Automatic control and error analysis method and system for glass measuring vessel capacity verification process

Abstract

The application provides a glass measuring container capacity verification process automatic control and error analysis method and system, relates to the technical field of automatic control, and comprises the following steps: calculating a compensation value through an environmental parameter, adopting a double-branch neural network to analyze a liquid level image, combining weight data of an electronic balance to perform multi-modal feature fusion, inputting error data into a recurrent neural network with a hierarchical memory structure to perform dynamic error analysis, obtaining a dynamic compensation value, and adjusting a control parameter to perform capacity verification. The application realizes high-accuracy automation of glass measuring container capacity verification, reduces human error, and improves verification efficiency and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automation control, in particular to a glass container capacity verification process automation control and error analysis method and system. BACKGROUND

[0002] Glass containers are basic measuring instruments commonly used in laboratories to measure capacity, including standard glass containers, commonly used glass containers and special glass containers, etc., and the capacity accuracy directly affects the accuracy of experimental and analysis results. The existing glass container capacity verification mainly adopts the weighing method, that is, the capacity is calculated by measuring the weight difference before and after filling water, and the influence of environmental temperature, air pressure, humidity and other factors on the measurement results needs to be considered. In practical application, the following defects and deficiencies still exist:

[0003] The compensation method of the existing verification system for the influence of environmental parameters is relatively simple, and linear models or table lookup methods are mainly used for simple correction, which cannot accurately reflect the complex nonlinear relationship between environmental factors and the characteristics of glass container materials, shapes and other characteristics, resulting in insufficient compensation accuracy.

[0004] The existing liquid level recognition technology is mainly based on static image processing, and it is difficult to accurately capture the dynamic change characteristics of the liquid surface, especially when the liquid surface is shaking or there are bubbles, the recognition result is easy to deviate, and high-accuracy liquid level interpretation cannot be realized.

[0005] The existing verification system lacks systematic error analysis and dynamic compensation mechanism, and it is difficult to evaluate and correct the cumulative error generated in the verification process in real time, resulting in a decrease in accuracy during a long verification process. Especially for multi-batch and large-scale container verification tasks, it is difficult to maintain high-accuracy consistency. SUMMARY

[0006] The embodiments of the present application provide a glass container capacity verification process automation control and error analysis method and system, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiments of the present application provide a glass container capacity verification process automation control and error analysis method, comprising:

[0008] According to the specification parameter information of the glass container to be tested and the environmental temperature, air pressure and humidity, the influence compensation value of the environment on the capacity of the glass container is calculated;

[0009] A double-branch neural network structure is used to dynamically capture and analyze the liquid level image of the glass container, wherein the first branch extracts the liquid surface static characteristics through a spatial attention module, and the second branch obtains the liquid surface dynamic characteristics through a time sequence attention module. The liquid surface static characteristics, liquid surface dynamic characteristics and weight data of the electronic balance are fused in a multi-modal feature fusion manner, and the compensated capacity value is calculated in combination with the influence compensation value;

[0010] performing real-time error evaluation on the compensated capacity value to obtain initial error data;

[0011] inputting the initial error data into a recurrent neural network with a hierarchical memory structure to perform dynamic error analysis, acquiring instantaneous change characteristics of liquid surface state and environmental parameters through a short-term memory unit, analyzing error accumulation characteristics through a long-term memory unit, and adopting a bidirectional attention mechanism to adaptively weight error characteristics of different time scales to obtain a dynamic compensation value;

[0012] adjusting verification control parameters according to the dynamic compensation value, performing capacity verification on the glass measuring cylinder, and obtaining verification result data.

[0013] In an optional implementation,

[0014] The step of calculating the compensation value of the influence of the environment on the capacity of the glass measuring cylinder according to the specification parameter information of the glass measuring cylinder to be verified and the environmental temperature, air pressure and humidity comprises:

[0015] The specification parameter information comprises a nominal capacity and a use temperature range of the glass measuring cylinder;

[0016] The volume thermal deformation amount of the glass measuring cylinder and the density change amount of the liquid to be measured are calculated according to the specification parameter information and the environmental temperature data, and the volume thermal deformation amount and the density change amount are coupled to calculate a temperature compensation coefficient; the liquid surface tension change value and the water vapor partial pressure are calculated according to the air pressure data and the relative humidity data, and the surface tension change value and the water vapor partial pressure are weighted to calculate an environmental compensation coefficient;

[0017] The temperature compensation coefficient and the environmental compensation coefficient are combined to calculate a comprehensive compensation coefficient; the compensation value of the influence of the environment on the capacity of the glass measuring cylinder is calculated according to the comprehensive compensation coefficient, the nominal capacity and the current measurement value; the deviation of the compensated capacity value from the nominal capacity is calculated, and when the deviation exceeds a preset deviation threshold, the weights of the temperature compensation coefficient and the environmental compensation coefficient are dynamically adjusted.

[0018] In an optional implementation,

[0019] The first branch extracts liquid surface static characteristics through a spatial attention module, and the second branch acquires liquid surface dynamic characteristics through a time series attention module; the liquid surface static characteristics, the liquid surface dynamic characteristics and the weight data of the electronic balance are subjected to multi-modal feature fusion, and the compensation value of the influence is calculated to obtain the compensated capacity value.

[0020] The liquid surface shape feature is extracted by a spatial attention module in the first branch, the spatial attention module enhances the liquid surface profile curvature and capillary effect feature based on channel attention weight and position attention weight, and a liquid surface static feature is obtained; the liquid surface oscillation feature is modeled by a time sequence attention module in the second branch, the time sequence attention module calculates the correlation degree between multiple time sequence states based on time window weight, and dynamically aggregates the correlation degree by a convolution long short-term memory network to obtain a liquid surface dynamic feature; the liquid surface static feature and the liquid surface dynamic feature are subjected to feature alignment and normalization processing to obtain a liquid level image feature;

[0021] The weight data of the electronic balance is mapped to a feature space to obtain weight data features;

[0022] The liquid level image feature and the weight data feature are subjected to multi-modal feature fusion to obtain a fusion feature, and a compensated capacity value of the glass measuring cylinder is calculated based on the fusion feature and the influence compensation value.

[0023] In an optional implementation,

[0024] The step of performing multi-modal feature fusion on the liquid level image feature and the weight data feature to obtain a fusion feature, and calculating a compensated capacity value of the glass measuring cylinder based on the fusion feature and the influence compensation value includes:

[0025] The liquid level image feature and the weight data feature are subjected to normalization processing to obtain normalized features, self-attention and cross-modal attention between the normalized features are calculated through a multi-head attention mechanism to obtain a feature interaction matrix;

[0026] A feature confidence evaluation network is constructed based on the feature interaction matrix, the feature confidence evaluation network obtains a feature confidence score by calculating the contribution degree of each modal feature to capacity prediction; and a dynamic fusion weight is obtained by normalizing the feature confidence score based on a temperature parameter;

[0027] The dynamic fusion weight is combined with the normalized features by weighting, and an enhanced fusion feature is obtained through a residual connection; a multi-objective loss function including capacity prediction, modal consistency and feature reconstruction is constructed, and the enhanced fusion feature is optimized;

[0028] The optimized fusion features are spliced with the influence compensation value and input into a full connection layer, a capacity prediction value is calculated through a multi-layer perception, a prediction error between the capacity prediction value and a real capacity value is calculated, a comprehensive weight update gradient is calculated based on the prediction error and the multi-objective loss function, and the dynamic fusion weight is adaptively adjusted by using the comprehensive weight update gradient; feature fusion is performed again by using the updated dynamic fusion weight, and the capacity value of the glass measuring cylinder is predicted in combination with the influence compensation value.

[0029] In an alternative embodiment,

[0030] The initial error data is input into a recurrent neural network with a hierarchical memory structure for dynamic error analysis, the instantaneous change characteristics of the liquid level state and the environmental parameters are obtained through a short-term memory unit, the error accumulation characteristics are analyzed through a long-term memory unit, and a bidirectional attention mechanism is used to adaptively weight the error characteristics of different time scales to obtain a dynamic compensation value, and the steps include:

[0031] The short-term memory unit models the instantaneous changes of the liquid level state parameters and the environmental parameters through a state vector, a forgetting gate and an input gate to obtain short-term state characteristics, and the long-term memory unit dynamically aggregates the short-term state characteristics based on error accumulation and a long-term state update function to obtain long-term state characteristics;

[0032] A bidirectional attention mechanism is constructed for the short-term state characteristics and the long-term state characteristics, the attention scores between the characteristics of different time scales are calculated through a query matrix and a key-value matrix, and the attention scores are normalized to obtain attention weights;

[0033] The short-term state characteristics and the long-term state characteristics are multi-scale feature fused based on the attention weights to obtain comprehensive state characteristics, and the comprehensive state characteristics are optimized through a dynamic weight update function to obtain optimized state characteristics;

[0034] The optimized state characteristics are separated by feature separation, the short-term component separated is processed through a convolutional neural network to obtain a short-term error feature, and the long-term component separated is processed through a fully connected network to obtain a long-term error feature;

[0035] The short-term error feature and the long-term error feature are weighted and combined based on an adaptive weight to obtain a dynamic compensation value; the dynamic compensation value is error compensated with the initial error data to obtain a compensated error, and the compensated capacity value is corrected based on the compensated error to obtain a final capacity value.

[0036] In an alternative embodiment,

[0037] The steps of constructing the bidirectional attention mechanism include:

[0038] constructing a forward query matrix according to the short-term state feature, constructing a backward query matrix according to the long-term state feature, and performing feature enhancement on the forward query matrix and the backward query matrix respectively through a nonlinear projection layer to obtain a forward enhanced query matrix and a backward enhanced query matrix;

[0039] performing dot product operation on the forward enhanced query matrix and the long-term state feature to obtain a forward attention score, and performing dot product operation on the backward enhanced query matrix and the short-term state feature to obtain a backward attention score;

[0040] calculating an adaptive temperature parameter based on an error gradient of the current moment, applying the adaptive temperature parameter to the forward attention score and the backward attention score respectively for dynamic scaling, and obtaining a forward attention weight and a backward attention weight through softmax normalization processing;

[0041] constructing a dynamic balance factor that is adaptively adjusted according to an error change rate, and performing nonlinear combination on the forward attention weight and the backward attention weight based on the dynamic balance factor to obtain an initial fusion weight;

[0042] applying adaptive L1 regularization constraint based on weight entropy to the initial fusion weight to obtain a sparse weight distribution, and performing cross-scale feature calibration on the sparse weight distribution through a residual connection mechanism to obtain the attention weight.

[0043] In an optional implementation,

[0044] adjusting a verification control parameter according to the dynamic compensation value, performing capacity verification on the glass measuring container, and obtaining verification result data, the step comprising:

[0045] determining an adjustment range of a liquid injection rate parameter and an adjustment range of a temperature control parameter according to the dynamic compensation value;

[0046] constructing an adaptive control strategy library based on verification historical data, extracting an optimal control parameter template from the control strategy library according to the type, specification and environmental conditions of the current glass measuring container, and dynamically fine-tuning the control parameter in combination with the dynamic compensation value to generate a verification control curve;

[0047] controlling the liquid injection unit to perform staged liquid injection operation according to the liquid injection rate parameter in the verification control curve, and adjusting the verification environment temperature according to the temperature control parameter;

[0048] Collect liquid level image and high-accuracy electronic balance data in the liquid injection process, record liquid injection speed and environmental temperature, obtain calibration process data; comprehensively analyze the calibration process data and the dynamic compensation value, generate a calibration report containing capacity measurement value, error analysis result and quality assessment information; feedback the calibration control parameters and calibration results to the adaptive control strategy library, update the optimal control parameter template.

[0049] In a second aspect of the embodiment of the present application, a glass measuring container capacity calibration process automation control and error analysis system is provided, comprising:

[0050] A first unit is configured to calculate an environmental influence compensation value on the capacity of the glass measuring container according to the specification parameter information of the glass measuring container to be calibrated and the environmental temperature, air pressure and humidity;

[0051] A second unit is configured to dynamically capture and analyze the liquid level image of the glass measuring container by using a double-branch neural network structure, wherein a first branch extracts liquid surface static features by a spatial attention module, and a second branch obtains liquid surface dynamic features by a time series attention module, and the liquid surface static features and the liquid surface dynamic features are fused with the weight data of the electronic balance to calculate a compensated capacity value in combination with the influence compensation value;

[0052] A third unit is configured to perform real-time error evaluation on the compensated capacity value to obtain initial error data;

[0053] A fourth unit is configured to input the initial error data into a recurrent neural network with a hierarchical memory structure for dynamic error analysis, obtain instantaneous change features of the liquid surface state and environmental parameters by a short-term memory unit, analyze error accumulation features by a long-term memory unit, and adaptively weight error features of different time scales by using a bidirectional attention mechanism to obtain a dynamic compensation value;

[0054] A fifth unit is configured to adjust calibration control parameters according to the dynamic compensation value, calibrate the capacity of the glass measuring container, and obtain calibration result data.

[0055] In a third aspect of the embodiment of the present application, an electronic device is provided, comprising:

[0056] A processor;

[0057] A memory for storing processor-executable instructions;

[0058] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0059] The fourth aspect of the embodiment of the present application provides a computer readable storage medium, which stores computer program instructions, and the computer program instructions are executed by a processor to realize the method described above.

[0060] The present application realizes the intelligentization and precision of the glass measuring container capacity verification process by adopting the double-branch neural network structure and the hierarchical memory recurrent neural network, significantly improves the verification efficiency and accuracy, and reduces the human operation error.

[0061] The multi-modal feature fusion method of the present application can simultaneously consider the liquid surface static features and dynamic features, combine the influence compensation value of the environmental parameters, make the capacity measurement more comprehensive and accurate, adapt to various complex detection environments, and solve the problem of insufficient accuracy of the traditional verification method when the environmental factors such as temperature and air pressure change.

[0062] The dynamic error analysis system of the present application realizes the real-time monitoring and self-adaptive adjustment of the error in the verification process by the combination of the short-term memory unit and the long-term memory unit, can automatically optimize the verification parameters for different types of glass measuring containers, reduces the operation difficulty, and improves the reliability and consistency of the verification results. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The flowchart of the glass measuring container capacity verification process automatic control and error analysis method of the embodiment of the present application is shown in

[0064] Figure 2 The flowchart of the dynamic compensation value obtained by the dynamic error analysis of the present application is shown in DETAILED DESCRIPTION

[0065] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme of the embodiment of the present application will be described clearly and completely in combination with the drawings of the embodiment of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0066] The technical scheme of the present application will be described in detail in specific embodiments. The following several specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0067] Figure 1 The flowchart of the glass measuring container capacity verification process automatic control and error analysis method of the embodiment of the present application is shown in Figure 1 As shown in the figure, the method comprises:

[0068] According to the specification parameter information of the glass measure to be detected and the environmental temperature, air pressure and humidity, an influence compensation value of the environment on the capacity of the glass measure is calculated;

[0069] The liquid level image of the glass measure is captured and analyzed by using a double-branch neural network structure, wherein a first branch extracts liquid surface static features by a spatial attention module, and a second branch obtains liquid surface dynamic features by a time series attention module; the liquid surface static features and the liquid surface dynamic features are fused with the weight data of the electronic balance, and a compensated capacity value is calculated in combination with the influence compensation value;

[0070] The compensated capacity value is subjected to real-time error evaluation to obtain initial error data;

[0071] The initial error data is input into a recurrent neural network with a hierarchical memory structure for dynamic error analysis; the instantaneous change features of the liquid surface state and the environmental parameters are obtained by a short-term memory unit, and the error accumulation features are analyzed by a long-term memory unit; a bidirectional attention mechanism is used to adaptively weight the error features of different time scales to obtain a dynamic compensation value;

[0072] The verification control parameters are adjusted according to the dynamic compensation value, the capacity of the glass measure is verified, and verification result data is obtained.

[0073] In an alternative embodiment, the step of calculating the influence compensation value of the environment on the capacity of the glass measure according to the specification parameter information of the glass measure to be detected and the environmental temperature, air pressure and humidity comprises:

[0074] The specification parameter information includes the nominal capacity and the use temperature range of the glass measure;

[0075] The volume thermal deformation amount of the glass measure and the density change amount of the liquid to be measured are calculated according to the specification parameter information and the environmental temperature data, and the volume thermal deformation amount and the density change amount are coupled to calculate a temperature compensation coefficient; the liquid surface tension change value and the water vapor partial pressure are calculated according to the air pressure data and the relative humidity data, and the surface tension change value and the water vapor partial pressure are weighted to calculate an environmental compensation coefficient;

[0076] The temperature compensation coefficient and the environmental compensation coefficient are combined to calculate a comprehensive compensation coefficient; the influence compensation value of the environment on the capacity of the glass measure is calculated according to the comprehensive compensation coefficient, the nominal capacity and the current measurement value; the deviation of the compensated capacity value from the nominal capacity is calculated, and when the deviation exceeds a preset deviation threshold, the weights of the temperature compensation coefficient and the environmental compensation coefficient are dynamically adjusted.

[0077] For example, by acquiring the nominal capacity of the glass measuring container and the specification parameter information such as the temperature range for use, and the environmental temperature, air pressure and humidity data, the influence compensation value of the environment on the capacity of the glass measuring container is calculated.

[0078] The influence of the environmental temperature on the capacity of the glass measuring container mainly embodies in two aspects: one is the volume thermal deformation of the glass material itself due to the temperature change, and the other is the influence of the density change of the liquid to be measured with the temperature change. When the actual measurement environmental temperature is different from the standard temperature (usually 20℃), the temperature compensation needs to be carried out. The volume thermal deformation amount of the glass measuring container needs to consider the linear expansion coefficient of the glass. Taking borosilicate glass as an example, its linear expansion coefficient is about 3.3×10 -6 / ℃. When the environmental temperature is 25℃ and the standard temperature is 20℃, for the 100mL nominal capacity volumetric flask, its volume thermal deformation amount is 100mL×3.3×10 -6 / ℃×(25-20)℃=0.00165mL.

[0079] The density change of the liquid to be measured (usually water) also needs to be considered. The density of water at 20℃ is about 0.9982g / mL, and at 25℃ is about 0.9970g / mL. This density change will cause the same mass of water to occupy different volumes at different temperatures. Based on the above data, the temperature compensation coefficient at 25℃ can be calculated. The volume thermal deformation amount and the density change amount are coupled to calculate the temperature compensation coefficient, which is 0.00205, that is, at 25℃ environment, the actual capacity of the 100mL nominal capacity glass measuring container should be 100.205mL.

[0080] The influence of the air pressure and humidity on the liquid surface tension and vapor pressure also needs to be considered. When the air pressure changes, the liquid surface tension will change accordingly. Taking the standard atmospheric pressure (101.325kPa) as the benchmark, when the air pressure rises to 102kPa, the surface tension of water increases by about 0.15%. For a 100mL nominal capacity measuring container, this change causes a capacity change of about 0.02mL.

[0081] Relative humidity affects the evaporation rate and water vapor partial pressure of water. When the relative humidity is 50% and the temperature is 25℃, the water vapor partial pressure is about 1.58kPa. When the relative humidity rises to 70%, the water vapor partial pressure increases to about 2.21kPa. This change will affect the evaporation amount and the surface meniscus shape of the liquid in the measuring container. Through weighted calculation, the environmental compensation coefficient can be obtained. Taking the above conditions as an example, the environmental compensation coefficient is calculated to be 0.00135.

[0082] The temperature compensation coefficient and the environmental compensation coefficient are combined to obtain a comprehensive compensation coefficient. The combination calculation can use a weighted average method, and the temperature factor usually dominates, and the weight can be set to 0.75, and the weight of the environmental factor is set to 0.25. Based on the foregoing case, the comprehensive compensation coefficient is 0.00205*0.75+0.00135*0.25=0.00188.

[0083] For a glass measuring container with a nominal capacity of 100 mL and an actual measured value of 99.95 mL, after applying the comprehensive compensation coefficient, the compensation value is calculated as 100 mL*0.00188=0.188 mL. The compensated capacity value is 99.95 mL+0.188 mL=100.138 mL. Compared with the nominal capacity, the deviation is 0.138 mL, i.e. 0.138%.

[0084] When the deviation exceeds the preset deviation threshold, the weights of the temperature compensation coefficient and the environmental compensation coefficient need to be dynamically adjusted. For example, set the deviation threshold to 0.1%, when the calculated deviation 0.138% exceeds this threshold, the weight distribution will be automatically adjusted. The adjustment method can be based on sensitivity analysis, if the temperature change has a more significant impact on the compensation value, the weight of the temperature compensation coefficient is increased; otherwise, the weight of the environmental compensation coefficient is increased. Through iterative calculation, the optimal weight configuration can be found. In this example, the adjusted weights may be temperature factor 0.65 and environmental factor 0.35, and the recalculated comprehensive compensation coefficient is 0.00181, and the compensated deviation is reduced to 0.095%, which meets the deviation threshold requirement.

[0085] The implementation of this method also requires high-accuracy environmental parameter measurement equipment. The temperature measurement accuracy should reach ±0.1℃, the air pressure measurement accuracy should reach ±0.1kPa, and the relative humidity measurement accuracy should reach ±2%. During the measurement process, the stability of the environmental parameters should be ensured to avoid large fluctuations in a short period of time affecting the calculation accuracy.

[0086] The present application provides a scientific compensation value calculation method by accurately calculating the influence of environmental factors on the capacity of the glass measuring container, effectively improving the accuracy and reliability of the glass measuring container measurement, and is particularly suitable for laboratories and production environments that require high-accuracy measurement.

[0087] In an optional implementation, the first branch extracts liquid surface static features through a spatial attention module, and the second branch extracts liquid surface dynamic features through a time sequence attention module. The liquid surface static features, liquid surface dynamic features, and weight data of the electronic balance are fused for multi-modal feature fusion, and the compensated capacity value is calculated based on the influence compensation value.

[0088] The liquid surface shape feature is extracted by a spatial attention module in the first branch, the spatial attention module enhances the liquid surface profile curvature and capillary effect feature based on channel attention weight and position attention weight to obtain a liquid surface static feature; the liquid surface oscillation feature is modeled by a time sequence attention module in the second branch, the time sequence attention module calculates the correlation degree between multiple time sequence states based on time window weight, and dynamically aggregates the correlation degree by a convolution long short-term memory network to obtain a liquid surface dynamic feature; the liquid surface static feature and the liquid surface dynamic feature are subjected to feature alignment and normalization processing to obtain a liquid level image feature;

[0089] The weight data of the electronic balance is mapped to a feature space to obtain weight data features;

[0090] The liquid level image feature and the weight data feature are subjected to multi-modal feature fusion to obtain a fusion feature, and a compensated capacity value of the glass measuring cylinder is calculated based on the fusion feature and the influence compensation value.

[0091] For example, a glass measuring cylinder liquid surface image sequence and corresponding electronic balance weight data are collected as input. The image sequence includes 20 frames of continuously collected liquid surface images, each frame of image has a resolution of 640x480 pixels, and is stored in an RGB format.

[0092] For the spatial attention module in the first branch, a pre-trained ResNet-50 network is used to extract features from each frame of liquid surface image to obtain an initial feature map with a dimension of 256x32x32. Then, a channel attention mechanism is applied to the feature map, which generates channel descriptors through global average pooling and global maximum pooling, respectively. After processing by a shared multi-layer perception (including a hidden layer of 16 neurons and a ReLU activation function), the two kinds of descriptors are added and normalized by a Sigmoid function to obtain the weight coefficients of each channel. Multiplying these coefficients with the original feature map realizes the enhancement of the liquid surface profile curvature feature.

[0093] The position attention mechanism enhances the capillary effect feature through attention calculation in the spatial dimension. The specific implementation is to reshape the feature map into a 256x1024 form, generate a query matrix, a key matrix and a value matrix through three independent 1x1 convolution layers respectively, perform matrix multiplication on the query matrix and the transpose of the key matrix, and normalize it by a Softmax function to generate a 32x32 spatial attention map. Multiply this attention map with the value matrix and reshape it back to the original dimension, and then add it to the input feature map to obtain the position-enhanced feature. By combining the outputs of the channel attention and the position attention, a 256-dimensional liquid surface static feature vector is finally generated, which contains key information such as liquid surface shape, curvature and capillary effect.

[0094] For the temporal attention module in the second branch, 20 consecutive liquid surface images were processed to capture the liquid surface oscillation characteristics. First, 128-dimensional basic features were extracted from each image, forming a 20x128 temporal feature sequence. The temporal attention module used a 5-frame sliding time window to calculate the correlation between frames within the window. Specifically, the features within each time window were processed through a self-attention mechanism to generate a 5x5 temporal correlation matrix, which represented the dependency between liquid surface states at different times.

[0095] ConvLSTM was used to model these temporal correlations. ConvLSTM contained 64 hidden units, and the convolution kernel size of the input gate, forget gate, and output gate was 3x3 with a padding of 1. This structure could capture dependencies in both spatial and temporal dimensions, making it particularly suitable for processing dynamic changes such as liquid surface oscillation. After ConvLSTM processing, a 128-dimensional liquid surface dynamic feature vector was output, which encoded the liquid surface oscillation frequency, amplitude, and decay characteristics.

[0096] In the feature alignment stage, the 256-dimensional static features and the 128-dimensional dynamic features were mapped to the same 192-dimensional feature space through two fully connected layers, respectively. Subsequently, feature normalization was performed using the Layer Normalization technique to adjust the feature values to a distribution with a mean of 0 and a variance of 1, ensuring that static and dynamic features are on the same scale for subsequent fusion. The normalized static and dynamic features were combined through concatenation to form a 384-dimensional liquid level image feature vector.

[0097] For electronic balance weight data processing, the accuracy of the original weight data collected was 0.001 grams, and the corresponding weight readings were recorded simultaneously with the liquid surface images. These weight data were mapped to the feature space through a three-layer fully connected network (with layer node numbers of 64, 128, and 192, respectively) to obtain 192-dimensional weight data features. Batch Normalization and ReLU activation functions were used after each fully connected layer to enhance the network's expression ability and training stability.

[0098] The multi-modal feature fusion stage used a cross-attention mechanism, taking the 384-dimensional liquid level image features and the 192-dimensional weight data features as input. Specifically, an attention matrix was calculated between the two features, and this matrix was used to weight and fuse the features. After fusion, a 256-dimensional multi-modal fusion feature vector was generated, which contained both visual and weight information.

[0099] The fusion feature vector is input into a regression network composed of two fully connected layers (128 nodes and 64 nodes) and an output layer. The output of the regression network is combined with a pre-calculated influence compensation value (calculated based on environmental factors such as laboratory temperature, atmospheric pressure, and relative humidity) to obtain the final compensated capacity value through weighted summation.

[0100] The spatial attention module of the present application can accurately capture subtle features such as liquid surface profile curvature and capillary effect, and the temporal attention module can effectively model the oscillation characteristics of the liquid surface. The combination of the two overcomes the defect that traditional liquid level recognition technology is difficult to handle both static and dynamic liquid surface states. This multi-modal feature fusion method takes into account the environmental influence compensation value, making the capacity value calculation more accurate and reliable, providing key technical support for high-accuracy glass measuring instrument calibration.

[0101] In an alternative embodiment, the liquid level image features and the weight data features are subjected to multi-modal feature fusion to obtain fusion features, and the step of calculating the compensated capacity value of the glass measuring instrument based on the fusion features and the influence compensation value comprises:

[0102] The liquid level image features and the weight data features are subjected to normalization processing to obtain normalized features, and the self-attention and cross-modal attention between the normalized features are calculated through a multi-head attention mechanism to obtain a feature interaction matrix;

[0103] A feature confidence evaluation network is constructed based on the feature interaction matrix, which calculates the feature confidence score by calculating the contribution of each modal feature to capacity prediction; and a dynamic fusion weight is obtained by normalizing the feature confidence score based on the temperature parameter;

[0104] The dynamic fusion weight and the normalized features are combined by weighting, and an enhanced fusion feature is obtained through residual connection; a multi-objective loss function including capacity prediction, modal consistency, and feature reconstruction is constructed, and the enhanced fusion feature is optimized;

[0105] The optimized fusion features and the influence compensation value are input into a fully connected layer, and a capacity prediction value is calculated through a multi-layer perceptron; a prediction error between the capacity prediction value and the true capacity value is calculated, a comprehensive weight update gradient is calculated based on the prediction error and the multi-objective loss function, and the dynamic fusion weight is adaptively adjusted using the comprehensive weight update gradient; the updated dynamic fusion weight is used to re-fuse the features, and the capacity value of the glass measuring instrument is predicted in combination with the influence compensation value.

[0106] Exemplarily, the liquid level image and weight data of the glass measuring cylinder are collected, and after extracting the liquid level image features and weight data features, the two modal features are normalized to map the feature values to the range of 0-1. For the liquid level image features, the maximum and minimum normalization method is used, that is, the feature value is calculated as (feature value-minimum value) / (maximum value-minimum value); for the weight data features, the Z-score normalization is used, that is, the feature value is calculated as (feature value-mean value) / standard deviation. For example, when the liquid level image feature value range is [100, 500], and a certain feature value is 300, the normalized value is 0.5; when the weight data feature mean value is 200, the standard deviation is 50, and a certain feature value is 250, the normalized value is 1.0.

[0107] After normalization, the self-attention and cross-modal attention between the features are calculated through the multi-head attention mechanism. In actual application, 8 attention heads can be set, and the output dimension of each head is 64. For self-attention calculation, the similarity between the features in the same modality is calculated, such as the correlation between the liquid level image features; for cross-modal attention, the mutual influence between the features of different modalities is calculated, such as the influence degree of the liquid level features on the weight features. Through the two attention mechanisms, a feature interaction matrix is generated, which contains the correlation scores between the features.

[0108] Based on the feature interaction matrix, a feature confidence evaluation network is constructed, which consists of three fully connected layers with hidden layer node numbers of 256, 128 and 64 respectively. The network analyzes the correlation patterns in the feature interaction matrix to evaluate the contribution of each modal feature to the capacity prediction and generate a feature confidence score. For example, in a certain measurement, the confidence score of the liquid level image feature is 0.75, and the confidence score of the weight data feature is 0.85. To prevent the confidence score from being too concentrated or dispersed, a temperature parameter is introduced to normalize the feature confidence score. The temperature parameter is set to 2.0, and by dividing by the temperature parameter and applying the softmax function, a dynamic fusion weight is obtained. In the above example, the normalized dynamic fusion weights are 0.47 and 0.53 respectively, indicating that the weight feature is slightly more important than the liquid level feature.

[0109] The calculated dynamic fusion weight is combined with the normalized feature by weighting, that is, each modal feature is multiplied by its corresponding weight and then added. In order to preserve the original feature information, a residual connection mechanism is used to add the original normalized feature and the weighted feature to obtain an enhanced fusion feature.

[0110] A multi-objective loss function is constructed to optimize the enhanced fusion features, which includes three parts: capacity prediction loss, modal consistency loss and feature reconstruction loss. The capacity prediction loss adopts mean square error to measure the difference between the predicted capacity and the true capacity; the modal consistency loss ensures the consistency of different modal feature expressions, which is realized by calculating the cosine similarity between the features of different modalities; the feature reconstruction loss ensures that the fused features can restore the original features of each modal, and the reconstruction error is calculated by the reconstruction network. The three parts of loss are combined in the proportion of 7:2:1 to form the final multi-objective loss function.

[0111] The optimized fusion features are spliced with the influence compensation value, which includes factors such as environmental temperature and humidity. The spliced feature vector is input into the fully connected layer, which includes three layers with node numbers of 512, 256 and 128 respectively, and the activation function is ReLU. The capacity prediction value is calculated by the multilayer perceptron, and the output layer of the prediction network is a single node representing the predicted capacity value.

[0112] The prediction error between the capacity prediction value and the true capacity value is calculated, and the root mean square error (RMSE) is used as the evaluation index. For example, when the predicted capacity is 99.5ml and the true capacity is 100ml, the prediction error is 0.5ml. Based on the prediction error and the multi-objective loss function, the comprehensive weight update gradient is calculated. The comprehensive weight update gradient combines the gradient of the capacity prediction error and the gradient of the multi-objective loss function, and is weighted and averaged in the proportion of 8:2.

[0113] The calculated comprehensive weight update gradient is used to adaptively adjust the dynamic fusion weight through the back propagation algorithm. The learning rate is set to 0.001, and the Adam optimizer is used to update the weight. For example, if it is found that the liquid level feature is more important for accurate prediction, its weight is increased; otherwise, the weight is decreased.

[0114] The updated dynamic fusion weight is used to re-fuse the features to generate new fusion features. The new fusion features are combined with the influence compensation value, and the final compensated capacity value of the glass measuring cylinder is calculated through the previously trained prediction network.

[0115] The present application calculates self-attention and cross-modal attention through multi-head attention mechanism, constructs a rich feature interaction matrix, and effectively solves the correlation modeling problem between different modal features. The feature confidence evaluation network is introduced to calculate the contribution degree of each modal feature to the capacity prediction, realizing the dynamic evaluation of feature importance and avoiding the limitations of traditional fixed weight fusion method. The adaptive weight updating mechanism of the method can dynamically adjust the fusion strategy according to the prediction error, forming a closed loop optimization, which significantly improves the accuracy and reliability of the capacity prediction, and is especially suitable for the calibration scene of different specifications of glass measuring cylinders, with strong universality and adaptability.

[0116] In an alternative embodiment, the initial error data is input into a recurrent neural network with a hierarchical memory structure for dynamic error analysis, the short-term memory unit is used to obtain the instantaneous change characteristics of the liquid level state and environmental parameters, the long-term memory unit is used to analyze the error accumulation characteristics, and a bidirectional attention mechanism is used to adaptively weight the error characteristics of different time scales to obtain a dynamic compensation value, and the steps include:

[0117] Comparing the compensated capacity value with the nominal capacity value to obtain initial error data, the initial error data including absolute error data, relative error data and standard deviation data;

[0118] The short-term memory unit models the instantaneous change of the liquid level state parameters and environmental parameters through the state vector, the forgetting gate and the input gate to obtain short-term state characteristics, and the long-term memory unit dynamically aggregates the short-term state characteristics based on error accumulation and long-term state update function to obtain long-term state characteristics;

[0119] A bidirectional attention mechanism is constructed for the short-term state characteristics and the long-term state characteristics, the attention scores between different time scale characteristics are calculated through the query matrix and the key-value matrix, and the attention scores are normalized to obtain attention weights;

[0120] The short-term state characteristics and the long-term state characteristics are fused based on the attention weights to obtain comprehensive state characteristics, and the comprehensive state characteristics are optimized through a dynamic weight update function to obtain optimized state characteristics;

[0121] The optimized state characteristics are separated by feature separation, the short-term component separated is processed by a convolutional neural network to obtain short-term error characteristics, and the long-term component separated is processed by a fully connected network to obtain long-term error characteristics;

[0122] The short-term error characteristics and the long-term error characteristics are weighted and combined based on adaptive weights to obtain a dynamic compensation value; the dynamic compensation value is error compensated with the initial error data to obtain a compensated error, and the compensated capacity value is corrected based on the compensated error to obtain a final capacity value.

[0123] For example, in combination with Figure 2The flow chart of acquiring dynamic compensation values through dynamic error analysis is described. The acquisition of initial error data is based on the comparison of the compensated capacity value and the nominal capacity value. Three types of key error data are calculated: absolute error data representing the absolute difference between the actual measurement value and the nominal value, relative error data representing the percentage of error to the nominal value, and standard deviation data reflecting the dispersion degree of multiple measurement results. These three types of data together constitute the initial error data set, which is used as the input for subsequent dynamic error analysis.

[0124] The hierarchical memory structure of the recurrent neural network includes two core components: short-term memory units and long-term memory units. The short-term memory unit adopts a long short-term memory network structure, which captures the instantaneous features of the liquid level state through a state vector. The state vector has a dimension of twice the number of liquid level parameters, ensuring the richness of information expression. The forgetting gate controls the retention proportion of historical state information. For rapidly changing liquid level state parameters, the threshold of the forgetting gate is set low, about 0.3, allowing the network to quickly update the state. For relatively stable environmental parameters, the threshold of the forgetting gate is set high, about 0.7, maintaining the continuity of the state. The input gate adjusts the degree of reception of new input information, dynamically adjusting according to the parameter change rate. When the change rate is high, the input gate is open, and when the change rate is low, the input gate is closed, ensuring the balance between sensitivity and stability of state updating. The short-term memory unit models the instantaneous changes of liquid level state parameters and environmental parameters through this mechanism, generating a short-term state feature vector with a dimension of 128.

[0125] The long-term memory unit is designed to capture the pattern of error accumulation over time, using an error accumulation function to weight and aggregate errors over multiple time steps. The error accumulation function uses an exponential decay weight, so that recent errors have a higher weight than distant errors. The decay rate is dynamically adjusted according to the autocorrelation of the error. When the autocorrelation is high, the decay rate is low, and when the autocorrelation is low, the decay rate is high. The long-term state update function takes the short-term state feature as input, combines the historical long-term state, and selectively updates the long-term state information through a gating mechanism. During the update process, features with high importance are retained at a high proportion, and features with low importance are retained at a low proportion. The importance is determined by the correlation between the feature and the historical error. This mechanism ensures that the long-term state feature can effectively express the long-term trend and accumulation pattern of the error, ultimately generating a long-term state feature vector with a dimension of 256.

[0126] The bidirectional attention mechanism establishes an information interaction channel between short-term and long-term state features, enhancing the expression ability of the features. The query matrix and the key-value matrix are used to calculate the correlation between features of different time scales. The query matrix has the same dimension as the state feature and is generated from the state feature through linear transformation. The attention score calculation adopts the scaled dot-product method, which divides the dot product result of the query matrix and the key-value matrix by the square root of the feature dimension to avoid the gradient problem caused by large dimension. The attention score is normalized by the softmax function to convert it into an attention weight with a value range of 0 to 1. The weight sum is 1, ensuring that the importance of different features is reasonably distributed. The attention weight reflects the relative importance of features of different time scales, with a high weight indicating that the feature has a high contribution to the current state understanding.

[0127] The multi-scale feature fusion process weights and sums the short-term state feature and the long-term state feature based on the attention weight to obtain the preliminary fusion feature. In the preliminary fusion feature, the proportion of important feature components is large, and the proportion of low-importance feature components is small, ensuring the effectiveness of information transmission. The dynamic weight update function further optimizes the fusion feature by adjusting the weight distribution according to the error change trend. When the error changes greatly, the short-term feature weight is increased, and when the error changes little, the long-term feature weight is increased. The optimization uses the gradient descent method, calculates the weight gradient based on the error prediction loss function, and the learning rate is initially set to 0.01 and decays with training. The decay coefficient is 0.95 every 1000 steps. After about 50 rounds of iteration optimization, the optimized state feature is obtained, and the dimension remains unchanged, but the feature distribution is more in line with the needs of the error prediction task.

[0128] The feature separation link decomposes the optimized state feature into short-term and long-term components. The frequency domain analysis method is used to treat the feature as a time series and decompose it into different frequency components through Fourier transform. High-frequency components correspond to short-term changes, and low-frequency components correspond to long-term trends. The separation threshold is determined according to the power spectral density of the feature, and the frequency at the inflection point of the power spectral density is used as the dividing point. The short-term component obtained by separation is processed through a convolutional neural network. The network contains three convolutional layers with kernel sizes of 3, 5, and 7, respectively. After each convolution, batch normalization and ReLU activation functions are added to extract local change patterns and obtain short-term error features. The long-term component is processed through a three-layer fully connected network with hidden layer node numbers of 128, 64, and 32, respectively. After each layer, batch normalization and ReLU activation functions are added to capture global change trends and obtain long-term error features.

[0129] The generation of the dynamic compensation value adopts an adaptive weight mechanism to combine the short-term error characteristics and the long-term error characteristics. The adaptive weight is calculated according to the recent error change rate, the short-term characteristic weight is high when the change rate is high, and the long-term characteristic weight is high when the change rate is low, thereby ensuring the adaptability of the compensation value to different error modes. In a typical case, the short-term characteristic weight varies between 0.3 and 0.7, and the long-term characteristic weight is 1 minus the short-term characteristic weight. The combined characteristics are mapped to the compensation value space through a fully connected layer to obtain the dynamic compensation value, and the compensation value range is usually between 0.01% and 0.5% of the nominal capacity, which varies according to the accuracy level of the glass container.

[0130] The error compensation process combines the dynamic compensation value with the initial error data to calculate the compensated error. The compensation method is direct subtraction, that is, the dynamic compensation value is subtracted from the initial error to obtain the compensated error. The compensated error is used to correct the compensated capacity value, and the correction method is to add the opposite of the compensated error to the compensated capacity value to obtain the final capacity value. For the case of large initial error, the compensation effect is more obvious, and the correction amplitude can be more than 70% of the initial error; for the case of small initial error, the compensation mainly targets the specific error component, and the correction amplitude is relatively small.

[0131] The present application captures error characteristics of different time scales through a hierarchical memory structure, and enhances the feature expression capability through a bidirectional attention mechanism, thereby realizing the differentiated identification of systematic errors and random errors. The dynamic compensation mechanism adaptively adjusts the compensation strategy according to the error characteristics, thereby significantly improving the accuracy and reliability of the capacity verification; and effectively eliminating the influence of external factors on the verification result, thereby providing strong technical support for high-accuracy glass container capacity verification.

[0132] In an optional implementation, the step of constructing the bidirectional attention mechanism comprises:

[0133] A forward query matrix is constructed according to the short-term state characteristics, a backward query matrix is constructed according to the long-term state characteristics, and the forward query matrix and the backward query matrix are respectively subjected to feature enhancement through a nonlinear projection layer to obtain a forward enhanced query matrix and a backward enhanced query matrix;

[0134] The forward enhanced query matrix is subjected to dot product operation with the long-term state characteristics to obtain a forward attention score, and the backward enhanced query matrix is subjected to dot product operation with the short-term state characteristics to obtain a backward attention score;

[0135] An adaptive temperature parameter is calculated based on the error gradient of the current time, the adaptive temperature parameter is applied to the forward attention score and the backward attention score for dynamic scaling, and forward attention weight and backward attention weight are obtained through softmax normalization processing;

[0136] a dynamic balance factor that is self-adaptively adjusted according to an error change rate, and performing nonlinear combination on the forward attention weight and the backward attention weight based on the dynamic balance factor to obtain an initial fusion weight;

[0137] applying a weight entropy-based adaptive L1 regularization constraint to the initial fusion weight to obtain a sparse weight distribution, and performing cross-scale feature calibration on the sparse weight distribution through a residual connection mechanism to obtain the attention weight.

[0138] For example, the construction process of the bidirectional attention mechanism starts from the construction of the query matrix. The short-term state feature is used to construct the forward query matrix, which has the same dimension as the short-term state feature, i.e., 128 dimensions. The construction method is to generate the query matrix by linear transformation of the short-term state feature. The transformation weight is initialized with a uniform distribution, and the value range is [-0.1, 0.1]. The long-term state feature is used to construct the backward query matrix, which has the same dimension as the long-term state feature, i.e., 256 dimensions. The same linear transformation is used to achieve this, and the parameters of the transformation matrix are initialized using the Xavier initialization method to ensure that the variances of the gradients during forward propagation and backward propagation are consistent. A nonlinear projection layer is used to enhance the features of the forward query matrix and the backward query matrix. The projection layer consists of two fully connected layers with an activation function inserted in between. For the forward query matrix, the first fully connected layer maps the 128-dimensional feature to 256 dimensions, and after the ReLU activation function, the second fully connected layer maps the feature back to 128 dimensions to obtain the forward enhanced query matrix. For the backward query matrix, the first fully connected layer maps the 256-dimensional feature to 512 dimensions, and after the ReLU activation function, the second fully connected layer maps the feature back to 256 dimensions to obtain the backward enhanced query matrix. The projection layer increases the nonlinear expression ability of the features, so that the query matrix can capture more complex feature relationships.

[0139] The dot product operation stage calculates the correlation between different time scale features. The forward enhanced query matrix is dot multiplied with the long-term state feature to obtain the forward attention score. Specifically, each vector of the forward enhanced query matrix is dot multiplied with each vector of the long-term state feature. If the shape of the forward enhanced query matrix is (T, 128) and the shape of the long-term state feature is (T, 256), the forward enhanced query matrix needs to be adjusted to 256 dimensions through linear mapping to match the dimension of the long-term state feature, and the shape of the adjusted forward enhanced query matrix is (T, 256). The dot product result has a shape of (T, T), representing the correlation between the short-term state feature of each time step and the long-term state feature of each time step. The backward enhanced query matrix is dot multiplied with the short-term state feature to obtain the backward attention score. The operation mode is similar, that is, each vector of the backward enhanced query matrix is dot multiplied with each vector of the short-term state feature, and the backward enhanced query matrix needs to be adjusted to 128 dimensions through linear mapping to match the dimension of the short-term state feature. The shape of the adjusted backward enhanced query matrix is (T, 128), and the shape of the dot product result is also (T, T), representing the correlation between the long-term state feature of each time step and the short-term state feature of each time step.

[0140] The adaptive temperature parameter is calculated based on the error gradient of the current time. The error gradient represents the rate of change of the capacity prediction error with respect to the model parameter, and a large gradient amplitude indicates that the current parameter has a significant impact on the error, requiring a more concentrated attention distribution; a small gradient amplitude indicates that the current parameter has a small impact on the error, and a more dispersed attention distribution can be used. The calculation method is to take the two-norm of the error gradient of the current batch, which is usually in the range of 0.001 to 10. The adaptive temperature parameter is inversely proportional to the gradient amplitude. When the gradient amplitude is large, the temperature parameter is small, making the attention distribution more concentrated; when the gradient amplitude is small, the temperature parameter is large, making the attention distribution more gentle. The value range of the temperature parameter is usually controlled between 0.5 and 5, and the default initial value is 1. The dynamic scaling process applies the adaptive temperature parameter to the attention score. Specifically, the forward attention score and the backward attention score are divided by the temperature parameter respectively. For example, if the temperature parameter is 2 and the attention score is 4, the scaled score is 2. The scaled attention score is normalized by the softmax function to convert it into a probability distribution with a sum of 1 to obtain the forward attention weight and the backward attention weight. The softmax function ensures that all weights are positive and the sum is 1, effectively preventing the problem of gradient explosion caused by some weights being too large.

[0141] The dynamic balance factor is constructed based on the error change rate. The error change rate represents the degree of error change between consecutive time steps, and the calculation method is to take the difference between the errors of two consecutive time steps divided by the error of the previous time step. The value range of the change rate is usually between -0.5 and 0.5, and the greater the absolute value, the more intense the error change. The dynamic balance factor is proportional to the absolute value of the error change rate. When the error changes dramatically, the balance factor value is large, increasing the weight of short-term information; when the error changes gently, the balance factor value is small, increasing the weight of long-term information. The value range of the balance factor is usually controlled between 0.3 and 0.7, and the default value is 0.5. The nonlinear combination process is based on the dynamic balance factor to weight and fuse the forward attention weight and the backward attention weight. The specific operation is to multiply the forward attention weight by the balance factor, multiply the backward attention weight by (1-balance factor), and then add the two parts to obtain the initial fusion weight. For example, if the balance factor is 0.6, the forward weight is 0.7, and the backward weight is 0.4, then the fused weight is 0.6x0.7+0.4x0.4=0.58. To increase nonlinearity, a hyperbolic tangent function is applied after the weighted sum in practical applications, so that the value range of the fused weight is controlled within [-1, 1], and then scaling and translation are performed to adjust the value range to [0, 1].

[0142] The adaptive L1 regularization constraint of weight entropy is used to generate a sparse weight distribution. Weight entropy represents the uncertainty of the weight distribution, and the calculation method is to calculate the entropy value by regarding the weight as a probability distribution. The higher the entropy value, the more uniform the weight distribution, and the lower the entropy value, the more concentrated the weight distribution. For glassware capacity verification tasks, it is expected that the attention weight distribution has a certain sparsity, that is, only a few key time steps are focused on, and irrelevant time steps are ignored. L1 regularization constraint is an effective means to achieve sparsity, which imposes a penalty on the sum of the absolute values of the weights, encouraging some weights to be zero. The coefficient of adaptive L1 regularization is dynamically adjusted according to the weight entropy. When the entropy value is high, the regularization coefficient is increased to strengthen the sparsity effect; when the entropy value is low, the regularization coefficient is reduced to weaken the sparsity effect. The value range of the regularization coefficient is usually between 0.001 and 0.1, which is adjusted according to the task complexity. After applying L1 regularization constraint, usually about 20% to 30% of the weights are close to zero, forming a sparse weight distribution.

[0143] The residual connection mechanism calibrates the sparse weight distribution across the scale features. The core idea of residual connection is to establish a shortcut between the transformation before and after, to ensure that the information can be transmitted without loss. In the bidirectional attention mechanism, the residual connection combines the original attention weight with the sparse weight, and the combination ratio is usually 0.7:0.3, that is, 70% of the sparse weight and 30% of the original weight are added. This design ensures that the key information in the original weight will not be lost while pursuing sparsity. The cross-scale feature calibration is a process of rebalancing the features of different time scales. For the attention weight corresponding to the long-term state feature, the calibration tends to retain the low-frequency feature; for the attention weight corresponding to the short-term state feature, the calibration tends to retain the high-frequency feature. The calibration method adjusts the weight through a frequency-selective filter, and the parameters of the filter are set according to the task characteristics. The final attention weight has good sparsity and retains cross-scale feature information, which can effectively guide the multi-scale feature fusion process.

[0144] The application significantly improves the perception ability of key error modes in the capacity verification process of glass containers, especially in the case of environmental condition fluctuations or interference in the measurement process, can accurately identify and extract key features related to errors, provide reliable feature representation for the calculation of dynamic compensation values, and ultimately improve the accuracy and robustness of capacity verification.

[0145] In an optional embodiment, the step of adjusting the verification control parameters according to the dynamic compensation value to perform capacity verification on the glass container to obtain verification result data comprises:

[0146] According to the dynamic compensation value, determine the adjustment range of the liquid injection rate parameter and the adjustment range of the temperature control parameter;

[0147] Construct an adaptive control strategy library based on verification historical data, extract the optimal control parameter template from the control strategy library according to the type, specification and environmental conditions of the current glass container, and dynamically fine-tune the control parameters combined with the dynamic compensation value to generate a verification control curve;

[0148] According to the liquid injection rate parameter in the verification control curve, control the liquid injection unit to perform staged liquid injection operation, and adjust the verification environment temperature according to the temperature control parameter;

[0149] Collect liquid surface images and high-accuracy electronic balance data during liquid injection, record the liquid injection speed and environmental temperature, and obtain verification process data; comprehensively analyze the verification process data and the dynamic compensation value to generate a verification report containing capacity measurement value, error analysis result and quality evaluation information; feedback the verification control parameters and verification results to the adaptive control strategy library to update the optimal control parameter template.

[0150] For example, the dynamic compensation value reflects the trend of error change, and its value is usually within the range of ±0.5%. When the dynamic compensation value is positive, it indicates that the measured value is too large, and the liquid injection rate needs to be reduced and the temperature control accuracy needs to be improved. When the compensation value is negative, it indicates that the measured value is too small, and the liquid injection rate can be appropriately increased and the temperature control accuracy can be relaxed. The adjustment range of the liquid injection rate parameter is determined according to the size of the dynamic compensation value. For a 100-milliliter measuring cylinder, the standard liquid injection rate is 2 milliliters per second, and when the compensation value is 0.2%, the adjusted liquid injection rate range is 1.6-1.8 milliliters per second. When the compensation value is -0.2%, the adjusted liquid injection rate range is 2.2-2.4 milliliters per second. The adjustment range of the temperature control parameter is also related to the dynamic compensation value. The standard temperature control accuracy is ±0.5°C, and when the compensation value is 0.2%, the adjusted temperature control accuracy is ±0.3°C. When the compensation value is -0.2%, the adjusted temperature control accuracy can be relaxed to ±0.7°C. This parameter adjustment strategy ensures the directionality and effectiveness of the compensation effect.

[0151] The construction of the adaptive control strategy library is based on a large amount of historical calibration data, including calibration parameters and result records of different types and specifications of glass measuring vessels under various environmental conditions. The data structure of the control strategy library includes three main dimensions: glass measuring vessel feature dimension, environmental condition dimension, and control parameter dimension. The glass measuring vessel feature dimension includes information such as measuring vessel type (pipette, measuring cylinder, burette, etc.), nominal capacity (1 milliliter to 5000 milliliters), and accuracy grade (A grade, B grade). The environmental condition dimension includes parameters such as temperature range (15°C to 25°C), relative humidity range (40% to 70%), and air pressure range (95 kPa to 105 kPa). The control parameter dimension stores the optimal liquid injection rate curve, temperature control parameters, and actual calibration results under corresponding conditions. The strategy library uses a multi-dimensional index structure to support fast retrieval and similar condition matching, and the initial capacity is 5000 records, which will be continuously expanded and optimized as the calibration work progresses.

[0152] The extraction process of the optimal control parameter template adopts a similarity-based retrieval method. Given the type, specification, and environmental conditions of the current glass measuring cylinder, the similarity with each record in the strategy library is calculated. The similarity calculation considers the weighted differences of the characteristics of the measuring cylinder and environmental conditions. The matching weight of the measuring cylinder type is 0.4, the nominal capacity difference weight is 0.3, the environmental temperature difference weight is 0.2, and the humidity and air pressure difference weights are 0.05 each. The top three records with the highest similarity are selected, and the initial control parameter template is generated by weighted averaging. Taking a 100-milliliter Class A measuring cylinder as an example, under the condition of an environmental temperature of 20°C and a relative humidity of 60%, the initial liquid injection rate in the optimal control parameter template obtained by retrieval is 2.1 milliliters per second, the liquid injection rate near the scale line is 0.5 milliliters per second, and the temperature control range is ±0.4°C. The parameters in the template also include the liquid injection rate change point (5 milliliters from the nominal scale line) and the temperature fluctuation monitoring frequency (every 5 seconds).

[0153] The dynamic fine-tuning process applies dynamic compensation values to the optimal control parameter template to generate control parameters that adapt to the current calibration task. Fine-tuning uses a piecewise linear mapping method to map dynamic compensation values to the adjustment amount of each control parameter. For the liquid injection rate, the adjustment formula is: actual liquid injection rate = template liquid injection rate × (1-dynamic compensation value × adjustment coefficient), and the adjustment coefficient is usually set to 5. For the temperature control accuracy, the adjustment formula is: actual control accuracy = template control accuracy × (1-dynamic compensation value × adjustment coefficient), and the adjustment coefficient is usually set to 3. The fine-tuned control parameter combination forms the calibration control curve, which describes the variation of the liquid injection rate and temperature control parameters with time or liquid level during the entire calibration process. The control curve is represented by a piecewise function. For a 100-milliliter measuring cylinder, the curve may contain three stages: the initial stage (0-80 milliliters) with a liquid injection rate of 1.8 milliliters per second, the intermediate stage (80-95 milliliters) with a liquid injection rate of 0.8 milliliters per second, and the final stage (95-100 milliliters) with a liquid injection rate of 0.3 milliliters per second.

[0154] The liquid injection operation is performed by the injection unit, and the injection speed is adjusted in real time according to the injection rate parameter in the verification control curve. The injection unit is composed of a precision syringe pump, a solenoid valve and a flow sensor, which can realize accurate flow control in the range of 0.1 ml / s to 10 ml / s, and the control accuracy is better than ±1%. The staged liquid injection strategy is usually divided into a fast injection stage, a transition injection stage and a fine injection stage. The fast injection stage uses a higher injection rate to quickly fill the main part of the volumeter; the transition injection stage reduces the injection rate to approach the nominal scale line; the fine injection stage uses the lowest injection rate to accurately control the liquid level to the nominal scale line position. Taking a 100 ml graduated cylinder as an example, the fast injection stage injects from 0 ml to 85 ml at a rate of 1.8 ml / s, which takes about 47 seconds; the transition injection stage injects from 85 ml to 95 ml at a rate of 0.8 ml / s, which takes about 12.5 seconds; the fine injection stage injects from 95 ml to 100 ml at a rate of 0.3 ml / s, which takes about 16.7 seconds. The total time of the entire injection process is about 76.2 seconds, which saves about 15% of the time compared to fixed rate injection, while improving the liquid level control accuracy.

[0155] The temperature control parameter adjustment adopts closed-loop control, including temperature sensor, heating / cooling element and PID controller. The temperature sensor uses PT100 platinum resistance, with an accuracy of ±0.1℃ and a sampling frequency of 1Hz. The target value of the ambient temperature is set to 20℃, and the control range is adjusted according to the dynamic compensation value, usually between ±0.3℃ and ±0.7℃. The parameters of the PID controller are also dynamically adjusted according to the temperature control accuracy requirements. When higher accuracy is required, increase the proportional and integral coefficients and reduce the differential coefficient; when the accuracy requirement is wider, reduce the proportional and integral coefficients and increase the differential coefficient. For the case of ±0.3℃ control accuracy, the typical PID parameters are: proportional coefficient 5.0, integral coefficient 0.1, differential coefficient 0.5; for the case of ±0.7℃ control accuracy, the parameters are adjusted to: proportional coefficient 2.0, integral coefficient 0.05, differential coefficient 1.0. The response time of temperature control is usually within 30 seconds, ensuring the stability of the ambient temperature during verification.

[0156] The acquisition of the test process data includes two key links: liquid surface image acquisition and electronic balance data acquisition. The liquid surface image acquisition uses a high-resolution industrial camera with a resolution of 1920x1080 pixels and a frame rate of 30 fps, equipped with a macro lens and a ring LED light source to ensure clear imaging of the liquid surface. The image acquisition is installed near the scale line of the measuring cylinder to capture the morphological characteristics of the meniscus. The electronic balance data acquisition uses a high-accuracy electronic balance with an accuracy of 0.1 mg and a sampling frequency of 10 Hz, which transmits weight data in real time through a data acquisition interface. The liquid injection speed record is based on the output of the flow sensor and the control signal of the syringe pump, with a recording frequency of 10 Hz. The environmental temperature record is based on the output of the temperature sensor, with a recording frequency of 1 Hz. All collected data have a unified time stamp for time synchronization in subsequent analysis. The typical amount of test process data is: about 2000 frames of liquid surface image, about 800 points of electronic balance data, about 800 points of liquid injection speed data, and about 80 points of environmental temperature data.

[0157] The comprehensive analysis of the test process data and the dynamic compensation value uses a multi-source data fusion method. The liquid surface image data are processed by image processing algorithms to extract liquid surface position and morphological parameters, including liquid surface height, meniscus curvature, and liquid surface oscillation amplitude. The electronic balance data are filtered to obtain a smooth weight curve, and combined with the density of water to calculate the volume data. The liquid injection speed data and environmental temperature data are used to monitor the stability of the test process. These processed data are combined with the dynamic compensation value to calculate the final capacity measurement value and error index. The capacity measurement value calculation formula is: actual capacity = electronic balance measured volume x (1 + dynamic compensation coefficient), where the dynamic compensation coefficient is obtained from the dynamic compensation value through nonlinear mapping. Error analysis includes the calculation of absolute error, relative error, and repeatability error. The absolute error is the difference between the actual capacity and the nominal capacity, the relative error is the absolute error percentage of the nominal capacity, and the repeatability error is the standard deviation of multiple measurement results. Quality assessment is based on error indicators and standard requirements. For a class A 100 ml measuring cylinder, the allowable error range is ±0.1 ml. If the measurement error is within the allowable range, the evaluation result is "qualified"; if it exceeds the range, the evaluation result is "unqualified" and the specific out-of-tolerance value is given.

[0158] The feedback updating mechanism of the verification control parameters and the verification results is the key to the continuous optimization of the adaptive control strategy library. After each verification is completed, the current glass container characteristics, environmental conditions, control parameters and verification results are combined to form a new record, and it is evaluated whether the strategy library needs to be updated. The update strategy is based on the quality of the verification results. When the verification error is less than the error of the existing similar condition record, the new record replaces the original record; when the verification error is between the existing records, the new record is added to the strategy library; when the verification error is greater than the existing record, the reason is analyzed and it is decided whether to add it as a counterexample. The strategy library will also be optimized regularly, deleting long-term unused or poor performance records to keep the strategy library simple and efficient. The update frequency of the strategy library is adjusted with the use time. In the early stage, it is updated after each verification, and after stable operation, it can be reduced to once every 10 verifications. Through this continuous learning and optimization mechanism, the control strategy library can adapt to different verification conditions and container characteristics, and provide more and more accurate control parameter suggestions.

[0159] The present application can provide personalized verification schemes for different types and specifications of glass containers by dynamically adjusting the liquid injection rate and temperature control parameters. The introduction of the adaptive control strategy library enables experience accumulation and self-optimization. The staged liquid injection operation and accurate environmental control ensure the stability and reliability of the verification process.

[0160] The second aspect of the embodiment of the present application provides a glass container capacity verification process automatic control and error analysis system, comprising:

[0161] The first unit is used for calculating the environmental influence compensation value on the capacity of the glass container according to the specification parameter information of the glass container to be verified and the environmental temperature, air pressure and humidity;

[0162] The second unit is used for dynamically capturing and analyzing the liquid level image of the glass container by adopting a double-branch neural network structure, wherein the first branch extracts liquid surface static features through a spatial attention module, the second branch obtains liquid surface dynamic features through a time series attention module, the liquid surface static features, the liquid surface dynamic features and the weight data of the electronic balance are subjected to multi-modal feature fusion, and the compensated capacity value is calculated in combination with the influence compensation value;

[0163] The third unit is used for real-time error evaluation of the compensated capacity value to obtain initial error data;

[0164] The fourth unit is used for inputting the initial error data into a recurrent neural network with a hierarchical memory structure for dynamic error analysis, acquiring instantaneous change features of the liquid surface state and environmental parameters through a short-term memory unit, analyzing error accumulation features through a long-term memory unit, and adaptively weighting error features of different time scales through a bidirectional attention mechanism to obtain a dynamic compensation value;

[0165] a fifth unit configured to adjust a verification control parameter according to the dynamic compensation value, perform a capacity verification on the glass measuring cylinder, and obtain verification result data.

[0166] In a third aspect, the present application provides an electronic device, comprising:

[0167] a processor;

[0168] a memory for storing processor-executable instructions;

[0169] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0170] In a fourth aspect, the present application provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, implement the method described above.

[0171] The present application can be a method, apparatus, system and / or computer program product. The computer program product can include a computer-readable storage medium having stored thereon computer-readable program instructions that, when executed by a computer, cause the computer to carry out various aspects of the present application.

[0172] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for automatic control and error analysis of the process of capacity verification of glass measuring vessels, characterized in that, The method comprises the following steps: According to the specification parameter information of the glass measuring container to be detected, the environmental temperature, the air pressure and the humidity, a compensation value of the influence of the environment on the capacity of the glass measuring container is calculated; A double-branch neural network structure is used to dynamically capture and analyze the liquid level image of the glass measuring container, wherein the first branch extracts liquid surface static features through a spatial attention module, and the second branch obtains liquid surface dynamic features through a time series attention module; multi-modal feature fusion is performed on the liquid surface static features, the liquid surface dynamic features and the weight data of the electronic balance; and a compensated capacity value is calculated in combination with the compensation value. Real-time error evaluation is performed on the compensated capacity value to obtain initial error data. The initial error data is input into a recurrent neural network with a hierarchical memory structure for dynamic error analysis; the instantaneous change features of the liquid surface state and the environmental parameters are obtained through a short-term memory unit; the error accumulation features are analyzed through a long-term memory unit; a bidirectional attention mechanism is used to adaptively weight the error features of different time scales to obtain a dynamic compensation value. The verification control parameters are adjusted according to the dynamic compensation value, the capacity of the glass measuring container is verified, and verification result data is obtained. The step of extracting liquid surface static features through a spatial attention module in the first branch, obtaining liquid surface dynamic features through a time series attention module in the second branch, and performing multi-modal feature fusion on the liquid surface static features, the liquid surface dynamic features and the weight data of the electronic balance to calculate a compensated capacity value of the glass measuring container comprises: The liquid surface shape features are extracted through the spatial attention module in the first branch, the spatial attention module enhances the liquid surface profile curvature and capillary effect features based on channel attention weights and position attention weights to obtain liquid surface static features; The liquid surface oscillation features are modeled through the time series attention module in the second branch, the time series attention module calculates the correlation degree between multiple time series states based on time window weights, and dynamically aggregates the correlation degree through a convolution long short-term memory network to obtain liquid surface dynamic features; The liquid level image features are obtained by aligning and normalizing the liquid surface static features and the liquid surface dynamic features; The weight data of the electronic balance is mapped to the feature space to obtain weight data features; Multi-modal feature fusion is performed on the liquid level image features and the weight data features to obtain fusion features, and a compensated capacity value of the glass measuring container is calculated based on the fusion features and the compensation value. The step of inputting the initial error data into a recurrent neural network with a hierarchical memory structure for dynamic error analysis, obtaining the instantaneous change features of the liquid surface state and the environmental parameters through a short-term memory unit, analyzing the error accumulation features through a long-term memory unit, and adaptively weighting the error features of different time scales through a bidirectional attention mechanism to obtain a dynamic compensation value comprises: The short-term memory unit models the instantaneous changes of the liquid surface state parameters and the environmental parameters through a state vector, a forgetting gate and an input gate to obtain short-term state features, and the long-term memory unit dynamically aggregates the short-term state features based on error accumulation and a long-term state update function to obtain long-term state features; A bidirectional attention mechanism is constructed for the short-term state features and the long-term state features, attention scores between different time scale features are calculated through a query matrix and a key-value matrix, and the attention scores are normalized to obtain attention weights; The short-term state features and the long-term state features are fused based on the attention weights to obtain comprehensive state features, and the comprehensive state features are optimized through a dynamic weight update function to obtain optimized state features; The optimized state features are separated by feature separation, the separated short-term components are processed through a convolutional neural network to obtain short-term error features, and the separated long-term components are processed through a fully connected network to obtain long-term error features; The short-term error features and the long-term error features are combined based on adaptive weights to obtain a dynamic compensation value; The dynamic compensation value is compensated with the initial error data to obtain compensated error, and the compensated capacity value is corrected based on the compensated error to obtain a final capacity value.

2. The method of claim 1, wherein, The step of calculating the compensation value of the influence of the environment on the capacity of the glass measuring cylinder according to the specification parameter information of the glass measuring cylinder to be inspected and the environmental temperature, air pressure and humidity comprises: The specification parameter information includes the nominal capacity and the use temperature range of the glass measuring cylinder; The volume thermal deformation amount of the glass measuring cylinder and the density change amount of the liquid to be measured are calculated according to the specification parameter information and the environmental temperature data, and the volume thermal deformation amount and the density change amount are coupled to calculate a temperature compensation coefficient; The liquid surface tension change value and the water vapor partial pressure are calculated according to the air pressure data and the relative humidity data, and the surface tension change value and the water vapor partial pressure are weighted to calculate an environmental compensation coefficient; The temperature compensation coefficient and the environmental compensation coefficient are combined to calculate a comprehensive compensation coefficient; The compensation value of the influence of the environment on the capacity of the glass measuring cylinder is calculated according to the comprehensive compensation coefficient, the nominal capacity and the current measurement value; The deviation between the compensated capacity value and the nominal capacity is calculated, and when the deviation exceeds a preset deviation threshold, the weights of the temperature compensation coefficient and the environmental compensation coefficient are dynamically adjusted.

3. The method of claim 1, wherein, The step of calculating the compensated capacity value of the glass measuring cylinder based on the liquid level image features and the weight data features comprises: The liquid level image features and the weight data features are normalized to obtain normalized features, and self-attention and cross-modal attention between the normalized features are calculated through a multi-head attention mechanism to obtain a feature interaction matrix; A feature confidence evaluation network is constructed based on the feature interaction matrix, and the feature confidence evaluation network calculates the contribution degree of each modal feature to the capacity prediction to obtain a feature confidence score; normalizing the feature confidence score based on a temperature parameter to obtain a dynamic fusion weight; combining the dynamic fusion weight and the normalized feature in a weighted manner, and obtaining an enhanced fusion feature through a residual connection; constructing a multi-objective loss function including capacity prediction, modal consistency and feature reconstruction to optimize the enhanced fusion feature; concatenating the optimized fusion feature and the influence compensation value into a full connection layer, calculating a capacity prediction value through a multi-layer perception, calculating a prediction error between the capacity prediction value and a true capacity value, calculating a comprehensive weight update gradient based on the prediction error and the multi-objective loss function, and adaptively adjusting the dynamic fusion weight using the comprehensive weight update gradient; re-performing feature fusion using the updated dynamic fusion weight, and predicting the capacity value of the glass measuring cylinder in combination with the influence compensation value.

4. The method of claim 1, wherein, The steps of constructing a bidirectional attention mechanism include: constructing a forward query matrix according to the short-term state feature and a backward query matrix according to the long-term state feature, and performing feature enhancement on the forward query matrix and the backward query matrix through a nonlinear projection layer to obtain a forward enhanced query matrix and a backward enhanced query matrix; performing dot product operation on the forward enhanced query matrix and the long-term state feature to obtain a forward attention score, and performing dot product operation on the backward enhanced query matrix and the short-term state feature to obtain a backward attention score; calculating an adaptive temperature parameter based on the error gradient of the current time, applying the adaptive temperature parameter to the forward attention score and the backward attention score for dynamic scaling, and obtaining a forward attention weight and a backward attention weight through softmax normalization processing; constructing a dynamic balance factor that adaptively adjusts with the error change rate, and nonlinearly combining the forward attention weight and the backward attention weight based on the dynamic balance factor to obtain an initial fusion weight; applying adaptive L1 regularization constraint based on weight entropy to the initial fusion weight to obtain a sparse weight distribution, and performing cross-scale feature calibration on the sparse weight distribution through a residual connection mechanism to obtain the attention weight.

5. The method of claim 1, wherein, The steps of adjusting the verification control parameters according to the dynamic compensation value, performing capacity verification on the glass measuring cylinder, and obtaining verification result data include: determining the adjustment range of the liquid injection rate parameter and the adjustment range of the temperature control parameter according to the dynamic compensation value; constructing an adaptive control strategy library based on verification historical data, extracting an optimal control parameter template from the control strategy library according to the type, specification and environmental conditions of the current glass measuring cylinder, dynamically fine-tuning the control parameters in combination with the dynamic compensation value, and generating a verification control curve; controlling the liquid injection unit to perform staged liquid injection operation according to the liquid injection rate parameter in the verification control curve, and adjusting the verification environment temperature according to the temperature control parameter; collecting liquid surface images and high-accuracy electronic balance data during the liquid injection process, recording the liquid injection speed and the environmental temperature, and obtaining verification process data; The test process data is analyzed in combination with the dynamic compensation value to generate a test report containing capacity measurement value, error analysis result and quality evaluation information; The control parameters and test results of the test are fed back to the adaptive control strategy library to update the optimal control parameter template.

6. System for the automatic control of the process of verification of the capacity of glass measures and error analysis, for the implementation of the method according to any of the previous claims 1 to 5, characterized in that, The method comprises the following steps: The first unit is configured to calculate an environmental impact compensation value on the capacity of the glass measuring container according to the specification parameter information of the glass measuring container to be tested and the environmental temperature, air pressure and humidity; The second unit is configured to dynamically capture and analyze the liquid level image of the glass measuring container by using a double-branch neural network structure, wherein the first branch extracts liquid surface static features by a spatial attention module, and the second branch obtains liquid surface dynamic features by a time series attention module, and the liquid surface static features and the liquid surface dynamic features are fused with the weight data of the electronic balance to calculate a compensated capacity value in combination with the impact compensation value; The third unit is configured to perform real-time error evaluation on the compensated capacity value to obtain initial error data; The fourth unit is configured to input the initial error data into a recurrent neural network with a hierarchical memory structure to perform dynamic error analysis, obtain instantaneous change features of the liquid surface state and the environmental parameters by a short-term memory unit, analyze error accumulation features by a long-term memory unit, and obtain a dynamic compensation value by adaptively weighting error features of different time scales by using a bidirectional attention mechanism; The fifth unit is configured to adjust the test control parameters according to the dynamic compensation value, perform capacity test on the glass measuring container, and obtain test result data.

7. An electronic device, comprising: The method comprises the following steps: A processor; A memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method of any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 5.

Images