Acid Value Determination Method and System Based on AI Image Recognition

Abstract

The present invention proposes a method and system for determining acid value based on AI image recognition, belonging to the field of AI image recognition. The method includes: performing real-time data collection during the acid value titration process, and preprocessing the real-time collected data; constructing a color gradient dynamic extraction model to generate a comprehensive feature matrix; calculating the feature importance weights of feature mutation points, quantifying the contribution degree of feature mutation points in key time periods, weighting the comprehensive feature matrix to generate a weighted feature matrix, extracting the titration endpoint to obtain the titration endpoint prediction result, and calculating the titrant volume according to the titration endpoint; constructing a confidence score model to perform confidence scoring on the titration endpoint prediction time, and calculating the final titrant volume according to the final endpoint time; calculating the acid value according to the final endpoint time and generating an experimental report. The present invention effectively overcomes the shortcomings of the prior art, provides a comprehensive, efficient and intelligent solution for the acid value detection field, and has extremely high application value.

Description

Technical Field

[0001] The present invention belongs to the field of AI image recognition, and particularly relates to a method and system for determining acid value based on AI image recognition. Background Art

[0002] In the field of acid value detection, titration methods are widely used in fields such as chemical engineering, food, pharmaceuticals, and environmental detection to evaluate acidic components in liquid or solid samples. Traditional titration methods mainly rely on manual judgment of the titration endpoint or monitor color changes through simple optical detection devices. However, these methods have obvious limitations:

[0003] Large manual judgment error: Manual judgment of the endpoint is based on the subjective perception of the operator and is easily affected by personal experience, reaction speed, and changes in ambient light, resulting in inaccurate results and poor experimental repeatability.

[0004] Low robustness of simple optical detection: Some automated devices use single light sources or color sensors to detect color changes, but they are not suitable for complex experimental conditions (such as uneven illumination and changes in sample transparency), and the detection results are prone to deviate from the actual titration endpoint.

[0005] Lack of modeling of the dynamic characteristics of chemical reactions: Existing titration monitoring systems are mainly based on simple algorithms of static color or optical changes and cannot capture the dynamic changes during titration (such as color gradient trends or spectral feature changes), showing significant limitations in titration experiments with complex reactions.

[0006] Single application of spectral data: As an important means of characterizing reactions, the application of spectral data in titration detection is often single and not fully explored. Only the absorbance threshold is used to judge the endpoint, and it is impossible to comprehensively analyze the changes in color and spectrum for accurate analysis.

[0007] In view of the above problems, there is an urgent need for a method that can fully utilize multimodal data to capture the dynamic changes during titration under complex experimental conditions and combine intelligent analysis algorithms to achieve accurate identification of the titration endpoint and rapid calculation of acid value. Summary of the Invention

[0008] The object of the present invention is to provide a method and system for determining acid value based on AI image recognition, which effectively overcomes the shortcomings of the prior art and also provides a comprehensive, efficient, and intelligent solution for the field of acid value detection, with extremely high application value.

[0009] To achieve the above object, in the first aspect of the present invention, a method for determining acid value based on AI image recognition is provided, and the method includes:

[0010] S1. During the acid value titration process, real-time data acquisition is carried out, and the real-time acquired data is preprocessed based on temporal consistency to obtain preprocessed data; wherein, the data includes image data and spectral data;

[0011] S2. Construct a color gradient dynamic extraction model to extract image dynamic features from the preprocessed image data, construct an absorbance change dynamic extraction model to extract spectral dynamic features from the preprocessed spectral data, and synthesize the image dynamic features and spectral dynamic features to generate a comprehensive feature matrix;

[0012] S3. Calculate the feature importance weights of the feature mutation points according to the comprehensive feature matrix to quantify the contribution degree of the feature mutation points in the critical time period, then weight the comprehensive feature matrix based on the feature importance weights to generate a weighted feature matrix. Within the critical time window, the titration endpoint is extracted through a non-linear determination model based on the weighted feature matrix to obtain the titration endpoint prediction result. At the same time, the volume of the titrant is calculated according to the titration endpoint to obtain the titration endpoint prediction time and the volume of the titrant;

[0013] S4. Construct a confidence score model to score the confidence of the titration endpoint prediction time, judge whether it is necessary to correct the titration endpoint prediction time, and recalculate the final titrant volume according to the final endpoint time:

[0014] If the confidence score is less than the preset threshold, it is considered that there is a potential error in the titration endpoint prediction time, then correction is carried out, and the corrected endpoint time is output as the final endpoint time. The correction is realized through a non-linear error correction model, and the non-linear error correction model is realized by combining the time series smoothness of the weighted feature matrix;

[0015] If the confidence score is greater than or equal to the preset threshold, it is considered that there is no potential error in the titration endpoint prediction time, and it is directly output as the final endpoint time;

[0016] S5. Calculate the acid value according to the final endpoint time and generate an experimental report.

[0017] Furthermore, the preprocessing further includes introducing a dynamic illumination normalization model , and performing adaptive light intensity correction on the image data :

[0018] ;

[0019] Wherein, is the corrected image; is the width and height of the image; and are the average light intensity and standard deviation of the image respectively; is a regularization parameter to prevent over-normalization, is the pixel value at the i-th row and j-th column in the image, where i is the row index of the image and j is the column index of the image;

[0020] Use Gaussian filtering based on the frequency domain for noise suppression:

[0021] ;

[0022] where, is the preprocessed image data, is the frequency domain component of the image, is the noise control parameter, dynamically optimized through experiments;

[0023] The preprocessing further includes using an adaptive baseline correction model to perform baseline correction and dynamic feature enhancement on the spectral data:

[0024] ;

[0025] where, is the corrected spectral data, i.e., the preprocessed spectral data; is the second-order dynamic fitting baseline; is the dynamic fitting parameter, optimized in real time to adapt to different experimental conditions;

[0026] In the corrected spectral data extract the dynamic change rate of absorbance :

[0027] ;

[0028] where, is the wavelength of the spectral data, is the time, is the corrected spectral data.

[0029] Furthermore, the color gradient dynamic extraction model is expressed as:

[0030] ;

[0031] where, is the image feature matrix, and each column represents the feature vector of time ; is the preprocessed image data;

[0032] and are the convolution kernel and bias, specifically designed to capture color changes; is the color gradient dynamic extraction model; is a color gradient regularization term used to emphasize color mutation points; is a regularization coefficient that controls the weight of mutation points in the model; is an activation function that retains significant features of color changes;

[0033] Dynamically extract the model through color gradient Process the image frames at each time point and extract the color change sequence matrix :

[0034] ;

[0035] where each column represents the color feature at a time point, and T is the total number of time points;

[0036] The absorbance change dynamic extraction model is expressed as:

[0037] ;

[0038] where is the spectral feature matrix, and each column represents the absorbance feature at time and is the preprocessed spectral data; represents the rate of change of absorbance with time and is used to capture dynamic changes; is the second-order characteristic of the change of absorbance with wavelength and is used to identify gradually changing or mutating bands; is a regularization coefficient used to balance the change information of time and wavelength;

[0039] Through the model Process the spectral data at each time point to generate a spectral change feature matrix :

[0040] ;

[0041] where each column represents the spectral feature at a certain time point.

[0042] Furthermore, by combining the image dynamic features and spectral dynamic features, a comprehensive feature matrix is generated, which is expressed as:

[0043] Introduce a modal weighted fusion model to fuse and into a comprehensive feature matrix :

[0044] ;

[0045] Among them, and are weight matrices, which are dynamically optimized according to the importance of modal data; is a modal alignment regularization term, which is used to reduce the modal difference between image and spectral features; is a regularization parameter that controls the degree of modal alignment.

[0046] Furthermore, calculating the feature importance weights of feature mutation points according to the comprehensive feature matrix is specifically as follows:

[0047] ;

[0048] Among them, is the feature importance weight at a specific time point ; controls the influence of the feature intensity ; represents a specific time point , representing the specific time point being currently processed; is the loop variable index of the time point; controls the influence of the time proximity change ; is the feature vector of the t-th column in the comprehensive feature matrix, representing the multimodal features at a specific time point t, is the feature vector of the t-th column in the comprehensive feature matrix, representing the multimodal features at a specific time point t - 1; normalizing the denominator to ensure ; represents the second norm;

[0049] Based on the feature importance weight at a specific time point to form the feature importance weight vector to weight the comprehensive feature matrix to generate a weighted feature matrix , among which, the comprehensive feature matrix is a matrix containing the features of all time points:

[0050] ;

[0051] Among them, is the feature importance weight vector, and the time point with the maximum feature intensity is extracted from the weighted feature matrix , as well as the nearby time window , where is the window length, reflecting the key change stage of the titration reaction.

[0052] Furthermore, in the key time window Based on the weighted feature matrix within Extract the predicted titration endpoint time through the non - linear determination model Meanwhile, calculate the titrant volume according to the predicted titration endpoint time : :

[0053] ;

[0054] Wherein, is the predicted titration endpoint time, calculated based on the changing trend of multi - modal features; is the titrant volume calculation function, deduced by combining the titration rate and time.

[0055] Furthermore, the confidence score model is expressed as:

[0056] ;

[0057] Wherein, is the confidence score of the predicted titration endpoint time , with a value range of ; represents the amplitude of multi - modal feature changes near the endpoint, used to evaluate the significance of endpoint changes, is the weighted multi - modal feature vector of the titration endpoint, is the weighted multi - modal feature vector at the previous moment before the titration endpoint, wherein, is at the specific value of the predicted titration endpoint time ; represents the titrant volume change rate, used to capture abnormal titration reactions; and are adjustment parameters, respectively controlling the influence of feature changes and volume changes on the confidence;

[0058] The non - linear error correction model , is expressed as:

[0059] ;

[0060] Wherein, is the corrected titration endpoint time; is the correction term of the prediction error, obtained by weighted averaging the time - series features of , is the weighted multi - modal feature vector at time point t; is the error correction factor, used to balance the amplitude of correction.

[0061] ​​Further, step S5 specifically includes:

[0062] Obtain experimental parameters, including: titrant concentration, sample weight, and conversion factor;

[0063] Calculate the acid value according to the experimental parameters:

[0064] ;

[0065] Wherein, is the acid value result, the first part , is the basic acid value calculation formula, is the titrant concentration, is the final titrant volume, F is the conversion factor, is the weight of the sample; the correction term dynamically adjusts the acid value result according to the confidence level. When the confidence level is low, a conservative deviation is increased to reduce the influence of experimental errors, is the adjustment coefficient, which controls the weight of the confidence correction term on the result, () is the confidence score for the prediction of the titration endpoint;

[0066] Generate an experimental report according to the acid value, including:

[0067] The acid value result ;

[0068] The record of the titration process, including details of endpoint correction and confidence score;

[0069] Graphical display;

[0070] The complete data of the report, ensuring that the report contains all experimental input and output data.

[0071] In another aspect of the present invention, an acid value determination system based on AI image recognition is provided. The system includes:

[0072] An acid value titration acquisition unit, which is used to collect data in real time during the acid value titration process, and perform preprocessing based on temporal consistency on the real-time collected data to obtain preprocessed data; wherein, the data includes image data and spectral data;

[0073] A feature extraction unit, which is used to construct a color gradient dynamic extraction model to extract image dynamic features from the preprocessed image data, construct an absorbance change dynamic extraction model to extract spectral dynamic features from the preprocessed spectral data, and generate a comprehensive feature matrix by integrating the image dynamic features and spectral dynamic features;

[0074] A titration analysis unit is used to calculate the contribution degree of the characteristic mutation point in the key time period according to the comprehensive characteristic matrix, and then weight the comprehensive characteristic matrix based on the characteristic importance weight to generate a weighted characteristic matrix. In the key time window, the titration end point is extracted based on the weighted characteristic matrix through a non-linear decision model to obtain the titration end point prediction result. At the same time, the volume of the titrant is calculated according to the titration end point to obtain the titration end point prediction time and the volume of the titrant;

[0075] An analysis and correction unit is used to construct a confidence score model to score the confidence of the titration end point prediction time, judge whether it is necessary to correct the titration end point prediction time, and recalculate the final titrant volume according to the final end time:

[0076] If the confidence score is less than the preset threshold, it is considered that there is a potential error in the titration end point prediction time, and then correction is carried out. The corrected end time is output as the final end time. The correction is realized through a non-linear error correction model, and the non-linear error correction model is realized by combining the time series smoothness of the weighted characteristic matrix;

[0077] If the confidence score is greater than or equal to the preset threshold, it is considered that there is no potential error in the titration end point prediction time, and it is directly output as the final end time;

[0078] A report generation unit is used to calculate the acid value according to the final end time and generate an experimental report.

[0079] The beneficial technical effects of the present invention are at least as follows:

[0080] (1) By introducing an improved adaptive multi-modal analysis model (a non-linear decision model based on support vector machine), the present invention can capture the dynamic change trend of the liquid color in the titration process in real time. This technology can not only identify the mutation points of the color, but also analyze the change speed and direction of the color gradient, so as to accurately locate the titration end point. Compared with the traditional static color analysis method, this technology has stronger adaptability to the interference of ambient light, and solves the problems of large manual judgment error and low robustness of simple optical detection.

[0081] (2) Through the multi-modal Transformer model, the present invention deeply fuses the image data and spectral absorbance data collected in the titration process. This technology makes full use of the cross characteristics of spectral features (such as the derivative of absorbance change and the position of the wave peak) and image color changes to comprehensively judge the titration end point. Compared with the single spectral or image analysis method, this method can significantly improve the accuracy of titration end point detection under complex conditions and solve the problem of single application of spectral data.

[0082] (3)Based on the end - point detection, the present invention designs an automatic acid - value calculation module, which directly uses the data at the detection end - point (titrant volume, sample mass) in the acid - value calculation formula to generate experimental results and record them automatically. This module can significantly improve the efficiency and accuracy of the experiment, adapt to high - throughput detection scenarios, and solve the problem of poor repeatability in traditional experiments. Brief Description of the Drawings

[0083] The present invention will be further described with reference to the accompanying drawings. However, the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the following drawings.

[0084] Figure 1 It is a flow chart of the acid - value determination method based on AI image recognition in the embodiment of the present invention.

[0085] Figure 2 It is a framework diagram of the acid - value determination system based on AI image recognition in the embodiment of the present invention. Detailed Embodiments

[0086] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation to the present invention.

[0087] As Figure 1 shown, the acid - value determination method based on AI image recognition provided by the embodiment of the present invention includes the following steps S1 - S5:

[0088] S1. During the acid - value titration process, data is collected in real - time, and the real - time collected data is pre - processed based on temporal consistency to obtain pre - processed data; wherein, the data includes image data and spectral data.

[0089] Specifically, during the titration process, real - time collection:

[0090] Image data : Through a high - definition camera, an RGB image sequence with a frame rate of per second is collected to record the color change process of the titrant.

[0091] Spectral data : Through a multi - channel spectral sensor, the dynamic change of absorbance is recorded in units of wavelength , with a resolution of .

[0092] Synchronization of image and spectral data: ​

[0093] ;

[0094] Among them, is the acquisition timestamp, is the multi-modal data sequence after synchronization. Innovative analysis: To overcome the challenges of fast and instantaneous changes in the titration reaction rate, a dynamic time synchronization method is adopted to ensure the temporal consistency of image and spectral data, thus avoiding the influence of asynchronous errors on feature analysis.

[0095] Furthermore, to address issues such as uneven illumination and complex backgrounds in the experimental environment and ensure the accurate capture of color changes in the titrant.

[0096] Introduce a dynamic illumination normalization model , and perform adaptive light intensity correction on the image data :

[0097] ;

[0098] Among them, is the corrected image; is the width and height of the image; and are the mean and standard deviation of the light intensity of the image respectively; is a regularization parameter (dynamically adjusted) to prevent over-normalization. Different from traditional normalization, the dynamic adjustment of

[0099] Noise suppression: Use Gaussian filtering based on the frequency domain :

[0100] ;

[0101] Among them, is the frequency domain component of the image, is the noise control parameter, dynamically optimized through experiments. While removing high-frequency noise, the color gradient characteristics of the titrant are retained.

[0102] Furthermore, solve the baseline drift problem of spectral data caused by experimental equipment offset and external interference during the titration process, and at the same time extract key dynamic absorbance change characteristics.

[0103] Model design: Use an adaptive baseline correction model :

[0104] ;

[0105] Among them, is the corrected spectral data; is the second-order dynamic fitting baseline; is the dynamic fitting parameter, which is optimized in real time to adapt to different experimental conditions.

[0106] In the corrected spectral data extract the dynamic change rate of absorbance:

[0107] ;

[0108] Capture the instantaneous characteristics of the absorbance change, providing dynamic feature input for subsequent endpoint recognition.

[0109] Furthermore, output the preprocessed data and , where:

[0110] is the image sequence after dynamic normalization and noise suppression, containing accurate color change characteristics;

[0111] is the spectral data after correction and enhancement, retaining the dynamic characteristics of absorbance during titration.

[0112] The output is the input for the subsequent steps, ensuring a unified data format and high-quality signals.

[0113] S2. Construct a dynamic color gradient extraction model to extract image dynamic features from the preprocessed image data, construct a dynamic absorbance change extraction model to extract spectral dynamic features from the preprocessed spectral data, and synthesize the image dynamic features and spectral dynamic features to generate a comprehensive feature matrix.

[0114] Specifically, receive the preprocessed data output in step 1 and :

[0115] is the image sequence after illumination normalization and noise suppression, containing the time-series change characteristics of the titrant color;

[0116] is the spectral data after baseline correction, containing the information of the absorbance of the titrant changing with time and wavelength.

[0117] Time synchronization condition: Ensure that and are aligned according to the timestamp, thus ensuring the dynamic consistency of subsequent feature extraction.

[0118] Furthermore, extract the time-series characteristics of the titrant color change, especially the dynamic change of the color gradient, to capture the information of color gradual change and mutation during titration.

[0119] Model Design: Define a Dynamic Color Gradient Extraction Model :

[0120] ;

[0121] Among them, is the image feature matrix, and each column represents the feature vector at a certain time ; and are the convolution kernel and bias, which are specially designed to capture color changes; is the color gradient regularization term, which is used to emphasize color mutation points; is the regularization coefficient, which controls the weight of mutation points in the model; is the activation function, which retains the significant features of color changes.

[0122] By passing the image frames at each time point through the model , a color change sequence matrix is extracted: :

[0123] ;

[0124] where each column is the color feature at a certain time point, and T is the total number of time points.

[0125] Furthermore, the goal is to extract the key features of absorbance changes in spectral data, including the dynamic change rate over time and the gradual change pattern within the wavelength range.

[0126] Model Design: Construct a Dynamic Absorbance Change Extraction Model :

[0127] ;

[0128] Among them, is the spectral feature matrix, and each column represents the absorbance feature at a certain time ; represents the change rate of absorbance over time and is used to capture dynamic changes; is the second-order characteristic of absorbance change with respect to wavelength and is used to identify the bands with gradual or sudden changes; is the regularization coefficient, which is used to balance the change information in time and wavelength.

[0129] By passing the spectral data at each time point through the model , a spectral change feature matrix is generated: :

[0130] ;

[0131] Each column represents the spectral characteristics at a certain time point.

[0132] Furthermore, by integrating the image and spectral characteristics, a unified feature matrix is generated to capture the dynamic change patterns during the titration process.

[0133] Fusion method: Introduce a modal weighted fusion model , and and are fused into a comprehensive feature matrix :

[0134] ;

[0135] wherein, and are weight matrices that are dynamically optimized according to the importance of modal data; is a modal alignment regularization term used to reduce the modal differences between the image and spectral characteristics; is a regularization parameter that controls the degree of modal alignment.

[0136] Finally, the output: the comprehensive feature matrix , each column represents the multimodal characteristics at a certain moment, and is used for the precise identification of the subsequent titration endpoint. By dynamically extracting and fusing the image and spectral characteristics, it is ensured that the feature matrix comprehensively and accurately reflects the dynamic process of the titration reaction, providing high-quality input for the core task of the patent.

[0137] S3. Calculate the feature importance weights of the feature mutation points based on the comprehensive feature matrix to quantify the contribution degree of the feature mutation points in the key time period, and then weight the comprehensive feature matrix based on the feature importance weights to generate a weighted feature matrix. Within the key time window, extract the titration endpoint through a non-linear decision model based on the weighted feature matrix to obtain the titration endpoint prediction result, and at the same time calculate the titrant volume according to the titration endpoint to obtain the titration endpoint prediction time and the titrant volume.

[0138] Specifically, receive the output comprehensive feature matrix of step 2 , and its time series has been synchronized:

[0139] is a matrix generated by the modal weighted and regularized fusion of the image feature and the spectral feature , and each column represents the multimodal characteristics at a certain moment, reflecting the dynamic changes in the color and absorbance of the titrant.

[0140] Furthermore, for the comprehensive feature matrix Assign importance weights to each column (i.e., the features at each time point) to capture the contribution degrees of feature mutation points and key time periods. Define the formula for assigning feature importance weights:

[0141] ;

[0142] where is the feature importance weight at a specific time point ; controls the influence of the feature intensity (i.e., the combined intensity of color and absorbance changes); represents a specific time point , representing the specific time point currently being processed; is the loop variable index of the time point; controls the influence of the time proximity change (i.e., feature mutation), emphasizing the key time points, is the feature vector of the t-th column in the matrix, representing the multi-modal features at a specific time point t, is the feature vector of the t-th column in the matrix, representing the multi-modal features at a specific time point t - 1; the denominator is normalized to ensure . This weight formula not only considers the feature intensity but also introduces the feature mutation term , which is specifically optimized for the mutation phenomenon near the titration endpoint; represents the two-norm;

[0143] Furthermore, based on the feature importance weight at the time point weight the feature matrix to generate the weighted feature matrix , where the feature matrix is the matrix containing all time point features;

[0144] ;

[0145] Extract the time point with the maximum feature intensity in the weighted feature matrix , as well as the nearby time window , where is the window length, reflecting the key change stage of the titration reaction. Through feature weight weighting and window extraction, it is possible to focus on the time period with significant feature changes during the titration process, thereby avoiding misjudgment and noise interference.

[0146] Furthermore, within the key time window , extract the titration endpoint through the non-linear determination model , and at the same time according to Calculating the volume of the titrant .

[0147] ;

[0148] Among them, is the predicted time of the titration endpoint, which is calculated based on the changing trend of multimodal features; is the titrant volume calculation function, which is deduced by combining the titration speed and time.

[0149] Among them, the nonlinear determination model uses the support vector machine as the nonlinear determination model. This nonlinear determination model uses the support vector machine (SVM) to predict the titration endpoint time according to the changing trend of multimodal features. By transforming the image and spectral features during the titration process into a high-dimensional feature space, the support vector machine can effectively capture the complex nonlinear relationships between features.

[0150] Specifically, the support vector machine (SVM) maps the multimodal features ( ) at each moment to a high-dimensional feature space to capture the nonlinear feature change patterns during the titration process. By introducing a nonlinear kernel function, the SVM can transform the linearly inseparable problem in the original feature space into a linearly separable situation in the high-dimensional space. In this high-dimensional space, the SVM finds an optimal hyperplane to distinguish different stages of the titration process, maximizes the distance from each category of data points to the decision boundary, and classifies through support vectors. The SVM performs excellently in dealing with complex nonlinear relationships. Therefore, it can identify the time points with the most drastic feature changes during the titration process and accurately predict the time of the titration endpoint. In this way, the SVM can not only identify the feature differences in each stage of the titration process but also effectively capture the complex nonlinear relationships between the image and spectral features, thereby improving the accuracy of titration endpoint prediction. Especially in the case of multiple factors such as light, sample composition, and reaction rate intertwining during the reaction process, it can make more accurate predictions.

[0151] In the training stage, first, it is necessary to collect training data containing multimodal features. Each data point will correspond to a label information (such as the time of the titration endpoint). The SVM maps these features to a high-dimensional space and searches for the optimal hyperplane in this space that can distinguish different titration stages or key time points. Through an optimization algorithm, the SVM adjusts its kernel function parameters (such as the parameters of the radial basis kernel function) and regularization parameters to ensure maximizing the interval between categories while avoiding overfitting, thereby improving the generalization ability of the model.

[0152] During the usage phase, the pre-trained SVM model will receive new multi-modal feature data. These data will be mapped to the same high-dimensional feature space as in the training phase and classified by the hyperplane determined in the model. Based on the decision boundary obtained in the training phase, the SVM will predict the key time points during the titration process, especially the time of the titration endpoint. In this way, the SVM can handle complex non-linear relationships, capture subtle changes between features, and give accurate predictions of the titration endpoint.

[0153] Finally, output the predicted time of the titration endpoint and the volume of the titrant , for use by the subsequent acid value calculation module. This step accurately extracts the titration endpoint through multi-modal feature analysis and a dynamic determination model, solving the complex determination problem of dynamic changes in the titration reaction.

[0154] S4. Construct a confidence score model to score the confidence of the predicted time of the titration endpoint, determine whether the predicted time of the titration endpoint needs to be corrected, and recalculate the final volume of the titrant according to the final endpoint time.

[0155] Specifically, receive the output of step 3: the preliminary predicted time of the titration endpoint and the corresponding volume of the titrant , as well as the comprehensive feature matrix :

[0156] is the predicted value of the titration endpoint time, obtained by multi-modal feature analysis;

[0157] is the volume of the titrant calculated based on the titration speed and ;

[0158] is a matrix containing time series features, describing the dynamic changes of color and absorbance during the titration process.

[0159] This data is used to further evaluate the reliability of the endpoint prediction and make necessary corrections.

[0160] Furthermore, define a confidence score model , used to quantify 's credibility. The core formula is:

[0161] ;

[0162] where is the confidence score of , with a value range of ; Indicates the amplitude of multimodal feature changes near the end point, used to evaluate the significance of end point changes; among them, is the specific value at the predicted titration end point time . Indicates the titrant volume change rate (the calculation formula is shown in Step 3), used to capture abnormal titration reactions; and are adjustment parameters, respectively controlling the influence of feature changes and volume changes on the confidence level.

[0163] The feature change term captures the significant features of the titration end point, while the volume change term further verifies the credibility of the predicted value through the smoothness of the reaction.

[0164] Furthermore, when is less than the preset threshold , it is considered that there are potential errors and enter the correction stage. The correction is achieved through a non-linear error correction model , and the core process combines the time series smoothness of the comprehensive feature matrix:

[0165] ;

[0166] Among them, is the corrected titration end point time; is the correction term of the prediction error, obtained by weighted averaging the time series features of ; is the error correction factor, used to balance the amplitude of the correction. Through the smoothing and weighting mechanisms, combined with the time distribution of significant features, the error of the preliminary prediction is corrected to avoid deviations caused by single-point features.

[0167] Furthermore, according to the corrected end point time , recalculate the final titrant volume :

[0168] ;

[0169] Among them is the flow rate model, calculated by combining the actual parameters of the experimental equipment.

[0170] Finally, output the corrected end point time and the titrant volume , and attach the confidence score for use by the subsequent acid value calculation module.

[0171] S5. Calculate the acid value based on the final end point time and generate an experimental report.

[0172] Specifically, receive the output of step 4:

[0173] End time : The final titration end time after confidence analysis and correction;

[0174] Titrant volume : Corresponding to the amount of titrant used;

[0175] Confidence score : Quantify the reliability of end point prediction for subsequent result adjustment.

[0176] Combine with known experimental parameters:

[0177] Titrant concentration (mol / L);

[0178] Sample weight (g);

[0179] Conversion factor (56.1, converting the amount of substance to acid value unit).

[0180] These data form the basic input for acid value calculation to ensure the accuracy of the calculation process.

[0181] Furthermore, define the acid value calculation formula, including the basic part and the innovative correction term:

[0182] ;

[0183] The first part: , is the basic acid value calculation formula;

[0184] Correction term : Dynamically adjust the acid value result according to the confidence. When the confidence is low, increase the conservative deviation to reduce the influence of experimental error;

[0185] is the adjustment coefficient to control the weight of the confidence correction term on the result.

[0186] Furthermore, the confidence in the titration end point prediction directly reflects the reliability of the end time . In the case of low confidence, by amplifying the adjustment amplitude, the error of the measurement result can be effectively reduced. Different from direct addition, here the overall acid value is dynamically adjusted by multiplication, which can better reflect the global influence of confidence on the experimental result.

[0187] All parameters in the formula maintain the corresponding relationship with the experimental data and physical processes to ensure that the calculation result has practical significance.

[0188] Further, acid value calculation and verification:

[0189] Calculate the acid value item by item according to the formula:

[0190] Amount of substance of the titrant: ;

[0191] Core acid value calculation: ;

[0192] Confidence correction: Multiply by , and dynamically adjust the final result.

[0193] Verification steps: Test low-confidence samples and high-confidence samples separately, and check the rationality and accuracy of the correction term for the result distribution.

[0194] Further, report content design:

[0195] Acid value result , with the complete calculation process and input parameters ( etc.);

[0196] Record of the titration process, including details of endpoint correction and confidence score;

[0197] Graphical display: Titrant volume change curve and acid value calculation curve, intuitively reflecting the titration process and results;

[0198] Complete data of the report, ensuring that the report contains all experimental input and output data.

[0199] Finally, output method:

[0200] Generate an exportable PDF report;

[0201] Store it in the experimental management system to support subsequent comparison and review;

[0202] Provide a visualization interface for users to view chart and numerical details.

[0203] This step ensures the reliability and adaptability of the measurement results through the innovative design and dynamic adjustment of the acid value calculation formula; the experimental report module provides detailed data records and visualization support, providing comprehensive guarantee for experimental analysis and optimization.

[0204] As Figure 2 shown, in another embodiment of the present invention, an acid value determination system based on AI image recognition is provided, and the system includes:

[0205] An acid value titration acquisition unit 501, configured to perform real-time data acquisition during the acid value titration process, and perform preprocessing based on time series consistency on the real-time acquired data to obtain preprocessed data; wherein, the data includes image data and spectral data;

[0206] A feature extraction unit 502 is configured to construct a dynamic color gradient extraction model to extract image dynamic features from the preprocessed image data, construct a dynamic absorbance change extraction model to extract spectral dynamic features from the preprocessed spectral data, and generate a comprehensive feature matrix by integrating the image dynamic features and the spectral dynamic features;

[0207] A titration analysis unit 503 is configured to calculate the contribution degree of the feature mutation point in the key time period by quantifying the feature importance weight according to the comprehensive feature matrix, then weight the comprehensive feature matrix based on the feature importance weight to generate a weighted feature matrix, and within the key time window, extract the titration end point through a non-linear decision model based on the weighted feature matrix to obtain a titration end point prediction result. At the same time, calculate the volume of the titrant according to the titration end point to obtain the titration end point prediction time and the volume of the titrant;

[0208] An analysis and correction unit 504 is configured to construct a confidence score model to perform a confidence score on the titration end point prediction time, determine whether it is necessary to correct the titration end point prediction time, and recalculate the final titrant volume according to the final end point time:

[0209] If the confidence score is less than a preset threshold, it is considered that there is a potential error in the titration end point prediction time, and then correction is performed. The corrected end point time is output as the final end point time. The correction is realized through a non-linear error correction model, and the non-linear error correction model is realized by combining the time series smoothness of the weighted feature matrix;

[0210] If the confidence score is greater than or equal to the preset threshold, it is considered that there is no potential error in the titration end point prediction time, and it is directly output as the final end point time;

[0211] A report generation unit 505 is configured to calculate the acid value according to the final end point time and generate an experimental report.

[0212] It should be noted that the above-described work flow is only illustrative and does not limit the protection scope of the present invention. In actual applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and no limitation is made here.

[0213] In addition, for the technical details not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and details are not described here again.

[0214] It should be noted that the above-described work process is only illustrative and does not limit the protection scope of the present invention. In actual applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and no limitation is made here.

[0215] In addition, for the technical details not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and details will not be repeated here.

[0216] It should be noted that in this article, the terms "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article or system. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or system including that element.

[0217] The serial numbers of the above embodiments of the present invention are only for description and do not represent the superiority or inferiority of the embodiments.

[0218] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory / random access memory, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

[0219] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for determining acid value based on AI image recognition, characterized in that: The method comprises: S1. During the acid value titration process, data is collected in real time, and the real-time collected data is preprocessed based on time series consistency to obtain preprocessed data; wherein the data includes image data and spectral data; S2, constructing a color gradient dynamic extraction model to extract image dynamic features from the preprocessed image data, constructing an absorbance change dynamic extraction model to extract spectral dynamic features from the preprocessed spectral data, and combining the image dynamic features and spectral dynamic features to generate a comprehensive feature matrix; S3, calculating the feature importance weight of the feature mutation point according to the comprehensive feature matrix, which is used to quantify the contribution of the feature mutation point in the key time period, and then weighting the comprehensive feature matrix based on the feature importance weight to generate a weighted feature matrix, and within the key time window, extracting the titration endpoint through a nonlinear judgment model based on the weighted feature matrix to obtain a titration endpoint prediction time, and calculating the titrant volume according to the titration endpoint prediction time; S4. Construct a confidence scoring model to score the confidence of the titration endpoint prediction time, determine whether the titration endpoint prediction time needs to be corrected, and recalculate the final titrant volume according to the final endpoint time: If the confidence score is less than a preset threshold, it is considered that there is a potential error in the titration endpoint prediction time, and correction is performed, and the corrected endpoint time is output as the final endpoint time. The correction is achieved through a nonlinear error correction model, and the nonlinear error correction model is achieved in combination with the time series smoothness of the weighted feature matrix; If the confidence score is greater than or equal to the preset threshold, it is considered that there is no potential error in the predicted titration endpoint time and it is directly output as the final endpoint time; S5. Calculate the acid value according to the final endpoint time and generate a test report; The feature importance weight of the feature mutation point is calculated according to the comprehensive feature matrix, specifically: ； in, A specific time point The feature importance weights of Controlling feature strength The impact of Represents a specific point in time that is currently being processed; is the loop variable index at the time point; Controlling temporal proximity changes The impact of is the eigenvector of the tth column in the comprehensive feature matrix, representing the multimodal features at a specific time point t. is the first The feature vector of the column represents the multimodal features at a specific time point t-1; the denominator is normalized to ensure ; represents the two-norm; Based on time point The feature importance weights The feature importance weight vector The comprehensive feature matrix Perform weighting to generate a weighted feature matrix , where the comprehensive feature matrix is a matrix containing features at all time points: ； in, is the feature importance weight vector, in the weighted feature matrix Extract the time point with the largest feature intensity , and critical time windows ,in is the window length, reflecting the key change stage of the titration reaction; Among them, in the critical time window Internally, based on the weighted feature matrix Through nonlinear decision model Extract titration endpoint prediction time , and predict the time based on the titration endpoint Calculating titrant volume : ； in, is the predicted time of the titration endpoint, given by Calculated based on the changing trend of multimodal features; It is a titrant volume calculation function, combining titration rate and time calculation; Wherein, the confidence scoring model is expressed as: ； in, Predicting the time for the titration endpoint The confidence score of ; Indicates the magnitude of multimodal feature changes near the endpoint, which is used to evaluate the significance of the endpoint change. is the weighted multimodal feature vector of the titration endpoint, is the weighted multimodal feature vector at the moment before the titration endpoint, where yes Predicting the time at the titration endpoint The specific value of Indicates the rate of change of titrant volume, used to capture abnormal titration reactions; and are adjustment parameters that control the impact of feature changes and volume changes on confidence respectively; The nonlinear error correction model , expressed as: ； in, is the corrected titration endpoint time; is the correction term for the prediction error, given by The time series features are weighted averaged to obtain: is the weighted multimodal feature vector at time point t; is the error correction factor, used to balance the amplitude of the correction; nonlinear determination model Support vector machine is used as the nonlinear judgment model.

2. The method for determining acid value based on AI image recognition according to claim 1, characterized in that: The preprocessing also includes introducing a dynamic illumination normalization model , for image data To perform adaptive light intensity correction: ； in, is the rectified image; is the image width and height; and are the mean and standard deviation of the light intensity of the image, respectively; is a regularization parameter to prevent over-normalization, is the pixel value of the i-th row and j-th column in the image, i is the row index of the image, and j is the column index of the image; Using Gaussian filtering based on frequency domain To perform noise suppression: ； in, is the preprocessed image data, is the frequency domain component of the image, is the noise control parameter, which is dynamically optimized through experiments; The preprocessing also includes using an adaptive baseline correction model Baseline correction and dynamic feature enhancement of spectral data: ； in, is the corrected spectral data, i.e. the preprocessed spectral data; is the second-order dynamic fitting baseline; It is a dynamic fitting parameter, which is optimized in real time to adapt to different experimental conditions; After correction, the spectral data Extract the dynamic change rate of absorbance : ； in, is the wavelength of the spectral data, For time, is the corrected spectral data; Represents the rate of change of absorbance over time and is used to capture dynamic changes.

3. The method for determining acid value based on AI image recognition according to claim 2, characterized in that: The color gradient dynamic extraction model is expressed as: ； in, is the image feature matrix, each column represents the time The eigenvector of is the preprocessed image data; and are convolution kernels and biases, specifically designed to capture color variations; It is a color gradient dynamic extraction model; is the color gradient regularization term, which is used to emphasize the color mutation point; is the regularization coefficient, which controls the weight of the mutation point in the model; is an activation function that retains the salient features of color changes; Dynamically extract models via color gradients Process the image frames at each time point , extract the color change sequence matrix : ； Among them, each column is the color feature of a time point, and T is the total number of time points; The absorbance change dynamic extraction model , expressed as: ； in, is the spectral feature matrix, each column represents the time The absorbance characteristics of is the preprocessed spectral data; It represents the rate of change of absorbance over time and is used to capture dynamic changes; It is the second-order characteristic of absorbance changing with wavelength, which is used to identify bands with gradual or sudden changes; is the regularization coefficient, which is used to balance the variation information of time and wavelength; By Model For each time point, the spectral data Processing to generate spectral change feature matrix : ； Each column represents the spectral characteristics at a certain time point.

4. The method for determining acid value based on AI image recognition according to claim 3, characterized in that: The image dynamic features and the spectrum dynamic features are integrated to generate a comprehensive feature matrix, which is expressed as: Introducing modal weighted fusion model ,Will and Fusion into comprehensive feature matrix : ； in, and is a weight matrix, which is dynamically optimized according to the importance of the modal data; is the modality alignment regularization term, which is used to reduce the modality difference between image and spectral features; is a regularization parameter that controls the degree of mode alignment.

5. The method for determining acid value based on AI image recognition according to claim 1, characterized in that: The S5 specifically includes: Obtain experimental parameters, including titrant concentration, sample weight, and conversion factor; Calculate the acid value according to the experimental parameters: ； in, For acid value results, part 1 , is the basic acid value calculation formula, is the titrant concentration, is the final titrant volume, F is the conversion factor, is the weight of the sample; correction term Dynamically adjust the acid value results according to the confidence level. When the confidence level is low, increase the conservative deviation to reduce the impact of experimental error. is the adjustment coefficient, which controls the weight of the confidence correction term on the result. Provide confidence scores for titration endpoint predictions; Generate a lab report based on the acid value.

6. The method for determining acid value based on AI image recognition according to claim 1, characterized in that: The contents of the acid value generation test report include: Acid value results ; Titration process records, including endpoint correction details and confidence scores; Graphical display; Complete data for the report, ensuring that the report contains all experimental input and output data.

7. A system for implementing the method for determining acid value based on AI image recognition as claimed in any one of claims 1 to 6, characterized in that: The system comprises: An acid value titration acquisition unit, used for real-time data acquisition during the acid value titration process, and preprocessing the real-time acquired data based on time sequence consistency to obtain preprocessed data; wherein the data includes image data and spectral data; A feature extraction unit is used to construct a color gradient dynamic extraction model to extract image dynamic features from the preprocessed image data, to construct an absorbance change dynamic extraction model to extract spectral dynamic features from the preprocessed spectral data, and to generate a comprehensive feature matrix by combining the image dynamic features and spectral dynamic features; A titration analysis unit, for calculating the feature importance weight according to the comprehensive feature matrix to quantify the contribution of the feature mutation point in the key time period, and then weighting the comprehensive feature matrix based on the feature importance weight to generate a weighted feature matrix, and within the key time window, extracting the titration endpoint through a nonlinear judgment model based on the weighted feature matrix to obtain a titration endpoint prediction time, and calculating the titrant volume according to the titration endpoint prediction time; The analysis and correction unit is used to construct a confidence scoring model to perform confidence scoring on the titration endpoint prediction time, determine whether the titration endpoint prediction time needs to be corrected, and recalculate the final titrant volume based on the final endpoint time: If the confidence score is less than a preset threshold, it is considered that there is a potential error in the titration endpoint prediction time, and correction is performed, and the corrected endpoint time is output as the final endpoint time. The correction is achieved through a nonlinear error correction model, and the nonlinear error correction model is achieved in combination with the time series smoothness of the weighted feature matrix; If the confidence score is greater than or equal to the preset threshold, it is considered that there is no potential error in the predicted titration endpoint time and it is directly output as the final endpoint time; The report generation unit is used to calculate the acid value according to the final endpoint time and generate an experimental report.

Images