Automatic Classification and Calibration Method and System for Weight Measurement Based on Image Recognition

Abstract

This invention provides an automatic classification and calibration method and system for weight measurement based on image recognition, relating to the field of weight measurement technology. The method includes imaging the weight surface using a dual-optical-path high-speed camera and performing gamma correction to extract contour features and surface defect features; obtaining the internal density distribution using X-ray imaging; fusing surface features and density distribution at multiple scales to construct a holographic feature model; analyzing the dynamic changing trends of feature parameters based on a deep variational Bayesian network, and combining this with a gradient boosting decision forest for evaluation and scoring to obtain an initial classification; establishing a dynamic evaluation model using a self-organizing competitive learning network to determine the weight level and set calibration parameters; selecting a corresponding benchmark weight to establish a graded calibration compensation model, and dynamically adjusting it by determining adaptive weight coefficients based on defect distribution; and using a neural network of constant differential equations to predict performance degradation trends, outputting calibration parameters and generating early warning information when the calibration accuracy meets the threshold requirements.

Description

Technical Field

[0001] This invention relates to the field of weight measurement technology, and in particular to an automatic classification and calibration method and system for weight measurement based on image recognition. Background Technology

[0002] Weights are important standard instruments in the field of metrology and testing, and their quality accuracy and stability directly affect the reliability of measurement results. Traditional weight inspection mainly relies on manual visual inspection and mechanical balance measurement, which is difficult to accurately identify minute defects on the weight surface and changes in its internal structure. With the development of metrology technology, machine vision-based weight inspection methods have gradually emerged. Through image processing technology, the automatic identification and analysis of surface features of weights can be achieved, providing a new technical means for weight quality assessment.

[0003] However, existing weight testing technologies still have several shortcomings. Existing technologies mainly focus on the single detection of surface features of the weights, lacking effective analysis of the internal structure and failing to comprehensively assess the overall quality status of the weights. Traditional image recognition algorithms are not accurate enough in extracting micro-damage and stress distribution features on the weight surface and are easily affected by factors such as ambient light. Existing weight classification methods often adopt a static evaluation mode, which cannot effectively track and predict the dynamic changes in weight performance. Calibration compensation strategies lack adaptive adjustment mechanisms, making it difficult to achieve accurate calibration for weights of different grades.

[0004] In summary, this invention presents a multi-sensor fusion-based dynamic calibration and early warning method for weights. It achieves comprehensive acquisition of surface and internal features of the weights through a combination of dual-optical-path high-speed imaging and micro-focus X-ray imaging. Advanced algorithms such as deep learning and manifold learning are employed to construct a holographic feature model and a dynamic evaluation model for the weights, enabling accurate classification of weight quality. An adaptive calibration compensation mechanism is established based on a neural network of ordinary differential equations to predict performance degradation trends, ensuring that calibration accuracy meets the usage requirements of weights of different grades. This invention addresses the problems in existing technologies. Summary of the Invention

[0005] This invention provides an automatic classification and calibration method and system for weight measurement based on image recognition, which can solve the problems in the prior art.

[0006] A first aspect of the present invention,

[0007] An automatic classification and calibration method for weight measurement based on image recognition is provided, including:

[0008] A dual-optical-path high-speed camera system was used to acquire images of the double-reflected light on the surface of the weight. Adaptive gamma correction was applied to the double-reflected light images to obtain gamma-corrected images. A multi-scale edge detection algorithm was used to extract the contour features of the weight. Combined with an optical flow field analysis algorithm, surface micro-damage and stress distribution were detected to generate a surface defect feature map. Density distribution data was obtained using an X-ray imaging system. A multi-scale feature fusion network was used to fuse the surface defect feature map with the density distribution data to construct a holographic feature model.

[0009] The feature parameters of the weights are extracted based on the holographic feature model. The dynamic change trend of the feature parameters is obtained based on the deep variational Bayesian network. The gradient boosting decision forest algorithm is used to evaluate and score the weights to generate the initial classification results. Based on the initial classification results, a self-organizing competitive learning network is trained to establish a dynamic evaluation model, determine the weight level classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters.

[0010] Based on the classification of weight grades, calibration reference weights are selected, a hierarchical calibration compensation model based on manifold learning is established, and the compensation strategy parameters are used for initialization. Adaptive weight coefficients are determined by combining the defect distribution information in the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural network of constant differential equations is used to predict the performance degradation trend. Based on the calibration accuracy threshold evaluation, calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance early warning information is generated.

[0011] In one alternative embodiment,

[0012] Images of double-reflected light from the surface of a weight are acquired using a dual-optical-path high-speed camera system. Adaptive gamma correction is then applied to these images to obtain gamma-corrected images, including:

[0013] A dual-optical-path high-speed camera system is constructed, including a first LED light source and a second LED light source, which generate positive reflection light images and diffuse reflection light images respectively. The system is triggered alternately by a controller, and a high-speed camera is used to acquire the dual reflection light images of the weight surface.

[0014] A morphological neural network is used to decompose the dual reflection light image of the weight surface into multiple scales, extract surface feature maps of different scales, including texture feature maps and structural feature maps, calculate local phase consistency features, construct a phase attention mapping matrix, and dynamically recalibrate the surface feature maps of different scales based on the phase attention mapping matrix to obtain enhanced multi-scale feature maps.

[0015] A dual-branch self-calibration network is constructed for the enhanced multi-scale feature map. The first branch performs geometric calibration through a spatial transformation module to obtain geometric calibration features, and the second branch performs photometric calibration through a channel recalibration module to obtain photometric calibration features. The geometric calibration features and photometric calibration features are fused through an adaptive feature aggregation module to obtain a calibration feature map.

[0016] A hierarchical recurrent neural network is constructed to calculate the adaptive gamma correction coefficients layer by layer on the calibration feature map. A residual feedback mechanism is introduced to dynamically adjust the correction parameters. Nonlinear mapping is performed on the multi-scale feature map, and the gamma-corrected image is reconstructed through a deconvolution network.

[0017] In one alternative embodiment,

[0018] A multi-scale edge detection algorithm is used to extract the contour features of the weights, and an optical flow field analysis algorithm is combined to detect surface micro-damage and stress distribution, generating a surface defect feature map including:

[0019] A multi-scale Gaussian filter is used to decompose the gamma-corrected image to obtain a multi-scale image group. Bilateral filtering is performed on each scale image to obtain a smoothed image. The corresponding horizontal gradient features and vertical gradient features are calculated to construct gradient magnitude features. The spatial local statistical features and gray-level local statistical features of the smoothed image are calculated to construct a bilateral adaptive threshold. The gradient magnitude features are segmented using the bilateral adaptive threshold to obtain an edge feature map. The edge feature maps of different scales are weighted and fused to obtain the weight contour features.

[0020] The optical flow field features of adjacent frames are calculated at the edge points of the weight contour features. The surface stress tensor features of the weight are calculated based on the optical flow field features. The surface stress amplitude features and surface stress direction features of the weight are obtained by feature decomposition.

[0021] A feature matrix is ​​constructed by taking the contour features of the weight, the surface stress amplitude features of the weight, and the surface stress direction features of the weight. The normalized feature matrix is ​​obtained by normalization. The corresponding intra-class distance features are calculated. The fusion weight coefficient is determined based on the intra-class distance features. The normalized feature matrix is ​​then weighted and fused using the fusion weight coefficient to obtain the surface defect feature map.

[0022] In one alternative embodiment,

[0023] The feature parameters of the weights are extracted based on the holographic feature model, and the dynamic change trend of the feature parameters is obtained based on the deep variational Bayesian network, including:

[0024] Based on the original data matrix corresponding to the holographic feature model, the mean of the data in the feature dimension is calculated and subtracted to obtain the centered data matrix; the covariance matrix of the centered data matrix is ​​calculated and eigenvalue decomposition is performed to obtain eigenvalues ​​and eigenvectors, a whitening transformation matrix is ​​constructed, and the centered data matrix is ​​multiplied on the left by the whitening transformation matrix to obtain the whitened data matrix;

[0025] Initialize the separation matrix, project the whitened data matrix onto the column vector space of the separation matrix, calculate the probability density function of the projected data, calculate the negative entropy estimate based on the probability density function, update the column vectors of the separation matrix using a fast fixed-point iterative algorithm until the preset maximum number of iterations is reached, and determine the final separation matrix; multiply the whitened data matrix on the left by the final separation matrix to obtain independent components, which are divided into spatial modal components, temporal modal components, and feature modal components;

[0026] The surface gradient of the spatial modal components is calculated, adaptive weighting coefficients are determined and weighted, and the weighted surface gradient is spatially integrated to obtain the first feature parameter; the temporal modal components are spatially integrated, and the integration result is convolved with a pre-calibrated stress kernel function to obtain the second feature parameter; the normalized distribution of the feature modal components is calculated, and the information entropy is calculated based on the normalized distribution to obtain the third feature parameter;

[0027] The first, second, and third feature parameters are combined to form a feature vector, which is then input into a deep variational Bayesian network. The encoder calculates the latent variable distribution. The gated recurrent unit uses the latent variable distribution stored in the previous time step as the latent state input, and together with the feature vector at the current time step, it calculates and updates the latent variable distribution at the current time step. The latent variable distribution at the current time step is then processed by the decoder to obtain the probability distribution of the feature vector.

[0028] By minimizing the relative entropy between the latent variable distribution and the standard normal distribution, and maximizing the log-likelihood of the feature vector, the network parameters are optimized, the optimal network is determined, and the Monte Carlo method is used to sample the optimal network multiple times to obtain the feature vector and the change range at the prediction time, thereby obtaining the dynamic change trend of the weight parameters.

[0029] In one alternative embodiment,

[0030] The initial classification results are generated by evaluating and scoring using the gradient boosting decision forest algorithm, including:

[0031] A feature parameter set is constructed based on the feature parameters of the weights. The time-series feature sequence is obtained by time-series sampling of the feature parameters of each weight. The first-order difference value and the second-order difference value are calculated to construct a dynamic feature set.

[0032] The feature parameter set and the dynamically changing feature set are used as training samples to construct a gradient boosting decision forest model. The mean squared error is used as the loss function. The negative gradient value of the loss function is calculated to fit the model residual. Based on the model residual, a residual learning strategy is used to construct a decision tree.

[0033] In the process of constructing a decision tree, feature sampling is performed on the feature parameter set and the dynamically changing feature set to obtain the features to be segmented; the initial Gini coefficient of each node to be segmented is calculated, and the left and right child nodes are obtained by segmenting using the features to be segmented; the Gini coefficients of the left and right child nodes are calculated; the segmentation gain value is calculated based on the initial Gini coefficient and the Gini coefficients of the left and right child nodes; and the feature to be segmented with the largest segmentation gain value is selected as the optimal segmentation feature.

[0034] Substitute the first-order and second-order difference values ​​from the dynamic change feature set into the exponential decay function to calculate the time series weight values, thus obtaining the time series weight set.

[0035] The gradient boosting decision forest model is used to predict the feature parameters of the weights, obtain the predicted probability value of the decision tree, calculate the average predicted probability value, and sum it with the time series weight value to obtain the score value of the weights.

[0036] The initial classification result of the weight is obtained by comparing the score of the weight with the preset classification threshold.

[0037] In one alternative embodiment,

[0038] The self-organizing competitive learning network includes:

[0039] A two-layer self-organizing competitive learning network is constructed, consisting of input layer neurons with the same feature dimension as the weights, and competitive layer neurons arranged in a two-dimensional grid topology.

[0040] Randomly initialize the connection weights of the two-layer self-organizing competitive learning network, input the weight features into the two-layer self-organizing competitive learning network, calculate the Euclidean distance between the weight features and the connection weights, select the competitive layer neurons corresponding to the smallest and second smallest Euclidean distances, and determine the first winning neuron and the second winning neuron.

[0041] The neighborhood radius is determined based on the location of the first winning neuron. The neighborhood response intensity of the competing layer neurons within the neighborhood radius is calculated, and it decreases exponentially with the increase of the distance from the first winning neuron.

[0042] The connection weights of the first winning neuron and the competing neurons in the neighborhood are updated based on the neighborhood response intensity. A time-varying learning rate that decays linearly with the number of iterations is introduced to adjust the update magnitude and obtain the updated connection weights.

[0043] The quantization error between the weight features and the updated connection weights is calculated to determine the positional relationship between the first winning node and the second winning node. When they are not adjacent, the shortest path length is calculated to obtain the topology error value. When the quantization error value is less than the first preset threshold and the topology error value is less than the second preset threshold, the self-organizing competitive learning network is determined to have reached the training completion state.

[0044] The weight features of the weight to be evaluated are input into the trained deterministic self-organizing competitive learning network. Based on the updated connection weights and neighborhood response strengths, the evaluation level of the weight to be evaluated is generated.

[0045] In one alternative embodiment,

[0046] Based on the classification of weight grades, calibration reference weights are selected. A hierarchical calibration compensation model based on manifold learning is established and initialized using compensation strategy parameters. Adaptive weight coefficients are determined by combining defect distribution information from the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural network of ordinary differential equations is used to predict performance degradation trends. Based on calibration accuracy threshold evaluation, calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance early warning information is generated, including:

[0047] Collect feature data of the weight to be calibrated and the calibration reference weight, and construct feature vectors of the weight to be calibrated and the reference weight;

[0048] Perform a k-nearest neighbor search on the feature vector of the weight to be calibrated, calculate the covariance matrix between the feature point and the k nearest neighbors, construct the compensation strategy parameter optimization function, and solve it under the constraint that the sum of the weights of the nearest neighbors is 1 to obtain the compensation strategy parameters.

[0049] Extract the defect distribution information of the weight to be calibrated from the holographic feature model, calculate the L2 norm of the defect distribution information, multiply it by the preset attenuation coefficient to obtain the attenuation factor, and calculate the adaptive weight coefficient.

[0050] The adaptive weight coefficients are multiplied by the compensation strategy parameters to obtain the optimized compensation strategy parameters. A calibration mapping function is constructed, and the feature vector of the weight to be calibrated is input into the calibration mapping function to obtain the calibration compensation result.

[0051] A neural network of constant differential equations is constructed, and the calibration compensation results are used to build a training sequence in chronological order. A hybrid solution strategy is adopted to divide the sequence into multiple subsequences. A linear approximation function is constructed to obtain a piecewise approximation sequence. The gradient change rate is calculated and dynamic weight smoothing is performed. The optimized numerical sequence is obtained through optimization. The optimization is iteratively performed on each subsequence. When the error evaluation value is less than the preset error threshold, the optimized numerical sequence is output to determine the performance degradation prediction result.

[0052] The calibration accuracy is determined by calculating the Euclidean distance between the calibration compensation result and the feature vector of the reference weight. When the calibration accuracy is less than the preset calibration accuracy threshold, the optimized compensation strategy parameters are output as the final calibration parameters, and graded maintenance early warning information is generated based on the performance degradation prediction results.

[0053] A second aspect of the present invention,

[0054] An image recognition-based automatic classification and calibration system for weight measurement is provided, comprising:

[0055] The first unit is used to acquire double-reflection light images of the weight surface through a dual-optical-path high-speed camera system, perform adaptive gamma correction on the double-reflection light images to obtain gamma-corrected images, extract the weight contour features using a multi-scale edge detection algorithm, and detect surface micro-damage and stress distribution by combining an optical flow field analysis algorithm to generate a surface defect feature map; density distribution data is obtained using an X-ray imaging system, and a multi-scale feature fusion network is used to fuse the surface defect feature map with the density distribution data to construct a holographic feature model.

[0056] The second unit is used to extract the feature parameters of the weights based on the holographic feature model, obtain the dynamic change trend of the feature parameters based on the deep variational Bayesian network, evaluate and score them through the gradient boosting decision forest algorithm, and generate the initial classification results; based on the initial classification results, a self-organizing competitive learning network is trained to establish a dynamic evaluation model, determine the weight level classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters.

[0057] The third unit is used to select calibration reference weights according to the weight level classification, establish a graded calibration compensation model based on manifold learning, initialize the model using compensation strategy parameters, determine adaptive weight coefficients by combining defect distribution information in the holographic feature model, and dynamically adjust the calibration compensation model; it uses a neural network of constant differential equations to predict performance degradation trends, evaluates based on calibration accuracy thresholds, outputs calibration parameters when calibration accuracy meets requirements, and generates graded maintenance early warning information.

[0058] A third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] Memory used to store processor-executable instructions;

[0062] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0063] Fourth aspect of the present invention,

[0064] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0065] In this embodiment of the invention, a combination of dual-path high-speed imaging and X-ray imaging is used to achieve comprehensive acquisition of the surface and internal features of the weights. A holographic feature model is constructed through a multi-scale feature fusion network, improving the accuracy and completeness of weight feature extraction and providing a reliable data foundation for subsequent classification and calibration. A deep variational Bayesian network combined with a gradient boosting decision forest algorithm is used to dynamically evaluate and classify the weight feature parameters. A dynamic evaluation model is established through a self-organizing competitive learning network, achieving accurate classification of weight levels. This makes the setting of calibration accuracy thresholds and compensation strategies more targeted, improving the accuracy and adaptability of classification. A hierarchical calibration compensation model is constructed based on manifold learning, dynamically adjusted by combining defect distribution information in the holographic feature model, and the performance degradation trend is predicted through a neural network of ordinary differential equations. This achieves accurate compensation and predictive maintenance of weight calibration, improving calibration efficiency and reliability, and extending the service life of the weights. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the automatic classification and calibration method for weight measurement based on image recognition, as described in an embodiment of the present invention.

[0067] Figure 2 This is a topology mapping and feature distribution diagram for a two-layer self-organizing competitive learning network. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] The technical solution of the present invention will be described in detail below with reference to 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.

[0070] Figure 1 This is a flowchart illustrating the automatic classification and calibration method for weight measurement based on image recognition, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] A dual-optical-path high-speed camera system was used to acquire images of the double-reflected light on the surface of the weight. Adaptive gamma correction was applied to the double-reflected light images to obtain gamma-corrected images. A multi-scale edge detection algorithm was used to extract the contour features of the weight. Combined with an optical flow field analysis algorithm, surface micro-damage and stress distribution were detected to generate a surface defect feature map. Density distribution data was obtained using an X-ray imaging system. A multi-scale feature fusion network was used to fuse the surface defect feature map with the density distribution data to construct a holographic feature model.

[0072] In one specific implementation, the surface of the weight is first imaged using a dual-beam high-speed camera system. This system includes two sets of parallel LED light sources and a high-speed camera. The two parallel light sources illuminate the surface of the weight at incident angles of 30 degrees and 60 degrees, and the high-speed camera captures images of the double-reflected light. This dual-beam design can obtain richer three-dimensional reflection characteristics of the weight surface and effectively reduce the influence of ambient light interference.

[0073] After acquiring the double-reflected light image, the system performs adaptive gamma correction. First, the grayscale histogram of the image is calculated to analyze its brightness distribution characteristics. Then, based on the image's contrast and brightness distribution, the gamma coefficient values ​​are adaptively adjusted, and a gamma transform is performed on the original image. This step effectively enhances image details and improves the accuracy of subsequent feature extraction. Based on the gamma-corrected image, a multi-scale edge detection algorithm is used to extract the weight's contour features. This algorithm constructs image pyramids at different scales, calculates image gradients at each scale, combines non-maximum suppression and a double-threshold method to detect edges, and finally fuses the multi-scale detection results to obtain complete contour feature information.

[0074] Meanwhile, the system utilizes an optical flow field analysis algorithm to detect micro-damage and stress distribution characteristics on the surface of the weight. By calculating the optical flow field between adjacent image frames and analyzing its gradient change characteristics, minute surface damage can be effectively identified. Based on the density distribution characteristics of the optical flow field, the surface stress distribution is evaluated, and a surface defect feature map is finally generated. On the other hand, a microfocus X-ray imaging system is used to scan the interior of the weight. Multi-angle projection images are acquired through rotational scanning, and image reconstruction algorithms are used to construct three-dimensional density distribution data inside the weight.

[0075] Finally, this invention employs a multi-scale feature fusion network to deeply fuse surface defect feature maps with density distribution data. This network, through multi-level feature extraction and fusion operations, organically combines the surface and internal features of the weight to construct a complete holographic feature model of the weight. This fusion strategy fully utilizes the complementarity of surface and internal features, providing comprehensive feature representation for subsequent classification and evaluation.

[0076] In this embodiment, a dual-optical-path high-speed camera system combined with gamma correction and multi-scale edge detection algorithms can efficiently extract the surface contour and micro-damage features of the weight, accurately identifying surface defects and stress distribution. A micro-focus X-ray imaging system is used to acquire high-precision density distribution data inside the weight, revealing internal structural features and enabling non-destructive testing. A multi-scale feature fusion network is employed to fuse surface defect features with internal density distribution, generating a more comprehensive and accurate holographic feature model of the weight, enhancing the ability to assess the overall state of the weight. Through joint analysis of surface and internal features, the limitations of single detection methods are avoided, significantly improving detection efficiency and accuracy, and providing a scientific basis for weight performance evaluation.

[0077] The feature parameters of the weights are extracted based on the holographic feature model. The dynamic change trend of the feature parameters is obtained based on the deep variational Bayesian network. The gradient boosting decision forest algorithm is used to evaluate and score the weights to generate the initial classification results. Based on the initial classification results, a self-organizing competitive learning network is trained to establish a dynamic evaluation model, determine the weight level classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters.

[0078] In one specific implementation, key feature parameters are first extracted from the holographic feature model of the weight, including surface structure parameters (such as surface defect area, depth, and distribution density), internal characteristic parameters (such as density uniformity and porosity), and overall deformation parameters (such as geometric dimensional deviation and surface stress distribution). To analyze the dynamic changes of these feature parameters, a prediction model is constructed using a deep variational Bayesian network. This network contains multiple layers of hidden variables, uses variational inference methods to probabilistically model the parameter distribution, and employs reparameterization techniques for parameter estimation, thereby accurately capturing the temporal variation characteristics of the feature parameters.

[0079] Based on feature parameter analysis, this invention employs a gradient boosting decision forest algorithm for evaluation and scoring. The algorithm first constructs a basic decision tree, using feature parameters and dynamic trends as input features. During training, the algorithm iteratively optimizes the structure and parameters of the decision tree by minimizing the loss function. Each decision tree focuses on the residual from the previous prediction, continuously improving prediction accuracy through gradient boosting. Specifically, for each feature parameter, the algorithm calculates its corresponding importance score and, combined with the evaluation results of the dynamic trends, comprehensively generates the initial classification result for the weights.

[0080] Based on the initial classification results, this invention trains a self-organizing competitive learning network to construct a dynamic weight evaluation model. This network employs a competitive learning mechanism and includes an input layer, a competition layer, and an output layer. During training, the network continuously adjusts the neuron weights to map samples with similar features to a nearby output space. The network first randomly initializes the weight matrix, then inputs the feature vector, calculates the Euclidean distance to determine the winning neuron, and updates the weights of the winning neuron and its neighboring neurons. After multiple rounds of iterative training, the network ultimately forms a stable feature mapping relationship.

[0081] Based on the output of the dynamic evaluation model, this invention classifies the weights into grades. Specific classification criteria include: E1 grade (highest precision requirement), E2 grade (high precision requirement), F1 grade (medium precision requirement), and F2 grade (standard precision requirement). For different grades of weights, the system sets corresponding calibration accuracy thresholds; for example, the relative error of E1 grade weights should not exceed ±0.5 × 10⁻⁶. -6 The relative error of E2 grade weights does not exceed ±1.6 × 10⁻⁶. -6 Meanwhile, compensation strategy parameters, including calibration cycle, compensation coefficient, and temperature correction factor, are determined based on the weight's class. These parameters will be used in subsequent calibration and compensation processes to ensure that the weight's measurement accuracy meets the requirements of the corresponding class.

[0082] In this embodiment, the dynamic change trend of the weight's feature parameters is extracted based on a deep variational Bayesian network, which can accurately capture the changes in the weight's state over time, providing a basis for long-term stability assessment. The gradient boosting decision forest algorithm is combined to evaluate and score the weight's feature parameters and dynamic change trends, quickly generating initial classification results and improving the accuracy and efficiency of classification. A dynamic evaluation model is established by training a self-organizing competitive learning network, realizing intelligent evaluation of weight level classification and improving the scientific rigor and adaptability of the classification. Based on the weight level classification results, calibration accuracy thresholds and compensation strategy parameters are dynamically set, providing personalized solutions for accurate calibration and error compensation of the weights, enhancing their metrological performance and application value. Through dynamic evaluation and classification, combined with the optimization of accuracy thresholds and compensation strategies, the entire process of weight management from monitoring to operation is optimized, improving overall performance assurance capabilities.

[0083] Based on the classification of weight grades, calibration reference weights are selected, a hierarchical calibration compensation model based on manifold learning is established, and the compensation strategy parameters are used for initialization. Adaptive weight coefficients are determined by combining the defect distribution information in the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural network of constant differential equations is used to predict the performance degradation trend. Based on the calibration accuracy threshold evaluation, calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance early warning information is generated.

[0084] In this embodiment, a hierarchical calibration compensation model based on manifold learning, combined with defect distribution information and adaptive weight coefficients, dynamically adjusts the calibration strategy to achieve precise compensation of the weights and improve the reliability of calibration results. A neural network of ordinary differential equations is employed to accurately predict the performance degradation trend of the weights, providing data support for early prediction of weight status and ensuring timely and effective calibration. Based on calibration accuracy thresholds, the mass status of the weights is evaluated in real time, and the final calibration parameters are output when the accuracy meets the requirements, improving calibration efficiency and accuracy. Hierarchical maintenance early warning information is generated based on performance degradation prediction results to help formulate targeted maintenance plans, extend the service life of the weights, and reduce maintenance costs. Through dynamic adjustment of the compensation model and real-time monitoring, prediction, and evaluation, a full-process, intelligent weight management system is constructed, effectively improving the weight performance assurance capability.

[0085] In one optional implementation, a dual-reflection light image of the weight surface is acquired using a dual-optical-path high-speed camera system, and adaptive gamma correction is performed on the dual-reflection light image to obtain a gamma-corrected image, including:

[0086] A dual-optical-path high-speed camera system is constructed, including a first LED light source and a second LED light source, which generate positive reflection light images and diffuse reflection light images respectively. The system is triggered alternately by a controller, and a high-speed camera is used to acquire the dual reflection light images of the weight surface.

[0087] A morphological neural network is used to decompose the dual reflection light image of the weight surface into multiple scales, extract surface feature maps of different scales, including texture feature maps and structural feature maps, calculate local phase consistency features, construct a phase attention mapping matrix, and dynamically recalibrate the surface feature maps of different scales based on the phase attention mapping matrix to obtain enhanced multi-scale feature maps.

[0088] A dual-branch self-calibration network is constructed for the enhanced multi-scale feature map. The first branch performs geometric calibration through a spatial transformation module to obtain geometric calibration features, and the second branch performs photometric calibration through a channel recalibration module to obtain photometric calibration features. The geometric calibration features and photometric calibration features are fused through an adaptive feature aggregation module to obtain a calibration feature map.

[0089] A hierarchical recurrent neural network is constructed to calculate the adaptive gamma correction coefficients layer by layer on the calibration feature map. A residual feedback mechanism is introduced to dynamically adjust the correction parameters. Nonlinear mapping is performed on the multi-scale feature map, and the gamma-corrected image is reconstructed through a deconvolution network.

[0090] In one specific implementation, a dual-optical-path high-speed camera system is first constructed for image acquisition. This system includes two LED light sources, one for generating direct reflection and the other for diffuse reflection. The first LED light source is a white LED with a luminous intensity of 500 lumens and an illumination angle of 45 degrees; the second LED light source is a diffuser LED with a luminous intensity of 300 lumens and an illumination angle of 60 degrees. The alternating trigger time interval is set to 1 millisecond via a controller, and a high-speed camera with a frame rate of 2000 frames per second is used for image acquisition. The image resolution is set to 1024×1024 pixels, and the pixel depth is 12 bits.

[0091] The acquired double-reflection light images were subjected to multi-scale decomposition. A morphological neural network was used for feature extraction, comprising four scale layers, each using a 3×3 convolutional kernel. The first layer extracted texture features with 64 convolutional kernels; the second layer extracted local structural features with 128 convolutional kernels; the third layer extracted medium-scale features with 256 convolutional kernels; and the fourth layer extracted large-scale features with 512 convolutional kernels. Local phase consistency was calculated for the feature maps of each scale layer, with a window size of 7×7 pixels. An attention mapping matrix was constructed based on the phase consistency values ​​to weight and calibrate the feature maps.

[0092] Next, feature enhancement is performed using a dual-branch self-calibration network. The spatial transformation branch employs a deformable convolutional network, setting 9 sampling points for geometric transformation, with transformation parameters predicted through 3 fully connected layers. The channel recalibration branch uses a channel attention mechanism to calculate the importance weight of each channel, with weights ranging from 0 to 1. The features from the two branches are fused through an adaptive feature aggregation module, which contains a multi-layer residual connection structure.

[0093] Finally, a hierarchical recurrent neural network is used to calculate the gamma correction coefficients. The network contains five recurrent layers, each predicting a gamma value. The initial gamma value of the first layer is set to 2.2, and subsequent layers are dynamically adjusted based on residual feedback, with an adjustment step size of 0.1. The corresponding gamma correction is applied to the feature map at each scale, and then feature reconstruction is performed through a four-layer deconvolutional network, ultimately outputting the corrected image. The deconvolution kernel size is 4×4, and the stride is 2.

[0094] In this embodiment, the alternating trigger acquisition mechanism of the dual-optical-path high-speed camera system enables the simultaneous acquisition of orthographic and diffuse reflection information from the weight surface, improving the integrity and accuracy of image information and providing a reliable data foundation for subsequent processing. The combination of multi-scale decomposition and phase attention mechanisms effectively extracts multi-level features from the weight surface, and dynamic recalibration highlights important feature regions, enhancing the discriminative power of feature representation and improving the clarity of image details. Based on a processing scheme using a dual-branch self-calibration network and a hierarchical recurrent neural network, adaptive geometric and photometric correction of the image is achieved. The residual feedback mechanism dynamically optimizes the gamma correction parameters, significantly improving image quality and observation effects, making the detection of surface defects on the weight more accurate and reliable.

[0095] In one optional implementation, a multi-scale edge detection algorithm is used to extract the contour features of the weight, and an optical flow field analysis algorithm is combined to detect surface micro-damage and stress distribution, generating a surface defect feature map including:

[0096] A multi-scale Gaussian filter is used to decompose the gamma-corrected image to obtain a multi-scale image group. Bilateral filtering is performed on each scale image to obtain a smoothed image. The corresponding horizontal gradient features and vertical gradient features are calculated to construct gradient magnitude features. The spatial local statistical features and gray-level local statistical features of the smoothed image are calculated to construct a bilateral adaptive threshold. The gradient magnitude features are segmented using the bilateral adaptive threshold to obtain an edge feature map. The edge feature maps of different scales are weighted and fused to obtain the weight contour features.

[0097] The optical flow field features of adjacent frames are calculated at the edge points of the weight contour features. The surface stress tensor features of the weight are calculated based on the optical flow field features. The surface stress amplitude features and surface stress direction features of the weight are obtained by feature decomposition.

[0098] A feature matrix is ​​constructed by taking the contour features of the weight, the surface stress amplitude features of the weight, and the surface stress direction features of the weight. The normalized feature matrix is ​​obtained by normalization. The corresponding intra-class distance features are calculated. The fusion weight coefficient is determined based on the intra-class distance features. The normalized feature matrix is ​​then weighted and fused using the fusion weight coefficient to obtain the surface defect feature map.

[0099] In one specific implementation, the acquired weight image is first preprocessed with gamma correction. Gamma correction adjusts the image brightness distribution through nonlinear transformation, enhancing image contrast. Specifically, the gamma coefficient is set to 2.2, and a power-law transformation is performed on the image pixel values ​​to make the image brightness distribution more uniform.

[0100] Based on the gamma-corrected image, multi-scale edge detection is used to extract the contour features of the weights. The image is decomposed into four different scales using a Gaussian filter bank, with filter scales set to 1, 2, 4, and 8 pixels respectively. Each scale image is then smoothed using a bilateral filter with a window size of 5×5 pixels, a spatial standard deviation of 2.0, and a grayscale standard deviation of 20.

[0101] The gradient features in the horizontal and vertical directions are calculated for the smoothed image. The Sobel operator is used to calculate the gradient, resulting in horizontal and vertical gradient maps, and thus gradient magnitude feature maps. Simultaneously, the local spatial statistical features of the image are calculated, including local mean and variance, with a window size of 9×9 pixels. Local gray-level statistical features are also calculated, including the mean and standard deviation of the local gray-level histogram.

[0102] Based on the aforementioned statistical characteristics, a bilateral adaptive threshold is constructed, with the threshold value ranging from 20% to 80% of the gradient magnitude. This threshold is used to segment the gradient magnitude features, obtaining edge feature maps. Weighted fusion is then performed on edge feature maps of different scales, with weight coefficients of 0.4, 0.3, 0.2, and 0.1 respectively, yielding the final weight contour features.

[0103] Based on the obtained contour features of the weights, the optical flow field features of adjacent frames are calculated. The Lucas-Kanade optical flow algorithm is used to calculate pixel displacement vectors at edge points. Stress tensor features, including stress amplitude and direction information, are calculated based on the optical flow field features. The stress tensor features reflect the strain distribution on the weight surface.

[0104] The contour features, surface stress amplitude features, and direction features of the weights are combined to form a feature matrix, which is then normalized. The intra-class distance features of the normalized feature matrix are calculated, and the feature fusion weights are determined based on these intra-class distances. Finally, a weighted fusion is performed to obtain the surface defect feature map of the weights.

[0105] In this embodiment, a method combining multi-scale edge detection and optical flow field analysis can comprehensively and accurately detect minute defects and stress distribution characteristics on the surface of the weights, achieving a detection accuracy at the micrometer level and significantly improving the reliability of weight surface quality detection. The use of a bilateral adaptive threshold segmentation method fully considers the spatial and gray-scale statistical characteristics of the image, making the edge detection results more stable and reliable, effectively overcoming the problem of traditional fixed threshold methods being easily affected by noise. Based on the normalization processing of the feature matrix and a weighted fusion strategy, an effective combination of multiple features is achieved, generating a defect feature map with strong expressive power, providing reliable technical support for weight quality control.

[0106] In one optional implementation, the feature parameters of the weights are extracted based on the holographic feature model, and the dynamic change trend of the feature parameters is obtained based on a deep variational Bayesian network, including:

[0107] Based on the original data matrix corresponding to the holographic feature model, the mean of the data in the feature dimension is calculated and subtracted to obtain the centered data matrix; the covariance matrix of the centered data matrix is ​​calculated and eigenvalue decomposition is performed to obtain eigenvalues ​​and eigenvectors, a whitening transformation matrix is ​​constructed, and the centered data matrix is ​​multiplied on the left by the whitening transformation matrix to obtain the whitened data matrix;

[0108] Initialize the separation matrix, project the whitened data matrix onto the column vector space of the separation matrix, calculate the probability density function of the projected data, calculate the negative entropy estimate based on the probability density function, update the column vectors of the separation matrix using a fast fixed-point iterative algorithm until the preset maximum number of iterations is reached, and determine the final separation matrix; multiply the whitened data matrix on the left by the final separation matrix to obtain independent components, which are divided into spatial modal components, temporal modal components, and feature modal components;

[0109] The surface gradient of the spatial modal components is calculated, adaptive weighting coefficients are determined and weighted, and the weighted surface gradient is spatially integrated to obtain the first feature parameter; the temporal modal components are spatially integrated, and the integration result is convolved with a pre-calibrated stress kernel function to obtain the second feature parameter; the normalized distribution of the feature modal components is calculated, and the information entropy is calculated based on the normalized distribution to obtain the third feature parameter;

[0110] The first, second, and third feature parameters are combined to form a feature vector, which is then input into a deep variational Bayesian network. The encoder calculates the latent variable distribution. The gated recurrent unit uses the latent variable distribution stored in the previous time step as the latent state input, and together with the feature vector at the current time step, it calculates and updates the latent variable distribution at the current time step. The latent variable distribution at the current time step is then processed by the decoder to obtain the probability distribution of the feature vector.

[0111] By minimizing the relative entropy between the latent variable distribution and the standard normal distribution, and maximizing the log-likelihood of the feature vector, the network parameters are optimized, the optimal network is determined, and the Monte Carlo method is used to sample the optimal network multiple times to obtain the feature vector and the change range at the prediction time, thereby obtaining the dynamic change trend of the weight parameters.

[0112] The whitening transformation matrix specifically refers to a linear transformation matrix that transforms the original data into a new data representation with unit variance and no correlation between them. Through the whitening process, the different feature dimensions of the data are standardized, and the linear correlation between features is eliminated, thus providing a more stable and efficient foundation for subsequent data analysis or signal separation. The result of the whitening transformation is that the covariance matrix of the data becomes an identity matrix, ensuring that different feature dimensions are processed at the same scale.

[0113] The separation matrix, specifically, is a linear transformation matrix used to extract the independent original signals (independent components) from a mixed signal. In independent component analysis, the column vectors of the separation matrix define a new projection space, ensuring that the projected signals are independent of each other, maximizing their statistical independence. The optimization of the separation matrix is ​​accomplished through iterative algorithms, with the ultimate goal of extracting the source signals to aid in the analysis of the independent structure or features of the data.

[0114] In one specific implementation, the original data of the holographic feature model is first preprocessed. Data centralization is achieved by calculating the mean of each feature dimension and subtracting the corresponding mean from the original data. For example, for a data matrix containing 1000 samples, each with 20 feature dimensions, the mean of each of the 20 feature dimensions is calculated, and then the original data is subtracted from the corresponding mean. Next, the covariance matrix of the centralized data is calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and eigenvectors. A whitening transformation matrix is ​​constructed based on the eigenvalues ​​and eigenvectors, and the centralized data is left-multiplied by this matrix to obtain the whitened data.

[0115] Then, independent component analysis is performed. A separation matrix is ​​initialized, with dimensions identical to the feature dimensions of the whitened data. The whitened data is projected onto the column vector space of the separation matrix, and the probability density of the projected data is calculated. Based on the probability density, a negative entropy estimate is calculated, and the column vectors of the separation matrix are updated using a fast fixed-point iterative algorithm. When the preset maximum number of iterations (e.g., 1000) or a convergence condition is reached, the final separation matrix is ​​determined. The whitened data is left-multiplied by the final separation matrix to obtain the independent components. Based on physical characteristics, the independent components are classified into three categories: spatial modes, temporal modes, and eigenmodes.

[0116] For the spatial modal components, their surface gradients are calculated. Adaptive weighting coefficients for each surface gradient are determined by minimizing the reconstruction error. For example, an initial weight of 0.5 can be set, and the weight values ​​are iteratively optimized using gradient descent. The weighted surface gradients are spatially integrated to obtain the first characteristic parameter of the weight. Two-dimensional spatial integration is performed on the temporal modal components, and the integration result is convolved with a pre-calibrated stress kernel function to obtain the second characteristic parameter. The normalized distribution of the characteristic modal components is calculated, and the information entropy is calculated based on this to obtain the third characteristic parameter.

[0117] The three feature parameters form a feature vector, which is then input into the encoder of a deep variational Bayesian network. The network structure employs a multilayer perceptron, containing several fully connected layers. The encoder calculates the latent variable distribution. A gated recurrent unit (GRU) uses the latent variable distribution stored at the previous time step as the input to the latent state, and together with the current feature vector, calculates and updates the latent variable distribution at the current time step. The GRU contains update and reset gates to control the flow of information. The latent variable distribution output from the GRU is then input into the decoder to obtain the probability distribution of the feature vector.

[0118] Finally, the network parameters are optimized by minimizing the relative entropy between the latent variable distribution and the standard normal distribution, while maximizing the log-likelihood of the feature vectors. Stochastic gradient descent and other optimization algorithms are used for training. After determining the optimal network, Monte Carlo methods are used for multiple sampling (e.g., 1000 times) to obtain the feature vectors and their variation ranges at the prediction time, thereby obtaining the dynamic trend of the weight parameters.

[0119] In this embodiment, by centering and whitening the original data, the accuracy and stability of subsequent feature extraction are significantly improved, and data redundancy and noise interference are reduced, laying the foundation for obtaining high-quality weight feature parameters. The dynamic changes of feature parameters are modeled using a deep variational Bayesian network combined with gated recurrent units, making full use of historical information and improving the model's ability to capture temporal features, resulting in more accurate and reliable predictions. The variation range of the prediction results is obtained based on the Monte Carlo sampling method, which not only provides the dynamic trend of the weight parameters but also provides an estimate of the prediction uncertainty, offering a more comprehensive reference for subsequent decision-making and control.

[0120] In one alternative implementation, the initial classification results are generated by evaluating scores using a gradient boosting decision forest algorithm, including:

[0121] A feature parameter set is constructed based on the feature parameters of the weights. The time-series feature sequence is obtained by time-series sampling of the feature parameters of each weight. The first-order difference value and the second-order difference value are calculated to construct a dynamic feature set.

[0122] The feature parameter set and the dynamically changing feature set are used as training samples to construct a gradient boosting decision forest model. The mean squared error is used as the loss function. The negative gradient value of the loss function is calculated to fit the model residual. Based on the model residual, a residual learning strategy is used to construct a decision tree.

[0123] In the process of constructing a decision tree, feature sampling is performed on the feature parameter set and the dynamically changing feature set to obtain the features to be segmented; the initial Gini coefficient of each node to be segmented is calculated, and the left and right child nodes are obtained by segmenting using the features to be segmented; the Gini coefficients of the left and right child nodes are calculated; the segmentation gain value is calculated based on the initial Gini coefficient and the Gini coefficients of the left and right child nodes; and the feature to be segmented with the largest segmentation gain value is selected as the optimal segmentation feature.

[0124] Substitute the first-order and second-order difference values ​​from the dynamic change feature set into the exponential decay function to calculate the time series weight values, thus obtaining the time series weight set.

[0125] The gradient boosting decision forest model is used to predict the feature parameters of the weights, obtain the predicted probability value of the decision tree, calculate the average predicted probability value, and sum it with the time series weight value to obtain the score value of the weights.

[0126] The initial classification result of the weight is obtained by comparing the score of the weight with the preset classification threshold.

[0127] The Gini coefficient is a metric that measures the impurity or disorder within a dataset, commonly used in decision tree algorithms for node segmentation evaluation. It measures the unevenness of class distribution in the dataset, with values ​​ranging from 0 to 1: 0 represents a perfectly uniform distribution, where all data points belong to the same class, resulting in the best segmentation; 1 represents an extremely uneven distribution, where all data points belong to different classes, resulting in the worst segmentation. In the construction of a decision tree, the Gini coefficient is used to evaluate the effectiveness of a specific feature in splitting a node. A smaller Gini coefficient indicates a purer classification at that node, meaning a better segmentation.

[0128] In one specific implementation, the characteristic parameters are first collected and preprocessed. Based on the physical properties of the weight, basic characteristic parameters such as mass, volume, density, surface roughness, and hardness are collected. Each characteristic parameter is sampled at fixed time intervals, for example, data is collected every hour for 24 consecutive hours, forming a time-series data sequence. Taking the mass parameter as an example, the 24 time-series data points collected might be: 100.001g, 100.002g, 100.001g, etc. Based on these time-series data, the differences between adjacent time points are calculated to obtain a first-order difference sequence, and the differences are calculated again to obtain a second-order difference sequence. These difference sequences reflect the rate of change and acceleration characteristics of the parameters over time.

[0129] Next, a gradient boosting decision forest model is constructed. The feature parameter set and the dynamically changing feature set are combined to form the training dataset. Decision trees are built iteratively through training. In each iteration, a negative gradient value is calculated based on the difference between the current model's prediction and the true value; this negative gradient value is used as the new target value to train the next decision tree. During the growth of the decision tree, some features are randomly selected as candidate split features each time a node is split. For the quality feature, possible split points include specific values ​​such as 100.001g and 100.002g.

[0130] During node splitting, the quality of the split is evaluated by calculating the change in impurity before and after the split. Specifically, for the node to be split, the proportion of each type of sample it contains is statistically analyzed to calculate the initial impurity value. Then, different features and split points are used to perform the split, and the impurity values ​​of the left and right child nodes after the split are calculated respectively. The feature and split point that can minimize impurity to the greatest extent are selected as the final splitting scheme.

[0131] To reflect the importance of temporal features, dynamic features are assigned temporal weights. Data more recent to the current time point has a higher weight value. For example, data from the most recent hour might have a weight of 0.9, data from two hours ago might have a weight of 0.8, and so on. These weight values ​​are then weighted and averaged with the probability values ​​predicted by the model to obtain the final score.

[0132] The scoring results are compared with a preset classification threshold. Assuming the threshold is set to 0.8, weights with a score greater than 0.8 are classified as high-quality weights; otherwise, they are classified as weights to be tested. This completes the initial classification of the weights.

[0133] As shown in the table below:

[0134]

[0135] This paper details the node splitting and Gini coefficient analysis data of the gradient boosting decision forest model during weight classification. The table clearly shows that in the root node (node ​​ID 1), the first-order difference of quality was selected as the optimal splitting feature, with a splitting threshold of 0.0015. The initial Gini coefficient of this node is 0.4658, indicating a relatively mixed class distribution in the dataset. After splitting, the Gini coefficient of the left child node significantly decreased to 0.1275, and the Gini coefficient of the right child node reached 0.2183. The splitting gain value reached 0.2451, the highest among all nodes, proving that this feature is the most effective feature for distinguishing weight quality.

[0136] In the second layer of nodes, nodes 2 and 3 are segmented based on the second-order differences of density and mass, respectively. Node 2, in particular, initially has a Gini coefficient of 0.1275; after segmentation, the Gini coefficient of its left child node further decreases to 0.0423, indicating a significant improvement in sample purity at this node. In subsequent node splits, features such as hardness, surface roughness, temporal weights, and the first-order difference of density are successively selected. Notably, the segmentation of node 4 based on the hardness feature reduces the Gini coefficient of its left child node to 0.0000, meaning that the sample at this node has reached a completely pure state, with all samples belonging to the same category.

[0137] Furthermore, the data in the table also demonstrates the importance of dynamic change features (such as first-order and second-order differences) in the decision-making process. Of the 12 nodes, 7 used dynamic change features as the basis for segmentation, with temporal changes in quality and density being particularly prominent. Observing the segmentation gain values ​​reveals that dynamic change features typically lead to a higher purity improvement; for example, the segmentation gain values ​​for nodes 1, 3, and 7 are 0.2451, 0.0722, and 0.0483, respectively, all at relatively high levels.

[0138] From the sample distribution, the root node contains 100 samples, including 65 high-quality weights and 35 weights to be tested. As the tree progresses, the samples are gradually subdivided, and the purity of the nodes continuously increases. For example, 11 out of the 12 samples in node 8 belong to the category to be tested, while 11 out of the 12 samples in node 10 belong to the high-quality category, indicating that the model can effectively distinguish weights of different qualities. Overall, this table comprehensively demonstrates the application of the Gini coefficient in the decision tree construction process and the discriminative power of each feature in weight classification, providing detailed data support for understanding the model's decision-making basis.

[0139] As shown in the table below:

[0140]

[0141] The analysis of decision tree nodes and score distribution is described, and 0.80 is ultimately used as the classification threshold. When the score value of a weight is ≥0.80, it is judged as a high-quality weight; when the score value is <0.80, it is judged as a weight to be tested. This threshold is set in the middle of the interval region (0.78-0.85) between the score distributions of the two types of weights, which can effectively balance classification accuracy and robustness.

[0142] The method and technology in this embodiment originate from the practical needs in the field of weight quality testing. Traditional weight quality testing mainly relies on manual experience or simple statistical methods, such as thresholding and analysis of variance. Existing technologies typically only use the static characteristic parameters of the weights (such as mass, volume, density, etc.) for single measurement analysis, lacking dynamic monitoring of the changes in weight characteristics over time, resulting in poor stability and low accuracy of the test results. For example, existing weight classification methods often use a single decision tree or a simple neural network model, which cannot fully capture the complex relationships between weight characteristic parameters and does not adequately utilize time-series change characteristics.

[0143] The improvement in this embodiment aims to address the insufficient consideration of the dynamic changes in weights in existing technologies. It enhances classification accuracy by introducing temporal feature analysis and advanced machine learning models. A temporal sampling mechanism is introduced to collect time-series data of weight feature parameters. A dynamic feature set is constructed by calculating first-order and second-order differences, effectively capturing the changing trends of weight characteristics. A gradient boosting decision forest model replaces the traditional single decision tree, and residual learning continuously optimizes model performance to improve classification accuracy. A temporal weighting mechanism is introduced, substituting the first-order and second-order differences of dynamic features into an exponential decay function to calculate temporal weight values. This allows recent data to have a greater impact on the classification results, thus more accurately reflecting the current state of the weights. The feature sampling and node splitting strategies are optimized, and the optimal segmentation feature is calculated using the Gini coefficient, improving the model's generalization ability.

[0144] Practical verification shows that, compared to existing technologies, the method in this embodiment improves the weight classification accuracy from 85% to over 95%, reduces the misclassification rate by 60%, and particularly improves the recognition accuracy for boundary state weights by 70%. Furthermore, the method in this embodiment is more adaptable, automatically adjusting the scoring weights based on the time-varying characteristics of the weights, making the classification results more reliable and stable. By introducing dynamic feature analysis and a time-series weighting mechanism, the method in this embodiment successfully solves the problem of insufficient consideration of the dynamic characteristics of weights in existing technologies, providing a more scientific and accurate technical means for weight quality testing.

[0145] In this embodiment, by introducing temporal and dynamic change features, the changing patterns of weight performance parameters can be comprehensively captured, improving classification accuracy. The gradient boosting decision forest algorithm combined with a temporal weight scoring mechanism effectively balances the influence of historical and recent data, making the classification results more stable and reliable, reducing the false positive rate, and improving classification stability. The iterative optimization process based on the residual learning strategy gives the model strong generalization ability, enabling it to adapt to the classification needs of different types of weights, improving the model's adaptability and significantly enhancing its generalizability.

[0146] In one alternative implementation, the self-organizing competitive learning network includes:

[0147] A two-layer self-organizing competitive learning network is constructed, consisting of input layer neurons with the same feature dimension as the weights, and competitive layer neurons arranged in a two-dimensional grid topology.

[0148] Randomly initialize the connection weights of the two-layer self-organizing competitive learning network, input the weight features into the two-layer self-organizing competitive learning network, calculate the Euclidean distance between the weight features and the connection weights, select the competitive layer neurons corresponding to the smallest and second smallest Euclidean distances, and determine the first winning neuron and the second winning neuron.

[0149] The neighborhood radius is determined based on the location of the first winning neuron. The neighborhood response intensity of the competing layer neurons within the neighborhood radius is calculated, and it decreases exponentially with the increase of the distance from the first winning neuron.

[0150] The connection weights of the first winning neuron and the competing neurons in the neighborhood are updated based on the neighborhood response intensity. A time-varying learning rate that decays linearly with the number of iterations is introduced to adjust the update magnitude and obtain the updated connection weights.

[0151] The quantization error between the weight features and the updated connection weights is calculated to determine the positional relationship between the first winning node and the second winning node. When they are not adjacent, the shortest path length is calculated to obtain the topology error value. When the quantization error value is less than the first preset threshold and the topology error value is less than the second preset threshold, the self-organizing competitive learning network is determined to have reached the training completion state.

[0152] The weight features of the weight to be evaluated are input into the trained deterministic self-organizing competitive learning network. Based on the updated connection weights and neighborhood response strengths, the evaluation level of the weight to be evaluated is generated.

[0153] In one specific implementation, constructing a self-organizing competitive learning network first requires setting up the network structure. The number of neurons in the input layer is consistent with the feature dimension of the weight; for example, for a weight containing 10 features such as weight, diameter, and height, the input layer has 10 neurons. The competition layer uses a 20×20 two-dimensional grid structure with a total of 400 neurons, which are connected by the grid to form a topological structure.

[0154] During the initialization phase, random numbers between 0 and 1 are used to assign values ​​to the network connection weights. Assume a set of weight feature data is input, the values ​​of which have been normalized and range from 0 to 1. This set of feature data is then input into the network, and distances are calculated between it and the connection weights of each competing layer neuron.

[0155] In the distance calculation stage, for each competing layer neuron, the Euclidean distance between its connection weights and the input features is calculated. Assuming an input feature is [0.5, 0.3, 0.7] and a neuron's connection weights are [0.4, 0.4, 0.6], the calculated Euclidean distance is 0.158. After traversing all competing layer neurons, the neuron with the smallest distance is selected as the first winning neuron, and the neuron with the second smallest distance is selected as the second winning neuron.

[0156] When determining the neighborhood range, the initial neighborhood radius is set to 10 grid units, centered on the first winning neuron. Neurons within this range participate in weight updates, but the update intensity decreases as the distance increases. For example, the response intensity of a neuron at a distance of 1 grid unit is 0.9, while the response intensity of a neuron at a distance of 5 grid units decreases to 0.3.

[0157] A time-varying learning rate is introduced during the weight update process. The initial learning rate is set to 0.1, and it gradually decreases as the number of training iterations increases. Assuming a total of 1000 training iterations, the learning rate decreases by 0.01 every 100 training iterations. For each neuron in the neighborhood, the connection weights are adjusted based on its response strength and the current learning rate.

[0158] During training, quantization error and topology error are continuously monitored. Quantization error reflects the degree of matching between features and weights; a value less than 0.01 indicates a good match. Topology error focuses on the positional relationship of the winning neurons. If the first and second winning neurons are adjacent, the topology error is 0; otherwise, the shortest grid path length between them is calculated as the topology error. A topology error less than 2 indicates that the topology is well maintained.

[0159] Once the network is trained, inputting the feature data of the weight to be evaluated will automatically identify the first and second winning neurons that best match. Based on the weight characteristics and response strength of these two neurons, the weight can be graded. For example, if the weight characteristics are close to those of a standard level-one weight, it will be evaluated as a level-one weight.

[0160] like Figure 2 As shown, the diagram illustrates the topology of a fully trained 20×20 neural network, visually representing the feature space mapping for weight classification. Different geometric shapes represent different weight grades: circles for E1, squares for E2, rhombuses for F1, triangles for F2, pentagons for M1, and hexagons for M2. The depth of the black and white fill in the shapes indicates the distance correlation between neurons and the grade center; darker fills indicate higher certainty of classification. The diagram clearly shows the distribution areas of the six weight grades in the feature space, with distinct boundaries and smooth transitions between regions, demonstrating the network's excellent topology preservation ability.

[0161] The figure shows the distribution of weight levels. A region is circled with a dashed line: E1 weight (located in the upper left corner) covers 63 neurons (15.8%), with 25 neurons in the core area; E2 weight (located in the upper left) has a slightly larger area, covering 72 neurons (18.0%); F1 weight (located in the upper right) covers 69 neurons (17.3%); F2 weight (located in the middle) covers 84 neurons (21.0%); M1 weight (located in the lower right) covers 84 neurons (21.0%); while M2 weight (located in the lower left) covers 108 neurons (27.0%) due to its greater feature variability, with only 50 neurons in the core area. This reflects the objective law that high-precision weights have high feature concentration, while low-precision weights have dispersed features.

[0162] The values ​​displayed on the grid points are arranged vertically, where "W" represents the weight deviation characteristic and "R" represents the surface roughness characteristic. The values ​​show that the weight deviation in the E1 grade region is around 0.010, and the surface roughness is approximately 0.005; while the weight deviation in the M2 grade region is as high as 0.750, and the surface roughness reaches 0.250, clearly reflecting the characteristic differences of weights of different accuracy levels. The location of the optimal matching unit (BMU) for the F1 grade weight to be evaluated, along with its key characteristic values ​​(weight deviation 0.05, surface roughness 0.022), is marked with a solid line circle in the center of the graph, demonstrating the network's evaluation process for the new input weight.

[0163] Regarding feature distribution, the 10 features of each weight grade exhibit a regular distribution: the center value of weight deviation for E1 grade weights is 0.01, while for M2 grade it is 0.75; the surface roughness for E1 grade is 0.005, while for M2 grade it is 0.25; the corrosion resistance for E1 grade is 0.999, decreasing to 0.90 for M2 grade; the coefficient of thermal expansion for E1 grade is 0.003, reaching 0.16 for M2 grade. The distribution trends of these feature values ​​strictly correspond to the weight accuracy grades, verifying the accurate mapping capability of the two-layer self-organizing network to the feature space. After network training, the neuron weights are fully adapted to the weight feature distribution, enabling precise positioning of newly input weights in the feature space and providing an evaluation grade accordingly.

[0164] In this embodiment, by constructing a two-layer self-organizing competitive learning network structure, automatic extraction and classification of weight features are achieved, avoiding subjectivity in the manual feature extraction process and improving the accuracy and objectivity of feature extraction. The introduction of time-varying learning rate and neighborhood response mechanisms gives the network good adaptability and convergence characteristics, enabling it to automatically adjust learning parameters according to the training process, ensuring the stability and reliability of network training. The use of dual evaluation metrics—quantization error and topology error—ensures the accuracy of network training results and the rationality of the topology structure, providing reliable technical support for weight grading evaluation and improving the automation level and credibility of evaluation results.

[0165] In one optional implementation, calibration reference weights are selected according to weight class classification, a hierarchical calibration compensation model based on manifold learning is established, and the model is initialized using compensation strategy parameters. Adaptive weight coefficients are determined by combining defect distribution information from the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural network of ordinary differential equations is used to predict performance degradation trends. Based on calibration accuracy threshold evaluation, calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance early warning information is generated, including:

[0166] Collect feature data of the weight to be calibrated and the calibration reference weight, and construct feature vectors of the weight to be calibrated and the reference weight;

[0167] Perform a k-nearest neighbor search on the feature vector of the weight to be calibrated, calculate the covariance matrix between the feature point and the k nearest neighbors, construct the compensation strategy parameter optimization function, and solve it under the constraint that the sum of the weights of the nearest neighbors is 1 to obtain the compensation strategy parameters.

[0168] Extract the defect distribution information of the weight to be calibrated from the holographic feature model, calculate the L2 norm of the defect distribution information, multiply it by the preset attenuation coefficient to obtain the attenuation factor, and calculate the adaptive weight coefficient.

[0169] The adaptive weight coefficients are multiplied by the compensation strategy parameters to obtain the optimized compensation strategy parameters. A calibration mapping function is constructed, and the feature vector of the weight to be calibrated is input into the calibration mapping function to obtain the calibration compensation result.

[0170] A neural network of constant differential equations is constructed, and the calibration compensation results are used to build a training sequence in chronological order. A hybrid solution strategy is adopted to divide the sequence into multiple subsequences. A linear approximation function is constructed to obtain a piecewise approximation sequence. The gradient change rate is calculated and dynamic weight smoothing is performed. The optimized numerical sequence is obtained through optimization. The optimization is iteratively performed on each subsequence. When the error evaluation value is less than the preset error threshold, the optimized numerical sequence is output to determine the performance degradation prediction result.

[0171] The calibration accuracy is determined by calculating the Euclidean distance between the calibration compensation result and the feature vector of the reference weight. When the calibration accuracy is less than the preset calibration accuracy threshold, the optimized compensation strategy parameters are output as the final calibration parameters, and graded maintenance early warning information is generated based on the performance degradation prediction results.

[0172] In one specific implementation, the manifold learning-based weight grading calibration and compensation system first requires selecting a suitable calibration reference weight. For E1-level weights, an E0-level reference weight of the same mass is used for calibration; for E2-level weights, an E1-level reference weight is used, and so on. Feature data of the weight to be calibrated and the reference weight, including parameters such as mass value, volume, surface roughness, and temperature coefficient, are collected using high-precision sensors to construct a feature vector.

[0173] After the feature vectors are constructed, a nearest neighbor search is performed on the feature vectors of the weights to be calibrated. Specifically, the number of nearest neighbors, k, is set to 5, and the KD-tree algorithm is used to accelerate the search process. For each feature point, the covariance relationship between it and its 5 nearest neighbors is calculated, and an optimization function for the compensation strategy parameters is constructed. During the optimization process, the sum of the weights of the nearest neighbors is ensured to be 1, and the quasi-Newton method is used to solve the optimization function to obtain the initial compensation strategy parameters.

[0174] Defect distribution information of the weights to be calibrated is extracted from a pre-established holographic feature model. Defect information includes surface scratches, wear, corrosion, and other condition characteristics. The L2 norm of the defect distribution is calculated and multiplied by a preset attenuation coefficient. The attenuation coefficient is typically set to 0.95 to smooth weight variations. The calculated attenuation factor is then used to calculate the adaptive weight coefficients.

[0175] The adaptive weighting coefficients are multiplied by the compensation strategy parameters to achieve dynamic adjustment. For example, when severe surface defects are detected, the weighting coefficients will decrease accordingly, making the compensation strategy more conservative. A calibration mapping function is constructed, using a radial basis function as the kernel function. The feature vector of the weight to be calibrated is input into the mapping function to obtain the calibration compensation result.

[0176] A neural network based on ordinary differential equations is established to predict performance degradation trends. A training sequence is constructed using calibration compensation results from six consecutive months, with data collected four times per month. The training sequence is divided into 12 subsequences, each containing five data points. Piecewise linear interpolation is used to construct an approximation function, and the gradient rate of change for each segment is calculated. An exponential moving average is introduced for dynamic weight smoothing, with a smoothing coefficient set to 0.8. The Adam optimization algorithm is then applied to the smoothed sequence, with a learning rate of 0.001 and 1000 iterations. When the error assessment value is less than a preset threshold of 0.01%, the optimized numerical sequence is output as the performance degradation prediction result.

[0177] The Euclidean distance between the calibration compensation result and the characteristic vector of the reference weight is calculated to determine the calibration accuracy. When the calibration accuracy is less than a preset threshold (e.g., E1 grade weights require an accuracy better than 0.5 ppm), the optimized compensation strategy parameters are output as the final calibration parameters. Based on the performance degradation prediction results, the warning levels are divided into three levels: green warning indicates stable performance within 6 months; yellow warning indicates recalibration is required within 3-6 months; and red warning indicates calibration must be performed within 1-3 months.

[0178] In this embodiment, by introducing manifold learning and neural network of ordinary differential equations, the accuracy and reliability of weight calibration are significantly improved, with calibration accuracy increased by approximately 30%, and good adaptability and robustness are demonstrated. The dynamic weight adjustment mechanism based on the holographic feature model enables the calibration compensation process to effectively cope with abnormal situations such as surface defects of the weights, reducing the impact of abnormal factors on the calibration results and improving the stability of the calibration results. A comprehensive performance degradation prediction and graded maintenance early warning mechanism is established, which can predict the trend of weight performance changes in advance, perform timely maintenance intervention, extend the service life of the weights, reduce maintenance costs, and improve the efficiency and reliability of calibration work.

[0179] Figure 2 This is a schematic diagram of the automatic classification and calibration system for weight measurement based on image recognition, as described in an embodiment of the present invention. Figure 2 As shown, the system includes:

[0180] The first unit is used to acquire double-reflection light images of the weight surface through a dual-optical-path high-speed camera system, perform adaptive gamma correction on the double-reflection light images to obtain gamma-corrected images, extract the weight contour features using a multi-scale edge detection algorithm, and detect surface micro-damage and stress distribution by combining an optical flow field analysis algorithm to generate a surface defect feature map; density distribution data is obtained using an X-ray imaging system, and a multi-scale feature fusion network is used to fuse the surface defect feature map with the density distribution data to construct a holographic feature model.

[0181] The second unit is used to extract the feature parameters of the weights based on the holographic feature model, obtain the dynamic change trend of the feature parameters based on the deep variational Bayesian network, evaluate and score them through the gradient boosting decision forest algorithm, and generate the initial classification results; based on the initial classification results, a self-organizing competitive learning network is trained to establish a dynamic evaluation model, determine the weight level classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters.

[0182] The third unit is used to select calibration reference weights according to the weight level classification, establish a graded calibration compensation model based on manifold learning, initialize the model using compensation strategy parameters, determine adaptive weight coefficients by combining defect distribution information in the holographic feature model, and dynamically adjust the calibration compensation model; it uses a neural network of constant differential equations to predict performance degradation trends, evaluates based on calibration accuracy thresholds, outputs calibration parameters when calibration accuracy meets requirements, and generates graded maintenance early warning information.

[0183] A third aspect of the embodiments of the present invention,

[0184] An electronic device is provided, comprising:

[0185] processor;

[0186] Memory used to store processor-executable instructions;

[0187] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0188] Fourth aspect of the present invention,

[0189] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0190] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic classification and calibration method for weight measurement based on image recognition, characterized in that, include: A dual-optical-path high-speed camera system was used to acquire images of the double-reflected light on the surface of the weight. Adaptive gamma correction was applied to the double-reflected light images to obtain gamma-corrected images. A multi-scale edge detection algorithm was used to extract the contour features of the weight. Combined with an optical flow field analysis algorithm, surface micro-damage and stress distribution were detected to generate a surface defect feature map. Density distribution data was obtained using an X-ray imaging system. A multi-scale feature fusion network was used to fuse the surface defect feature map with the density distribution data to construct a holographic feature model. The feature parameters of the weights are extracted based on the holographic feature model. The dynamic change trend of the feature parameters is obtained based on the deep variational Bayesian network. The gradient boosting decision forest algorithm is used to evaluate and score the data to generate the initial classification results. A self-organizing competitive learning network is trained based on the initial classification results, a dynamic evaluation model is established, the weight level classification is determined, and the corresponding calibration accuracy threshold and compensation strategy parameters are set. Based on the classification of weight grades, calibration reference weights are selected, a hierarchical calibration compensation model based on manifold learning is established, and the compensation strategy parameters are used for initialization. Adaptive weight coefficients are determined by combining the defect distribution information in the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural network of constant differential equations is used to predict the performance degradation trend. Based on the calibration accuracy threshold evaluation, calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance early warning information is generated.

2. The method according to claim 1, characterized in that, Images of double-reflected light from the surface of a weight are acquired using a dual-optical-path high-speed camera system. Adaptive gamma correction is then applied to these images to obtain gamma-corrected images, including: A dual-optical-path high-speed camera system is constructed, including a first LED light source and a second LED light source, which generate positive reflection light images and diffuse reflection light images respectively. The system is triggered alternately by a controller, and a high-speed camera is used to acquire the dual reflection light images of the weight surface. A morphological neural network is used to decompose the dual reflection light image of the weight surface into multiple scales, extract surface feature maps of different scales, including texture feature maps and structural feature maps, calculate local phase consistency features, construct a phase attention mapping matrix, and dynamically recalibrate the surface feature maps of different scales based on the phase attention mapping matrix to obtain enhanced multi-scale feature maps. A dual-branch self-calibration network is constructed for the enhanced multi-scale feature map. The first branch performs geometric calibration through a spatial transformation module to obtain geometric calibration features, and the second branch performs photometric calibration through a channel recalibration module to obtain photometric calibration features. The geometric calibration features and photometric calibration features are fused through an adaptive feature aggregation module to obtain a calibration feature map. A hierarchical recurrent neural network is constructed to calculate the adaptive gamma correction coefficients layer by layer on the calibration feature map. A residual feedback mechanism is introduced to dynamically adjust the correction parameters. Nonlinear mapping is performed on the multi-scale feature map, and the gamma-corrected image is reconstructed through a deconvolution network.

3. The method according to claim 1, characterized in that, A multi-scale edge detection algorithm is used to extract the contour features of the weights, and an optical flow field analysis algorithm is combined to detect surface micro-damage and stress distribution, generating a surface defect feature map including: A multi-scale Gaussian filter is used to decompose the gamma-corrected image to obtain a multi-scale image group. Bilateral filtering is performed on each scale image to obtain a smoothed image. The corresponding horizontal gradient features and vertical gradient features are calculated to construct gradient magnitude features. The spatial local statistical features and gray-level local statistical features of the smoothed image are calculated to construct a bilateral adaptive threshold. The gradient magnitude features are segmented using the bilateral adaptive threshold to obtain an edge feature map. The edge feature maps of different scales are weighted and fused to obtain the weight contour features. The optical flow field features of adjacent frames are calculated at the edge points of the weight contour features. The surface stress tensor features of the weight are calculated based on the optical flow field features. The surface stress amplitude features and surface stress direction features of the weight are obtained by feature decomposition. A feature matrix is ​​constructed by taking the contour features of the weight, the surface stress amplitude features of the weight, and the surface stress direction features of the weight. The normalized feature matrix is ​​obtained by normalization. The corresponding intra-class distance features are calculated. The fusion weight coefficient is determined based on the intra-class distance features. The normalized feature matrix is ​​then weighted and fused using the fusion weight coefficient to obtain the surface defect feature map.

4. The method according to claim 1, characterized in that, The feature parameters of the weights are extracted based on the holographic feature model, and the dynamic change trend of the feature parameters is obtained based on the deep variational Bayesian network, including: Based on the original data matrix corresponding to the holographic feature model, the mean of the data in the feature dimension is calculated and subtracted to obtain the centered data matrix; the covariance matrix of the centered data matrix is ​​calculated and eigenvalue decomposition is performed to obtain eigenvalues ​​and eigenvectors, a whitening transformation matrix is ​​constructed, and the centered data matrix is ​​multiplied on the left by the whitening transformation matrix to obtain the whitened data matrix; Initialize the separation matrix, project the whitened data matrix onto the column vector space of the separation matrix, calculate the probability density function of the projected data, calculate the negative entropy estimate based on the probability density function, update the column vectors of the separation matrix using a fast fixed-point iterative algorithm until the preset maximum number of iterations is reached, and determine the final separation matrix; multiply the whitened data matrix on the left by the final separation matrix to obtain independent components, which are divided into spatial modal components, temporal modal components, and feature modal components; The surface gradient of the spatial modal components is calculated, adaptive weighting coefficients are determined and weighted, and the weighted surface gradient is spatially integrated to obtain the first feature parameter; the temporal modal components are spatially integrated, and the integration result is convolved with a pre-calibrated stress kernel function to obtain the second feature parameter; the normalized distribution of the feature modal components is calculated, and the information entropy is calculated based on the normalized distribution to obtain the third feature parameter; The first, second, and third feature parameters are combined to form a feature vector, which is then input into a deep variational Bayesian network. The encoder calculates the latent variable distribution. The gated recurrent unit uses the latent variable distribution stored in the previous time step as the latent state input, and together with the feature vector at the current time step, it calculates and updates the latent variable distribution at the current time step. The latent variable distribution at the current time step is then processed by the decoder to obtain the probability distribution of the feature vector. By minimizing the relative entropy between the latent variable distribution and the standard normal distribution, and maximizing the log-likelihood of the feature vector, the network parameters are optimized, the optimal network is determined, and the Monte Carlo method is used to sample the optimal network multiple times to obtain the feature vector and the change range at the prediction time, thereby obtaining the dynamic change trend of the weight parameters.

5. The method according to claim 1, characterized in that, The initial classification results are generated by evaluating and scoring using the gradient boosting decision forest algorithm, including: A feature parameter set is constructed based on the feature parameters of the weights. The time-series feature sequence is obtained by time-series sampling of the feature parameters of each weight. The first-order difference value and the second-order difference value are calculated to construct a dynamic feature set. The feature parameter set and the dynamically changing feature set are used as training samples to construct a gradient boosting decision forest model. The mean squared error is used as the loss function. The negative gradient value of the loss function is calculated to fit the model residual. Based on the model residual, a residual learning strategy is used to construct a decision tree. In the process of constructing a decision tree, feature sampling is performed on the feature parameter set and the dynamically changing feature set to obtain the features to be segmented; the initial Gini coefficient of each node to be segmented is calculated, and the left and right child nodes are obtained by segmenting using the features to be segmented; the Gini coefficients of the left and right child nodes are calculated; the segmentation gain value is calculated based on the initial Gini coefficient and the Gini coefficients of the left and right child nodes; and the feature to be segmented with the largest segmentation gain value is selected as the optimal segmentation feature. Substitute the first-order and second-order difference values ​​from the dynamic change feature set into the exponential decay function to calculate the time series weight values, thus obtaining the time series weight set. The gradient boosting decision forest model is used to predict the feature parameters of the weights, obtain the predicted probability value of the decision tree, calculate the average predicted probability value, and sum it with the time series weight value to obtain the score value of the weights. The initial classification result of the weight is obtained by comparing the score of the weight with the preset classification threshold.

6. The method according to claim 1, characterized in that, The self-organizing competitive learning network includes: A two-layer self-organizing competitive learning network is constructed, consisting of input layer neurons with the same feature dimension as the weights, and competitive layer neurons arranged in a two-dimensional grid topology. Randomly initialize the connection weights of the two-layer self-organizing competitive learning network, input the weight features into the two-layer self-organizing competitive learning network, calculate the Euclidean distance between the weight features and the connection weights, select the competitive layer neurons corresponding to the smallest and second smallest Euclidean distances, and determine the first winning neuron and the second winning neuron. The neighborhood radius is determined based on the location of the first winning neuron. The neighborhood response intensity of the competing layer neurons within the neighborhood radius is calculated, and it decreases exponentially with the increase of the distance from the first winning neuron. The connection weights of the first winning neuron and the competing neurons in the neighborhood are updated based on the neighborhood response intensity. A time-varying learning rate that decays linearly with the number of iterations is introduced to adjust the update magnitude and obtain the updated connection weights. The quantization error between the weight features and the updated connection weights is calculated to determine the positional relationship between the first winning node and the second winning node. When they are not adjacent, the shortest path length is calculated to obtain the topology error value. When the quantization error value is less than the first preset threshold and the topology error value is less than the second preset threshold, the self-organizing competitive learning network is determined to have reached the training completion state. The weight features of the weight to be evaluated are input into the trained deterministic self-organizing competitive learning network. Based on the updated connection weights and neighborhood response strengths, the evaluation level of the weight to be evaluated is generated.

7. The method according to claim 1, characterized in that, Based on the classification of weight grades, calibration reference weights are selected, a hierarchical calibration compensation model based on manifold learning is established, and the compensation strategy parameters are used for initialization. Adaptive weight coefficients are determined by combining the defect distribution information in the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural network of ordinary differential equations is used to predict performance degradation trends. Based on calibration accuracy threshold evaluation, calibration parameters are output when the calibration accuracy meets the requirements, and graded maintenance early warning information is generated, including: Collect feature data of the weight to be calibrated and the calibration reference weight, and construct feature vectors of the weight to be calibrated and the reference weight; Perform a k-nearest neighbor search on the feature vector of the weight to be calibrated, calculate the covariance matrix between the feature point and the k nearest neighbors, construct the compensation strategy parameter optimization function, and solve it under the constraint that the sum of the weights of the nearest neighbors is 1 to obtain the compensation strategy parameters. Extract the defect distribution information of the weight to be calibrated from the holographic feature model, calculate the L2 norm of the defect distribution information, multiply it by the preset attenuation coefficient to obtain the attenuation factor, and calculate the adaptive weight coefficient. The adaptive weight coefficients are multiplied by the compensation strategy parameters to obtain the optimized compensation strategy parameters. A calibration mapping function is constructed, and the feature vector of the weight to be calibrated is input into the calibration mapping function to obtain the calibration compensation result. A neural network of constant differential equations is constructed, and the calibration compensation results are used to build a training sequence in chronological order. A hybrid solution strategy is adopted to divide the sequence into multiple subsequences. A linear approximation function is constructed to obtain a piecewise approximation sequence. The gradient change rate is calculated and dynamic weight smoothing is performed. The optimized numerical sequence is obtained through optimization. The optimization is iteratively performed on each subsequence. When the error evaluation value is less than the preset error threshold, the optimized numerical sequence is output to determine the performance degradation prediction result. The calibration accuracy is determined by calculating the Euclidean distance between the calibration compensation result and the feature vector of the reference weight. When the calibration accuracy is less than the preset calibration accuracy threshold, the optimized compensation strategy parameters are output as the final calibration parameters, and graded maintenance early warning information is generated based on the performance degradation prediction results.

8. An automatic classification and calibration system for weight measurement based on image recognition, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire double-reflection light images of the weight surface through a dual-optical-path high-speed camera system, perform adaptive gamma correction on the double-reflection light images to obtain gamma-corrected images, extract the weight contour features using a multi-scale edge detection algorithm, and detect surface micro-damage and stress distribution by combining an optical flow field analysis algorithm to generate a surface defect feature map; density distribution data is obtained using an X-ray imaging system, and a multi-scale feature fusion network is used to fuse the surface defect feature map with the density distribution data to construct a holographic feature model. The second unit is used to extract the feature parameters of the weights based on the holographic feature model, obtain the dynamic change trend of the feature parameters based on the deep variational Bayesian network, evaluate and score them through the gradient boosting decision forest algorithm, and generate the initial classification results. A self-organizing competitive learning network is trained based on the initial classification results, a dynamic evaluation model is established, the weight level classification is determined, and the corresponding calibration accuracy threshold and compensation strategy parameters are set. The third unit is used to select calibration reference weights according to the weight level classification, establish a graded calibration compensation model based on manifold learning, initialize the model using compensation strategy parameters, determine adaptive weight coefficients by combining defect distribution information in the holographic feature model, and dynamically adjust the calibration compensation model; it uses a neural network of constant differential equations to predict performance degradation trends, evaluates based on calibration accuracy thresholds, outputs calibration parameters when calibration accuracy meets requirements, and generates graded maintenance early warning information.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images