CN114295704B - A micro-concentration gradient solution electrochemical determination method based on feature parameter extraction - Google Patents

Abstract

The invention discloses a micro-concentration gradient solution electrochemical determination method based on characteristic parameter extraction. The method is as follows: 1. Perform cyclic voltammetry detection, chronoamperometry detection and differential pulse voltammetry detection on multiple tested solutions with micro-concentration gradients, and obtain cyclic voltammetry detection data and timing of different multiple tested solutions. Current detection data and differential pulse voltammetry detection data. 2. Automatically extract graphic features from the obtained detection data. 3. Use the feature values and corresponding concentration labels to obtain the input data set. 4. Construct a concentration determination model and use the input data set obtained in step three to train the concentration determination model. The trained concentration determination model is used to detect the concentration of the tested solution with unknown concentration. Aiming at the problems of overlapping electrochemical detection curves and high noise of micro-concentration gradient detection substances, the present invention adopts multiple detection methods for parallel comparison to achieve effective information mining of detection signals and accurate concentration determination.

Description

A micro-concentration gradient solution electrochemical determination method based on feature parameter extraction

Technical field

The patent of this invention belongs to a rapid detection method for micro-concentration gradient solutions, and specifically relates to a rapid detection method for micro-concentration gradient solutions based on an integrated model of electrochemical detection technology and feature parameter extraction.

Background technique

In recent years, fields such as chemicals, pharmaceutical production, and food safety require on-site concentration determination of various micro-concentration gradient solutions, such as antibiotics, bacterial inhibitors, and food additives. Electrochemical detection technology is increasingly used in the determination of characteristic substance concentrations due to its advantages of simple operation and high sensitivity. It is a technology that collects electrochemical response signals and converts them into electrical signals that can be identified and detected, and finally analyzes and processes these electrical signals. Commonly used electrochemical detection methods include cyclic voltammetry and chronoamperometry. method and differential pulse voltammetry, etc. In the application process of electrochemical detection, the portable potentiostat is an indispensable instrument in electrochemical field testing. It can control the electrode potential to a set value to achieve the purpose of constant potential polarization. Compared with traditional electrochemical workstations, while realizing rapid detection of cyclic voltammetry of trace features based on a portable potentiostat system, it also needs to solve the problem of low sensitivity coefficient of functional modules. The low sensitivity of the portable potentiostat system based on on-site detection results in high background noise in the detection signal. It also causes overlapping detection curves of micro-concentration gradient characteristic solution solutions and unclear peak shape and peak height characteristics during the operation of different detection personnel. And other issues. The above problems bring great challenges to the rapid identification of trace feature solutions.

In order to solve the above problems and achieve on-site rapid measurement of micro-concentration gradient solutions, an integrated model for extracting characteristic parameters of electrochemical detection signals is proposed. This model integrates data preprocessing, feature extraction, feature dimensionality reduction and concentration determination. Compared with the traditional peak height and concentration fitting method, this method improves the analysis efficiency of electrochemical detection signals and obtains more detections. The information contained in the signal improves the accuracy of concentration determination. Currently, there are many studies on the concentration determination of electrochemical detection data signals, but there are still many quantitative analyzes that need to be improved on information mining of on-site rapid electrochemical detection signals and rapid electrochemical determination in micro-concentration gradient solutions.

Contents of the invention

The main goal of the patent of this invention is to realize a method for on-site rapid detection of micro-concentration gradient solutions based on an integrated model of electrochemical detection technology and feature extraction.

The micro-concentration gradient solution determination method based on the integrated feature extraction model includes the following steps:

Step 1: Perform cyclic voltammetry, chronoamperometry and differential pulse voltammetry on multiple test solutions with micro-concentration gradients to obtain cyclic voltammetry and chronoamperometry detection data for different multiple test solutions. and differential pulse voltammetry detection data.

Step 2: Extract graphic features from the cyclic voltammetry detection data, chronoamperometric detection data and differential pulse voltammetry detection data obtained in step one. The features extracted from the cyclic voltammetry curve include oxidation peak current I _op , reduction peak current I _rp , oxidation peak potential E _op , reduction peak potential E _rp , oxidation curve baseline slope K _o , reduction curve baseline slope K _r , oxidation peak potential E op , reduction peak potential E rp Peak area S _o , reduction peak area S _r , initial reduction potential V _ir and initial oxidation potential V _io . The features extracted from the chronoamperometry curve include the initial steady-state current time t _m and the steady-state current I _s . Features extracted from the differential pulse voltammetry curve include peak potential E _p and peak current I _p .

Step 3: Use the feature values and corresponding concentration labels extracted in step 2 to organize into an initial data set; after dimensionality reduction of the obtained initial data set, the input data set is obtained.

Step 4: Construct a concentration measurement model, and use the input data set obtained in step 3 to train the concentration measurement model. The trained concentration determination model is used to detect the concentration of the tested solution with unknown concentration.

Preferably, among the features extracted from the cyclic voltammetry curve, the oxidation peak current I _op is the height from the anode current peak on the cyclic voltammetry IV curve to the oxidation curve baseline; the reduction peak current I _rp is the height from the cathode current peak to the reduction curve. The height of the baseline; the oxidation peak potential E _op is the peak position of the anode current; the reduction peak potential E _rp is the peak position of the cathode current; the baseline slope of the oxidation curve K _o is the slope of the oxidized baseline; the baseline slope of the reduction curve K _r is the reduction curve The slope of the baseline; the oxidation peak area S _o is the area enclosed by the oxidation curve, the baseline and the distance from the oxidation peak to the baseline; the reduction peak area S _r is the area enclosed by the reduction curve, the baseline and the distance from the reduction peak to the baseline; The initial reduction potential V _ir is the starting potential of the reduction reaction of the reactant; the starting oxidation potential V _io is the starting potential of the oxidation reaction of the reactant.

Among the features extracted from the chronoamperometry curve, the initial steady-state current time t _m is the initial time to reach the steady-state current on the chronoamperometry It curve; the steady-state current I _s is the median after the initial steady-state current on the It curve. .

Among the features extracted from the differential pulse voltammetry curve, the peak potential E _p is the peak position of the anode current on the IV curve of the differential pulse voltammetry; the peak current I _p is the height from the peak value of the anode current to the baseline of the curve.

As an option, in step 2, the process of automatically extracting features from the cyclic voltammetry data is as follows:

Step 1: Divide the cyclic voltammetry detection data into upper and lower parts, where the upper part is the oxidation process data and the lower half is the reduction process data.

Step 2: Calculate the fitting functions f _o (x) and f _r (x) for the oxidation process data and reduction process data respectively to obtain the oxidation curve and reduction curve; take the oxidation curve and reduction curve respectively within the preset range. Maximum and minimum values, the maximum potential of the oxidation curve is the oxidation peak potential I _op , and the minimum potential of the reduction curve is the reduction peak potential I _rp .

Step 3: Find the first-order derivative of the fitting function before the peak potential in the oxidation curve and the reduction curve respectively. Take the median point of the two first-order derivatives as the baseline slope, which are the baseline slope K _o of the oxidation curve and the reduction curve respectively. Baseline slope K _r . At the same time, the original potentials of the two median points are the initial response potentials, which are the initial oxidation potential V _io and the initial reduction potential V _ir respectively.

Step 4: Taking the initial response potential as the tangent point, plus the baseline slope obtained in the third step, determine the oxidation curve baseline L _o (x) and the reduction curve baseline L _r (x) within the smooth curve.

Step 5: Take the height from the peak position point to the baseline as the peak current, respectively, as the oxidation peak current I _op and the reduction peak current I _rp .

Step 6: Take the range from the initial response potential to the peak position, and integrate the difference between the fitting function and the baseline as the peak area, which are the oxidation peak area S _o and the reduction peak area S _r respectively.

As a preference, in step 2, the process of automatically extracting features from the chronoamperometric data is as follows:

The first step: Find the fitting function f(x) for the chronocurrent detection data.

Step 2: Find the first derivative of the fitting function and obtain the differential array.

Step 3: Use the differential value of the next data point to subtract the previous one to form a differential difference array.

Step 4: Obtain the first negative value in the differential difference array. This point indicates that the growth rate of the detection data has slowed down and that the current value has begun to reach a steady state. The position of this data point is determined to be the initial steady-state current time.

Step 5: Take the median current value in the range from the initial steady-state current time to the end of the response as the steady-state current.

As an option, in step 2, the process of automatically extracting features from differential pulse voltammetry data is as follows:

The first step: Find the fitting function f(x) for the differential pulse voltammetry detection data.

Step 2: Take the corresponding position of the maximum current value within the preset range as the peak potential.

Step 3: Find the first-order derivative of the fitting function before the peak potential, take the median point of the first-order derivative as the baseline slope, and the original point position as the tangent point between the baseline and the fitting function.

Step 4: Determine the baseline within the smooth curve, and take the height from the peak position point to the baseline as the peak current.

Preferably, the tested solution is potassium ferricyanide/potassium ferrocyanide solution.

As a preferred method, in step 1, a three-electrode system is constructed in the solution to be measured; the three-electrode system uses a nanogold modified electrode as the working electrode, an Ag/AgCl electrode as the reference electrode, and a platinum wire electrode as the auxiliary electrode, in conjunction with a portable potentiostat. It constitutes an on-site rapid electrochemical detection platform to realize cyclic voltammetry detection, chronoamperometry detection and differential pulse voltammetry detection.

As a preferred method, the cyclic voltammetry I-V curve, chronoamperometry I-t curve and differential pulse voltammetry I-V curve obtained in step 1 are all smoothed and denoised using the recursive average filtering method to improve the signal-to-noise ratio of the original data. , to facilitate subsequent feature extraction operations.

As a preference, the dimensionality reduction method in step three specifically uses the nonlinear dimensionality reduction algorithm t-SNE.

Preferably, the concentration determination model described in step 4 is constructed through the XGBoost algorithm.

The beneficial effects of the present invention are:

The present invention proposes an integrated feature extraction model based on electrochemical detection technology. Aiming at the problems of overlapping electrochemical detection curves and high noise of micro-concentration gradient detection substances, multiple detection methods are used for parallel comparison to achieve effective information of detection signals. Excavation and concentration are accurately determined. The electrochemical detection signal based on the on-site portable electrochemical workstation has high background noise, which reduces the detection sensitivity. Therefore, the recursive average filtering method is used to smooth and denoise the original detection signal to improve the signal-to-noise ratio of the signal, which is beneficial to subsequent detection data. analyze. Automatically extract the graphic features of the smoothed original signal and mine more useful detection signal data information based on previous research. Use t-SNE to perform a dimensionality reduction operation on the feature set, and finally input the dimensionally reduced data set into the prediction model to accurately measure the concentration of the micro-concentration gradient solution. Compared with traditional electrochemical analysis methods, this method innovatively quantifies data from the testing process of multiple electrochemical methods and integrates signal detection, signal preprocessing, signal feature extraction, feature dimensionality reduction and concentration determination. Feature extraction and visual analysis tools under multi-electrochemical methods improve the analysis efficiency of electrochemical detection signals, obtain more useful information contained in detection signals, and achieve rapid and accurate determination of micro-concentration gradient solutions.

Description of the drawings

Figure 1 is an overall framework diagram of the present invention;

Figure 2a is the flow chart of the cyclic voltammetry (CV) algorithm;

Figure 2b is the flow chart of the chronoamperometry (CA) algorithm;

Figure 2c is a flow chart of the differential pulse voltammetry (DPV) algorithm;

Figure 3 is the flow chart of the recursive average filtering algorithm;

Figure 4a is a diagram showing the characteristic parameters of the cyclic voltammetry curve;

Figure 4b is a diagram showing the characteristic parameters of the chronocurrent curve;

Figure 4c is a diagram showing the characteristic parameters of the differential pulse voltammetry curve;

Figure 5 is a flow chart of the automatic extraction algorithm for electrochemical signal features;

Figure 6 is a schematic diagram of the CV curve of a micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution;

Figure 7a is a schematic diagram before smoothing processing;

Figure 7b is a schematic diagram after smoothing processing;

Figure 8 shows the visualization effect of t-SNE dimensionality reduction;

Figure 9 shows the prediction results of the XGBoost model on the concentration of micro-concentration gradient solutions;

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings:

The present invention proposes a method for on-site rapid detection of micro-concentration gradient solutions based on an integrated model of electrochemical detection technology and feature extraction.

As shown in Figure 1, this integrated feature extraction model performs a rapid determination method for micro-concentration gradient solutions, including the following steps:

Step 1: Use the nano-gold modified electrode as the working electrode, the Ag/AgCl electrode as the reference electrode, and the platinum wire electrode as the auxiliary electrode. The three constitute a three-electrode system, and cooperate with a portable potentiostat to form an on-site rapid electrochemical detection platform.

Step 2: Using 0.1M PBS buffer as the supporting electrolyte, perform repeated cross-cyclic voltammetry detection, chronoamperometry detection and differential pulse voltammetry detection on the micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution. to obtain experimental data.

Step 3: For the detection signal, use the recursive average filtering (MAF) method to perform smoothing and denoising to improve the signal-to-noise ratio of the original data and facilitate subsequent feature extraction operations.

Step 4: Automatically extract graphic features from the smoothed original data to include as much data information as possible while ensuring relative independence between features.

Step 5: Organize the extracted feature values and corresponding concentrations into a data set, use the nonlinear dimensionality reduction algorithm t-SNE to reduce the feature dimension, and organize it into a model input data set.

Step 6: Use the data set processed in Step 5 to train the concentration measurement model, and conduct accurate concentration measurement based on the trained model.

The specific process of the present invention:

Step 1: The cleaned glassy carbon electrode (GCE) is immersed in 0.1MH ₂ SO ₄ electrolyte solution containing 0.1% HAuCl ₄ , and electrochemical deposition is performed in -200mV (vs.Ag/AgCl) single potential mode. The deposition time is 30s. After taking it out, rinse it with water and dry it for later use to obtain the gold nanoparticle modified electrode (AuNPs/GCE).

Step 2: Use 0.1M PBS buffer as the supporting electrolyte to prepare a micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution with a gradient interval of 1mM. Using nano-gold modified electrode as the working electrode, Ag/AgCl electrode as the reference electrode, and platinum wire electrode as the auxiliary electrode, the three constitute a three-electrode system. Based on the portable potentiostat electrochemical detection platform, micro-concentration gradient potassium ferricyanide is detected. /Potassium ferrocyanide solution was subjected to repeated cyclic voltammetry, chronoamperometry and differential pulse voltammetry. Each concentration was measured repeatedly 50 times under the same conditions.

As shown in Figure 2, it is the algorithm flow chart of cyclic voltammetry, chronoamperometry and differential pulse voltammetry. The specific algorithm steps for using this portable potentiostat system to perform cyclic voltammetry response are:

1) Select the cyclic voltammetry method on the PC host computer interface, and input relevant parameters including: initial potential Init E, upper limit potential High E, lower limit potential Low E, scanning speed Scan Rate, number of scanning segments Sweep Segments, sampling interval Sample Interval, etc.;

2) Apply a linear alternating initial voltage to both ends of the working electrode and auxiliary electrode through the DAC module;

3) Perform forward scanning on the redox reaction in the E. coli solution at the preset scanning speed V (V/s), and record the current and potential information of the redox reaction through the ADC module.

4) At the same time, continue to judge whether the current potential has reached the upper limit potential. If the current potential has not reached the upper limit potential, continue to record the current and potential information of the redox reaction; if it is detected that the current potential has reached the upper limit potential, scan at the scanning speed V ( V/s) reversely scans the redox reaction in the detection solution, and records the current and potential information of the redox reaction through the ADC module.

5) Continue to determine whether the current potential has reached the preset lower limit potential. If the current potential has not reached the lower limit potential, continue to record the current and potential information of the redox reaction; if it is detected that the current potential has reached the lower limit potential, save the data and exit. detection;

6) Display the detected CV curve on the PC host computer and save the detection data.

The algorithm steps for using chronoamperometry to conduct concentration testing and composition analysis of E. coli solutions are:

1) Select the chronoamperometry method on the PC host computer interface, and input relevant parameters including: initial potential Init E, upper limit potential High E, lower limit potential Low E, number of steps Stepnum, pulse width Pulse Width, sampling interval SampleInterval, etc.;

2) A constant potential, that is, the upper limit potential, is applied to the working electrode through the DAC module, which promotes a redox reaction on the electrode surface, the reactants are gradually consumed, and the thickness of the diffusion layer near the electrode increases, resulting in a larger current density;

3) During the redox process, the reaction current and potential are sampled n times with the sampling interval unchanged, and the current and potential information of the redox reaction are recorded through the ADC module;

4) After the sampling in the first step pulse is completed, the number of steps is reduced by one at the same time. If the number of steps is not zero at this time, the lower limit potential is applied to the working electrode through the DAC module, and oxidation and reduction will also occur on the electrode surface. React and generate a large current density, sample the reaction current and potential n times at the same time, the sampling interval remains unchanged, and record the current and potential information of the redox reaction through the ADC module;

5) At this time, continue to judge whether the number of steps is zero. If it is not zero, restart the detection from 2) until the number of steps is zero; if the number of steps is zero, save the data and exit the detection;

6) Display the measured CA curve on the PC host computer and analyze the CA curve.

The algorithm steps for using differential pulse voltammetry to conduct concentration testing and composition analysis of E. coli solutions are:

1) Select differential pulse voltammetry on the PC host computer interface, and input relevant parameters include: Select chronoamperometry on the PC host computer interface, and enter relevant parameters including: initial potential Init E, endpoint potential Final E, potential increment Incr E, Amplitude, pulse width Pulse Width, sampling interval Sample Width, sampling period Pulse Period, etc.;

2) During the detection process, the ADC module will measure the Faradaic current generated by the redox reaction at the end of each pulse voltage. An initial potential is applied to the working electrode through the DAC module, and the Faraday current and potential are sampled after a timer delay of (PulsePeriod-Pulse Width-Sample Width)s. The sampling time is (Sample Width)s. After the sampling is completed, the DAC module applies a step potential to the working electrode. The step potential is formed by adding a preset amplitude potential to the initial potential;

3) Use a timer to delay (Pulse Width-Sample Width) s with a step potential, and measure the Faradaic reaction current at a later stage and perform sampling. The sampling time is (Sample Width) s. After the sampling is completed, the DAC module applies a new step potential to the working electrode. This step potential adds a preset potential increment to the initial potential to form a new initial potential;

4) Determine whether the new initial potential is equal to the end potential at this time. If not, restart detection from step 2); if equal, save the detection data and exit detection.

Step 3: After obtaining the original electrochemical detection signal data, since the portable potentiostat system itself has shortcomings such as insufficient sensitivity and susceptibility to environmental interference, the recursive average filtering algorithm is used for smoothing and denoising. In the continuous domain, the expression of the recursive average filtering algorithm is:

Among them, T _w is the length of the sliding window, which is an important parameter that affects the performance of smoothing filtering.

From the above formula, we can get the transfer function of the recursive average filtering algorithm as:

From equation (2), the amplitude-frequency expression of the recursive average filter is obtained:

Use the window to divide the discrete electrochemical detection data into N small intervals, and perform average calculations within the intervals. The specific recursive average filtering algorithm flow chart is shown in Figure 3.

The smoothing filtering operation will improve the signal-to-noise ratio of the original data and facilitate subsequent feature extraction operations.

Step 4: Obtain the smoothed electrochemical detection data and automatically extract features. Figure 4 shows the characteristic parameters of the electrochemical measurement curve.

The selection of characteristic parameters follows the following principles: (1) reflects the information difference of the experimental results; (2) is easy to calculate and analyze; (3) the parameters are independent of each other. According to the above requirements, extract features such as peak current, peak potential, peak area, baseline slope and initial response potential in the cyclic voltammetry curve, extract the initial steady-state current time and steady-state current in the chronoamperometry, and extract the differential pulse voltammetry. Mid-peak potential and peak current characteristics. Table 1 shows the extracted feature parameters and detailed description.

Table 1 Feature parameters extracted by three detection methods and detailed descriptions

An automatic feature extraction algorithm is constructed based on the feature representation diagram. The algorithm flow chart is shown in Figure 5.

The automatic extraction process of cyclic voltammetry curve features is expanded to the following steps:

Step 1: Divide the denoised and smoothed cyclic voltammetry detection data into two parts: the upper part is the oxidation curve, and the lower part is the reduction curve.

Step 2: Obtain the fitting functions f _o (x) and _fr (x) for the oxidation curve and reduction curve respectively. Take the maximum and minimum values within a certain range in the oxidation curve and the reduction curve respectively. The maximum potential of the oxidation curve is the oxidation peak potential (I _op ), and the minimum potential of the reduction curve is the reduction peak potential (I _rp ).

Step 3: Find the first-order derivative of the fitting function before the peak position, and take the median point of the first-order derivative as the baseline slope, which is the baseline slope of the oxidation curve (K _o ) and the baseline slope of the reduction curve (K _r ). At the same time, the original potential at this point is the initial response potential, which are the initial oxidation potential (V _io ) and the initial reduction potential (V _ir ) respectively.

Step 4: Taking the initial response potential as the tangent point, plus the baseline slope obtained in the third step, determine the oxidation curve baseline L _o (x) and the reduction curve baseline L _r (x) within the smooth curve.

Step 5: Take the height from the peak position point to the baseline as the peak current, and determine it as the oxidation peak current (I _op ) and reduction peak current (I _rp ).

Step 6: Take the range from the initial response potential to the peak position, and integrate the difference between the fitting function and the baseline as the peak area, which are the oxidation peak area (S _o ) and the oxidation peak area (S _r ) respectively.

The automatic extraction process of chronocurrent curve features is expanded to the following steps:

The first step: Obtain the fitting function f(x) for the denoised and smoothed chronocurrent detection data.

Step 2: Find the first derivative of the fitting function and obtain the differential array.

Step 3: Use the differential value of the next data point to subtract the previous one to form a differential difference array.

Step 4: Obtain the first negative value in the differential difference array. This point indicates that the growth rate of the detection data has slowed down, and it also indicates that the current value has begun to reach a steady state. The position of this data point is determined to be the initial steady-state current time.

Step 5: Take the median current value in the range from the initial steady-state current time to the end of the response as the steady-state current.

The automatic extraction process of differential pulse voltammetry curve features is expanded to the following steps:

The first step: Obtain the fitting function f(x) for the denoised and smoothed differential pulse voltammetry detection data.

Step 2: Take the corresponding position of the maximum current value within the range as the peak potential.

Step 3: Find the first-order derivative of the fitting function before the peak potential, take the median point of the first-order derivative as the baseline slope, and the original point position as the tangent point between the baseline and the fitting function.

Step 4: Determine the baseline within the smooth curve, and the height from the peak position point to the baseline is the peak current.

According to the above steps, all important features of cyclic voltammetry, chronocurrent and differential pulse voltammetry curves can be automatically extracted and a feature set can be generated.

Step 5: Use the t-SNE algorithm to perform dimensionality reduction and visualization on the feature parameter set generated in Step 4. The specific dimensionality reduction process can be expanded into the following steps:

The first step: Calculate the Euclidean distance between the features of different detection samples. Assume that the feature set of all samples is an m×n two-dimensional matrix, expressed as:

Where m represents the number of detection samples, and n is the number of features.

According to the formula ||x _i -x _j || ² , calculate the Euclidean distance between each two groups of samples and obtain a new m×n two-dimensional matrix. The expression is:

Among them, d _ij represents the Euclidean distance of the feature row vector between the i-th sample and the j-th sample.

Step 2: Convert Euclidean distance to characteristic conditional distribution probability. Convert the Euclidean distance between feature vectors obtained in the first step into the conditional probability p _i|j representing similarity. The calculation formula of p _i|j is:

Among them, λ _i is the Gaussian variance centered on _xi .

Step 3: Sum and normalize the conditional probabilities p _i|j to obtain the symmetric joint probability p _ij . The conversion formula is as follows:

Step 4: Calculate the low-dimensional feature conditional distribution probability q _ij , and use the t distribution in the low-dimensional subspace to represent the similarity between samples _yi and y _j . The calculation formula of q _ij is as follows:

Step 5: Use the gradient descent method to solve the matching cost function of t-SNE. The expression is:

C=∑ _i KL(P _i |Q _i ) (7)

Finally, the low-dimensional subspace corresponding detection sample Y = (y ₁ , y ₂ ,..., y _N ) is obtained.

Step 6: Train the concentration measurement model based on the dimensionally reduced data set in step 5, and conduct accurate concentration measurement based on the trained model. The XGBoost algorithm is used to train the concentration measurement model, and some data sets are extracted from it to test the trained model and verify the model performance.

The objective function of the XGBoost algorithm is:

in, is the loss function, ∑ _k Ω(f _k ) is a complex function, and Ω(f _k )=γT+1/2γ||ω|| ² .

Using the second-order Taylor expansion of the objective function, we get:

in, is the prediction error of the first t-1 trees, which is a constant.

remember, Then the objective function can be converted into:

Obtainable objective function:

in, The value of the final objective function is the score of a CART tree, and the prediction result of the entire XGBoost is the sum of the scores of all CARTs.

Example:

Using the portable potentiostat system as the detection platform, based on the integrated model of electrochemical detection technology and feature extraction parameters, the on-site rapid measurement of 51-60mM micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution is realized. The process is as follows:

(1) First, perform cyclic voltammetry, chronoamperometry and differential pulse voltammetry on potassium ferricyanide/potassium ferrocyanide solutions with equal concentration intervals of 51 to 60mM and 1mM under a portable potentiostat system. Detection, a total of 1500 sets of experimental data were obtained. Figure 6 is a schematic diagram of a set of CV curves of a micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution. The inset shows a ten-fold magnification of the lower peak. The detection curve has periodic signal interference, so under certain conditions In this case, the detection curve aliasing phenomenon is caused.

(2) The obtained detection data are automatically transferred to the smoothing processing module, and the electrochemical detection data is smoothed using the recursive average filtering module. The cyclic voltammetry curve of the micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution is: For example, as shown in Figure 7, the detection data after recursive average filtering is smoother and the gradient is more obvious.

(3) Then use the automatic feature extraction module in the integrated model to extract corresponding features from the smoothed detection signal. The cyclic voltammetry curve automatically extracts features such as peak current, peak potential, peak area, baseline slope and initial response potential, extracts the initial steady-state current time and steady-state current in chronoamperometry, and extracts peak potential in differential pulse voltammetry. and peak current characteristics, and organized into feature sets.

(4) In order to improve the accuracy of concentration measurement and reduce the redundancy of the feature set, two-dimensional t-SNE dimensionality reduction was performed on the generated feature set. Figure 8 shows the visualization result after dimensionality reduction. It can be seen that the same concentration The detection data are gathered together, and there is a clear dividing line between concentrations.

(5) Use the XGBoost algorithm to build a concentration prediction model. 80% of the feature set after t-SNE dimensionality reduction processing is the training set, and the remaining data set is the test set. The data sets with concentrations of 53mmol/L, 56mmol/L, and 59mmol/L are set as blind without training. Concentration prediction set. The prediction results of the test set are shown in Figure 9. The predicted values of the samples at each concentration are basically consistent with the actual values. Among them, the three untrained blind concentration sets also have good prediction effects (as pointed by the arrows). The actual concentration The degree of fit (R-Squared) to the predicted concentration was 0.957.

In summary, this example provides a method for rapid determination of micro-concentration gradient solutions based on an integrated model of electrochemical detection technology and feature extraction parameters. The operation steps of the integrated model are: first, perform repetitive cross-detection of cyclic voltammetry, chronoamperometry and differential pulse voltammetry on the micro-concentration gradient potassium ferricyanide/potassium ferrocyanide solution, and obtain a total of 1500 sets of data; secondly, , use the recursive average filtering module to smooth the detection data; next, perform automatic feature extraction operations on the detection data of the three electrochemical detection methods, in which the peak current, peak potential, peak area, and baseline are automatically extracted from the cyclic voltammetry curve. Characteristics such as slope and initial response potential are extracted, and the initial steady-state current time and steady-state current in chronoamperometry are extracted. Peak potential and peak current characteristics in differential pulse voltammetry are extracted and organized into feature sets; then, using t- SNE dimensionality reduction reduces the feature set to two dimensions, performs visual analysis, and reduces the complexity of the prediction model; finally, the processed feature set is input into the prediction model to achieve better concentration prediction effects, and the simulation between the actual concentration and the predicted concentration is achieved. The R-Squared degree is 0.957. Therefore, this method realizes rapid feature analysis of electrochemical detection data and on-site rapid determination of micro-concentration gradient solutions.

Claims (3)

1. An electrochemical measurement method for micro-concentration gradient solutions based on feature parameter extraction, which is characterized by: Step 1. Perform cyclic voltammetry detection, chronoamperometry detection and differential pulse voltammetry on multiple measured solutions in micro-concentration gradients. Amperometric detection, obtain cyclic voltammetry detection data, chronoamperometry detection data and differential pulse voltammetry detection data of different multiple solutions to be tested; Step 2: Extract graphical features from the cyclic voltammetry detection data, chronoamperometry detection data and differential pulse voltammetry detection data obtained in step 1; the features extracted from the cyclic voltammetry curve include oxidation peak current I _op and reduction peak current I _rp , oxidation peak potential E _op , reduction peak potential E _rp , oxidation curve baseline slope K _o , reduction curve baseline slope K _r , oxidation peak area S _o , reduction peak area S _r , initial reduction potential V _ir and starting Oxidation potential V _io ; the features extracted from the chronoamperometric curve include the initial steady-state current time t _m and the steady-state current I _s ; the features extracted from the differential pulse voltammetry curve include the peak potential E _p and the peak current I _p ; Step 3: Use the feature values extracted in Step 2 and the corresponding concentration labels to organize into an initial data set; after dimensionality reduction of the initial data set, the input data set is obtained; Step 4: Construct a concentration determination model, and use the input data set obtained in step 3 to train the concentration determination model; the trained concentration determination model is used to detect the concentration of the measured solution of unknown concentration; Among the features extracted from the cyclic voltammetry curve, the oxidation peak current I _op is the height from the anode current peak on the cyclic voltammetry IV curve to the oxidation curve baseline; the reduction peak current I _rp is the height from the cathode current peak to the reduction curve baseline; The oxidation peak potential E _op is the peak position of the anode current; the reduction peak potential E _rp is the peak position of the cathode current; the oxidation curve baseline slope K _o is the slope of the oxidation baseline; the reduction curve baseline slope K _r is the slope of the baseline of the reduction curve; The oxidation peak area S _o is the area surrounded by the oxidation curve, the baseline and the distance from the oxidation peak to the baseline; the reduction peak area S _r is the area surrounded by the reduction curve, the baseline and the distance from the reduction peak to the baseline; the initial reduction potential V _ir is the starting potential of the reduction reaction of the reactant; the starting oxidation potential V _io is the starting potential of the oxidation reaction of the reactant; Among the features extracted from the chronoamperometry curve, the initial steady-state current time t _m is the initial time to reach the steady-state current on the chronoamperometry It curve; the steady-state current I _s is the median after the initial steady-state current on the It curve; Among the features extracted from the differential pulse voltammetry curve, the peak potential E _p is the peak position of the anode current on the IV curve of the differential pulse voltammetry; the peak current I _p is the height from the peak value of the anode current to the baseline of the curve; For the cyclic voltammetry I-V curve, chronoamperometry I-t curve and differential pulse voltammetry I-V curve obtained in step 1, the recursive average filtering method is used for smoothing and denoising to improve the signal-to-noise ratio of the original data and facilitate follow-up. Feature extraction operations; In step two, the process of automatic feature extraction from cyclic voltammetry data is as follows: Step 1: Divide the cyclic voltammetry detection data into upper and lower parts, where the upper part is the oxidation process data and the lower half is the reduction process data; Step 2: Calculate the fitting functions f _o (x) and f _r (x) for the oxidation process data and reduction process data respectively to obtain the oxidation curve and reduction curve; take the oxidation curve and reduction curve respectively within the preset range. Maximum and minimum values, the maximum potential of the oxidation curve is the oxidation peak potential I _op , and the minimum potential of the reduction curve is the reduction peak potential I _rp ; Step 3: Find the first-order derivative of the fitting function before the peak potential in the oxidation curve and the reduction curve respectively. Take the median point of the two first-order derivatives as the baseline slope, which are the baseline slope K _o of the oxidation curve and the reduction curve respectively. Baseline slope K _r ; at the same time, the original potentials of the two median points are the initial response potentials, which are the initial oxidation potential V _io and the initial reduction potential V _ir respectively; Step 4: Taking the initial response potential as the tangent point, plus the baseline slope obtained in the third step, determine the oxidation curve baseline L _o (x) and the reduction curve baseline L _r (x) within the smooth curve respectively; Step 5: Take the height from the peak position point to the baseline as the peak current, respectively, as the oxidation peak current I _op and the reduction peak current I _rp ; Step 6: Take the range from the initial response potential to the peak position, and integrate the difference between the fitting function and the baseline as the peak area, which are the oxidation peak area S _o and the reduction peak area S _r respectively; In step two, the process of automatically extracting features from the chronocurrent data is as follows: Step 1: Find the fitting function f(x) for the chronocurrent detection data; Step 2: Find the first derivative of the fitting function and obtain the differential array; Step 3: Use the differential value of the next data point to subtract the previous one to form a differential difference array; Step 4: Obtain the first negative value in the differential difference array, which indicates that the growth rate of the detection data has slowed down, and also indicates that the current value has begun to reach a steady state, and the position of this data point is determined to be the initial steady-state current time; Step 5: Take the median current value in the range from the initial steady-state current time to the end of the response as the steady-state current; In step two, the process of automatic feature extraction from differential pulse voltammetry data is as follows: Step 1: Find the fitting function f(x) for the differential pulse voltammetry detection data; Step 2: Take the corresponding position of the maximum current value within the preset range as the peak potential; Step 3: Find the first-order derivative of the fitting function before the peak potential, take the median point of the first-order derivative as the baseline slope, and the original point position as the tangent point between the baseline and the fitting function; Step 4: Determine the baseline within the smooth curve, and take the height from the peak position point to the baseline as the peak current; The dimensionality reduction method in step three specifically uses the nonlinear dimensionality reduction algorithm t-SNE; the concentration measurement model described in step four is constructed through the XGBoost algorithm. 2. A micro-concentration gradient solution electrochemical determination method based on characteristic parameter extraction according to claim 1, characterized in that: the measured solution is potassium ferricyanide/potassium ferrocyanide solution. 3. A micro-concentration gradient solution electrochemical determination method based on characteristic parameter extraction according to claim 1, characterized in that: in step one, a three-electrode system is constructed in the measured solution; the three-electrode system is modified with nano-gold The electrode is the working electrode, the Ag/AgCl electrode is the reference electrode, and the platinum wire electrode is the auxiliary electrode. Together with the portable potentiostat, it forms an on-site rapid electrochemical detection platform to realize cyclic voltammetry detection, chronoamperometry detection and differential pulse voltammetry. Legal testing.