Performance trend prediction method, device, equipment and system of metering equipment

Abstract

The application provides a performance trend prediction method, device, equipment and system of a metrological device. The performance trend prediction method comprises the following steps: obtaining a metrological device image; performing image recognition processing on the metrological device image to obtain nameplate text information of the metrological device; dynamically adjusting the fusion weight ratio of a full connection layer in a detection model according to the nameplate text information to obtain a target detection model; and performing defect detection on the metrological device image based on the target detection model to obtain a performance trend prediction result. The application solves the problems of low efficiency of manual inspection and poor adaptability of a general visual model to different types of metrological devices, improves the pertinence and stability of defect recognition of different types of metrological devices, improves the reliability of performance trend prediction, and provides support for equipment early warning and operation and maintenance decision-making.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance and image recognition technology for power equipment, and in particular to a method, device, equipment and system for predicting the performance trend of metering equipment. Background Technology

[0002] In power systems, low-voltage current transformers are core devices for energy metering, and their performance directly affects the stability of the power grid and the accuracy of metering data. Therefore, there is an urgent need for a performance trend prediction method that can adapt to differences in equipment models and integrate time-series characteristics to improve the intelligence level and predictive reliability of power equipment operation and maintenance.

[0003] In existing technologies, deep learning-based image recognition technology has been introduced into the field of equipment anomaly detection. For example, convolutional neural networks and other target detection algorithms are used to locate and classify defects in equipment images, thereby providing technical support for automated operation and maintenance.

[0004] However, existing general vision models are not adaptable to different types of equipment, and their use of a fixed-parameter general architecture reduces the accuracy of predicting the performance trends of metering equipment. Summary of the Invention

[0005] This application provides a method, apparatus, device, and system for predicting the performance trend of metering equipment, in order to improve the adaptability, stability, and reliability of performance trend judgment in metering equipment anomaly identification.

[0006] In a first aspect, this application provides a method for predicting the performance trend of a metering device, including:

[0007] Acquire images of the metering equipment;

[0008] The image of the measuring device is processed by image recognition to obtain the nameplate text information of the measuring device;

[0009] Based on the nameplate text information, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted to obtain the target detection model;

[0010] Based on the target detection model, defect detection is performed on the image of the measuring equipment to obtain the performance trend prediction result of the measuring equipment.

[0011] In one possible implementation, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted based on the nameplate text information to obtain the target detection model, including:

[0012] Based on the nameplate text information, determine the data sufficiency coefficient of the metering device;

[0013] If the data sufficiency coefficient meets the preset coefficient threshold, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted according to the first network layer of the detection model to obtain the target detection model.

[0014] If the data sufficiency coefficient does not meet the preset coefficient threshold, the fusion weight ratio is dynamically adjusted according to the second network layer of the detection model to obtain the target detection model.

[0015] In one possible implementation, after determining the data sufficiency coefficient of the metering device based on the nameplate text information, the method further includes:

[0016] Acquire multiple sets of images and multiple sets of abnormal appearance features for multiple device types of abnormal metering devices;

[0017] The multiple sets of abnormal metering device images and the multiple sets of abnormal appearance features are mapped and associated to obtain multiple mapping and association training sample sets;

[0018] Discriminative training is performed based on the multiple mapping-related training sample sets to obtain the first and second network layers of the detection model.

[0019] In one possible implementation, image recognition processing is performed on the image of the measuring device to obtain the nameplate text information of the measuring device, including:

[0020] Based on edge detection and contour analysis, the nameplate area image of the metering equipment is located;

[0021] The nameplate area image is subjected to text recognition processing to obtain the initial recognition text of the metering device;

[0022] Based on a preset rule library of keywords for power equipment nameplates, the initial identification text is parsed to obtain the nameplate text information of the metering equipment.

[0023] In one possible implementation, locating the nameplate region image of the metering device image based on edge detection and contour analysis includes:

[0024] Edge detection is performed on the image of the measuring device to obtain the edge image of the measuring device;

[0025] Contour analysis is performed on the edge image to obtain a set of closed contours;

[0026] Based on the preset rules of quadrilateral geometric features and texture features, the closed contour set is filtered to obtain the nameplate area image of the measuring device image.

[0027] In one possible implementation, based on the target detection model, defect detection is performed on the image of the measuring equipment to obtain a performance trend prediction result of the measuring equipment, including:

[0028] Based on the target detection model, the appearance anomaly identification of the metering equipment image is performed to obtain image anomaly features;

[0029] Based on the target fusion weight ratio of the target detection model, the abnormal features of the image are subjected to weighted fusion processing to obtain the performance trend prediction result of the metering equipment.

[0030] In one possible implementation, the image anomaly features are weighted and fused according to the target fusion weight ratio of the target detection model. The method then further includes:

[0031] Obtain the historical metering device fusion feature sequence within the historical window;

[0032] The performance trend of the metering equipment is predicted based on the fusion feature sequence of the historical metering equipment, and the performance trend prediction result is obtained.

[0033] Secondly, this application proposes a performance trend prediction device for metering equipment, comprising:

[0034] The image acquisition module is used to acquire images of the metering equipment;

[0035] The text information acquisition module is used to perform image recognition processing on the image of the metering device to obtain the nameplate text information of the metering device.

[0036] The detection model acquisition module is used to dynamically adjust the fusion weight ratio of the fully connected layer in the detection model based on the nameplate text information to obtain the target detection model;

[0037] The performance trend prediction module is used to perform defect detection on the image of the measuring equipment based on the target detection model, and obtain the performance trend prediction result of the measuring equipment.

[0038] Thirdly, this application provides a performance trend prediction device for metering equipment, including: a memory and a processor;

[0039] The memory stores computer-executed instructions;

[0040] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0041] Fourthly, embodiments of this application provide a performance trend prediction system for a metering device, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0042] The method, apparatus, equipment, and system for predicting the performance trend of metering equipment provided in this application acquire images of the metering equipment; perform image recognition processing on the metering equipment images to obtain the nameplate text information of the metering equipment; dynamically adjust the fusion weight ratio of the fully connected layer in the detection model based on the nameplate text information to obtain a target detection model; and perform defect detection on the metering equipment images based on the target detection model to obtain the performance trend prediction results of the metering equipment. This solves the problems of low efficiency of existing manual inspections, poor adaptability of general vision models to different models of metering equipment, and difficulty in reliably predicting performance change trends based on a single image. It improves the pertinence, stability, and generalization ability of defect identification for different models of metering equipment, enhances the reliability of performance trend prediction, and provides support for equipment early warning and operation and maintenance decisions. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0044] Figure 1 A flowchart illustrating an embodiment of the performance trend prediction method for metering equipment provided in this application;

[0045] Figure 2 A flowchart illustrating Embodiment 2 of the method for predicting the performance trend of metering equipment provided in this application;

[0046] Figure 3 A flowchart illustrating Embodiment 3 of the method for predicting the performance trend of metering equipment provided in this application;

[0047] Figure 4 A schematic diagram of the performance trend prediction device for the metering equipment provided in this application;

[0048] Figure 5 A schematic diagram of the performance trend prediction device for the metering equipment provided in this application.

[0049] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0051] This application applies to intelligent operation and maintenance scenarios for low-voltage current transformers in power systems. Operators use portable smart terminals (such as explosion-proof industrial mobile phones or smart safety helmets) to capture images of the equipment's exterior and upload them to the intelligent operation and maintenance system. The system uses image recognition technology to identify nameplate information, dynamically adjusts the fusion weights of the target detection module, and combines this with historical equipment performance data to generate performance trend prediction results. This solution is deployed on edge computing nodes or cloud servers, communicating with terminal devices via wireless networks (cellular networks or local area networks) to achieve real-time monitoring and early warning of equipment status.

[0052] There are two major contradictions in the existing technology for power equipment operation and maintenance: (1) Insufficient model adaptability: general visual models (such as convolutional neural networks) have low sensitivity to the differences in defect features of different equipment models, and fixed parameter models cannot dynamically adjust the detection strategy, resulting in fluctuations in detection accuracy. (2) Lack of time series modeling: existing methods only perform static analysis on a single image and do not combine historical performance data of the equipment (such as crack propagation rate and corrosion area change rate), which cannot predict the long-term degradation trend of equipment performance, resulting in delayed operation and maintenance decisions.

[0053] To address the aforementioned technical challenges, this application dynamically adjusts the fusion weights of a deep learning model using equipment nameplate information. By combining a dual-path network structure with historical time-series features, it achieves accurate prediction of the performance trends of power metering equipment. Specifically, it utilizes information such as the model number and serial number on the equipment nameplate to dynamically infer the fusion strategy of the model parameters, enabling the model to adapt to the differences in defect characteristics among different equipment types. Simultaneously, through historical fusion feature sequence modeling, it combines single image analysis with equipment performance evolution trends, thereby improving prediction reliability. This technical concept overcomes the fixed parameter limitations of general models on equipment models and solves the problem of existing solutions lacking dynamic time-series modeling.

[0054] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0055] Figure 1 This is a flowchart illustrating an embodiment of the performance trend prediction method for metering equipment provided in this application. Figure 1 As shown, it includes:

[0056] S101. Obtain an image of the metering equipment.

[0057] In this application, the measuring equipment is the object being tested, and the image of the measuring equipment is used as the basic input data for subsequent image recognition processing and defect detection. The digital image file is formed by the acquisition device taking pictures of the appearance of the equipment.

[0058] In one possible implementation, on-site personnel use a portable smart terminal to take a picture of the appearance of the low-voltage current transformer, obtain an image of the metering equipment, and transmit the image of the metering equipment to the intelligent operation and maintenance system.

[0059] Among them, portable smart terminals refer to mobile devices with high-definition image acquisition and wireless communication functions, such as explosion-proof industrial mobile phones, smart safety helmets, or dedicated inspection terminals equipped with cameras.

[0060] During inspections or maintenance, on-site personnel use portable smart terminals to photograph low-voltage current transformers. For example, they can capture images of the overall appearance of the transformer and images of key areas where cracks may exist. The obtained images of the metering equipment are stored digitally in the smart terminal. Subsequently, the images of the metering equipment are uploaded to the intelligent operation and maintenance system via wireless communication networks, such as cellular networks or local area networks, so that the image data enters a unified data processing flow.

[0061] For example, the inspection personnel took a picture of the A-phase LMZ-0.66 model current transformer at tower T-102, obtained a JPEG image with a resolution of 4032x3024, and uploaded it to the system through the operator's IoT card.

[0062] By acquiring images, the technical effect of providing basic data for subsequent analysis of the performance of metering equipment and prediction of performance trends was achieved.

[0063] Based on the above analysis, it can be seen that the images of the metering equipment can be acquired by the on-site acquisition terminal and uploaded to the back-end processing platform, or they can be acquired by a fixed acquisition device and transmitted to the back-end processing platform, as long as the images of the metering equipment can enter the subsequent processing flow.

[0064] After the metering equipment images are acquired and received in this step, they can be converted into the input format used by subsequent algorithms.

[0065] This step establishes an input link for the metering equipment image, providing an input foundation for subsequent extraction of nameplate text information from the image and performance of defect detection.

[0066] S102. Perform image recognition processing on the image of the measuring equipment to obtain the nameplate text information of the measuring equipment.

[0067] In this application, image recognition processing is used to extract textual information that can characterize the device identity from an image of a metering device.

[0068] The nameplate text information serves as the basis for subsequent dynamic adjustments to the testing model. It should include at least one of the equipment model and equipment number, and may also include fields such as manufacturer, rated parameters, manufacturing date, or accuracy class.

[0069] In this step, image recognition processing is performed on the metering equipment image, and the recognition results are organized to output standardized nameplate text information. Specifically, after receiving the image output by S101, the backend processing platform recognizes the relevant content of the equipment nameplate, obtains the corresponding text information, and generates structured nameplate text information.

[0070] In another implementation, if the local area recognition effect is not ideal, the relevant text information in the image of the measuring equipment can be extracted directly, as long as the nameplate text information that can be used for subsequent model adjustment can be obtained.

[0071] This step converts the device identity information in the original image into computable text information, enabling subsequent processing to incorporate device-related information as a basis for model adjustment.

[0072] S103. Based on the nameplate text information, dynamically adjust the fusion weight ratio of the fully connected layer in the detection model to obtain the target detection model.

[0073] In this application, the detection model is a model for defect identification, which has a fully connected layer inside for feature fusion processing.

[0074] The target detection model is a detection model adapted to the current metering equipment after adjustment by fusion weight ratio.

[0075] In one implementation of this application, the target detection model is a target detection model based on deep learning algorithms deployed in an intelligent operation and maintenance system, used to perform appearance anomaly analysis on metering equipment images.

[0076] Specifically, this step dynamically adjusts the feature weights involved in fusion in the fully connected layer based on the device information represented by the nameplate text information, thereby adapting the detection model to the current metering equipment.

[0077] In practice, the system reads the nameplate text information output by S102 and adjusts the fusion weight ratio of the fully connected layer in the detection model based on the nameplate text information to generate the corresponding target detection model.

[0078] After the parameters are adjusted, the system marks the current model instance as an object detection model and binds it to the current image task for use in the next step.

[0079] This step allows the model to use a more suitable parameter configuration when processing images from different metrology devices by changing the fusion weight ratio of the fully connected layer according to the nameplate text information.

[0080] S104. Based on the target detection model, perform defect detection on the image of the measuring equipment to obtain the performance trend prediction results of the measuring equipment.

[0081] In this embodiment of the application, defect detection is performed by using the target detection model generated in S103 to identify abnormalities in the metering equipment image.

[0082] Performance trend prediction results are characterizations of subsequent equipment state changes based on identified defect information.

[0083] Based on the above analysis, it can be seen that after receiving the image of the measuring equipment, the target detection model detects the defects in the image and outputs the performance trend prediction result based on the detection result.

[0084] In practice, the system inputs the image of the measuring equipment into the target detection model, which then outputs the corresponding defect detection results and further obtains the performance trend prediction results of the measuring equipment. The output results can be stored or sent in a structured record format.

[0085] In this application, when the object detection model is set as an end-to-end, post-processing-free, single-stage real-time object detection convolutional network, the structure of the object detection model is determined before dynamic inference is performed. Specifically, it includes a first network layer and a second network layer connected in parallel, as well as a fully connected layer connected to the outputs of the first network layer and the second network layer.

[0086] The first and second network layers are connected in parallel, and their output features are weighted and fused through a shared fully connected layer.

[0087] In practice, the first network layer is activated based on the base class cue vector corresponding to the device type, and its feature extraction focuses more on the typical defect patterns of that model; the second network layer is trained through a mixed dataset and has generalization ability. The output features of the two paths are weighted and converged in the fully connected layer according to the target fusion weight ratio (e.g., [0.8, 0.2]) to generate the metering device fusion features.

[0088] The weighted convergence method involves performing a linear transformation (such as matrix multiplication) on each image anomaly feature component (e.g., cracks, corrosion), and then performing a weighted sum based on the target fusion weight ratio (e.g., [0.8, 0.2]). For example, the weight of the crack feature component is 0.8, the weight of the corrosion feature component is 0.2, and the final fused feature is a linear combination of the two.

[0089] The first network layer is activated based on any one of the multiple base class cue vectors stored in the first semantic coding layer, and each base class cue vector corresponds to a device type.

[0090] In other words, the first network layer can perform targeted anomaly detection for each device type because its parameters are determined based on the base class hint vector corresponding to different device types.

[0091] The second network layer is trained using data from multiple device types, exhibiting high generalization ability and achieving good detection results even for device types with insufficient data.

[0092] For the determination of the first and second network layers, this application can train the initial network architecture based on multiple mapping-associated training sample sets until convergence, and obtain multiple first network state parameters. The multiple first network state parameters refer to multiple independent network weight parameters obtained after training the initial network with a dataset specific to each device type.

[0093] The multiple device types are used to identify the multiple first network state parameters to obtain multiple base class hint vectors;

[0094] The multiple base class cue vectors are stored in the semantic encoding layer of the initial network architecture to obtain the first network layer, and the semantic encoding layer storing the multiple base class cue vectors is used as the first semantic encoding layer.

[0095] For example, abnormal images from N different types of low-voltage current transformers are collected and organized to form N independent image sets, where N is a positive integer, and the images in each set are labeled with the common abnormal categories and locations of that type.

[0096] For example, set A contains 500 images of cracks in LMZ-0.66, and set B contains 300 images of rust and overheat discoloration in LZZBJ9-10.

[0097] Next, pre-training of expert networks for each model was performed, using a standard architecture as the initial network architecture, and each of the N datasets was trained independently. Set A was used to train a network weight WA that is adept at detecting LMZ-0.66 cracks. Set B was used to train a network weight WB that is adept at detecting LZZBJ9-10 corrosion. The training objective was to ensure that each network achieved high detection accuracy on its specific model data, such as an output accuracy greater than or equal to 85%.

[0098] For each network that has converged during training, its parameter states are extracted from its key layers, such as the last convolutional layer or a feature bottleneck layer, and compressed into a fixed-length vector by an encoder network. This vector is then used as the base class cue vector for the corresponding device type.

[0099] In practice, for the network weight parameters trained on the dedicated dataset for each device type, the parameter states of its key layers (such as the last convolutional layer or feature bottleneck layer) are extracted and vectorized through the encoder network (such as a fully connected layer or dimensionality reduction module), and finally stored in the semantic encoding layer of the object detection model for subsequent activation of expert paths.

[0100] Finally, the semantic encoding layer storing multiple base class cue vectors is used as the first semantic encoding layer. These multiple base class cue vectors are stored in the semantic encoding layer of the initial network architecture, thus obtaining the first network layer. The first semantic encoding layer is a queryable vector database that stores the previously generated N base class cue vectors.

[0101] When the network is running, by inputting a device type identifier, the corresponding cue vector can be retrieved from this layer and used as a conditional input to modulate the network's forward propagation process.

[0102] In other words, an expert knowledge base was prepared offline, and the exclusive detection experience of different device models was solidified and stored in the form of base class prompt vectors. This allows the online first network layer to instantly switch to the expert mode for a specific model by calling different prompt vectors, without having to save a complete and huge neural network model for each model, which greatly improves the flexibility and efficiency of the system.

[0103] In one possible implementation, anomaly recognition of the metering equipment image can be performed based on the target detection model to obtain image anomaly features; then, according to the target fusion weight ratio of the target detection model, the image anomaly features are weighted and fused to obtain the performance trend prediction result of the metering equipment.

[0104] Specifically, the images of the metering equipment can be acquired by the on-site inspection terminal at different shooting angles and distances. After the images are input into the target detection model, the model first locates the equipment body, wiring terminals, nameplate edges and shell surfaces, and then extracts features of texture abrupt changes, edge damage and color anomalies in each area to form multiple image anomaly feature components.

[0105] Subsequently, the fully connected layer performs weighted aggregation of the abnormal feature components of the image according to the target fusion weight ratio obtained in the pre-training, and maps the fused features to the performance trend prediction result. The result can characterize the tendency of the metering device to change state in subsequent operation.

[0106] If different image anomaly features have different indicative strengths of state degradation in the training samples, the corresponding weights are set to different values ​​so that image anomaly features with strong correlations have a higher contribution ratio in the fusion result. In practical applications, the target fusion weight ratio can also be adjusted according to the device model and sample distribution.

[0107] In this application, the image anomaly feature set may include multiple sets of images of multiple abnormal metering devices of multiple device types and multiple sets of abnormal appearance features.

[0108] Among them, multiple sets of abnormal measurement equipment images refer to image datasets grouped by equipment type and labeled with appearance anomalies (such as crack locations and rust areas), such as LMZ-0.66_crack set and LZZBJ9-10_rust set. Multiple sets of abnormal appearance features are sets of deep feature vectors extracted from these images that can characterize abnormal patterns.

[0109] In one possible implementation, the image of the measuring device is simultaneously input into the dual paths of a pre-configured target detection model.

[0110] In the first network layer, based on the previously determined device type, such as LMZ-0.66, the corresponding base class cue vector is retrieved from the first semantic coding layer and injected into the network to guide the network forward propagation, making its feature extraction more focused on the typical defect patterns of this model.

[0111] In the second network layer, the image undergoes standard forward propagation without receiving specific cues. The two paths run independently, each ultimately producing a set of high-dimensional feature maps or feature vectors. These two sets of features are then fed into a shared fully connected layer.

[0112] This layer performs the fusion operation based on the target fusion weight ratio dynamically determined in the previous step. For example, for LMZ-0.66 with sufficient data, the ratio is [0.8, 0.2]. The fully connected layer first performs linear transformations on the two input features respectively, and then performs weighted summation to obtain the fused features of the metering equipment.

[0113] In this way, the target detection model can not only output appearance anomaly information in the image of the metering equipment, but also jointly represent the abnormal features of multiple types of images according to the target fusion weight ratio, so that the output result can be expanded from single defect identification to a comprehensive judgment on the direction of equipment performance change.

[0114] Because the feature contribution difference is introduced during the fusion process, the prediction results are less sensitive to local noise and fluctuations in single defects, and can more stably reflect the performance trend of the metering equipment.

[0115] In addition, after weighted fusion processing of image anomaly features, the historical fusion feature sequence of the metering equipment within the historical window can be obtained; the performance trend of the metering equipment can be predicted based on the historical fusion feature sequence of the metering equipment to obtain the performance trend prediction result.

[0116] In practical implementation, the image anomaly features output by the target detection model can include local features, location features, and confidence features of appearance anomalies such as cracks, corrosion, dirt, deformation, and insulation aging. The target fusion weight ratio is used to weight and integrate image anomaly features from different sources or at different levels.

[0117] Weighted fusion processing can be performed by concatenating vectors and then mapping the weights, or by using attention weighting to recalibrate image anomaly features to form unified metrology equipment fusion features.

[0118] The historical metering equipment fusion feature sequence can be organized by the metering equipment fusion features corresponding to the same equipment at different inspection times in chronological order and stored in the backend operation and maintenance database. The time interval can be set by day, week or month, and the historical window length can be configured according to the equipment operation and maintenance cycle and defect evolution cycle.

[0119] Performance trend prediction can be accomplished by recurrent neural networks, temporal convolutional networks, or attention-based temporal prediction models. After inputting the fused feature sequences of historical metering equipment, the output is a trend result characterizing the direction of equipment performance change, degradation rate, or abnormal risk level. The prediction model used can be configured to be co-trained with the target detection model. In practical applications, other models of this component can also be selected, and this application does not limit this.

[0120] For example, the identified equipment asset number is used as the key to query a dedicated time-series feature library. This library records the metering equipment fusion features obtained from each analysis of each equipment, and each feature is timestamped.

[0121] The system extracts relevant features based on a preset historical window, such as the most recent year or the most recent 10 times, to form a historical sequence. For example, for the CT-101 equipment, the feature vectors of the past five inspections [F1, F2, F3, F4, F5] are retrieved.

[0122] Next, sequential modeling and inference are performed. The currently generated feature F6 is concatenated with the historical sequences [F1 to F5] in chronological order to form the complete input sequence [F1, F2, F3, F4, F5, F6]. This sequence is then fed into a pre-trained time series prediction model.

[0123] This model typically employs structures such as Long Short-Term Memory networks or temporal convolutional networks, and has learned the complex correlation between the evolution patterns of such fused feature sequences and subsequent equipment failures. Analyzing this sequence, the model may, for example, identify a continuous and accelerating upward trend in the relevant feature dimensions reflecting insulation crack propagation and metal oxidation over the most recent three cycles.

[0124] Based on this pattern recognition, the model performs calculations. Finally, it generates actionable predictions. The model outputs a quantified prediction value, for example, an 82% probability of failure within the next 90 days, which is used as a performance trend prediction. By concatenating discrete, single-item visual inspection points and performing trend prediction, the technical effect of improving prediction reliability is achieved.

[0125] This implementation method, after completing the weighted fusion of image anomaly features, introduces temporal fusion features within the historical window for trend modeling, so that the current anomaly representation and historical evolution information jointly participate in the prediction, thereby forming a continuous performance judgment.

[0126] Since the prediction results no longer rely solely on static image anomaly features from a single image, but instead combine historical metrology equipment fusion feature sequences for comprehensive analysis, they can output more stable performance trend prediction results and improve the ability to characterize equipment degradation evolution.

[0127] This application's embodiment constructs a continuous processing chain consisting of image acquisition, nameplate text extraction, weight adjustment, defect detection, and trend output. The nameplate text information directly participates in the fusion control of the fully connected layer of the detection model, enabling different metrology devices to complete model adaptation before entering defect identification. Therefore, the performance trend prediction results, based on the processing of nameplate recognition, feature fusion, and defect detection, can form a unified analysis output corresponding to the metrology devices during on-site inspections.

[0128] Figure 2 This is a flowchart illustrating Embodiment Two of the performance trend prediction method for metering equipment provided in this application. Figure 2 As shown, in Figure 1Based on the implementation examples, the fusion weight ratio of the fully connected layers in the detection model is dynamically adjusted according to the nameplate text information to obtain the target detection model, including:

[0129] S201. Determine the data sufficiency coefficient of the metering equipment based on the nameplate text information.

[0130] In this step, after obtaining the nameplate text information, the system first performs structured parsing on the model identifier, manufacturer identifier, capacity parameters, and installation category in the text, then compares the parsing results with the sample index library to obtain the sample coverage of the corresponding model, and calculates the data sufficiency coefficient accordingly.

[0131] In practice, this step can be implemented by first archiving abnormal metering equipment images from different equipment types and labeling them according to the abnormality category, equipment model, and defect manifestation, and then performing feature extraction on the images to obtain a set of abnormal appearance features corresponding to the image content.

[0132] Subsequently, the set of abnormal measurement equipment images and the set of abnormal appearance features are associated according to a preset mapping rule. The mapping rule can be established based on the abnormal identifier, defect location or feature similarity under the same equipment type, so that each image sample corresponds to a set of abnormal appearance feature samples, thereby forming multiple mapping association training sample sets.

[0133] When performing discriminative training based on this training sample set, the parameters of the first network layer and the second network layer can be updated separately, so that the two network layers can form differentiated feature representation capabilities for different device types and different anomaly distributions. Among them, the first network layer can learn more fine-grained anomaly representations for device types with more sufficient samples, while the second network layer can learn more general anomaly representations for device types with relatively insufficient samples.

[0134] The specific methods for discriminative training include: constraining the ability of the first and second network layers to distinguish features between abnormal and normal images by using a contrastive loss function. During training, abnormal and normal images of the same device type are compared in terms of features, making the first network layer more sensitive to defect features of the target device type, and the second network layer more robust to general defect features.

[0135] The two network paths of the detection model can be set in parallel within the same model framework, and their ability to distinguish between abnormal measurement equipment images and abnormal appearance features can be constrained by a loss function.

[0136] Through this step, the detection model can simultaneously absorb image information and abnormal appearance feature information of abnormal metering equipment of multiple equipment types, and form network layer parameters that are adapted to different types of metering equipment during discriminative training.

[0137] Therefore, when selecting the corresponding network layer for fusion weight adjustment based on the data sufficiency coefficient, the model output can match the detection results with the device type and sample sufficiency, thereby improving the stability of abnormal appearance recognition and the model's adaptability.

[0138] S202. If the data sufficiency coefficient meets the preset coefficient threshold, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted according to the first network layer of the detection model to obtain the target detection model.

[0139] Specifically, the mechanism for dynamically adjusting the fusion weight ratio is based on the device type and data sufficiency coefficient in the nameplate text information. When there are sufficient device samples, the system calls the high weight (e.g., 0.8) of the first network layer (expert path) to make its feature extraction more focused on the typical defect patterns of that model, thereby obtaining the target detection model.

[0140] The dynamic adjustment strategy proposed in this application drives model adaptation through nameplate information, thus solving the limitation of fixed parameters of the general model on differences in equipment models.

[0141] S203. If the data sufficiency coefficient does not meet the preset coefficient threshold, the fusion weight ratio is dynamically adjusted according to the second network layer of the detection model to obtain the target detection model.

[0142] Specifically, if the coefficient is lower than a preset threshold, the second network layer is invoked, and general features are fed into the fusion weight calculation unit. A more stable feature response is used to generate the fusion weight ratio of the fully connected layer, thus forming the target detection model. This target detection model is then used for defect detection in images of metering equipment and outputs performance trend prediction results related to the equipment's appearance.

[0143] In one implementation, the system receives the nameplate text information parsed by the OCR (Optical Character Recognition) image recognition module, for example, model number: LMZ-0.6. By querying the device model knowledge base, it maps this information to the target device type TA. Next, the system calculates the data sufficiency coefficient S for this type TA. The calculation of S can be based on various factors, such as: the number of image samples used to train the base class cue vector for this type (Ntrain=500 images), the number of valid records successfully identified for this type of device in the historical maintenance database, and the stability measure of the historical fused feature sequences for this type of device, etc.

[0144] For example, S = min(1, log(Ntrain+1) / log(Nthreshold+1)), where Nthreshold is a preset sufficiency threshold, such as 1000. If Ntrain = 500, then S ≈ 0.85. Then, threshold judgment and weight selection are performed. The calculated coefficient S is compared with the preset coefficient threshold, such as Sthreshold = 0.7. Because 0.85 > 0.7, the data is judged to be sufficient.

[0145] At this point, the system retrieves the first fusion weight ratio preset for this scene from the object detection model's configuration. Assume this ratio specifies that, during fusion at the fully connected layer, the feature weight from the first network layer (the expert path activated by the LMZ-0.66 base class cue vector) is 0.8, and the feature weight from the second network layer (the general path) is 0.2. This ratio is determined as the target fusion weight ratio for this inference.

[0146] Conversely, if the target device is a new model with a training sample size Ntrain=10 and the calculated S=0.2<0.7, it is considered that the data is insufficient. The system will then apply a second fusion weight allocation, such as an expert path weight of 0.2 and a general path weight of 0.8. This achieves an adaptive on-demand fusion strategy, dynamically deciding whether to rely more on specialized experience specific to that model or on general knowledge in the final decision based on the depth of understanding of the device being tested. This effectively solves the problem of insufficient adaptability of general models to different device models. For familiar models, higher weights are given to the expert path to achieve accurate detection; for unfamiliar models or models with scarce data, the general path is relied upon to ensure basic and robust recognition capabilities, avoiding serious misjudgments or missed detections due to insufficient expert knowledge.

[0147] The embodiments of this application ensure that the target detection model can maintain stronger model adaptability when there are sufficient samples and maintain better robustness when there are insufficient samples, thereby improving the adaptability of the target detection model to different measuring equipment and the consistency of detection results.

[0148] Figure 3 This is a flowchart illustrating Embodiment 3 of the performance trend prediction method for metering equipment provided in this application. Figure 3 As shown, in Figure 1 Based on the embodiments, image recognition processing is performed on the image of the measuring equipment to obtain the nameplate text information of the measuring equipment, including:

[0149] S301. Based on edge detection and contour analysis, locate the nameplate area image of the measuring equipment.

[0150] The nameplate area image refers to the rectangular region in the image that contains key textual information such as the device model and serial number. In this step, traditional image processing techniques, namely Canny edge detection and contour finding, are used to locate the nameplate area image within the image.

[0151] In practical implementation, when performing edge detection on the image of the metering equipment, the input image can first be grayscaled and denoised, and then the Canny edge detection operator can be used to extract edge information. The edge threshold can be set as a dual threshold adaptive form according to the image resolution and the ambient lighting conditions to enhance the prominence of the nameplate border against a complex background.

[0152] When performing contour analysis on the edge image, the outer and inner contours can be extracted based on connected component tracing, and the closure of the first and last points of the contours can be verified to obtain a set of closed contours. For each contour in the set of closed contours, its minimum bounding rectangle, quadrilateral approximation result, aspect ratio, corner angle, and the ratio of the contour area to the original image area can be calculated. Combined with the internal gray-level variance of the contour, local texture gradient, and character stroke density, it can be determined whether it meets the rectangular flat plate characteristics of the nameplate.

[0153] During the screening process, when the outline is approximately quadrilateral and the number of the four vertices meets the preset threshold, the included angle between adjacent sides is close to a right angle, the ratio of the lengths of opposite sides falls within the preset range, and the texture inside the outline shows relatively stable character arrangement features, the outline is determined as a candidate area for the nameplate; after further sorting the confidence of multiple candidate areas, the nameplate area image is cropped.

[0154] If multiple contours that meet the criteria exist in the same image, the final nameplate area can be determined by comprehensively considering the contour area, the spatial relationship between the contours and the upper or middle parts of the front of the device, and the text density.

[0155] This step effectively separates the nameplate area from complex backgrounds, reflective surfaces, and equipment structural edges in the metering equipment image, resulting in higher regional accuracy for subsequent text recognition input. By simultaneously introducing quadrilateral geometric and texture feature constraints, interference from the equipment casing, mounting holes, labels, and other rectangular structures on nameplate localization is reduced, thereby improving the stability and consistency of the nameplate extraction results.

[0156] S302. Perform text recognition processing on the nameplate area image to obtain the initial recognition text of the metering equipment.

[0157] In this step, the image of the nameplate area is input into the text recognition unit of the OCR module. The OCR module can output the initial recognized text using a combination of convolutional neural networks and sequence recognition models.

[0158] Specifically, the text recognition unit performs preprocessing such as binarization and perspective correction on the nameplate area image, and then outputs the initial recognition text, such as "Model LMZ-0.66, Serial Number 202305001, Manufacturer ABC Electric".

[0159] S303. Based on the preset power equipment nameplate keyword rule library, the initial identification text is parsed to obtain the nameplate text information of the metering equipment.

[0160] Among them, the preset power equipment nameplate keyword rule base is a structured knowledge base that contains text patterns defined by regular expressions for matching key fields such as "model", "number", and "manufacturer".

[0161] In the specific implementation, the initial recognition text is input into the preset power equipment nameplate keyword rule library. It is parsed according to field tags, colon separators, fixed phrases and context field associations, noise characters are removed and easily confused characters are corrected, and finally the nameplate text information of the metering equipment is obtained.

[0162] This application's embodiments accurately locate the nameplate area through edge detection and contour analysis, then obtain the original text through character recognition, and complete field parsing using a keyword rule base, enabling the nameplate text information to be output in a structured form. Therefore, content related to the equipment model and rated parameters in the nameplate text information can be reliably extracted, providing reliable input for subsequent adjustments to the detection model's weight allocation based on the nameplate text information, thereby improving the compatibility consistency of different models of metering equipment and the stability of the recognition results.

[0163] Figure 4 A schematic diagram illustrating the structure for predicting the performance trend of the metering equipment provided in this application. (See attached diagram.) Figure 4 As shown, the performance trend prediction 40 of the metering equipment provided in this embodiment includes:

[0164] Image acquisition module 401 is used to acquire images of the metering equipment;

[0165] The text information acquisition module 402 is used to perform image recognition processing on the image of the metering equipment to obtain the nameplate text information of the metering equipment.

[0166] The detection model obtains module 403, which is used to dynamically adjust the fusion weight ratio of the fully connected layer in the detection model according to the nameplate text information, so as to obtain the target detection model;

[0167] The performance trend prediction module 404 is used to perform defect detection on the image of the measuring equipment based on the target detection model, and obtain the performance trend prediction result of the measuring equipment.

[0168] In one possible implementation, the detection model obtaining module 403 is further specifically used for:

[0169] Determine the data sufficiency coefficient of the metering equipment based on the information on the nameplate text;

[0170] If the data sufficiency coefficient meets the preset coefficient threshold, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted according to the first network layer of the detection model to obtain the target detection model.

[0171] If the data sufficiency coefficient does not meet the preset coefficient threshold, the fusion weight ratio is dynamically adjusted according to the second network layer of the detection model to obtain the target detection model.

[0172] In one possible implementation, the detection model obtaining module 403 is further specifically used for:

[0173] Acquire multiple sets of images and multiple sets of abnormal appearance features for multiple device types of abnormal metering devices;

[0174] Multiple sets of images of abnormal metering devices and multiple sets of abnormal appearance features are mapped and associated to obtain multiple mapping and association training sample sets;

[0175] Discriminative training is performed based on multiple mapping-related training sample sets to obtain the first and second network layers of the detection model.

[0176] In one possible implementation, the text information obtaining module 402 is further specifically used for:

[0177] Based on edge detection and contour analysis, the nameplate area image of the metering equipment is located.

[0178] The nameplate area image is processed for text recognition to obtain the initial recognition text of the metering device.

[0179] Based on a pre-defined rule base for keywords on power equipment nameplates, the initial identification text is parsed to obtain the nameplate text information of the metering equipment.

[0180] In one possible implementation, the text information obtaining module 402 is further specifically used for:

[0181] Edge detection is performed on the image of the measuring equipment to obtain the edge image of the measuring equipment;

[0182] Contour analysis is performed on the edge image to obtain a set of closed contours;

[0183] Based on the preset rules of quadrilateral geometric features and texture features, the closed contour set is filtered to obtain the nameplate area image of the measuring equipment.

[0184] In one possible implementation, the performance trend prediction module 404 is further specifically used for:

[0185] Based on the target detection model, appearance anomalies are identified in the images of metering equipment to obtain image anomaly features;

[0186] Based on the target fusion weight ratio of the target detection model, the abnormal features of the image are weighted and fused to obtain the performance trend prediction results of the metering equipment.

[0187] In one possible implementation, the performance trend prediction module 404 is further specifically used for:

[0188] Obtain the historical metering device fusion feature sequence within the historical window;

[0189] The performance trend of the metering equipment is predicted based on the fusion feature sequence of historical metering equipment, and the performance trend prediction results are obtained.

[0190] The performance trend prediction device for metering equipment proposed in this application acquires on-site images of low-voltage current transformers through an image acquisition module, and extracts model, specification, and parameter information from the nameplate through a text information acquisition module, enabling the system to determine the current equipment category. The detection model acquisition module dynamically adjusts the fusion weight ratio of the fully connected layer of the detection model based on the nameplate text information, making the target detection model more closely match the appearance and structural differences and defect distribution characteristics of the corresponding equipment model. The performance trend prediction module then uses this target detection model to detect defects such as cracks, corrosion, and insulation aging in the images, maintaining good identification stability even with imbalanced samples for different equipment models, thereby improving the targeting and accuracy of defect judgment. Therefore, the adaptive adjustment of the model driven by nameplate information can improve the reliability of the performance trend prediction results of metering equipment and provide a more effective basis for operation and maintenance early warning and maintenance decisions.

[0191] Figure 5 A schematic diagram of the performance trend prediction device for the metering equipment provided in this application. Figure 5 As shown, the performance trend prediction device 50 for metering equipment provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the performance trend prediction device 50 for metering equipment further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0192] In the specific implementation process, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to execute the above-mentioned performance trend prediction method for metering equipment.

[0193] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0194] The memory can be used to store programs and data related to image recognition, nameplate text extraction, model weight adjustment, and defect detection. After the processor calls the computer to execute instructions, it can recognize the collected images of the measuring equipment, obtain the nameplate text information, and dynamically adjust the fusion weight ratio of the fully connected layer in the detection model based on the nameplate text information to form a target detection model adapted to the current equipment model. Then, the target detection model is used to complete defect detection and performance trend prediction.

[0195] This enables the equipment-side or back-end platform to perform adaptive analysis based on the structural and sample differences of different models of low-voltage current transformers, thereby improving the targeting and stability of anomaly identification and enhancing the reliability of performance change judgment. As a result, it is beneficial to achieve early warning and optimization of operation and maintenance decisions for metering equipment.

[0196] In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc.

[0197] The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the invention can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0198] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0199] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc.

[0200] Buses can be categorized into address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0201] This application also provides a performance trend prediction system for metering equipment, including a computer program that, when executed by a processor, implements the aforementioned performance trend prediction method for metering equipment.

[0202] This system integrates image acquisition, nameplate text recognition, dynamic adjustment of model weights, defect detection, and trend analysis through the execution of computer programs on the processor, enabling the on-site inspection terminal, communication link, and back-end platform to form a collaborative closed loop.

[0203] During program execution, the fusion weight ratio of the fully connected layer in the detection model can be adaptively adjusted based on the text information on the nameplate of the metering equipment. This enhances the model's adaptability to structural differences, defect manifestation differences, and sample distribution differences among different types of low-voltage current transformers, making the defect identification results more stable and targeted.

[0204] Furthermore, the system does not only output a single static identification conclusion, but also supports the prediction of equipment performance change trends based on the detection results, thereby improving the reliability of early warning and operation and maintenance decisions. Therefore, it can meet the application needs of intelligent inspection and refined condition assessment of large-scale metering equipment.

[0205] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned method for predicting the performance trend of a metering device.

[0206] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0207] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0208] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0209] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0210] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0211] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0212] Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0213] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0214] Finally, it should be noted that other embodiments of the invention will readily come to mind for those skilled in the art upon consideration of the specification and practice of the invention disclosed herein.

[0215] This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art that are not disclosed in this invention. It is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting the performance trend of a metering device, characterized in that, include: Acquire images of the metering equipment; The image of the measuring device is processed by image recognition to obtain the nameplate text information of the measuring device; Based on the nameplate text information, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted to obtain the target detection model; Based on the target detection model, defect detection is performed on the image of the measuring equipment to obtain the performance trend prediction result of the measuring equipment.

2. The performance trend prediction method according to claim 1, characterized in that, Based on the nameplate text information, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted to obtain the target detection model, including: Based on the nameplate text information, determine the data sufficiency coefficient of the metering device; If the data sufficiency coefficient meets the preset coefficient threshold, the fusion weight ratio of the fully connected layer in the detection model is dynamically adjusted according to the first network layer of the detection model to obtain the target detection model. If the data sufficiency coefficient does not meet the preset coefficient threshold, the fusion weight ratio is dynamically adjusted according to the second network layer of the detection model to obtain the target detection model.

3. The performance trend prediction method according to claim 2, characterized in that, After determining the data sufficiency coefficient of the metering device based on the nameplate text information, the method further includes: Acquire multiple sets of images and multiple sets of abnormal appearance features for multiple device types of abnormal metering devices; The multiple sets of abnormal metering device images and the multiple sets of abnormal appearance features are mapped and associated to obtain multiple mapping and association training sample sets; Discriminative training is performed based on the multiple mapping-related training sample sets to obtain the first and second network layers of the detection model.

4. The performance trend prediction method according to claim 1, characterized in that, Image recognition processing is performed on the image of the measuring device to obtain the nameplate text information of the measuring device, including: Based on edge detection and contour analysis, the nameplate area image of the metering equipment is located; The nameplate area image is processed for text recognition to obtain the initial recognition text of the metering device; Based on a preset rule library of keywords for power equipment nameplates, the initial identification text is parsed to obtain the nameplate text information of the metering equipment.

5. The performance trend prediction method according to claim 4, characterized in that, Based on edge detection and contour analysis, the nameplate area image of the measuring equipment is located, including: Edge detection is performed on the image of the measuring device to obtain the edge image of the measuring device; Contour analysis is performed on the edge image to obtain a set of closed contours; Based on the preset rules of quadrilateral geometric features and texture features, the closed contour set is filtered to obtain the nameplate area image of the measuring device image.

6. The performance trend prediction method according to claim 1, characterized in that, Based on the target detection model, defect detection is performed on the image of the measuring equipment to obtain the performance trend prediction result of the measuring equipment, including: Based on the target detection model, the appearance anomaly identification of the metering equipment image is performed to obtain image anomaly features; Based on the target fusion weight ratio of the target detection model, the abnormal features of the image are subjected to weighted fusion processing to obtain the performance trend prediction result of the metering equipment.

7. The performance trend prediction method according to claim 6, characterized in that, After weighted fusion processing of the image anomaly features based on the target fusion weight ratio of the target detection model, the method further includes: Obtain the historical metering device fusion feature sequence within the historical window; The performance trend of the metering equipment is predicted based on the fusion feature sequence of the historical metering equipment, and the performance trend prediction result is obtained.

8. A performance trend prediction device for metering equipment, characterized in that, include: The image acquisition module is used to acquire images of the metering equipment; The text information acquisition module is used to perform image recognition processing on the image of the metering device to obtain the nameplate text information of the metering device. The detection model acquisition module is used to dynamically adjust the fusion weight ratio of the fully connected layer in the detection model based on the nameplate text information to obtain the target detection model; The performance trend prediction module is used to perform defect detection on the image of the measuring equipment based on the target detection model, and obtain the performance trend prediction result of the measuring equipment.

9. A performance trend prediction device for metering equipment, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A performance trend prediction system for metering equipment, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images