Weight metering automatic classification and calibration method and system based on image recognition

Abstract

The invention provides a weight metering automatic classification and calibration method and system based on image recognition, and relates to the technical field of weight metering, and the method comprises the steps: carrying out the imaging of the surface of a weight through dual-light-path high-speed camera shooting, carrying out the gamma correction, and extracting contour features and surface defect features; internal density distribution is obtained through X-ray imaging; performing multi-scale feature fusion on the surface features and the density distribution to construct a holographic feature model; analyzing the dynamic change trend of the characteristic parameters based on a deep variational Bayesian network, and performing evaluation and scoring in combination with a gradient boosting decision forest to obtain an initial classification; establishing a dynamic evaluation model by adopting a self-organizing competitive learning network, determining a weight grade and setting calibration parameters; selecting a corresponding reference weight to establish a grading calibration compensation model, and determining an adaptive weight coefficient for dynamic adjustment by combining defect distribution; and predicting a performance degradation trend by adopting a Shenchang differential equation network, and outputting calibration parameters and generating early warning information when the calibration precision meets a threshold value requirement.

Description

Technical Field

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

[0002] Weights are important standard instruments in the field of metrology and testing. Their quality, accuracy, and stability directly impact the reliability of measurement results. Traditional weight testing relies primarily on manual visual inspection and mechanical balance measurement, which makes it difficult to accurately identify minor surface defects and internal structural changes. With the advancement of metrology technology, machine vision-based weight testing methods have gradually emerged. Using image processing technology, they can automatically identify and analyze weight surface features, providing a new technical means for weight quality assessment.

[0003] However, existing weight detection technology still has several shortcomings. The existing technology mainly focuses on the single detection of weight surface features, lacks effective analysis of internal structure, and cannot comprehensively evaluate the overall quality status of the weight; traditional image recognition algorithms have insufficient accuracy in extracting micro-damage and stress distribution characteristics on the weight surface, and are easily affected by factors such as ambient light; existing weight classification methods often adopt static evaluation modes, which cannot effectively track and predict the dynamic changes in weight performance; calibration compensation strategies lack an adaptive adjustment mechanism, making it difficult to achieve accurate calibration for weights of different grades.

[0004] In summary, a dynamic calibration and early warning method for weights based on multi-sensor fusion is proposed. This method combines dual-path high-speed videography with microfocus X-ray imaging to achieve comprehensive acquisition of weight surface and internal features. Advanced algorithms such as deep learning and manifold learning are used to construct a holographic feature model and dynamic assessment model for accurate weight quality classification. Furthermore, a neural ordinary differential equation network is used to predict performance degradation trends, establishing an adaptive calibration compensation mechanism to ensure that calibration accuracy meets the requirements for the use of weights of different grades. This invention can address the problems of the prior art. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for automatic classification and calibration of weight measurement based on image recognition, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provided is a method for automatic classification and calibration of weight measurement based on image recognition, comprising:

[0008] A dual-optical high-speed camera system is used to capture dual-reflected light images on the weight surface. Adaptive gamma correction is performed on the dual-reflected light images to obtain gamma-corrected images. A multi-scale edge detection algorithm is used to extract the weight contour features. Surface micro-damage and stress distribution are detected in combination with an optical flow field analysis algorithm to generate a surface defect feature map. Density distribution data is obtained using an X-ray imaging system. 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.

[0009] The weight characteristic parameters are extracted based on the holographic feature model, and the dynamic change trend of the characteristic parameters is obtained based on the deep variational Bayesian network. The gradient boosting decision forest algorithm is used to evaluate and score 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 grade classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters.

[0010] Calibration reference weights are selected according to the weight grade classification, and a hierarchical calibration compensation model based on manifold learning is established. The compensation strategy parameters are used for initialization, and the adaptive weight coefficients are determined in combination with the defect distribution information in the holographic feature model to dynamically adjust the calibration compensation model. A neural ordinary differential equation network is used to predict the performance degradation trend. Based on the calibration accuracy threshold evaluation, the calibration parameters are output when the calibration accuracy meets the requirements, and graded maintenance warning information is generated.

[0011] In an optional embodiment,

[0012] The dual-light path high-speed camera system is used to collect the double-reflected light images on the weight surface, and the double-reflected light images are adaptively gamma corrected to obtain the gamma-corrected images including:

[0013] A dual-light-path high-speed camera system was constructed, including a first LED light source and a second LED light source, which respectively generate regular reflection light images and diffuse reflection light images. The dual-reflection light images of the weight surface were acquired using a high-speed camera through alternating triggering of the controller.

[0014] A morphological neural network is used to perform multi-scale decomposition on the double-reflected light image of the weight surface, and surface feature maps of different scales are extracted, including texture feature maps and structural feature maps. Local phase consistency features are calculated, and a phase attention mapping matrix is constructed. The surface feature maps of different scales are dynamically recalibrated based on the phase attention mapping matrix to obtain an enhanced multi-scale feature map.

[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 recursive neural network is constructed to calculate the adaptive gamma correction coefficients of the calibration feature map layer by layer. 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 an optional embodiment,

[0018] A multi-scale edge detection algorithm is used to extract the weight contour features, and an optical flow field analysis algorithm is used to detect surface micro-damage and stress distribution. The surface defect feature map is generated, including:

[0019] A multi-scale Gaussian filter is used to decompose the gamma-corrected image to obtain a multi-scale image group. Each scale image is bilaterally filtered to obtain a smoothed image. The corresponding horizontal gradient features and vertical gradient features are calculated to construct the gradient amplitude features. The spatial local statistical features and grayscale local statistical features of the smoothed image are calculated to construct a bilateral adaptive threshold. The bilateral adaptive threshold is used to segment the gradient amplitude features to obtain an edge feature map, and the edge feature maps of different scales are weightedly fused to obtain the weight contour feature.

[0020] Calculate the optical flow field features of adjacent frame images at the edge point position of the weight contour feature, calculate the weight surface stress tensor features based on the optical flow field features, and decompose the features to obtain the weight surface stress amplitude features and weight surface stress direction features;

[0021] The weight contour features, weight surface stress amplitude features and weight surface stress direction features are used to construct a feature matrix, which is normalized to obtain a normalized feature matrix. The corresponding intra-class distance features are calculated, and the fusion weight coefficient is determined based on the intra-class distance features. The normalized feature matrix is weightedly fused using the fusion weight coefficient to obtain a surface defect feature map.

[0022] In an optional embodiment,

[0023] The weight characteristic parameters are extracted based on the holographic characteristic model, and the dynamic change trends of the characteristic parameters are obtained based on the deep variational Bayesian network, including:

[0024] Based on the original data matrix corresponding to the holographic feature model, the data mean of the feature dimension is calculated and subtracted to obtain the centralized data matrix; the covariance matrix of the centralized data matrix is calculated and eigenvalue decomposition is performed to obtain the eigenvalues and eigenvectors, and a whitening transformation matrix is constructed. The centralized 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, and use a fast fixed point iterative algorithm to update the column vectors of the separation matrix until a preset maximum number of iterations is reached to determine the final separation matrix; multiply the whitened data matrix by the final separation matrix to obtain independent components, and divide them into spatial modal components, temporal modal components, and eigenmodal components;

[0026] Calculate the surface gradient of the spatial modal component, determine the adaptive weight coefficient and perform weighted calculation, and spatially integrate the weighted surface gradient to obtain the first characteristic parameter; perform a two-dimensional spatial integration on the time modal component, and convolve the integration result with a pre-calibrated stress kernel function to obtain the second characteristic parameter; calculate the normalized distribution of the characteristic modal component, and calculate the information entropy based on the normalized distribution to obtain the third characteristic parameter;

[0027] The first, second, and third feature parameters are combined into a feature vector, which is input into a deep variational Bayesian network. The encoder calculates the latent variable distribution. The gated recurrent unit uses the latent variable distribution stored at the previous moment as the hidden state input, and calculates and updates it together with the feature vector at the current moment to obtain the latent variable distribution at the current moment. The latent variable distribution at the current moment is passed through 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 value of the eigenvector, the network parameters are optimized and the optimal network is determined. The Monte Carlo method is used to sample the optimal network multiple times to obtain the eigenvector and change interval at the prediction moment, and the dynamic change trend of the weight parameters is obtained.

[0029] In an optional embodiment,

[0030] The gradient boosting decision forest algorithm is used to evaluate and score the initial classification results, including:

[0031] A characteristic parameter set is constructed based on the characteristic parameters of the weights, and a time series characteristic sequence is obtained by sampling the characteristic parameters of each weight. The first-order difference value and the second-order difference value are calculated to construct a dynamic change characteristic 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 square error is used as the loss function, and 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 building 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 features to be segmented are used to segment the left and right child nodes, and 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 difference value and the second-order difference value in the dynamic change feature set into the exponential decay function to calculate the time series weight value and obtain the time series weight set;

[0035] The gradient boosting decision forest model is used to predict the weight feature parameters, and the prediction probability value of the decision tree is obtained. The average prediction probability value is calculated and weighted summed with the time series weight value to obtain the weight score value;

[0036] Compare the weight's score with the preset classification threshold to obtain the initial classification result of the weight.

[0037] In an optional embodiment,

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

[0039] Construct a two-layer self-organizing competitive learning network, including input layer neurons with the same dimension as the weight feature, and competition layer neurons arranged in a two-dimensional grid topology;

[0040] The connection weights of the two-layer self-organizing competitive learning network are randomly initialized, the weight features are input into the two-layer self-organizing competitive learning network, the Euclidean distance between the weight features and the connection weights is calculated, and the competition layer neurons corresponding to the minimum Euclidean distance and the second minimum Euclidean distance are selected to determine the first winning neuron and the second winning neuron;

[0041] The neighborhood radius is determined based on the position of the first winning neuron, and the neighborhood response strength of the competitive layer neurons within the neighborhood radius is calculated, which decays exponentially with the distance from the first winning neuron.

[0042] Based on the neighborhood response strength, the connection weights of the first winning neuron and the competition layer neurons in the neighborhood are updated. A time-varying learning rate that decays linearly with the number of iterations is introduced to adjust the update amplitude to obtain the updated connection weights.

[0043] Calculate the quantization error value between the weight feature and the updated connection weight, determine the positional relationship between the first winning node and the second winning node, and when they are not adjacent, calculate the shortest path length to obtain the topological error value; when the quantization error value is less than a first preset threshold and the topological error value is less than a second preset threshold, determine that the self-organizing competitive learning network has reached a training completion state;

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

[0045] In an optional embodiment,

[0046] Calibration reference weights are selected according to weight grade classification, and a hierarchical calibration compensation model based on manifold learning is established. The compensation strategy parameters are initialized, and the adaptive weight coefficients are determined in combination with the defect distribution information in the holographic feature model. The calibration compensation model is dynamically adjusted. A neural ordinary differential equation network is used to predict performance degradation trends. Based on the calibration accuracy threshold assessment, the calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance warning information is generated, including:

[0047] Collecting characteristic data of the weight to be calibrated and the calibration reference weight, and constructing a characteristic vector of the weight to be calibrated and a characteristic vector of the reference weight;

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

[0049] Extracting the defect distribution information of the weight to be calibrated from the holographic feature model, calculating the second norm of the defect distribution information, and multiplying it with the preset attenuation coefficient to obtain the attenuation factor, and calculating the adaptive weight coefficient;

[0050] Multiplying the adaptive weight coefficient and the compensation strategy parameter to obtain the optimized compensation strategy parameter, constructing a calibration mapping function, inputting the characteristic vector of the weight to be calibrated into the calibration mapping function, and obtaining a calibration compensation result;

[0051] A neural network of ordinary differential equations is constructed, and the calibration compensation results are used to construct a training sequence in chronological order. A hybrid solution strategy is used to divide the training sequence into multiple subsequences, and 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 optimization process is performed to obtain an optimized numerical sequence. The optimization is iteratively performed on each subsequence. When the error evaluation value is less than a 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 characteristic 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 warning information is generated based on the performance degradation prediction results.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provided is an automatic classification and calibration system for weight measurement based on image recognition, comprising:

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

[0056] The second unit is used to extract the characteristic parameters of the weights based on the holographic feature model, obtain the dynamic change trend of the characteristic 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 grade 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 weight grade classification, establish a hierarchical calibration compensation model based on manifold learning, and initialize it using compensation strategy parameters. It determines the adaptive weight coefficients in combination with the defect distribution information in the holographic feature model, and dynamically adjusts the calibration compensation model. It uses a neural ordinary differential equation network to predict performance degradation trends, evaluates based on the calibration accuracy threshold, outputs calibration parameters when the calibration accuracy meets the requirements, and generates hierarchical maintenance warning information.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

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

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0065] In an embodiment of the present invention, a comprehensive collection of surface and internal features of the weight is achieved by combining dual-path high-speed photography and X-ray imaging, and a holographic feature model is constructed through a multi-scale feature fusion network, which improves the accuracy and completeness of weight feature extraction and provides a reliable data basis for subsequent classification and calibration; a deep variational Bayesian network is combined with a gradient boosting decision forest algorithm to dynamically evaluate and classify the weight feature parameters, and a dynamic evaluation model is established through a self-organizing competitive learning network to achieve accurate classification of weight grades, making the setting of calibration accuracy thresholds and compensation strategies more targeted, and improving the accuracy and adaptability of classification; a hierarchical calibration compensation model is constructed based on manifold learning, and dynamic adjustment is performed in combination with the defect distribution information in the holographic feature model, and the performance degradation trend is predicted through a neural ordinary differential equation network, which achieves accurate compensation and predictive maintenance of weight calibration, improves calibration efficiency and reliability, and extends the service life of the weight. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Schematic diagram of the process of the automatic classification and calibration method of weight measurement based on image recognition according to an embodiment of the present invention;

[0067] Figure 2 Topology mapping and feature distribution map for a two-layer self-organizing competitive learning network. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0069] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of an automatic classification and calibration method for weight measurement based on image recognition according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] A dual-optical high-speed camera system is used to capture dual-reflected light images on the weight surface. Adaptive gamma correction is performed on the dual-reflected light images to obtain gamma-corrected images. A multi-scale edge detection algorithm is used to extract the weight contour features. Surface micro-damage and stress distribution are detected in combination with an optical flow field analysis algorithm to generate a surface defect feature map. Density distribution data is obtained using an X-ray imaging system. 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.

[0072] In one specific embodiment, the weight surface is first imaged using a dual-optical-path high-speed camera system. This system comprises two sets of parallel LED light sources and a high-speed camera. The two parallel light sources illuminate the weight surface at angles of incidence of 30 and 60 degrees, and the high-speed camera captures the dual reflected light images. This dual-optical-path design captures richer three-dimensional reflective features from the weight surface and effectively reduces the impact of ambient light interference.

[0073] After acquiring the dual-reflection light image, the system performs adaptive gamma correction on the image. First, the image's grayscale histogram is calculated and its brightness distribution characteristics are analyzed. Then, based on the image's contrast and brightness distribution, the gamma coefficient is adaptively adjusted, and a gamma transform is performed on the original image. This step effectively enhances image detail 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 an image pyramid at different scales, calculates image gradients at each scale, and detects edges using a combination of non-maximum suppression and a dual-thresholding method. Finally, the multi-scale detection results are integrated to obtain complete contour feature information.

[0074] At the same time, the system uses an optical flow analysis algorithm to detect micro-damage and stress distribution characteristics on the weight's surface. By calculating the optical flow field between adjacent image frames and analyzing its gradient variation, it can effectively identify minor surface damage. Based on the density distribution characteristics of the optical flow field, the surface stress distribution is evaluated, ultimately generating a surface defect feature map. Furthermore, a microfocus X-ray imaging system is used to scan the weight's interior, acquiring multi-angle projection images through rotational scanning. An image reconstruction algorithm is then used to construct three-dimensional density distribution data within the weight.

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

[0076] In this embodiment, a dual-optical path high-speed camera system is used in combination with gamma correction and a multi-scale edge detection algorithm to efficiently extract the surface contour and micro-damage features of the weight, and accurately identify surface defects and stress distribution; a micro-focus X-ray imaging system is used to obtain high-precision density distribution data inside the weight, which can reveal the internal structural characteristics and achieve non-destructive testing; with the help of a multi-scale feature fusion network, the surface defect characteristics are fused with the internal density distribution to generate a more comprehensive and accurate holographic feature model of the weight, thereby enhancing the ability to evaluate the overall state of the weight; through the joint analysis of surface and internal features, the limitations of a single detection method are avoided, the efficiency and accuracy of detection are greatly improved, and a scientific basis is provided for weight performance evaluation.

[0077] The weight characteristic parameters are extracted based on the holographic feature model, and the dynamic change trend of the characteristic parameters is obtained based on the deep variational Bayesian network. The gradient boosting decision forest algorithm is used to evaluate and score 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 grade classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters.

[0078] In one specific implementation, key characteristic parameters are first extracted from the weight holographic feature model, including surface structural 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 dimension deviation and surface stress distribution). To analyze the dynamic changes in these characteristic parameters, a deep variational Bayesian network is used to construct a prediction model. This network comprises multiple layers of latent variables, uses variational inference methods to probabilistically model the parameter distribution, and utilizes reparameterization techniques for parameter estimation, thereby accurately capturing the temporal variation characteristics of the characteristic parameters.

[0079] Based on feature parameter analysis, the present invention employs a gradient boosting decision forest algorithm for evaluation and scoring. This algorithm first constructs a basic decision tree, using the feature parameters and dynamic change 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 considers the residuals 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 dynamic change trend assessment results, comprehensively generates an initial classification result for the weights.

[0080] Based on the initial classification results, the present invention trains a self-organizing competitive learning network to construct a dynamic weight evaluation model. This network uses a competitive learning mechanism and comprises an input layer, a competition layer, and an output layer. During training, the network continuously adjusts neuron weights so that samples with similar features are mapped to similar output spaces. The network first randomly initializes the weight matrix, then inputs the feature vector, determines the winning neuron by calculating the Euclidean distance, 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, the present invention classifies weights into different grades. Specific classification standards include: E1 (highest precision requirement), E2 (high precision requirement), F1 (medium precision requirement), and F2 (standard precision requirement). For weights of different grades, the system sets corresponding calibration accuracy thresholds. For example, the relative error of E1 weights does not exceed ±0.5×10 -6 The relative error of E2 weights does not exceed ±1.6×10 -6 At the same time, the compensation strategy parameters are determined according to the weight grade, including calibration cycle, compensation coefficient, temperature correction factor, etc. These parameters will be used in the subsequent calibration and compensation process to ensure that the measurement accuracy of the weight meets the requirements of the corresponding grade.

[0082] In this embodiment, the dynamic change trend of the characteristic parameters of the weights is extracted based on the deep variational Bayesian network, which can accurately capture the changing characteristics of the weight state over time and provide a basis for long-term stability evaluation; the gradient boosting decision forest algorithm is combined to evaluate and score the characteristic parameters and dynamic change trends of the weights, quickly generate initial classification results, and improve the accuracy and efficiency of classification; a dynamic evaluation model is established by training a self-organizing competitive learning network to realize intelligent evaluation of the weight grade classification, thereby improving the scientificity and adaptability of the classification; based on the weight grade classification results, the calibration accuracy threshold and compensation strategy parameters are dynamically set to provide personalized solutions for the accurate calibration and error compensation of the weights, thereby enhancing the metrological performance and application value of the weights; through dynamic evaluation and classification, combined with the optimization of accuracy threshold and compensation strategy, the whole process optimization of weights from monitoring to management is realized, and the overall performance guarantee capability is improved.

[0083] Calibration reference weights are selected according to the weight grade classification, and a hierarchical calibration compensation model based on manifold learning is established. The compensation strategy parameters are used for initialization, and the adaptive weight coefficients are determined in combination with the defect distribution information in the holographic feature model to dynamically adjust the calibration compensation model. A neural ordinary differential equation network is used to predict the performance degradation trend. Based on the calibration accuracy threshold evaluation, the calibration parameters are output when the calibration accuracy meets the requirements, and graded maintenance warning information is generated.

[0084] In this embodiment, a hierarchical calibration compensation model based on manifold learning is used, combined with defect distribution information and adaptive weight coefficients, to dynamically adjust the calibration strategy, achieve accurate compensation for weights, and improve the reliability of calibration results; a neural ordinary differential equation network is used to accurately predict the performance degradation trend of weights, provide data support for early prediction of weight status, and ensure that calibration work is timely and effective; based on the calibration accuracy threshold, the quality status of the weights is evaluated in real time, and the final calibration parameters are output when the accuracy meets the requirements, thereby improving the efficiency and accuracy of calibration; hierarchical maintenance warning information is generated based on the performance degradation prediction results to help formulate targeted maintenance plans, extend the service life of weights, and reduce operation and maintenance costs; by dynamically adjusting the compensation model and real-time monitoring, prediction, and evaluation, a full-process, intelligent weight management system is constructed to effectively improve the weight performance guarantee capability.

[0085] In an optional embodiment, collecting dual-reflected light images of the weight surface by a dual-light-path high-speed camera system, and performing adaptive gamma correction on the dual-reflected light images to obtain a gamma-corrected image includes:

[0086] A dual-light-path high-speed camera system was constructed, including a first LED light source and a second LED light source, which respectively generate regular reflection light images and diffuse reflection light images. The dual-reflection light images of the weight surface were acquired using a high-speed camera through alternating triggering of the controller.

[0087] A morphological neural network is used to perform multi-scale decomposition on the double-reflected light image of the weight surface, and surface feature maps of different scales are extracted, including texture feature maps and structural feature maps. Local phase consistency features are calculated, and a phase attention mapping matrix is constructed. The surface feature maps of different scales are dynamically recalibrated based on the phase attention mapping matrix to obtain an enhanced multi-scale feature map.

[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 recursive neural network is constructed to calculate the adaptive gamma correction coefficients of the calibration feature map layer by layer. 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 embodiment, a dual-optical path high-speed camera system is first constructed for image acquisition. The system includes two LED light sources, one for generating regular reflection light and the other for generating diffuse reflection light. The first LED light source is a white light LED with a light intensity of 500 lumens and an illumination angle of 45 degrees; the second LED light source is a diffuse light LED with a light intensity of 300 lumens and an illumination angle of 60 degrees. A controller sets the alternating trigger interval to 1 millisecond, 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, with a pixel depth of 12 bits.

[0091] The collected dual-reflection images were subjected to multi-scale decomposition. A morphological neural network was used for feature extraction. The network consists of four scale layers, each using a 3×3 convolution kernel. The first layer extracts texture features with 64 convolution kernels; the second layer extracts local structural features with 128 convolution kernels; the third layer extracts medium-scale features with 256 convolution kernels; and the fourth layer extracts large-scale features with 512 convolution kernels. Local phase congruence was calculated for the feature maps at each scale layer, with a window size of 7×7 pixels. An attention map matrix was constructed based on the phase congruence values, and the feature maps were weighted and calibrated.

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

[0093] Finally, a layered recurrent neural network is used to calculate the gamma correction coefficients. The network consists of five recurrent layers, each of which predicts a gamma value. The initial gamma value of the first layer is set to 2.2, and subsequent layers dynamically adjust it based on residual feedback, with a step size of 0.1. The corresponding gamma correction is applied to the feature maps at each scale. Feature reconstruction is then performed through a four-layer deconvolution network, ultimately outputting the corrected image. The deconvolution kernel size is 4×4, with a step size of 2.

[0094] In this embodiment, through the alternating triggering acquisition mechanism of the dual-optical path high-speed camera system, the regular reflection and diffuse reflection information of the weight surface can be obtained simultaneously, which improves the integrity and accuracy of the image information and provides a reliable data basis for subsequent processing; the combination of multi-scale decomposition and phase attention mechanism can effectively extract the multi-level features of the weight surface, and highlight the important feature areas through dynamic recalibration, thereby improving the discriminability of feature expression and enhancing the clarity of image details; based on the processing scheme of the dual-branch self-calibration network and the hierarchical recursive neural network, adaptive geometric correction and photometric correction of the image are realized, and the gamma correction parameters are dynamically optimized through the residual feedback mechanism, which significantly improves the image quality and observation effect, making the detection of weight surface defects more accurate and reliable.

[0095] In an optional embodiment, a multi-scale edge detection algorithm is used to extract the weight contour features, and an optical flow field analysis algorithm is combined to detect surface micro-damage and stress distribution to generate 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. Each scale image is bilaterally filtered to obtain a smoothed image. The corresponding horizontal gradient features and vertical gradient features are calculated to construct the gradient amplitude features. The spatial local statistical features and grayscale local statistical features of the smoothed image are calculated to construct a bilateral adaptive threshold. The bilateral adaptive threshold is used to segment the gradient amplitude features to obtain an edge feature map, and the edge feature maps of different scales are weightedly fused to obtain the weight contour feature.

[0097] Calculate the optical flow field features of adjacent frame images at the edge point position of the weight contour feature, calculate the weight surface stress tensor features based on the optical flow field features, and decompose the features to obtain the weight surface stress amplitude features and weight surface stress direction features;

[0098] The weight contour features, weight surface stress amplitude features and weight surface stress direction features are used to construct a feature matrix, which is normalized to obtain a normalized feature matrix. The corresponding intra-class distance features are calculated, and the fusion weight coefficient is determined based on the intra-class distance features. The normalized feature matrix is weightedly fused using the fusion weight coefficient to obtain a 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. In a specific implementation, the gamma coefficient is set to 2.2, and the image pixel values are subjected to a power transformation to achieve a more uniform image brightness distribution.

[0100] Based on the gamma-corrected image, multi-scale edge detection was used to extract the weight's contour features. The image was decomposed at multiple scales using a Gaussian filter bank, with the filter scales set to 1, 2, 4, and 8 pixels, resulting in four images at different scales. Each scaled image was smoothed using bilateral filtering with a window size of 5×5 pixels, a spatial standard deviation of 2.0, and a grayscale standard deviation of 20.

[0101] Calculate the horizontal and vertical gradient features of the smoothed image. Use the Sobel operator to calculate the gradient, generating horizontal and vertical gradient maps, which in turn generate the gradient magnitude feature map. Simultaneously, calculate the image's spatial local statistical features, including the local mean and variance, using a 9×9 pixel window size. Calculate the grayscale local statistical features, including the mean and standard deviation of the local grayscale histogram.

[0102] Based on these statistical features, a bilateral adaptive threshold is constructed, ranging from 20% to 80% of the gradient amplitude. This threshold is used to segment the gradient amplitude features, generating an edge feature map. Edge feature maps at different scales are weighted and fused, with weight coefficients of 0.4, 0.3, 0.2, and 0.1, to obtain the final weight contour feature.

[0103] Based on the obtained weight contour features, 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. Based on the optical flow field features, stress tensor features are calculated, including stress amplitude and direction information. The stress tensor features reflect the strain distribution on the weight surface.

[0104] The weight contour features, surface stress amplitude features, and directional features are combined into a feature matrix, which is then normalized. The intra-class distance feature of the normalized feature matrix is calculated, and feature fusion weights are determined based on the intra-class distances. Finally, a weight surface defect feature map is obtained through weighted fusion.

[0105] In this embodiment, through a method that combines multi-scale edge detection and optical flow field analysis, it is possible to comprehensively and accurately detect tiny defects and stress distribution characteristics on the surface of the weight, with a detection accuracy reaching the micron level, significantly improving the reliability of the surface quality detection of the weight; a bilateral adaptive threshold segmentation method is adopted, which fully considers the statistical characteristics of the spatial and grayscale domains of the image, making the edge detection results more stable and reliable, and effectively overcoming the problem that the traditional fixed threshold method is easily affected by noise; based on the normalization processing and weighted fusion strategy of the feature matrix, an effective combination of multiple features is achieved, and the generated defect feature map has a strong expressive ability, providing reliable technical support for weight quality control.

[0106] In an optional embodiment, extracting weight characteristic parameters according to the holographic characteristic model and obtaining the dynamic change trend of the characteristic parameters based on the deep variational Bayesian network include:

[0107] Based on the original data matrix corresponding to the holographic feature model, the data mean of the feature dimension is calculated and subtracted to obtain the centralized data matrix; the covariance matrix of the centralized data matrix is calculated and eigenvalue decomposition is performed to obtain the eigenvalues and eigenvectors, and a whitening transformation matrix is constructed. The centralized 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, and use a fast fixed point iterative algorithm to update the column vectors of the separation matrix until a preset maximum number of iterations is reached to determine the final separation matrix; multiply the whitened data matrix by the final separation matrix to obtain independent components, and divide them into spatial modal components, temporal modal components, and eigenmodal components;

[0109] Calculate the surface gradient of the spatial modal component, determine the adaptive weight coefficient and perform weighted calculation, and spatially integrate the weighted surface gradient to obtain the first characteristic parameter; perform a two-dimensional spatial integration on the time modal component, and convolve the integration result with a pre-calibrated stress kernel function to obtain the second characteristic parameter; calculate the normalized distribution of the characteristic modal component, and calculate the information entropy based on the normalized distribution to obtain the third characteristic parameter;

[0110] The first, second, and third feature parameters are combined into a feature vector, which is input into a deep variational Bayesian network. The encoder calculates the latent variable distribution. The gated recurrent unit uses the latent variable distribution stored at the previous moment as the hidden state input, and calculates and updates it together with the feature vector at the current moment to obtain the latent variable distribution at the current moment. The latent variable distribution at the current moment is passed through 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 value of the eigenvector, the network parameters are optimized and the optimal network is determined. The Monte Carlo method is used to sample the optimal network multiple times to obtain the eigenvector and change interval at the prediction moment, and the dynamic change trend of the weight parameters is obtained.

[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 normalized 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 to make the covariance matrix of the data the unit matrix, ensuring that different feature dimensions are processed at the same scale.

[0113] The separation matrix specifically refers to a linear transformation matrix used to extract 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 so that the projected signals are independent of each other, maximizing their statistical independence. The optimization of the separation matrix is completed through an iterative algorithm, and its ultimate goal is to extract the source signal and help analyze the independent structure or characteristics of the data.

[0114] In one specific embodiment, the raw data of the holographic feature model is first preprocessed. Data centering is achieved by calculating the mean of each feature dimension and subtracting the corresponding mean from the raw data. For example, for a data matrix containing 1,000 samples, each with 20 feature dimensions, the mean of each of the 20 feature dimensions is calculated, and then the corresponding mean is subtracted from the raw data. The covariance matrix of the centered data is then calculated, and the eigenvalue decomposition of the covariance matrix is performed to obtain the eigenvalues and eigenvectors. A whitening transformation matrix is constructed based on the eigenvalues and eigenvectors, and the centered data is left-multiplied by this matrix to obtain the whitened data.

[0115] Then perform independent component analysis. Initialize a separation matrix whose dimension is the same as the characteristic dimension of the whitened data. Project the whitened data into the column vector space of the separation matrix and calculate the probability density of the projected data. Calculate the negative entropy estimate based on the probability density and use the fast fixed point iteration algorithm to update the column vector of the separation matrix. When the preset maximum number of iterations (such as 1000 times) or the convergence condition is reached, the final separation matrix is determined. Multiply the whitened data by the final separation matrix to obtain independent components. According to the physical characteristics, the independent components are divided into three categories: spatial modes, time modes and characteristic modes.

[0116] For the spatial modal component, calculate its surface gradient. Determine the adaptive weight coefficient for each surface gradient by minimizing the reconstruction error. For example, set the initial weight to 0.5, and then iteratively optimize the weight value through gradient descent. Perform spatial integration on the weighted surface gradient to obtain the first characteristic parameter of the weight. Perform a two-dimensional spatial integration on the temporal modal component, and convolve the integration result with a pre-calibrated stress kernel function to obtain the second characteristic parameter. Calculate the normalized distribution of the characteristic modal component and calculate the information entropy based on this to obtain the third characteristic parameter.

[0117] The three feature parameters form a feature vector and are input into the encoder of a deep variational Bayesian network. The network architecture uses a multilayer perceptron, consisting of several fully connected layers. The encoder calculates the latent variable distribution. A gated recurrent unit uses the latent variable distribution stored at the previous moment as the hidden state input. This is combined with the current feature vector to calculate and update the current latent variable distribution. The gated recurrent unit contains an update gate and a reset gate to control the flow of information. The latent variable distribution output by the gated recurrent unit is 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 eigenvector. Training is performed using optimization algorithms such as stochastic gradient descent. After determining the optimal network, the Monte Carlo method is used to perform multiple sampling (e.g., 1,000 times) to obtain the eigenvectors and their variation range at the prediction moment, thereby determining the dynamic variation trend of the weight parameters.

[0119] In this embodiment, by performing centralization and whitening preprocessing on the original data, the accuracy and stability of subsequent feature extraction are significantly improved, data redundancy and noise interference are reduced, and the foundation is laid for obtaining high-quality weight feature parameters; the deep variational Bayesian network combined with the gated recurrent unit is used to model the dynamic changes of feature parameters, which makes full use of historical information, improves the model's ability to capture time series features, and makes the prediction results more accurate and reliable; the change range of the prediction results is obtained based on the Monte Carlo sampling method, which not only gives the dynamic change trend of the weight parameters, but also provides an uncertainty estimate of the prediction, providing a more comprehensive reference basis for subsequent decision-making and control.

[0120] In an optional embodiment, generating an initial classification result by evaluating and scoring using a gradient boosting decision forest algorithm includes:

[0121] A characteristic parameter set is constructed based on the characteristic parameters of the weights, and a time series characteristic sequence is obtained by sampling the characteristic parameters of each weight. The first-order difference value and the second-order difference value are calculated to construct a dynamic change characteristic 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 square error is used as the loss function, and 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 building 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 features to be segmented are used to segment the left and right child nodes, and 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 difference value and the second-order difference value in the dynamic change feature set into the exponential decay function to calculate the time series weight value and obtain the time series weight set;

[0125] The gradient boosting decision forest model is used to predict the weight feature parameters, and the prediction probability value of the decision tree is obtained. The average prediction probability value is calculated and weighted summed with the time series weight value to obtain the weight score value;

[0126] Compare the weight's score with the preset classification threshold to obtain the initial classification result of the weight.

[0127] The Gini coefficient specifically refers to an indicator that measures the internal impurity or disorder of a dataset and is commonly used to evaluate node segmentation in decision tree algorithms. It measures the unevenness of the category distribution in a dataset, with a value ranging from 0 to 1: 0 indicates a perfect uniform distribution, where all data points belong to the same category, resulting in the best segmentation effect; 1 indicates an extremely uneven distribution, where all data points belong to different categories, resulting in the worst segmentation effect. In the process of building a decision tree, the Gini coefficient is used to evaluate the effectiveness of a specific feature in dividing a node. A smaller Gini coefficient indicates that the classification of the node is purer, meaning that the segmentation effect is better.

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

[0129] Next, a gradient boosting decision forest model is constructed. The feature parameter set and the dynamically changing feature set are combined to form a training dataset. A decision tree is gradually constructed through iterative training. In each iteration, a negative gradient is calculated based on the difference between the current model's prediction and the true value. This negative gradient serves as the new target value for training the next decision tree. As the decision tree grows, each time a node is split, some features are randomly selected as candidate split features. For the quality feature, possible split points include specific values such as 100.001g and 100.002g.

[0130] During the node splitting process, 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 initial impurity value is calculated by counting the proportion of each type of sample contained in it. Then, different features and splitting points are tried, and the impurity values of the left and right child nodes after the split are calculated. The feature and splitting point that minimizes impurity are selected as the final splitting solution.

[0131] To reflect the importance of time series features, dynamically changing features are assigned time series weights. Data closer to the current time point has a greater weight. For example, data from the last hour might have a weight of 0.9, data from two hours ago might have a weight of 0.8, and so on. These weights are then weighted averaged with the model's predicted probability to produce the final score.

[0132] The scoring results are compared with the preset classification threshold. For example, if the classification threshold is set to 0.8, weights with a score greater than 0.8 are classified as high-quality, and those with scores greater than 0.8 are classified as pending. This completes the initial classification of the weights.

[0133] As shown in the following table:

[0134]

[0135] The node splitting and Gini coefficient analysis data of the gradient boosting decision forest model in the weight classification process are recorded in detail. It can be clearly seen from the table that in the root node (node ID is 1), the first-order difference of mass is selected as the optimal segmentation feature, and the segmentation threshold is 0.0015. The initial Gini coefficient of this node is 0.4658, indicating that the category distribution in the data set is relatively mixed. After segmentation, the Gini coefficient of the left child node is significantly reduced to 0.1275, and the Gini coefficient of the right child node is 0.2183. The segmentation gain value reaches 0.2451, which is the highest gain value among all nodes, proving that this feature is the most effective feature for distinguishing the quality of weights.

[0136] In the second-layer nodes, nodes 2 and 3 were split based on the second-order difference of density and mass, respectively. In particular, node 2, whose initial Gini coefficient was 0.1275, saw its left child's Gini coefficient further reduced to 0.0423 after the split, indicating a significant improvement in the sample purity of this node. In subsequent node splits, features such as hardness, surface roughness, temporal weight, and first-order difference of density were successively selected. Notably, the splitting of node 4 based on the hardness feature reduced the Gini coefficient of its left child to 0.0000, indicating that the samples at this node were completely pure, with all samples belonging to the same category.

[0137] In addition, the data in the table also demonstrates the importance of dynamic change features (such as first-order differences and second-order differences) in the decision-making process. Of the 12 nodes, seven used dynamic change features as the basis for segmentation, with the temporal variation features of quality and density being particularly prominent. By observing the segmentation gain values, it can be seen that dynamic change features generally lead to higher purity improvements. For example, the segmentation gain values for nodes 1, 3, and 7 were 0.2451, 0.0722, and 0.0483, respectively, all of which are at relatively high levels.

[0138] Looking at the sample distribution, the root node contains 100 samples, including 65 high-quality weights and 35 untested weights. As the tree progresses, the samples are gradually refined, and the node purity continues to improve. For example, 11 of the 12 samples in node 8 fall into the untested category, while 11 of the 12 samples in node 10 fall into the high-quality category, demonstrating 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 discriminatory power of various features in weight classification, providing detailed data support for understanding the basis of the model's decisions.

[0139] As shown in the following table:

[0140]

[0141] The decision tree node analysis and score distribution are described, ultimately adopting a classification threshold of 0.80. Weights with a score ≥ 0.80 are classified as high-quality; those with a score < 0.80 are considered untested. This threshold, set midway between the score distributions of the two weight categories (0.78-0.85), effectively balances classification accuracy and robustness.

[0142] The method technology of the present embodiment is derived from the actual demand in the field of weight quality detection. Traditional weight quality detection mainly relies on manual experience judgment or simple statistical methods, such as threshold method, variance analysis, etc. The prior art usually only uses the static characteristic parameters (such as mass, volume, density, etc.) of the weight to carry out single measurement analysis, lacks the dynamic monitoring of the 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 relationship between the weight characteristic parameters and underutilizes the time series variation characteristics.

[0143] The starting point of the method improvement of this embodiment is to solve the problem of insufficient consideration of the dynamic change characteristics of weights in the existing technology, and to improve the classification accuracy by introducing time series feature analysis and advanced machine learning models. A time series sampling mechanism is introduced to collect time series data of weight feature parameters, and a dynamic change feature set is constructed by calculating the first-order difference value and the second-order difference value, which effectively captures the changing trend of weight characteristics; a gradient boosting decision forest model is used to replace the traditional single decision tree, and the model performance is continuously optimized through the residual learning strategy to improve the classification accuracy; a time series weight mechanism is introduced to substitute the first-order difference value and the second-order difference value of the dynamic change feature into the exponential decay function to calculate the time series weight value, so that the recent data has a greater impact on the classification result, thereby more accurately reflecting the current state of the weight; feature sampling and node splitting strategies are optimized, and the optimal segmentation feature is calculated through the Gini coefficient to improve the generalization ability of the model.

[0144] After actual verification, the method of this embodiment has increased the accuracy of weight classification from the original 85% to over 95% compared to the existing technology, and the misclassification rate has been reduced by 60%. In particular, the recognition accuracy of weights in borderline states has increased by 70%. In addition, the method of this embodiment has stronger adaptability and can automatically adjust the scoring weights according to the time changes of weight characteristics, making the classification results more reliable and stable. By introducing dynamic feature analysis and a time-series weighting mechanism, the method of this embodiment successfully solves the problem of insufficient consideration of the dynamic change characteristics of weights in the existing technology, providing a more scientific and accurate technical means for weight quality detection.

[0145] In this embodiment, by introducing time series features and dynamic change features, the changing patterns of weight performance parameters can be fully captured, thereby improving classification accuracy; the gradient boosting decision forest algorithm combined with the time series weight scoring mechanism can effectively balance the influence of historical data and the latest data, making the classification results more stable and reliable, reducing the misjudgment rate, and improving the stability of classification; the iterative optimization process based on the residual learning strategy enables the model to have a strong generalization ability, which can adapt to the classification requirements of different types of weights, improve the adaptability of the model, and significantly enhance its scalability.

[0146] In an optional embodiment, the self-organizing competitive learning network includes:

[0147] Construct a two-layer self-organizing competitive learning network, including input layer neurons with the same dimension as the weight feature, and competition layer neurons arranged in a two-dimensional grid topology;

[0148] The connection weights of the two-layer self-organizing competitive learning network are randomly initialized, the weight features are input into the two-layer self-organizing competitive learning network, the Euclidean distance between the weight features and the connection weights is calculated, and the competition layer neurons corresponding to the minimum Euclidean distance and the second minimum Euclidean distance are selected to determine the first winning neuron and the second winning neuron;

[0149] The neighborhood radius is determined based on the position of the first winning neuron, and the neighborhood response strength of the competitive layer neurons within the neighborhood radius is calculated, which decays exponentially with the distance from the first winning neuron.

[0150] Based on the neighborhood response strength, the connection weights of the first winning neuron and the competition layer neurons in the neighborhood are updated. A time-varying learning rate that decays linearly with the number of iterations is introduced to adjust the update amplitude to obtain the updated connection weights.

[0151] Calculate the quantization error value between the weight feature and the updated connection weight, determine the positional relationship between the first winning node and the second winning node, and when they are not adjacent, calculate the shortest path length to obtain the topological error value; when the quantization error value is less than a first preset threshold and the topological error value is less than a second preset threshold, determine that the self-organizing competitive learning network has reached a training completion state;

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

[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 matches the weight's feature dimensions. For example, for a weight with 10 features, such as weight, diameter, and height, the input layer would have 10 neurons. The competition layer uses a 20×20 two-dimensional grid structure with a total of 400 competitive layer neurons, connected by a grid to form a topological structure.

[0154] During the initialization phase, random numbers between 0 and 1 are assigned to the network connection weights. Assume that a set of weight feature data is input, and the values have been normalized to range between 0 and 1. This set of feature data is input into the network, and the distance between it and the connection weights of each competitive layer neuron is calculated.

[0155] During the distance calculation phase, the Euclidean distance between the connection weight and the input feature is calculated for each competitive layer neuron. For example, if the input features are [0.5, 0.3, 0.7] and the connection weights of a neuron are [0.4, 0.4, 0.6], the calculated Euclidean distance is 0.158. After traversing all competitive layer neurons, the neuron with the smallest distance is selected as the first winner, and the neuron with the second smallest distance is selected as the second winner.

[0156] When determining the neighborhood range, we set the initial neighborhood radius to 10 grid units, centered around the first winning neuron. All neurons within this radius participate in weight updates, but the intensity of the updates decreases with distance. For example, the response strength of a neuron at a distance of 1 grid unit is 0.9, while the response strength 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 gradually decreases as the number of training cycles increases. Assuming a total of 1000 training cycles, the learning rate decreases by 0.01 every 100 training cycles. 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 topological error are continuously monitored. Quantization error reflects the degree of match between features and weights; a value less than 0.01 is considered a good match. Topological error focuses on the positional relationship between the winning neurons. If the first and second winning neurons are adjacent, the topological error is 0. If not, the shortest grid path length between them is calculated as the topological error. A topological error less than 2 indicates a good topological structure.

[0159] Once the network training is complete, the weight's characteristic data is input, and the network automatically identifies the first and second best-matching winning neurons. 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 first-class weight, it will be evaluated as a first-class weight.

[0160] like Figure 2 The figure shows the topology of a fully trained 20×20 neural network, visually illustrating the feature space mapping for weight classification. Different geometric shapes are used to represent different weight grades: circles represent E1, squares represent E2, diamonds represent F1, triangles represent F2, pentagons represent M1, and hexagons represent M2. The depth of the black and white fill of the shape represents the distance correlation between the neuron and the grade center, with darker colors indicating higher certainty of classification. The figure clearly illustrates the distribution of the six weight grades in the feature space, with distinct boundaries and smooth transitions between the regions, demonstrating the network's excellent topology preservation capabilities.

[0161] The figure shows the distribution of areas of each weight grade, with a dotted circle used to circle an area: E1-level weights (located in the upper left corner) cover 63 neurons (15.8%), and the core area contains 25 neurons; E2-level weights (located in the upper left) have a slightly larger area, covering 72 neurons (18.0%); F1-level weights (located in the upper right) cover 69 neurons (17.3%); F2-level weights (located in the middle) cover 84 neurons (21.0%); M1-level weights (located in the lower right) cover 84 neurons (21.0%); and M2-level weights (located in the lower left) cover 108 neurons (27.0%) due to their large feature variability, with only 50 neurons in the core area, reflecting the objective law that high-precision weights have high feature concentration and low-precision weights have dispersed features.

[0162] The values displayed at the grid points are arranged vertically, with "W" representing the weight deviation characteristic and "R" representing the surface roughness characteristic. As can be seen from the values, the weight deviation in the E1-level region is around 0.010, and the surface roughness is approximately 0.005; while the weight deviation in the M2-level region is as high as 0.750, and the surface roughness reaches 0.250, clearly reflecting the characteristic differences between weights of different precision levels. The solid circle in the center of the figure marks the location of the best matching unit (BMU) of the F1-level weight to be evaluated, as well as its key characteristic values (weight deviation 0.05, surface roughness 0.022), demonstrating the network's evaluation process for a new input weight.

[0163] In terms of 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 that for M2-grade weights is 0.75; the surface roughness for E1-grade weights is 0.005, while that for M2-grade weights is 0.25; the corrosion resistance for E1-grade weights is 0.999, while that for M2-grade weights drops to 0.90; and the coefficient of thermal expansion is 0.003 for E1-grade weights, while that for M2-grade weights reaches 0.16. The distribution trends of these feature values closely correspond to the weight accuracy grades, validating the ability of the two-layer self-organizing network to accurately map the feature space. After network training, the neuron weights have fully adapted to the weight feature distribution, enabling them to accurately locate the position of newly input weights in the feature space and assign 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, the subjectivity in the manual feature extraction process is avoided, and the accuracy and objectivity of feature extraction are improved; the time-varying learning rate and neighborhood response mechanism are introduced to enable the network to have good adaptability and convergence characteristics, and can automatically adjust the learning parameters according to the training process, thereby ensuring the stability and reliability of network training; the dual evaluation indicators of quantization error and topological error are adopted to ensure the accuracy of the network training results and the rationality of the topological structure, providing reliable technical support for weight grade evaluation, and improving the automation level of weight evaluation and the credibility of evaluation results.

[0165] In an optional embodiment, calibration reference weights are selected according to weight grade classification, a hierarchical calibration compensation model based on manifold learning is established, and the compensation strategy parameters are used for initialization. The adaptive weight coefficients are determined in combination with the defect distribution information in the holographic feature model, and the calibration compensation model is dynamically adjusted. A neural ordinary differential equation network is used to predict performance degradation trends. Based on the calibration accuracy threshold assessment, the calibration parameters are output when the calibration accuracy meets the requirements, and hierarchical maintenance warning information is generated, including:

[0166] Collecting characteristic data of the weight to be calibrated and the calibration reference weight, and constructing a characteristic vector of the weight to be calibrated and a characteristic vector of the reference weight;

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

[0168] Extracting the defect distribution information of the weight to be calibrated from the holographic feature model, calculating the second norm of the defect distribution information, and multiplying it with the preset attenuation coefficient to obtain the attenuation factor, and calculating the adaptive weight coefficient;

[0169] Multiplying the adaptive weight coefficient and the compensation strategy parameter to obtain the optimized compensation strategy parameter, constructing a calibration mapping function, inputting the characteristic vector of the weight to be calibrated into the calibration mapping function, and obtaining a calibration compensation result;

[0170] A neural network of ordinary differential equations is constructed, and the calibration compensation results are used to construct a training sequence in chronological order. A hybrid solution strategy is used to divide the training sequence into multiple subsequences, and 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 optimization process is performed to obtain an optimized numerical sequence. The optimization is iteratively performed on each subsequence. When the error evaluation value is less than a 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 characteristic 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 warning information is generated based on the performance degradation prediction results.

[0172] In one specific implementation, a manifold learning-based weight-level calibration and compensation system first requires selecting appropriate calibration reference weights. 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 for calibration, and so on. High-precision sensors collect characteristic data from the calibrated and reference weights, including parameters such as mass, volume, surface roughness, and temperature coefficient, to construct a feature vector.

[0173] After constructing the feature vector, a nearest neighbor search is performed on the calibration weight feature vector. Specifically, the number of nearest neighbors, k, is set to 5, and a KD tree algorithm is used to accelerate the search process. For each feature point, the covariance relationship between it and its five nearest neighbors is calculated to construct a compensation strategy parameter optimization function. During the optimization process, the sum of the neighboring weights is ensured to be 1. The quasi-Newton method is used to solve the optimization function and obtain the initial compensation strategy parameters.

[0174] The defect distribution information of the calibration weight is extracted from a pre-established holographic feature model. This defect information includes surface characteristics such as scratches, wear, and corrosion. The bi-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 coefficient.

[0175] Multiplying the adaptive weight coefficients with the compensation strategy parameters enables dynamic adjustment. For example, when a severe surface defect is detected, the weight coefficients are reduced 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 ordinary differential equation network was established to predict performance degradation trends. A training sequence was constructed using calibration compensation results from six consecutive months, with data collected four times each month. The training sequence was divided into 12 subsequences, each containing five data points. Piecewise linear interpolation was used to construct an approximation function, and the gradient change rate of each segment was calculated. An exponential moving average was introduced for dynamic weight smoothing, with a smoothing coefficient set to 0.8. The Adam optimization algorithm was applied to the smoothed sequence, with a learning rate of 0.001 and 1000 iterations. When the error evaluation value was less than the preset threshold of 0.01%, the optimized numerical sequence was output as the performance degradation prediction result.

[0177] The Euclidean distance between the calibration compensation result and the reference weight's characteristic vector is calculated to determine the calibration accuracy. If the calibration accuracy falls below a preset threshold (e.g., an accuracy better than 0.5 ppm for E1-grade weights), the optimized compensation strategy parameters are output as the final calibration parameters. Based on the performance degradation prediction results, the system categorizes warning levels into three levels: green indicates stable performance within six months; yellow indicates recalibration is required within three to six months; and red indicates calibration is required within one to three months.

[0178] In this embodiment, by introducing manifold learning and neural ordinary differential equation networks, the accuracy and reliability of weight calibration are significantly improved, the calibration accuracy is improved by about 30%, and it has good adaptability and robustness; the dynamic weight adjustment mechanism based on the holographic feature model enables the calibration compensation process to effectively deal with abnormal situations such as surface defects of the weights, reduces the impact of abnormal factors on the calibration results, and improves the stability of the calibration results; a complete performance degradation prediction and graded maintenance early warning mechanism is established, which can predict the performance change trend of the weights in advance, carry out maintenance intervention in time, extend the service life of the weights, reduce maintenance costs, and improve the efficiency and reliability of the calibration work.

[0179] Figure 2 FIG. 1 is a structural diagram of an automatic classification and calibration system for weight measurement based on image recognition according to an embodiment of the present invention. Figure 2 As shown, the system includes:

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

[0181] The second unit is used to extract the characteristic parameters of the weights based on the holographic feature model, obtain the dynamic change trend of the characteristic 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 grade 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 weight grade classification, establish a hierarchical calibration compensation model based on manifold learning, and initialize it using compensation strategy parameters. It determines the adaptive weight coefficients in combination with the defect distribution information in the holographic feature model, and dynamically adjusts the calibration compensation model. It uses a neural ordinary differential equation network to predict performance degradation trends, evaluates based on the calibration accuracy threshold, outputs calibration parameters when the calibration accuracy meets the requirements, and generates hierarchical maintenance warning information.

[0183] According to a third aspect of the embodiments of the present invention,

[0184] An electronic device is provided, comprising:

[0185] processor;

[0186] a memory for storing processor-executable instructions;

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

[0188] According to a fourth aspect of the embodiments of the present invention,

[0189] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0190] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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 high-speed camera system is used to capture dual-reflected light images on the weight surface. Adaptive gamma correction is performed on the dual-reflected light images to obtain gamma-corrected images. A multi-scale edge detection algorithm is used to extract the weight contour features. Surface micro-damage and stress distribution are detected in combination with an optical flow field analysis algorithm to generate a surface defect feature map. Density distribution data is obtained using an X-ray imaging system. 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 weight characteristic parameters are extracted based on the holographic feature model, and the dynamic change trend of the characteristic parameters is obtained based on the deep variational Bayesian network. The gradient boosting decision forest algorithm is used to evaluate and score 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 grade classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters; Calibration reference weights are selected according to the weight grade classification, and a hierarchical calibration compensation model based on manifold learning is established. The compensation strategy parameters are used for initialization, and the adaptive weight coefficients are determined in combination with the defect distribution information in the holographic feature model to dynamically adjust the calibration compensation model. A neural ordinary differential equation network is used to predict the performance degradation trend. Based on the calibration accuracy threshold evaluation, the calibration parameters are output when the calibration accuracy meets the requirements, and graded maintenance warning information is generated.

2. The method according to claim 1, characterized in that The dual-light path high-speed camera system is used to collect the double-reflected light images on the weight surface, and the double-reflected light images are adaptively gamma corrected to obtain the gamma-corrected images including: A dual-light-path high-speed camera system was constructed, including a first LED light source and a second LED light source, which respectively generate regular reflection light images and diffuse reflection light images. The dual-reflection light images of the weight surface were acquired using a high-speed camera through alternating triggering of the controller. A morphological neural network is used to perform multi-scale decomposition on the double-reflected light image of the weight surface, and surface feature maps of different scales are extracted, including texture feature maps and structural feature maps. Local phase consistency features are calculated, and a phase attention mapping matrix is constructed. The surface feature maps of different scales are dynamically recalibrated based on the phase attention mapping matrix to obtain an enhanced multi-scale feature map. 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 recursive neural network is constructed to calculate the adaptive gamma correction coefficients of the calibration feature map layer by layer. 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 weight contour features, and an optical flow field analysis algorithm is used to detect surface micro-damage and stress distribution. The surface defect feature map is generated, including: A multi-scale Gaussian filter is used to decompose the gamma-corrected image to obtain a multi-scale image group. Each scale image is bilaterally filtered to obtain a smoothed image. The corresponding horizontal gradient features and vertical gradient features are calculated to construct the gradient amplitude features. The spatial local statistical features and grayscale local statistical features of the smoothed image are calculated to construct a bilateral adaptive threshold. The bilateral adaptive threshold is used to segment the gradient amplitude features to obtain an edge feature map, and the edge feature maps of different scales are weightedly fused to obtain the weight contour feature. Calculate the optical flow field features of adjacent frame images at the edge point position of the weight contour feature, calculate the weight surface stress tensor features based on the optical flow field features, and decompose the features to obtain the weight surface stress amplitude features and weight surface stress direction features; The weight contour features, weight surface stress amplitude features and weight surface stress direction features are used to construct a feature matrix, which is normalized to obtain a normalized feature matrix. The corresponding intra-class distance features are calculated, and the fusion weight coefficient is determined based on the intra-class distance features. The normalized feature matrix is weightedly fused using the fusion weight coefficient to obtain a surface defect feature map.

4. The method according to claim 1, wherein The weight characteristic parameters are extracted based on the holographic characteristic model, and the dynamic change trends of the characteristic parameters are obtained based on the deep variational Bayesian network, including: Based on the original data matrix corresponding to the holographic feature model, the data mean of the feature dimension is calculated and subtracted to obtain the centralized data matrix; the covariance matrix of the centralized data matrix is calculated and eigenvalue decomposition is performed to obtain the eigenvalues and eigenvectors, and a whitening transformation matrix is constructed. The centralized 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, and use a fast fixed point iterative algorithm to update the column vectors of the separation matrix until a preset maximum number of iterations is reached to determine the final separation matrix; multiply the whitened data matrix by the final separation matrix to obtain independent components, and divide them into spatial modal components, temporal modal components, and eigenmodal components; Calculate the surface gradient of the spatial modal component, determine the adaptive weight coefficient and perform weighted calculation, and spatially integrate the weighted surface gradient to obtain the first characteristic parameter; perform a two-dimensional spatial integration on the time modal component, and convolve the integration result with a pre-calibrated stress kernel function to obtain the second characteristic parameter; calculate the normalized distribution of the characteristic modal component, and calculate the information entropy based on the normalized distribution to obtain the third characteristic parameter; The first, second, and third feature parameters are combined into a feature vector, which is input into a deep variational Bayesian network. The encoder calculates the latent variable distribution. The gated recurrent unit uses the latent variable distribution stored at the previous moment as the hidden state input, and calculates and updates it together with the feature vector at the current moment to obtain the latent variable distribution at the current moment. The latent variable distribution at the current moment is passed through 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 value of the eigenvector, the network parameters are optimized and the optimal network is determined. The Monte Carlo method is used to sample the optimal network multiple times to obtain the eigenvector and change interval at the prediction moment, and the dynamic change trend of the weight parameters is obtained.

5. The method according to claim 1, wherein The gradient boosting decision forest algorithm is used to evaluate and score the initial classification results, including: A characteristic parameter set is constructed based on the characteristic parameters of the weights, and a time series characteristic sequence is obtained by sampling the characteristic parameters of each weight. The first-order difference value and the second-order difference value are calculated to construct a dynamic change characteristic 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 square error is used as the loss function, and 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 building 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 features to be segmented are used to segment the left and right child nodes, and 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 difference value and the second-order difference value in the dynamic change feature set into the exponential decay function to calculate the time series weight value and obtain the time series weight set; The gradient boosting decision forest model is used to predict the weight feature parameters, and the prediction probability value of the decision tree is obtained. The average prediction probability value is calculated and weighted summed with the time series weight value to obtain the weight score value; Compare the weight's score with the preset classification threshold to obtain the initial classification result of the weight.

6. The method according to claim 1, characterized in that The self-organizing competitive learning network includes: Construct a two-layer self-organizing competitive learning network, including input layer neurons with the same dimension as the weight feature, and competition layer neurons arranged in a two-dimensional grid topology; The connection weights of the two-layer self-organizing competitive learning network are randomly initialized, the weight features are input into the two-layer self-organizing competitive learning network, the Euclidean distance between the weight features and the connection weights is calculated, and the competition layer neurons corresponding to the minimum Euclidean distance and the second minimum Euclidean distance are selected to determine the first winning neuron and the second winning neuron; The neighborhood radius is determined based on the position of the first winning neuron, and the neighborhood response strength of the competitive layer neurons within the neighborhood radius is calculated, which decays exponentially with the distance from the first winning neuron. Based on the neighborhood response strength, the connection weights of the first winning neuron and the competition layer neurons in the neighborhood are updated. A time-varying learning rate that decays linearly with the number of iterations is introduced to adjust the update amplitude to obtain the updated connection weights. Calculate the quantization error value between the weight feature and the updated connection weight, determine the positional relationship between the first winning node and the second winning node, and when they are not adjacent, calculate the shortest path length to obtain the topological error value; when the quantization error value is less than a first preset threshold and the topological error value is less than a second preset threshold, determine that the self-organizing competitive learning network has reached a training completion state; The weight features of the weight to be evaluated are input into the trained deterministic self-organizing competitive learning network, and the evaluation level of the weight to be evaluated is generated based on the updated connection weights and the neighborhood response strength.

7. The method according to claim 1, characterized in that Calibration reference weights are selected according to weight grade classification, and a hierarchical calibration compensation model based on manifold learning is established. The compensation strategy parameters are initialized, and the adaptive weight coefficients are determined in combination with the defect distribution information in the holographic feature model to dynamically adjust the calibration compensation model. A neural network of ordinary differential equations is used to predict performance degradation trends. Based on the calibration accuracy threshold assessment, calibration parameters are output when the calibration accuracy meets the requirements, and graded maintenance warning information is generated, including: Collecting characteristic data of the weight to be calibrated and the calibration reference weight, and constructing a characteristic vector of the weight to be calibrated and a characteristic vector of the reference weight; Perform a k-nearest neighbor search on the calibration weight feature vector, calculate the covariance matrix between the feature point and the k nearest neighbor points, construct a compensation strategy parameter optimization function, and solve it under the constraint that the sum of the neighbor point weights is 1 to obtain the compensation strategy parameters; Extracting the defect distribution information of the weight to be calibrated from the holographic feature model, calculating the second norm of the defect distribution information, and multiplying it with the preset attenuation coefficient to obtain the attenuation factor, and calculating the adaptive weight coefficient; Multiplying the adaptive weight coefficient and the compensation strategy parameter to obtain the optimized compensation strategy parameter, constructing a calibration mapping function, inputting the characteristic vector of the weight to be calibrated into the calibration mapping function, and obtaining a calibration compensation result; A neural network of ordinary differential equations is constructed, and the calibration compensation results are used to construct a training sequence in chronological order. A hybrid solution strategy is used to divide the training sequence into multiple subsequences, and 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 optimization process is performed to obtain an optimized numerical sequence. The optimization is iteratively performed on each subsequence. When the error evaluation value is less than a 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 characteristic 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 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 according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect double-reflected light images of the weight surface through a dual-optical path high-speed camera system, perform adaptive gamma correction on the double-reflected light images to obtain gamma-corrected images, use a multi-scale edge detection algorithm to extract weight contour features, combine with an optical flow field analysis algorithm to detect surface micro-damage and stress distribution, and generate a surface defect feature map; use an X-ray imaging system to obtain density distribution data, use a multi-scale feature fusion network to fuse the surface defect feature map with the density distribution data, and construct a holographic feature model; The second unit is used to extract the weight characteristic parameters based on the holographic feature model, obtain the dynamic change trend of the characteristic 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 grade classification, and set the corresponding calibration accuracy threshold and compensation strategy parameters; The third unit is used to select calibration reference weights according to weight grade classification, establish a hierarchical calibration compensation model based on manifold learning, and initialize it using compensation strategy parameters. It determines the adaptive weight coefficients in combination with the defect distribution information in the holographic feature model, and dynamically adjusts the calibration compensation model. It uses a neural ordinary differential equation network to predict performance degradation trends, evaluates based on the calibration accuracy threshold, outputs calibration parameters when the calibration accuracy meets the requirements, and generates hierarchical maintenance warning information.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the 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 a processor, the method according to any one of claims 1 to 7 is implemented.

Images