A precision evaluation method and system of a measurement method based on deep learning

Abstract

This invention discloses a precision evaluation method and system for measurement methods based on deep learning. The method includes inputting measurement source data into a deep feature extraction model to obtain multimodal measurement features; determining measurement statistical indicators based on measured values; comparing the measured values ​​with calibration values ​​to obtain a first precision rating coefficient; calculating a second precision rating coefficient based on the measurement statistical indicators and measurement standard indicators; calculating an environmental state coefficient based on measurement state parameters; inputting the measurement data and environmental state coefficient into a precision evaluation index prediction model to obtain a precision evaluation index; constructing a precision evaluation model based on the precision evaluation index and the precision markers; and inputting the measurement data and measurement state parameters of the measurement method to be evaluated into the precision evaluation model to obtain a precision evaluation result. This method not only improves the efficiency and accuracy of precision evaluation of measurement methods but also has good interpretability and can be directly applied to precision evaluation systems for measurement methods.

Description

Technical Field

[0001] This invention relates to the field of precision assessment, and more particularly to a precision assessment method and system based on deep learning measurement methods. Background Technology

[0002] In modern industrial manufacturing, scientific research, and quality inspection, precise measurement is a crucial link in ensuring product quality and driving technological innovation. With the rapid development of technologies such as intelligent manufacturing and automated testing, higher demands are being placed on the precision evaluation of measurement methods. High-precision measurement methods can reduce fluctuations and errors in measurement results, providing solid data support for precise control of production processes, accurate revelation of scientific laws, and effective management of complex systems.

[0003] Traditional measurement precision assessment techniques often rely on human experience or simple statistical analysis. This approach fails to fully uncover the latent features within measurement data and is ill-suited to complex and ever-changing measurement environments. Furthermore, when dealing with multimodal measurement data, traditional methods cannot effectively integrate features from different data types, resulting in low accuracy and reliability of the assessment results. With the powerful advantages of deep learning in data processing and feature extraction, its application in measurement precision assessment has become a new technological development direction. Therefore, this invention proposes a deep learning-based measurement precision assessment method and system. This method mines deep features from multimodal measurement data using a deep feature extraction model, comprehensively assesses environmental impact by combining measurement state parameters, and constructs a multi-level assessment model to achieve accurate prediction. This overcomes the shortcomings of existing evaluation models, significantly improving the accuracy and adaptability of precision assessment. It provides reliable technical support for high-precision measurements in industrial production, scientific research, and other fields, and is of great significance for promoting the intelligent development of measurement technology. Summary of the Invention

[0004] The purpose of this invention is to provide a precision evaluation method and system for measurement methods based on deep learning.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0006] This invention includes the following steps:

[0007] Historical measurement data and historical calibration markers are acquired. The historical measurement data is divided into measurement source data and measurement values. The measurement source data is input into a deep feature extraction model to obtain multimodal measurement features. The historical calibration markers include calibration values ​​and accuracy markers.

[0008] Based on the measured values, a measurement statistical index is determined, and the measured values ​​are compared with the calibration values ​​to obtain a first accuracy rating coefficient. Based on the measurement statistical index and the measurement standard index depth features, a second accuracy rating coefficient is extracted.

[0009] The environmental state coefficient is calculated based on the measured state parameters. The measured data and the environmental state coefficient are then input into the accuracy evaluation index prediction model to obtain the accuracy evaluation index. The measured state parameters include the sensor usage status and environmental status corresponding to the measurement method.

[0010] A precision evaluation model is constructed based on the precision evaluation index and the precision marker. The measurement data and measurement state parameters of the measurement method to be evaluated are input into the precision evaluation model to obtain the precision evaluation result.

[0011] Furthermore, the method for obtaining multimodal measurement features includes:

[0012] Historical measurement data is acquired and divided into source data and measured values ​​using an SVM classifier. The source data includes image data, text data, and sensor data acquired using the corresponding measurement method. The text data is related to the operation records of the corresponding measurement method. The sensor data is related to the working status of the sensor used in the corresponding measurement method. The measured values ​​are obtained by processing the source data through specific operations of the corresponding measurement method.

[0013] The measurement source data is input into a deep feature extraction model to obtain multimodal measurement features;

[0014] The deep feature extraction model includes a specific feature extraction layer, an adaptive gated fusion module, a cross-modal attention layer, and a feature refinement layer;

[0015] The specific feature extraction layer includes an image quality assessment unit, a sensor operating status analysis unit, and an operation semantic extraction unit; the image quality assessment unit analyzes image data to obtain an image quality vector; the sensor operating status analysis unit performs deformable temporal convolution on sensor data through a multi-scale temporal convolution network to obtain a dynamic change vector of sensor data operation; the operation semantic extraction unit processes text data through an operation key information extraction network to obtain a key operation vector.

[0016] The adaptive gated fusion module is used to dynamically adjust the weights of each modality based on the input image quality vector, dynamic change vector, and key operation vector to perform gated fusion and obtain fused measurement features. The expression is:

[0017] ;

[0018] ;

[0019] in For the first Dynamic weights of each modality It is the sigmoid activation function. This is the weight matrix. This is the GeLU activation function. For the projection matrix, For the number of modes, For the first The feature vector of a modality To integrate measurement features, For modality-specific projection matrices, For the first Quality score vector for each modality;

[0020] The cross-modal attention layer injects fused measurement features into temporal information through positional encoding and outputs quality-guided measurement features through quality-sensitive feature interaction with the fused measurement features via quality-guided attention, expressed as:

[0021] ;

[0022] ;

[0023] in To guide attention to quality, For quality query matrix, The mass bond matrix, This is the mass value matrix. The dimension of the key vector. The learnable quality adjustment coefficient. For the quality rating matrix, For the first The position and the first The relative position encoding vector between positions. This is a learnable position offset vector. Let be the mass decay function. , For position , Feature quality information at the location;

[0024] The feature refinement layer enhances the quality of the quality-guided measurement features through quality-aware residual refinement and noise suppression to obtain multimodal measurement features, expressed as:

[0025] ;

[0026] ;

[0027] in For multimodal measurement characteristics, The feature vector is the refined feature vector of the quality-perceived residual. For noise estimator, For quality-guided measurement characteristics. These are learnable weight coefficients. For layer normalization operation, This is the overall quality vector.

[0028] Furthermore, the method for calculating the first accuracy rating coefficient includes:

[0029] The same measurement method is used to perform n measurements on the same object, and the average of the measurement results is taken as the measured value. The calibration result of the corresponding object is taken as the calibration value. The deviation between the measured value and the calibration value of the same object is calculated as the measurement deviation. The average of the measurement deviations of the same measurement method for different objects is calculated as the first accuracy rating coefficient of the corresponding measurement method.

[0030] Furthermore, the method for calculating the second precision rating coefficient includes:

[0031] The same measurement method is used to perform n measurements on the same object to obtain measurement results, and measurement statistical indicators are calculated based on the measurement results; the measurement statistical indicators include the standard deviation of measurement, the standard uncertainty of the measurement mean, the degrees of freedom, the confidence interval, the range, and the coefficient of variation;

[0032] The second accuracy rating coefficient is calculated based on the measurement statistical indicators and measurement standard indicators of different objects measured using the same measurement method. The expression is as follows:

[0033] ;

[0034] in This is the second accuracy rating coefficient. For a collection of measured objects, To measure the quantity of objects, As an industry-sensitive weight, for The standard deviation of the measurement results of the measured object. The standard deviation is the nominal standard deviation. The linear response exponent with uncertainty for Uncertainty of the measurement result of the measured object For standard uncertainty, As the baseline degree of freedom, for The degrees of freedom of the measurement results of the measured object. for The width of the measurement confidence interval for the measurement results of the measured object. To measure the width of the confidence interval for the standard measurement, The stability sensitivity coefficient, for The measurement range of the measurement results of the measured object. For standard range, The coefficient of variation is the linear response exponent. for The coefficient of variation of the measurement results of the measured object. is the standard coefficient of variation.

[0035] Furthermore, the method for calculating the environmental state coefficient includes:

[0036] The environmental state coefficient is calculated based on the sensor usage status and environmental status corresponding to the measurement method. The expression is as follows:

[0037] ;

[0038] ;

[0039] ;

[0040] in This is the environmental state coefficient. As a weight for sensor health, For sensor health, As the environmental disturbance degree weight, For environmental interference, , , Assign coefficients to sensor health status. , , , Assigning coefficients to environmental disturbance levels. The fatigue coefficient of the material. This represents the average daily usage time of the sensor. This represents the number of repairs conducted in the past six months. The continuous working time of the sensor. To measure the ambient noise, This represents the temperature fluctuation value. This is the rated allowable temperature fluctuation value. This represents humidity fluctuation values. This is the rated allowable humidity fluctuation value. For frequency band coefficients, This is the effective value of vibration acceleration. The standard deviation of uniformity in RGB images. This refers to the color temperature of the light source.

[0041] Furthermore, the method for obtaining the accuracy evaluation index includes:

[0042] The multimodal measurement features, the first accuracy rating coefficient, the second accuracy rating coefficient, and the environmental state coefficient are combined to form an accuracy prediction comprehensive set. The accuracy prediction comprehensive set is randomly divided into a first training set and a first test set in a ratio of 6:4. An accuracy evaluation index prediction model is constructed. The accuracy evaluation index prediction model is trained using the first training set and evaluated using the first test set.

[0043] The accuracy evaluation index prediction model includes a feature enhancement module, an environment interaction module, an accuracy modality fusion layer, and an accuracy evaluation index prediction layer. The feature enhancement module filters the first subset of features most relevant to accuracy evaluation from multimodal measurement features and refines these features to output enhanced multimodal measurement features. The environment interaction module expands environmental state coefficients into multidimensional environmental features and transforms these multidimensional features into environmental features through an environmental feature encoder. The accuracy modality fusion layer fuses the enhanced multimodal measurement features and environmental features through an environment-aware gating fusion unit to obtain environmental measurement features. The accuracy evaluation index prediction layer receives the environmental measurement features, outputs a predicted first accuracy rating coefficient through the P1 prediction branch, a predicted second accuracy rating coefficient through the P2 prediction branch, and a predicted confidence score through the confidence prediction branch. The predicted confidence score is used to quantify the reliability of the model's prediction results.

[0044] The measurement data and environmental state coefficients are input into the accuracy evaluation index prediction model to obtain the accuracy evaluation index; the accuracy evaluation index includes the prediction first accuracy rating coefficient, the prediction second accuracy rating coefficient, and the prediction confidence score.

[0045] Furthermore, the method for obtaining the precision assessment results includes:

[0046] The accuracy evaluation index and environmental state coefficient corresponding to the historical measurement data are combined into a condition vector. The multimodal measurement features, condition vector and accuracy label of the historical measurement data with different accuracy labels are combined into a comprehensive accuracy evaluation set. Random forest is used to randomly divide the comprehensive accuracy evaluation set into a second training set and a second test set in a ratio of 7:3. A precision evaluation model is constructed. The precision evaluation model is trained using the second training set and evaluated using the second test set.

[0047] The precision evaluation model includes a condition generator, a condition discriminator, and a precision level classifier.

[0048] The condition generator is used to generate noise vectors. It uses a cross-attention mechanism and dynamic gating to fuse and filter noise vectors, multimodal measurement features and condition vectors to obtain fused features, and performs regression prediction to output an accuracy score.

[0049] The cross-attention mechanism includes a fusion of quality perception, content-conditional interaction, and context modulation, and is expressed as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] in For quality perception attention, For content-conditional interactive attention, To modulate and integrate attention to the environment To query the projection matrix Acting on the quality query matrix , Key projection matrix Acting on the conditional key matrix , Value projection matrix Acting on the condition value matrix, For the condition key dimension, For the credibility gating function, For credibility score, For the number of attention heads, For the first One point of attention, This is a 3D convolution operation. For content query submatrix, To enhance the key matrix, For conditional submatrices, As an environmental scaling factor, the environmental state coefficients are... enter Function to obtain, To merge the query matrix, For the environment key matrix, The original value matrix, For the environment key dimension, Encode the location of the environment;

[0054] The fusion feature expression is:

[0055] ;

[0056] ;

[0057] in As a feature of fusion, For dynamic gating vectors, For layer normalization operation, Scaling networks to fit the environment For environmental offset networks, To assign credibility scores With environmental state coefficient Perform conditional concatenation;

[0058] The conditional discriminator performs conditional discrimination based on accuracy score, multimodal measurement features, and conditional vector, and outputs the probability of authenticity.

[0059] The accuracy classifier predicts the accuracy level probability distribution of the accuracy score and outputs the accuracy level and probability.

[0060] The measurement data and measurement status parameters of the measurement method to be evaluated are input into the precision evaluation model to obtain the accuracy level and probability.

[0061] Secondly, a precision evaluation system based on a deep learning-based measurement method includes:

[0062] Feature module: Used to input measurement source data into the deep feature extraction model to obtain multimodal measurement features;

[0063] Accuracy rating coefficient module: used to determine the measurement statistical index based on the measured value, compare the measured value with the calibration value to obtain the first accuracy rating coefficient, and calculate the second accuracy rating coefficient based on the measurement statistical index and the measurement standard index;

[0064] Condition coefficient module: used to calculate environmental state coefficients based on measured state parameters, and input the measured data and environmental state coefficients into the accuracy evaluation index prediction model to obtain the accuracy evaluation index;

[0065] Evaluation model module: used to construct a precision evaluation model based on the precision evaluation index and the precision mark, and input the measurement data and measurement state parameters of the measurement method to be evaluated into the precision evaluation model to obtain the precision evaluation result;

[0066] Intelligent monitoring module: used to store, view and manage the precision assessment results, and select a suitable measurement method based on the precision assessment results.

[0067] The beneficial effects of this invention are:

[0068] This invention is a precision evaluation method and system based on deep learning for measurement. Compared with the prior art, this invention has the following technical advantages:

[0069] This invention, through deep feature extraction, calculation of accuracy rating coefficients, calculation of environmental state coefficients, obtaining accuracy evaluation indicators, and model construction steps, can improve data preprocessing capabilities and enhance model adaptability in the precision evaluation of measurement methods, thereby improving the efficiency and accuracy of measurement method precision evaluation. Optimizing the precision evaluation technology of measurement methods can significantly save resources and improve work efficiency. It enables the evaluation of the precision of measurement methods, providing a scientific basis for the selection of measurement methods, and is of great significance for promoting the intelligent development of measurement technology. It can adapt to different precision evaluation systems based on deep learning measurement methods and the precision evaluation needs of different users based on deep learning measurement methods, possessing a certain degree of universality. Attached Figure Description

[0070] Figure 1 This is a flowchart illustrating the steps of a precision evaluation method based on deep learning, as described in this invention. Detailed Implementation

[0071] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0072] The present invention provides a precision evaluation method and system for measurement based on deep learning, comprising the following steps:

[0073] like Figure 1 As shown, this embodiment includes the following steps:

[0074] Historical measurement data and historical calibration markers are acquired. The historical measurement data is divided into measurement source data and measurement values. The measurement source data is input into a deep feature extraction model to obtain multimodal measurement features. The historical calibration markers include calibration values ​​and accuracy markers.

[0075] Based on the measured values, a measurement statistical index is determined, and the measured values ​​are compared with the calibration values ​​to obtain a first accuracy rating coefficient. Based on the measurement statistical index and the measurement standard index depth features, a second accuracy rating coefficient is extracted.

[0076] The environmental state coefficient is calculated based on the measured state parameters. The measured data and the environmental state coefficient are then input into the accuracy evaluation index prediction model to obtain the accuracy evaluation index. The measured state parameters include the sensor usage status and environmental status corresponding to the measurement method.

[0077] A precision evaluation model is constructed based on the precision evaluation index and the precision marker. The measurement data and measurement state parameters of the measurement method to be evaluated are input into the precision evaluation model to obtain the precision evaluation result.

[0078] In this embodiment, the method for obtaining multimodal measurement features includes:

[0079] Historical measurement data is acquired and divided into source data and measured values ​​using an SVM classifier. The source data includes image data, text data, and sensor data acquired using the corresponding measurement method. The text data is related to the operation records of the corresponding measurement method. The sensor data is related to the working status of the sensor used in the corresponding measurement method. The measured values ​​are obtained by processing the source data through specific operations of the corresponding measurement method.

[0080] The measurement source data is input into a deep feature extraction model to obtain multimodal measurement features;

[0081] The deep feature extraction model includes a specific feature extraction layer, an adaptive gated fusion module, a cross-modal attention layer, and a feature refinement layer;

[0082] The specific feature extraction layer includes an image quality assessment unit, a sensor operating status analysis unit, and an operation semantic extraction unit; the image quality assessment unit analyzes image data to obtain an image quality vector; the sensor operating status analysis unit performs deformable temporal convolution on sensor data through a multi-scale temporal convolution network to obtain a dynamic change vector of sensor data operation; the operation semantic extraction unit processes text data through an operation key information extraction network to obtain a key operation vector.

[0083] The adaptive gated fusion module is used to dynamically adjust the weights of each modality based on the input image quality vector, dynamic change vector, and key operation vector to perform gated fusion and obtain fused measurement features. The expression is:

[0084] ;

[0085] ;

[0086] in For the first Dynamic weights of each modality It is the sigmoid activation function. This is the weight matrix. This is the GeLU activation function. For the projection matrix, For the number of modes, For the first The feature vector of a modality To integrate measurement features, For modality-specific projection matrices, For the first Quality score vector for each modality;

[0087] The cross-modal attention layer injects fused measurement features into temporal information through positional encoding and outputs quality-guided measurement features through quality-sensitive feature interaction with the fused measurement features via quality-guided attention, expressed as:

[0088] ;

[0089] ;

[0090] in To guide attention to quality, For quality query matrix, The mass bond matrix, This is the mass value matrix. The dimension of the key vector. The learnable quality adjustment coefficient. For the quality rating matrix, For the first The position and the first The relative position encoding vector between positions. This is a learnable position offset vector. Let be the mass decay function. , For position , Feature quality information at the location;

[0091] The feature refinement layer enhances the quality of the quality-guided measurement features through quality-aware residual refinement and noise suppression to obtain multimodal measurement features, expressed as:

[0092] ;

[0093] ;

[0094] in For multimodal measurement characteristics, The feature vector is the refined feature vector of the quality-perceived residual. For noise estimator, For quality-guided measurement characteristics. These are learnable weight coefficients. For layer normalization operation, This is the overall quality vector;

[0095] In the actual evaluation, in the image quality evaluation unit, the sharpness score of each pixel in the image data is calculated using Laplacian variance and three layers of sharpness feature maps at different scales (32×32, 16×16, 8×8) are output. The image data is converted to the Lab color space and compared block by block with the pre-stored standard color card to generate a color deviation heatmap. Color deviation features are extracted through a convolutional network. Multi-scale LBP feature extraction is performed on the image data and local binary mode entropy is calculated. Texture consistency analysis is performed to obtain texture quality features. The sharpness features, color deviation features and texture quality features at different scales are spliced ​​and fused to output an image quality vector.

[0096] In the sensor operating status analysis unit, the sensor time series data (length N×channel C) is input into a multi-scale temporal convolutional network, and a dynamic change vector is output through three deformable temporal convolutional modules (dilation coefficients of 1, 3 and 5 respectively).

[0097] In the adaptive gating fusion module, the original features of each modality are quantized by an MLP network to generate quality scores. A modality-specific fully connected (FC) layer is used to unify the original features of each modality into a unified feature space and output projected features. A learnable gating method is used to dynamically allocate and fuse the quality scores and the original features of each modality to obtain modal weights. Element-wise multiplication is used to perform quality weighting on the projected features and the quality scores to output quality weighted features. The fused measurement features are obtained by weighting and summing the quality weighted features based on the modal weights and the weighted features.

[0098] In the cross-modal attention layer, the fused measurement features are split into triples (query matrix) through linear projection. Key matrix Sum matrix The sequence position is obtained by learning relative encoding of the fused measurement features, and temporal information is injected according to the positional deviation. A learnable quality adjustment coefficient is then selected. The quality-guided attention is calculated using the quality matrix addition method, which incorporates time-series information into the fused measurement features. The quality-guided attention and the fused measurement features are then connected through a residual connection unit to output the quality-guided measurement features.

[0099] In the feature refinement layer, learnable weight coefficients are selected. The quality-guided measurement features are concatenated with the quality scores from the adaptive gating fusion module to obtain concatenated features. Then, the concatenated features are refined using residuals through a multilayer perceptron (MLP) and a learnable mechanism to obtain intermediate features. An adaptive gating mechanism is used to control intermediate features. Noise suppression and layer normalization are performed to obtain multimodal measurement characteristics.

[0100] In this embodiment, the method for calculating the first accuracy rating coefficient includes:

[0101] The same measurement method is used to perform n measurements on the same object, and the average of the measurement results is taken as the measured value. The calibration result of the corresponding object is taken as the calibration value. The deviation between the measured value and the calibration value of the same object is calculated as the measurement deviation. The average of the measurement deviations of the same measurement method for different objects is calculated as the first accuracy rating coefficient of the corresponding measurement method.

[0102] In actual assessment, taking the calculation of the first accuracy rating coefficient of historical data as an example, the A1 measurement method is used to perform 20 measurements on the b-measured object, and the average of the measurement results is taken as the measured value. , For measuring dimensions, The measurement dimension set is set to 3 (including size, sphericity, and surface roughness), and the calibration result of the measured object b is taken as the calibration value. Calculate measurement deviation Measurement method A is used to measure the set of measurement objects B (standard spheres, standard cubes, and standard cylinders) and calculate the first accuracy rating coefficient. .

[0103] In this embodiment, the method for calculating the second precision rating coefficient includes:

[0104] The same measurement method is used to perform n measurements on the same object to obtain measurement results, and measurement statistical indicators are calculated based on the measurement results; the measurement statistical indicators include the standard deviation of measurement, the standard uncertainty of the measurement mean, the degrees of freedom, the confidence interval, the range, and the coefficient of variation;

[0105] The second accuracy rating coefficient is calculated based on the measurement statistical indicators and measurement standard indicators of different objects measured using the same measurement method. The expression is as follows:

[0106] ;

[0107] in This is the second accuracy rating coefficient. For a collection of measured objects, To measure the quantity of objects, As an industry-sensitive weight, for The standard deviation of the measurement results of the measured object. The standard deviation is the nominal standard deviation. The linear response exponent with uncertainty for Uncertainty of the measurement result of the measured object For standard uncertainty, As the baseline degree of freedom, for The degrees of freedom of the measurement results of the measured object. for The width of the measurement confidence interval for the measurement results of the measured object. To measure the width of the confidence interval for the standard measurement, The stability sensitivity coefficient, for The measurement range of the measurement results of the measured object. For standard range, The coefficient of variation is the linear response exponent. for The coefficient of variation of the measurement results of the measured object. The standard coefficient of variation;

[0108] In practical assessments, taking the calculation of the second-precision rating coefficient based on historical data as an example, the A1 measurement method is used to measure the set of measured objects B (industrial testing industry). Each measured object is measured 20 times. Measurement statistics are calculated based on the measurement results, and industry-sensitive weights are taken. =0.4, uncertainty linear response exponent Stability sensitivity coefficient Coefficient of variation and linear response exponent Rated standard deviation Standard uncertainty Reference degrees of freedom Standard measurement confidence interval width Standard range Standard coefficient of variation The second accuracy rating coefficient is calculated by combining the measurement standard indicators. .

[0109] In this embodiment, the method for calculating the environmental state coefficient includes:

[0110] The environmental state coefficient is calculated based on the sensor usage status and environmental status corresponding to the measurement method. The expression is as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] in This is the environmental state coefficient. As a weight for sensor health, For sensor health, As the environmental disturbance degree weight, For environmental interference, , , Assign coefficients to sensor health status. , , , Assigning coefficients to environmental disturbance levels. The fatigue coefficient of the material. This represents the average daily usage time of the sensor. This represents the number of repairs conducted in the past six months. The continuous working time of the sensor. To measure the ambient noise, This represents the temperature fluctuation value. This is the rated allowable temperature fluctuation value. This represents humidity fluctuation values. This is the rated allowable humidity fluctuation value. For frequency band coefficients, This is the effective value of vibration acceleration. The standard deviation of uniformity in RGB images. Color temperature of the light source;

[0115] In practical assessments, taking the calculation of the environmental state coefficient using the A1 measurement method as an example, the sensor health weight is taken. Environmental interference weight Sensor health factor allocation coefficient is 0.4 / 0.4 / 0.2; environmental interference factor allocation coefficient is 0.3 / 0.3 / 0.2 / 0.2; material fatigue factor... Frequency band coefficient The environmental condition coefficient is calculated to be 0.57 based on the usage and environmental conditions.

[0116] In this embodiment, the method for obtaining the accuracy evaluation index includes:

[0117] The multimodal measurement features, the first accuracy rating coefficient, the second accuracy rating coefficient, and the environmental state coefficient are combined to form an accuracy prediction comprehensive set. The accuracy prediction comprehensive set is randomly divided into a first training set and a first test set in a ratio of 6:4. An accuracy evaluation index prediction model is constructed. The accuracy evaluation index prediction model is trained using the first training set and evaluated using the first test set.

[0118] The accuracy evaluation index prediction model includes a feature enhancement module, an environment interaction module, an accuracy modality fusion layer, and an accuracy evaluation index prediction layer. The feature enhancement module filters the first subset of features most relevant to accuracy evaluation from multimodal measurement features and refines these features to output enhanced multimodal measurement features. The environment interaction module expands environmental state coefficients into multidimensional environmental features and transforms these multidimensional features into environmental features through an environmental feature encoder. The accuracy modality fusion layer fuses the enhanced multimodal measurement features and environmental features through an environment-aware gating fusion unit to obtain environmental measurement features. The accuracy evaluation index prediction layer receives the environmental measurement features, outputs a predicted first accuracy rating coefficient through the P1 prediction branch, a predicted second accuracy rating coefficient through the P2 prediction branch, and a predicted confidence score through the confidence prediction branch. The predicted confidence score is used to quantify the reliability of the model's prediction results.

[0119] The measurement data and environmental state coefficients are input into the accuracy evaluation index prediction model to obtain the accuracy evaluation index; the accuracy evaluation index includes a first accuracy rating coefficient, a second accuracy rating coefficient, and a prediction confidence score.

[0120] In actual assessment, historical measurement data is used to obtain the corresponding multimodal measurement characteristics, the first accuracy rating coefficient (predicted by the accuracy evaluation index prediction model), the second accuracy rating coefficient (predicted by the accuracy evaluation index prediction model), and the environmental state coefficient, and these are combined to form an accuracy prediction comprehensive set.

[0121] In the feature enhancement module, an attention-based feature gating algorithm is used to process multimodal measurement features, and residual feature transformation is used to refine the selected multimodal measurement features and output enhanced multimodal measurement features.

[0122] The environmental state coefficient is included in the environmental interaction module. Expanding to multidimensional environmental features The environmental feature encoder uses the Swish activation function;

[0123] The expression for the environment perception-gated fusion algorithm in the precision modality fusion layer is:

[0124] ;

[0125] ;

[0126] in For environmental measurement characteristics, To integrate weights, To enhance multimodal measurement characteristics The weight matrix, Environmental characteristics The weight matrix, Environmental gating factor;

[0127] In the accuracy evaluation index prediction layer, the P1 prediction branch uses the Huber Loss loss function, the P2 prediction branch uses the Focal Loss loss function, and the confidence prediction branch uses the prediction error as a supervision signal.

[0128] During model training, feature noise injection (Gaussian noise) and ECC perturbation (set to ±20% random fluctuation) are used for data augmentation, and AdamW (learning rate is scheduled by cosine annealing, batch size is 64) is used to adjust the model hyperparameters.

[0129] The multimodal measurement features of the A2 measurement method to be evaluated (obtained through a deep feature extraction model) and environmental state coefficients (calculated based on usage and environmental conditions) are input into the accuracy evaluation index prediction model to obtain the accuracy evaluation indexes: prediction first accuracy rating coefficient 0.0129, prediction second accuracy rating coefficient 0.098, and prediction confidence score 0.95.

[0130] In this embodiment, the method for obtaining the precision evaluation result includes:

[0131] The accuracy evaluation index and environmental state coefficient corresponding to the historical measurement data are combined into a condition vector. The multimodal measurement features, condition vector and accuracy label of the historical measurement data with different accuracy labels are combined into a comprehensive accuracy evaluation set. Random forest is used to randomly divide the comprehensive accuracy evaluation set into a second training set and a second test set in a ratio of 7:3. A precision evaluation model is constructed. The precision evaluation model is trained using the second training set and evaluated using the second test set.

[0132] The precision evaluation model includes a condition generator, a condition discriminator, and a precision level classifier.

[0133] The condition generator is used to generate noise vectors. It uses a cross-attention mechanism and dynamic gating to fuse and filter noise vectors, multimodal measurement features and condition vectors to obtain fused features, and performs regression prediction to output an accuracy score.

[0134] The cross-attention mechanism includes a fusion of quality perception, content-conditional interaction, and context modulation, and is expressed as follows:

[0135] ;

[0136] ;

[0137] ;

[0138] in For quality perception attention, For content-conditional interactive attention, To modulate and integrate attention to the environment To query the projection matrix Acting on the quality query matrix , Key projection matrix Acting on the conditional key matrix , Value projection matrix Acting on the condition value matrix, For the condition key dimension, For the credibility gating function, For credibility score, For the number of attention heads, For the first One point of attention, This is a 3D convolution operation. For content query submatrix, To enhance the key matrix, For conditional submatrices, As an environmental scaling factor, the environmental state coefficients are... enter Function to obtain, To merge the query matrix, For the environment key matrix, The original value matrix, For the environment key dimension, Encode the location of the environment;

[0139] The fusion feature expression is:

[0140] ;

[0141] ;

[0142] in As a feature of fusion, For dynamic gating vectors, For layer normalization operation, Scaling networks to fit the environment For environmental offset networks, To assign credibility scores With environmental state coefficient Perform conditional concatenation;

[0143] The conditional discriminator performs conditional discrimination based on accuracy score, multimodal measurement features, and conditional vector, and outputs the probability of authenticity.

[0144] The accuracy classifier predicts the accuracy level probability distribution of the accuracy score and outputs the accuracy level and probability.

[0145] The measurement data and measurement status parameters of the measurement method to be evaluated are input into the precision evaluation model to obtain the precision evaluation result; the precision evaluation result includes the accuracy level and probability.

[0146] In actual assessment, the corresponding accuracy evaluation index (predicted by the accuracy evaluation index prediction model) is obtained based on historical measurement data with different accuracy markers, and a condition vector is formed with the environmental state coefficient. The condition vector, multimodal measurement features and accuracy markers are then combined to form a comprehensive accuracy evaluation set.

[0147] In the condition generator, the multimodal measurement features are first decoupled to obtain multimodal measurement features represented by concatenated content features and quality features. The condition vector is then encoded using conditional pyramid to obtain a multi-scale embedded condition vector. The noise vector is then encoded using noise. The noise vector, multimodal measurement features, and condition vector are fused and filtered to obtain fused features. Based on the fused features, regression prediction is performed to output an accuracy score.

[0148] In the conditional discriminator, the accuracy score S is first expanded from a scalar to an accuracy feature vector. The conditional projection adversarial loss is used to conditionally project the conditional vector. The feature pyramid is used to fuse multimodal measurement features. The precision feature vector, conditional vector and multimodal measurement features are concatenated to obtain the discriminative concatenated features. The discriminative concatenated features are convolved to output the authenticity probability (0 or 1, 0 indicates that the output result is not true, 1 indicates that the output result is true).

[0149] The conditional projection adversarial loss expression is as follows:

[0150] ;

[0151] in Let the conditional projection adversarial loss function be... The expected value is used to calculate the average loss under the data distribution. For the discriminator model, This is a concatenated vector of conditional vectors and multimodal measurement features. The actual score corresponding to the precision marker. This is a generator model used to generate vectors. and noise vector Generate fake accuracy scores;

[0152] In the accuracy-level classifier, the accuracy score S is first expanded from a scalar to an accuracy feature vector. The precision feature vector and the confidence score are concatenated and then input into the confidence attention mechanism to obtain the confidence attention weight. The precision feature vector and the confidence attention weight are then input into the classification layer to output the precision level and precision probability. The expression of the confidence attention mechanism is as follows:

[0153] ;

[0154] in This is the weighted precision feature vector. For learnable weight matrix, For accuracy scoring and credibility score Concatenate vectors, For precision feature vectors;

[0155] The accuracy evaluation index of the measurement method to be evaluated (prediction first accuracy rating coefficient 0.0129, prediction second accuracy rating coefficient 0.098, and prediction confidence score 0.95), environmental state coefficient 0.57, and multimodal measurement features are input into the precision evaluation model. The condition generator outputs an accuracy score of 0.93, and the condition discriminator outputs a truth probability of 1 (indicating that the output of the accuracy level classifier is true and reliable). The output of the accuracy level classifier (accuracy level: high precision, probability 0.9) is used as the precision evaluation result of the A1 measurement method.

[0156] The correspondence between the accuracy score of 0.93 and the accuracy score range is as follows: Ultra-precision: (0.95, 1], High precision: (0.9, 0.95], Precision: (0.8, 0.9], Pass: (0.65, 0.8], Fail: [0, 0.65].

[0157] Secondly, a precision evaluation system based on a deep learning-based measurement method includes:

[0158] Feature module: Used to input measurement source data into the deep feature extraction model to obtain multimodal measurement features;

[0159] Accuracy rating coefficient module: used to determine the measurement statistical index based on the measured value, compare the measured value with the calibration value to obtain the first accuracy rating coefficient, and calculate the second accuracy rating coefficient based on the measurement statistical index and the measurement standard index;

[0160] Condition coefficient module: used to calculate environmental state coefficients based on measured state parameters, and input the measured data and environmental state coefficients into the accuracy evaluation index prediction model to obtain the accuracy evaluation index;

[0161] Evaluation model module: used to construct a precision evaluation model based on the precision evaluation index and the precision mark, and input the measurement data and measurement state parameters of the measurement method to be evaluated into the precision evaluation model to obtain the precision evaluation result;

[0162] Intelligent monitoring module: used to store, view and manage the precision assessment results, and select a suitable measurement method based on the precision assessment results.

[0163] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the precision of a measurement method based on deep learning, characterized by, The method comprises the following steps: S1, obtaining historical measurement data and historical calibration marks, dividing the historical measurement data into measurement source data and measurement values, and inputting the measurement source data into a deep feature extraction model to obtain multi-modal measurement features; The historical calibration marks comprise calibration values and precision marks; S2, determining measurement statistical indicators according to the measurement values, comparing the measurement values with the calibration values to obtain a first precision rating coefficient, and calculating a second precision rating coefficient according to the measurement statistical indicators and measurement standard indicators; S3, calculating an environmental state coefficient according to measurement state parameters, inputting the measurement data and the environmental state coefficient into a precision evaluation index prediction model to obtain a precision evaluation index; the measurement state parameters comprise a sensor usage state and an environmental state corresponding to a measurement method; S4, constructing a precision evaluation model according to the precision evaluation index and the precision marks, inputting measurement data and measurement state parameters of a measurement method to be evaluated into the precision evaluation model to obtain a precision evaluation result; The measurement source data comprises image data, text data and sensor data obtained by using a corresponding measurement method; the text data is related to operation records of the corresponding measurement method; and the sensor data is related to a sensor working state used by the corresponding measurement method; The measurement values are obtained by processing the measurement source data through specific operations of the corresponding measurement method; The method for calculating the environmental state coefficient comprises: The environmental state coefficient is calculated according to the sensor usage state and the environmental state corresponding to the measurement method, and the expression is as follows: ； ； ； wherein is an environmental state coefficient, is a sensor health degree weight, is a sensor health degree, is an environmental interference degree weight, is an environmental interference degree, , , is a sensor health degree distribution coefficient, , , , is an environmental interference degree distribution coefficient, is a material fatigue coefficient, is a sensor daily average usage time, is a maintenance frequency in the past half year, is a sensor continuous working time, is a measured environmental noise, is a temperature fluctuation value, is a rated allowable temperature fluctuation value, is a humidity fluctuation value, is a rated allowable humidity fluctuation value, is a frequency band coefficient, is a vibration acceleration effective value, is an RGB image uniformity standard deviation, is a light source color temperature The construction method of the precision evaluation index prediction model is as follows: The multi-modal measurement features, the first precision rating coefficient, the second precision rating coefficient and the environmental state coefficient are combined into a precision prediction comprehensive set, the precision prediction comprehensive set is randomly divided into a first training set and a first test set according to a ratio of 6:4, the precision evaluation index prediction model is constructed, the first training set is used to train the precision evaluation index prediction model, and the first test set is used to evaluate the performance of the precision evaluation index prediction model; The construction method of the precision evaluation model is as follows: The precision evaluation index corresponding to the historical measurement data and the environmental state coefficient are combined into a condition vector, the multi-modal measurement features, the condition vector and the precision marks of the historical measurement data with different precision marks are combined into a precision evaluation comprehensive set, the precision evaluation comprehensive set is randomly divided into a second training set and a second test set according to a ratio of 7:3 by using a random forest, the precision evaluation model is constructed, the second training set is used to train the precision evaluation model, and the second test set is used to evaluate the precision evaluation model.

2. The method for evaluating the precision of a measurement method based on deep learning according to claim 1, characterized in that, The method for obtaining the multi-modal measurement features comprises: Obtaining historical measurement data, and dividing the historical measurement data into measurement source data and measurement values by using an SVM classifier; Inputting the measurement source data into a deep feature extraction model to obtain multi-modal measurement features; The deep feature extraction model comprises a specific feature extraction layer, an adaptive gate fusion module, a cross-modal attention layer and a feature refining layer; The specific feature extraction layer comprises an image quality evaluation unit, a sensor working state analysis unit and an operation semantic extraction unit; the image quality evaluation unit analyzes image data to obtain an image quality vector; the sensor working state analysis unit obtains a dynamic change vector of sensor data working through deformable time sequence convolution of the sensor data by a multi-scale time sequence convolution network; and the operation semantic extraction unit obtains a key operation vector by processing text data through an operation key information extraction network; The adaptive gating fusion module is used for dynamically adjusting the weights of each modality to obtain a fusion measurement feature through gating fusion according to the input image quality vector, the dynamic change vector and the key operation vector, and the expression is: ； ； wherein is a dynamic weight for the th modality, is a sigmoid activation function, is a weight matrix, is a GeLU activation function, is a projection matrix, is a number of modalities, is a feature vector for the th modality, is a fused measurement feature, is a modality-specific projection matrix, is a quality score vector for the th modality; The cross-modal attention layer injects the fusion measurement feature into time sequence information through position encoding, and outputs quality-guided measurement features through quality-sensitive feature interaction of the fusion measurement feature through quality-guided attention, and the expression is: ； ； wherein is a quality attention, is a quality query matrix, is a quality key matrix, is a quality value matrix, is a key vector dimension, is a learnable quality adjustment coefficient, is a quality score matrix, is a relative position encoding vector between the th position and the th position, is a learnable position offset vector, is a quality decay function, , is a feature quality information at position , . The feature refining layer performs quality enhancement on the quality-guided measurement features through quality-aware residual refining and noise suppression to obtain a multi-modal measurement feature, and the expression is: ； ； wherein is a multi-modal measurement feature, is a quality-aware residual refined feature vector, is a noise estimator, is a quality-guided measurement feature, is a learnable weight coefficient, is a layer normalization operation, is a comprehensive quality vector.

3. The method of claim 1, wherein the method of precision evaluation is based on deep learning. The method for calculating the first accuracy rating coefficient comprises: The same measurement method is used to measure the same measurement object n times to obtain a measurement result, and the measurement result is taken as a measurement value; a calibration result of the corresponding measurement object is taken as a calibration value; a measurement deviation of the same measurement object is calculated as a measurement deviation of the same measurement method; and a measurement deviation of different measurement objects is calculated as a first accuracy rating coefficient of the corresponding measurement method.

4. The method for evaluating the precision of a measurement method based on deep learning according to claim 1, characterized in that, The method for calculating the second accuracy rating coefficient comprises: The same measurement method is used to measure the same measurement object n times to obtain a measurement result, and a measurement statistical index is calculated according to the measurement result; the measurement statistical index comprises a measurement standard deviation, a standard uncertainty of a measurement mean value, a degree of freedom, a confidence interval, a range and a coefficient of variation; A second accuracy rating coefficient is calculated according to the measurement statistical index and the measurement standard index of different measurement objects measured by the same measurement method, and the expression is: ； wherein is a second precision rating coefficient, is a set of measurands, is a number of measurands, is an industry sensitivity weight, is a measurement standard deviation of the measurand measurement, is a nominal standard deviation, is an uncertainty linear response index, is an uncertainty of the measurand measurement, is a standard uncertainty, is a reference degrees of freedom, is a degrees of freedom of the measurand measurement, is a measurement confidence interval width of the measurand measurement, is a standard measurement confidence interval width, is a stability sensitivity coefficient, is a measurement range of the measurand measurement, is a standard range, is a coefficient of variation linear response index, is a coefficient of variation of the measurand measurement, is a standard coefficient of variation.

5. The method for evaluating the precision of a measurement method based on deep learning according to claim 1, characterized in that, The method for obtaining the accuracy evaluation index comprises: The measurement data and the environmental state coefficient are input into an accuracy evaluation index prediction model to obtain an accuracy evaluation index; the accuracy evaluation index comprises a predicted first accuracy rating coefficient, a predicted second accuracy rating coefficient and a predicted confidence score. The precision evaluation index prediction model comprises a feature enhancement module, an environment interaction module, a precision modal fusion layer, and a precision evaluation index prediction layer; the feature enhancement module is used for screening a first feature subset most relevant to precision evaluation from multi-modal measurement features, and performing feature refining to output enhanced multi-modal measurement features; the environment interaction module expands an environment state coefficient into an environment multi-dimensional feature, and converts the environment multi-dimensional feature into an environment feature through an environment feature encoder; the precision modal fusion layer fuses the enhanced multi-modal measurement features and the environment feature through an environment perception gating fusion unit to obtain an environment measurement feature; the precision evaluation index prediction layer receives the environment measurement feature, outputs a predicted first precision rating coefficient through a P1 prediction branch, outputs a predicted second precision rating coefficient through a P2 prediction branch, and outputs a predicted credibility score through a credibility prediction branch; the predicted credibility score is used for quantifying the reliability of the model prediction result.

6. The method for evaluating the precision of a measurement method based on deep learning according to claim 1, characterized in that, The method for obtaining the precision evaluation result comprises: inputting measurement data and measurement state parameters of a measurement method to be evaluated into the precision evaluation model to obtain a precision evaluation result; the precision evaluation result comprises a precision level and a probability; the precision evaluation model comprises a condition generator, a condition discriminator, and a precision level classifier; the condition generator is used for generating a noise vector, fusing and screening the noise vector, multi-modal measurement features, and a condition vector through a cross-attention mechanism and a dynamic gating to obtain fusion features, and performing regression prediction to output a precision score; the cross-attention mechanism comprises quality perception, content-condition interaction, and environment modulation fusion, and the expression is: ； ； ； wherein is a quality-aware attention, is a content-condition interaction attention, is an environment modulation fusion attention, is a query projection matrix acts on a quality query matrix , is a key projection matrix acts on a conditional key matrix , is a value projection matrix acts on a conditional value matrix, is a conditional key dimension, is a confidence gating function, is a confidence score, is a number of attention heads, is an attention head, is a 3D convolution operation, is a content query sub-matrix, is an augmented key sub-matrix, is a conditional value sub-matrix, is an environment scaling factor that scales the environment state coefficient is input to a function, is a fusion query matrix, is an environment key matrix, is an original value matrix, is an environment key dimension, is an environment positional encoding; the expression of the fusion features is: ； ； wherein is a fusion feature, is a dynamic gating vector, is a layer normalization operation, is an environment scaling network, is an environment offset network, is to combine the confidence scores with the environment state coefficients to conditionally concatenate; the condition discriminator performs condition identification according to the precision score, the multi-modal measurement features, and the condition vector to output a truth probability; the precision level classifier performs precision level probability distribution prediction on the precision score to output a precision level and a probability.

7. A precision evaluation system of a measurement method based on deep learning, used to perform the method of any one of claims 1-6, comprising: a feature module: used for inputting measurement source data into a deep feature extraction model to obtain multi-modal measurement features; a precision rating coefficient module: used for determining measurement statistical indicators according to measurement values, obtaining a first precision rating coefficient by comparing the measurement values with calibration values, and calculating a second precision rating coefficient according to the measurement statistical indicators and measurement standard indicators; a condition coefficient module: used for calculating an environment state coefficient according to measurement state parameters, and inputting measurement data and the environment state coefficient into a precision evaluation index prediction model to obtain precision evaluation indexes; an evaluation model module: used for constructing a precision evaluation model according to the precision evaluation indexes and the precision labels, inputting measurement data and measurement state parameters of a measurement method to be evaluated into the precision evaluation model to obtain a precision evaluation result; an intelligent supervision module: used for storing, viewing, and managing the precision evaluation result, and selecting a measurement method meeting a condition according to the precision evaluation result.

Images