. Through convolutional activation functions, such as ReLU, the convolutional layer can not only extract features but also provide nonlinear transformations for subsequent layers. The shallow feature map generated by this process contains the basic features of the input data and is low-level visual information.

[0060] Residual modules are a core component of deep convolutional neural networks. Through skip connections, they allow information to bypass some layers, thus solving the gradient vanishing problem in deep networks and ensuring effective training even for deeper layers. Each residual module contains several convolutional layers and activation functions, enabling the extraction of more complex features. Cascading means that multiple residual modules are linked together to form a deeper network structure. In this cascaded structure, each module further extracts and enhances information from the input features, gradually capturing more abstract and complex features. The features obtained after processing by multiple cascaded residual modules are called residual features. Residual features are more abstract than shallow features, capable of capturing deep-level patterns in the data. They contain more complex information and can effectively represent the relationship between assembly quality and environmental fluctuations.

[0061] Skip connections are a neural network technique that allows certain layers in a network to directly connect to deeper layers. This allows the network to maintain information transfer during training, avoiding the vanishing gradient problem and helping to capture multi-layered information. Here, the skip connection structure combines the original environment-adaptive assembly features with residual features processed by residual modules. This fusion allows the network to utilize environmental features and deep feature information from the assembly process, improving model performance. The generated high-level semantic feature map represents the high-level semantic information of the environment-adaptive assembly features after processing through multiple convolutional and residual modules. This high-level semantic feature map can capture deep-seated patterns and regularities in the assembly process, such as the relationship between assembly quality and environmental changes.

[0062] To reduce computational complexity and extract key features, global dimensionality reduction is performed on the high-level semantic feature map. Dimensionality reduction methods, such as global average pooling or global max pooling, are used to transform high-dimensional data into low-dimensional data. Pooling operations can effectively reduce feature dimensionality while retaining important information. The generated pooled feature vector, obtained through dimensionality reduction, contains key information about assembly quality and environmental impact, which is used for subsequent linear transformations and classification operations. After linear transformation, the pooled feature vector yields the final mapped feature vector, which represents the assembly quality of the energy meter and provides input for subsequent classification.

[0063] Based on the mapped feature vector, a classification algorithm, such as the Softmax function, is used to calculate the predicted probability of each assembly quality level, representing the assembly quality level of the electricity meter, such as qualified, good, or poor. Based on the classification calculation results, a predicted assembly quality level is set for the electricity meter. This predicted level is used to assess whether the assembly quality of the electricity meter meets the standard requirements. If the predicted value is lower than a preset threshold, it indicates that the assembly quality is substandard and further processing is required.

[0064] Furthermore, when the predicted assembly quality level is lower than a preset threshold, root cause analysis is performed based on the temporal correlation between the dynamic environmental parameters and the assembly status data to determine key environmental disturbance factors. The method includes: When the predicted assembly quality level is detected to be lower than a preset threshold, a root cause tracing command is automatically triggered. The unique identifier of the target energy meter is obtained through the root cause tracing command. Based on the unique identifier, the data storage record library is traversed to retrieve the time-series record dataset of the target energy meter. The dynamic environmental parameters are matched with the time-series record dataset to extract the time-varying sequences of ambient temperature, ambient humidity, and electrostatic field strength, constructing a dynamic environmental parameter time series. The assembly status data is matched with the time-series record dataset to extract the image quality change sequence and torque feature change sequence, constructing an assembly status data time series. A time-series correlation analysis is performed between the dynamic environmental parameter time series and the assembly status data time series to obtain the time-series correlation relationship for root cause tracing, calculating the assembly quality anomaly impact coefficient. The impact is then filtered based on the assembly quality anomaly impact coefficient to determine the key environmental disturbance factors.

[0065] If the predicted assembly quality level is lower than the preset threshold, such as below the qualified level, it is considered that there is a problem with the assembly quality of the electricity meter. At this time, the root cause tracing instruction is automatically triggered. Root cause tracing refers to tracing the cause of the quality problem and locating the key factors that caused the problem. Through tracing the root cause, it can be determined whether environmental changes, errors in the assembly process, or other factors caused the quality problem.

[0066] Each electricity meter has a unique identifier, such as a serial number. This identifier allows for the location of a specific meter and all recorded data related to its assembly process. The unique identifier is used to search the data repository for all records associated with the target meter, including environmental data, various sensor data from the assembly process, and image data. The extracted time-series dataset records environmental parameters and assembly status data at each point in time during the meter assembly process. This time-series data provides detailed information for each stage of the assembly process, aiding in the analysis and diagnosis of quality issues.

[0067] Based on the time-series dataset of the assembly process, environmentally relevant data, particularly dynamic environmental parameters such as temperature, humidity, and electric field strength, are extracted. These data reflect the fluctuations in environmental conditions during assembly. A time series of temperature changes is extracted to represent the temperature variations at each point in time during assembly. Similarly, a time series of humidity changes is extracted to reflect how humidity changes over time during assembly, and a time series of electrostatic field strength changes is extracted to reflect how fluctuations in electric field strength affect assembly quality. The time series of temperature, humidity, and electric field strength are combined to construct a complete time series of dynamic environmental parameters for subsequent analysis and model training.

[0068] The time-series dataset includes not only environmental data but also state data related to assembly quality. By analyzing image data collected during the assembly process, a sequence of image quality changes over time is extracted. These changes include factors such as focus and sharpness, which affect the visual inspection quality during assembly. Torque data reflects the forces applied during component assembly; extracting the time series of torque features indicates the mechanical state during assembly. Combining the image quality change sequence and the torque feature change sequence constructs a time series of assembly state data, indicating how the assembly state changes over time under different environmental conditions and how it interacts with environmental factors.

[0069] Time-series analysis is performed on the time series of dynamic environmental parameters and assembly status data to identify the correlation between them. This analysis reveals the temporal relationship between changes in environmental factors and assembly status, such as whether temperature fluctuations are related to changes in torque characteristics. Through this analysis, root cause analysis is conducted to trace the environmental disturbances that lead to assembly quality problems. For example, if fluctuations in certain environmental parameters coincide with the occurrence of assembly quality problems, it indicates that these environmental factors are the root cause of the quality issues. Root cause analysis calculates the assembly quality anomaly impact coefficient, quantifying the degree of influence of environmental factors on assembly quality. This coefficient reflects the actual impact of environmental parameter fluctuations on assembly quality; a high impact coefficient indicates a significant influence of environmental factors on quality, while a low coefficient indicates a relatively small effect.

[0070] Based on the impact coefficient of assembly quality anomalies, key environmental disturbance factors were identified, namely, the environmental factors that have the greatest impact on assembly quality anomalies. These factors include excessive temperature fluctuations, unstable humidity changes, and excessively high electrostatic field strength. These key factors can serve as key targets for optimizing the assembly process and adjusting the production environment, thereby helping to improve assembly quality.

[0071] Furthermore, a time-series correlation analysis is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to obtain the time-series correlation relationship for root cause tracing and to calculate the impact coefficient of assembly quality anomalies. The method includes: A causal test is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to calculate the characteristic causal strength value; a correlation analysis is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to calculate the sequence correlation coefficient; the characteristic causal strength value and the sequence correlation coefficient are weighted and fused to generate a comprehensive influence weight, which is then accumulated to determine the impact coefficient of the assembly quality anomaly.

[0072] Causality tests aim to determine whether one variable directly affects another. For environmental and assembly state data during the assembly process, causality tests reveal which environmental factors directly cause changes in the assembly state. For example, does temperature fluctuation directly affect torque characteristics, or does a change in electric field strength lead to changes in image quality? Specifically, causal relationships are tested by traversing the time series of dynamic environmental parameters and the time series of assembly state data. For example, statistical methods such as the Granger causality test are used to determine whether one variable can predict changes in another. Based on the results of the causality tests, a characteristic causality strength value is calculated. This value represents the degree of influence of environmental factors on the assembly state. A larger causality strength value indicates that the environmental factor has a significant impact on assembly quality, while a smaller value indicates a smaller impact.

[0073] Correlation analysis is used to measure the strength and direction of the relationship between two variables. Here, we analyze the correlation between environmental factors and assembly status. By analyzing these time series data, we can clarify the extent to which environmental changes affect the assembly status. The correlation coefficient is measured using the Pearson correlation coefficient, ranging from -1 to 1. A value of 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no linear correlation. By calculating the correlation coefficient, we can clarify the linear relationship between environmental parameters and assembly status. The higher the correlation coefficient, the stronger the influence of environmental factors on the assembly status.

[0074] A weighted fusion method is applied to the characteristic causal strength value and the sequence correlation coefficient. The weighting method assigns different weights based on the importance of the causal strength value and the correlation coefficient; for example, if the causal strength value is more important than the correlation coefficient, it is assigned a larger weight, and vice versa. By weighted fusion of the causal strength value and the correlation coefficient, a comprehensive impact weight is obtained. This weight represents the overall impact of environmental factors on assembly quality. This weight integrates causal and correlation relationships, comprehensively reflecting the role of environmental factors in assembly quality. Based on the comprehensive impact weight, an assembly quality anomaly impact coefficient is determined. This coefficient quantifies the degree of influence of environmental factors on assembly quality anomalies. A large impact coefficient indicates that the environmental factor significantly contributes to the assembly quality anomaly and needs to be prioritized for control and optimization.

[0075] Example 2 is based on the same inventive concept as the electricity meter assembly quality inspection method based on production environment control in the previous examples, such as... Figure 2 As shown in the figure, this application provides an assembly quality inspection system for electricity meters based on production environment control. The system includes: The real-time acquisition module 10 is used to collect data on the assembly process of the electricity meter in real time through a multi-source sensor network, and obtain dynamic environmental parameters and assembly status data of key processes of the electricity meter. The environmental noise compensation module 20 is used to compensate for environmental noise based on the dynamic environmental parameters and the assembly status data, and generate environmentally adaptive assembly features. The forward calculation module 30 is used to construct an assembly quality twin model based on a deep residual network, and synchronize the environmentally adaptive assembly features to the assembly quality twin model for forward calculation to obtain the predicted assembly quality level of the electricity meter. The root cause tracing module 40 is used to perform root cause tracing based on the temporal correlation between the dynamic environmental parameters and the assembly status data when the predicted assembly quality level is lower than a preset threshold, and determine the key environmental disturbance factors. The local environmental control module 50 is used to generate closed-loop control commands based on the key environmental disturbance factors, and perform local environmental control of the electricity meter through the closed-loop control commands to construct an electricity meter assembly quality optimization strategy.

[0076] Furthermore, the environmental noise compensation module 20 is used to perform the following operation steps: Based on the dynamic environmental parameters, time-series analysis is performed to obtain time-series environmental parameter vectors. Fluctuation features of the production environment are extracted to construct an environmental fluctuation feature map. A generative adversarial network (GAN) is constructed, comprising a generator and a discriminator. High-resolution image data of the production environment is collected. The generator performs environmental noise removal on the high-resolution image data according to the environmental fluctuation feature map, generating a denoised image set. The discriminator performs real-world discrimination on the denoised image set, generating image discrimination results. Adversarial training is performed based on the denoised image set and the image discrimination results to obtain an image feature set. Real-time torque data is introduced, and the environmental fluctuation feature map is fused with the real-time torque data to construct an environment-torque mapping relationship for temperature influence analysis, determining the torque drift influence coefficient. The real-time torque data is nonlinearly corrected according to the torque drift influence coefficient to generate torque compensation features. The image feature set and the torque compensation features are multimodal feature concatenated to generate the environmental adaptive assembly features.

[0077] Furthermore, the environmental noise compensation module 20 is used to perform the following operation steps: Based on the time-series environmental parameter vector, environmental temperature time-series analysis is performed to determine the environmental temperature time series. Temperature change is calculated based on the environmental temperature time series to construct temperature fluctuation characteristics, which include temperature change rate characteristics and temperature fluctuation amplitude characteristics. Based on the time-series environmental parameter vector, environmental humidity time-series analysis is performed to determine the environmental humidity time series. Humidity change is calculated based on the environmental humidity time series to construct humidity fluctuation characteristics, which include humidity change rate characteristics and humidity fluctuation amplitude characteristics. Based on the time-series environmental parameter vector, electrostatic electric field strength time-series analysis is performed to determine the electrostatic electric field strength time series. Intensity change is calculated based on the electrostatic electric field strength time series to construct electric field fluctuation characteristics, which include electric field strength change rate and electric field strength peak value. The temperature change rate characteristics, temperature fluctuation amplitude characteristics, humidity change rate characteristics, humidity fluctuation amplitude characteristics, electric field strength change rate, and electric field strength peak value are fused to construct the environmental fluctuation feature map.

[0078] Furthermore, the environmental noise compensation module 20 is used to perform the following operation steps: A generator is constructed based on an encoder-decoder structure. The generator includes multiple convolutional layers, a feature embedding layer, a bottleneck layer, and multiple deconvolutional layers. The high-resolution image data is synchronized to the encoder and downsampled through the multiple convolutional layers to extract multi-scale image features. The environmental fluctuation feature map is fully mapped to the environmental fluctuation feature map through the feature embedding layer to obtain an environmental modulation vector. The environmental modulation vector is fused with the multi-scale image features through the bottleneck layer to obtain environmental adaptive image features. The environmental adaptive image features are synchronized to the decoder and upsampled through the multiple deconvolutional layers to generate the denoised image set.

[0079] Furthermore, the forward computation module 30 is used to perform the following operation steps: An initial deep residual network is constructed, and a historical assembly dataset is introduced to iteratively train the initial deep residual network to generate training results. The initial deep residual network is updated according to the training results to obtain a deep residual network. Backpropagation is performed based on the deep residual network to obtain an initial assembly quality twin model. The initial assembly quality twin model is validated and evaluated to generate evaluation indicators. When the evaluation indicators reach a preset indicator threshold, the assembly quality twin model is generated.

[0080] Furthermore, the forward computation module 30 is used to perform the following operation steps: Extract the feature dimension information of the environment-adaptive assembly features, the feature dimension information including feature vector length and feature channel number; read the preset expected input dimension of the assembly quality twin model input layer, the preset expected input dimension including expected vector length and expected channel number; compare the feature dimension information with the expected input dimension: when the feature vector length meets the expected vector length and the feature channel number meets the expected channel number, generate a format verification pass signal; when the feature vector length does not meet the expected vector length and the feature channel number meets the expected channel number, generate a first feature alignment instruction, and perform feature interpolation on the feature vector length using the first feature alignment instruction to update the feature vector length until it meets the requirements. The expected vector length is specified. When the feature vector length meets the expected vector length but the number of feature channels does not meet the expected number of feature channels, a second feature alignment instruction is generated. The second feature alignment instruction is used to perform feature dimensionality reduction on the number of feature channels, and the number of feature channels is updated until it meets the expected number of channel channels. When the feature vector length does not meet the expected vector length and the number of feature channels does not meet the expected number of channel channels, a third feature alignment instruction is generated. The third feature alignment instruction is used to perform feature interpolation on the feature vector length, and the feature vector length is updated until it meets the expected vector length. The third feature alignment instruction is used to perform feature dimensionality reduction on the number of feature channels, and the number of feature channels is updated until it meets the expected number of channel channels.

[0081] Furthermore, the forward computation module 30 is used to perform the following operation steps: The environmentally adaptive assembly features are passed from the input layer of the assembly quality twin model to the first convolutional layer. The environmentally adaptive assembly features are then activated by convolution in the first convolutional layer to generate a shallow feature map. The shallow feature map is then passed to multiple cascaded residual modules for feature extraction to generate residual features. The environmentally adaptive assembly features and the residual features are fused through a skip connection structure to generate a high-level semantic feature map. Global dimensionality reduction is performed based on the high-level semantic feature map to generate a pooled feature vector, which is then linearly transformed to generate a mapped feature vector. Classification calculations are performed according to the mapped feature vector to set the predicted assembly quality level of the energy meter.

[0082] Furthermore, the root cause tracing module 40 is used to perform the following operation steps: When the predicted assembly quality level is detected to be lower than a preset threshold, a root cause tracing command is automatically triggered. The unique identifier of the target energy meter is obtained through the root cause tracing command. Based on the unique identifier, the data storage record library is traversed to retrieve the time-series record dataset of the target energy meter. The dynamic environmental parameters are matched with the time-series record dataset to extract the time-varying sequences of ambient temperature, ambient humidity, and electrostatic field strength, constructing a dynamic environmental parameter time series. The assembly status data is matched with the time-series record dataset to extract the image quality change sequence and torque feature change sequence, constructing an assembly status data time series. A time-series correlation analysis is performed between the dynamic environmental parameter time series and the assembly status data time series to obtain the time-series correlation relationship for root cause tracing, calculating the assembly quality anomaly impact coefficient. The impact is then filtered based on the assembly quality anomaly impact coefficient to determine the key environmental disturbance factors.

[0083] Furthermore, the root cause tracing module 40 is used to perform the following operation steps: A causal test is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to calculate the characteristic causal strength value; a correlation analysis is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to calculate the sequence correlation coefficient; the characteristic causal strength value and the sequence correlation coefficient are weighted and fused to generate a comprehensive influence weight, which is then accumulated to determine the impact coefficient of the assembly quality anomaly.

[0084] Through the foregoing detailed description of the method for detecting the assembly quality of electricity meters based on production environment control, those skilled in the art can clearly understand the electricity meter assembly quality detection system based on production environment control in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to in the method section.

[0085] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for inspecting the assembly quality of electricity meters based on production environment control, characterized in that, The method includes: The assembly process of the electricity meter is collected in real time through a multi-source sensor network to obtain dynamic environmental parameters and assembly status data of key processes of the electricity meter. Based on the dynamic environmental parameters, environmental noise compensation is performed on the assembly state data to generate environmentally adaptive assembly features. An assembly quality twin model is constructed based on a deep residual network. The environmental adaptive assembly features are synchronized to the assembly quality twin model for forward calculation to obtain the predicted assembly quality level of the electricity meter. When the predicted assembly quality level is lower than the preset threshold, the root cause is traced according to the temporal correlation between the dynamic environmental parameters and the assembly status data to determine the key environmental disturbance factors. Based on the key environmental disturbance factors, a closed-loop control command is generated, and the electricity meter is locally regulated through the closed-loop control command to construct an optimization strategy for the assembly quality of the electricity meter.

2. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 1, characterized in that, Based on the dynamic environmental parameters, environmental noise compensation is performed on the assembly state data to generate environmentally adaptive assembly features. The method includes: Based on the dynamic environmental parameters, time series analysis is performed to obtain a time series environmental parameter vector, and the fluctuation characteristics of the production environment are extracted to construct an environmental fluctuation feature map. Construct a generative adversarial network, which includes a generator and a discriminator; High-resolution image data of the production environment is collected, and environmental noise is eliminated from the high-resolution image data by a generator based on the environmental fluctuation feature map to generate a noise-reduced image set. The denoised image set is realistically judged by a discriminator to generate image discrimination results; Adversarial training is performed based on the denoised image set and the image discrimination results to obtain an image feature set; Real-time torque data is introduced, and the environmental fluctuation characteristic spectrum is fused with the real-time torque data to construct an environment-torque mapping relationship for temperature influence analysis and to determine the torque drift influence coefficient. The real-time torque data is nonlinearly corrected according to the torque drift influence coefficient to generate torque compensation features; The image feature set and the torque compensation feature are concatenated using multimodal features to generate the environment-adaptive assembly feature.

3. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 2, characterized in that, Based on the dynamic environmental parameters, time-series analysis is performed to obtain a time-series environmental parameter vector. Fluctuation characteristics of the production environment are then extracted, and an environmental fluctuation feature map is constructed. The method includes: Based on the time-series environmental parameter vector, environmental temperature time-series analysis is performed to determine the environmental temperature time series. Temperature change is calculated based on the environmental temperature time series to construct temperature fluctuation characteristics, which include temperature change rate characteristics and temperature fluctuation amplitude characteristics. Based on the time-series environmental parameter vector, environmental humidity time-series analysis is performed to determine the environmental humidity time series. Based on the environmental humidity time series, humidity change is calculated to construct humidity fluctuation characteristics, which include humidity change rate characteristics and humidity fluctuation amplitude characteristics. Based on the time-series environmental parameter vector, electrostatic electric field strength time-series analysis is performed to determine the electrostatic electric field strength time series. According to the electrostatic electric field strength time series, the intensity change is calculated to construct electric field fluctuation characteristics, which include the electric field strength change rate and the electric field strength peak value. The environmental fluctuation feature map is constructed by fusing the temperature change rate feature, the temperature fluctuation amplitude feature, the humidity change rate feature, the humidity fluctuation amplitude feature, the electric field intensity change rate, and the electric field intensity peak value.

4. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 2, characterized in that, The method involves using a generator to remove environmental noise from the high-resolution image data based on environmental fluctuation feature maps to generate a denoised image set. A generator is constructed based on an encoder-decoder structure, which includes multiple convolutional layers, feature embedding layers, bottleneck layers, and multiple deconvolutional layers. The high-resolution image data is synchronized to the encoder and downsampled through the multiple convolutional layers to extract multi-scale image features. The environmental fluctuation feature map is fully connected and mapped to the environmental modulation vector by the feature embedding layer. The environment modulation vector is fused with multi-scale image features through the bottleneck layer to obtain environment-adaptive image features; The environment-adaptive image features are synchronized to the decoder and upsampled through the multiple deconvolution layers to generate the denoised image set.

5. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 1, characterized in that, The method for constructing an assembly quality twin model based on deep residual networks includes: An initial deep residual network is constructed, and a historical assembly dataset is introduced to iteratively train the initial deep residual network to generate training results. The initial deep residual network is then updated based on the training results to obtain a deep residual network. Backpropagation is performed based on the deep residual network to obtain an initial assembly quality twin model. The initial assembly quality twin model is validated and evaluated to generate evaluation indicators. When the evaluation indicators reach a preset threshold, the assembly quality twin model is generated.

6. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 1, characterized in that, The method for synchronizing the environmental adaptive assembly features to the assembly quality twin model before forward computation includes: Extract the feature dimension information of the environment adaptive assembly feature, wherein the feature dimension information includes the feature vector length and the number of feature channels; Read the preset expected input dimension of the assembly quality twin model input layer, wherein the preset expected input dimension includes the expected vector length and the expected number of channels; The feature dimension information is compared with the expected input dimension: When the feature vector length meets the expected vector length and the number of feature channels meets the expected number of channels, a format verification pass signal is generated. When the feature vector length does not meet the expected vector length, but the number of feature channels meets the expected number of channels, a first feature alignment instruction is generated. The feature vector length is then interpolated using the first feature alignment instruction, and the feature vector length is updated until it meets the expected vector length. When the feature vector length meets the expected vector length, but the number of feature channels does not meet the expected number of channels, a second feature alignment instruction is generated. The second feature alignment instruction is used to perform feature dimensionality reduction on the number of feature channels, and the number of feature channels is updated until it meets the expected number of channels. When the feature vector length does not conform to the expected vector length, and the number of feature channels does not conform to the expected number of channels, a third feature alignment instruction is generated. The feature vector length is then interpolated using the third feature alignment instruction, and the feature vector length is updated until it conforms to the expected vector length. The feature channel number is then reduced using the third feature alignment instruction, and the number of feature channels is updated until it conforms to the expected number of channels.

7. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 1, characterized in that, The method involves synchronizing the environmental adaptive assembly features to the assembly quality twin model for forward calculation to obtain the predicted assembly quality level of the energy meter. The environment-adaptive assembly features are passed from the input layer of the assembly quality twin model to the first convolutional layer, and the environment-adaptive assembly features are convolved and activated by the first convolutional layer to generate a shallow feature map. The shallow feature map is passed to multiple cascaded residual modules for feature extraction to generate residual features; By fusing the environment-adaptive assembly features with the residual features through a skip connection structure, a high-level semantic feature map is generated. Global dimensionality reduction is performed based on the high-level semantic feature map to generate pooled feature vectors, which are then linearly transformed to generate mapped feature vectors. The predicted assembly quality level of the energy meter is set by performing classification calculations based on the mapped feature vector.

8. The method for inspecting the assembly quality of electricity meters based on production environment control as described in claim 1, characterized in that, When the predicted assembly quality level is lower than a preset threshold, root cause analysis is performed based on the temporal correlation between the dynamic environmental parameters and the assembly status data to determine key environmental disturbance factors. The method includes: When the predicted assembly quality level is detected to be lower than the preset threshold, a root cause tracing command is automatically triggered. The unique identifier of the target energy meter is obtained through the root cause tracing instruction. The data storage record database is traversed according to the unique identifier to retrieve the time-series record dataset of the target energy meter. Match the dynamic environmental parameters with the time-series record dataset, extract the time-series changes in ambient temperature, ambient humidity, and electrostatic field strength, and construct a time series of dynamic environmental parameters. Match the assembly status data with the time-series record dataset, extract the image quality change sequence and torque feature change sequence, and construct the assembly status data time series. A time-series correlation analysis is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to obtain the time-series correlation relationship for root cause tracing and to calculate the impact coefficient of assembly quality anomalies. The key environmental disturbance factors are determined by screening the impact of the assembly quality anomaly coefficient.

9. The method for inspecting the assembly quality of an energy meter based on production environment control as described in claim 8, characterized in that, A time-series correlation analysis is performed on the time series of the dynamic environmental parameters and the time series of the assembly status data to obtain the time-series correlation relationship for root cause tracing and to calculate the impact coefficient of assembly quality anomalies. The method includes: Perform causality tests on the time series of the dynamic environment parameters and the time series of the assembly status data, and calculate the characteristic causality strength value. Correlation analysis was performed on the time series of the dynamic environmental parameters and the time series of the assembly status data, and the correlation coefficient of the sequences was calculated. The feature causal strength value and the sequence correlation coefficient are weighted and fused to generate a comprehensive influence weight, which is then accumulated to determine the assembly quality anomaly influence coefficient.

10. An energy meter assembly quality inspection system based on production environment control, characterized in that, For implementing the method for inspecting the assembly quality of electricity meters based on production environment control as described in any one of claims 1-9, the system comprises: The real-time acquisition module is used to collect data on the assembly process of the electricity meter in real time through a multi-source sensor network, and obtain dynamic environmental parameters and assembly status data of key processes of the electricity meter. An environmental noise compensation module is used to perform environmental noise compensation on the assembly state data based on the dynamic environmental parameters, and generate environmental adaptive assembly features. The forward computation module is used to construct an assembly quality twin model based on a deep residual network, synchronize the environmental adaptive assembly features to the assembly quality twin model for forward computation, and obtain the predicted assembly quality level of the energy meter. The root cause tracing module is used to perform root cause tracing based on the temporal correlation between the dynamic environmental parameters and the assembly status data when the predicted assembly quality level is lower than a preset threshold, and to determine the key environmental disturbance factors. The local environment control module is used to generate closed-loop control commands based on the key environmental disturbance factors, and to perform local environment control on the electricity meter through the closed-loop control commands to construct an optimization strategy for the assembly quality of the electricity meter.

Images