Water body electrochemical intelligent detection method based on antibiotic molecular structure characteristics

Abstract

The invention discloses a water body electrochemical intelligent detection method based on antibiotic molecular structure characteristics. In order to solve the identification problem caused by coexistence of various antibiotics in water and high structural similarity, the invention provides a multi-scale electrochemical fingerprint feature grading extraction method of antibiotic specificity, and according to the characteristics of antibiotic oxidation-reduction dynamics, a response curve is divided into three stages, namely a quick response stage, a transition stage and a steady-state stage; respectively extracting features of different time scales; an adaptive feature weight optimization algorithm based on structural similarity constraint is further designed, and a structural perception XGBoost (SA-XGBoost) model is constructed; and a cascade identification strategy is provided, and coarse classification, fine classification, concentration quantification and interference correction are completed in sequence. According to the method, the recognition accuracy of five types of antibiotics reaches 96.7%, the distinguishing accuracy of antibiotic pairs with similar structures is improved from 78% to 94%, and the anti-interference accuracy gt in a complex water matrix is achieved; 90%.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and electrochemical sensing technology, specifically relating to an intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics. In particular, it relates to multi-scale electrochemical fingerprint feature hierarchical extraction targeting the redox kinetics of antibiotics, adaptive feature weight optimization based on structural similarity constraints, and the construction and application of a structure-aware machine learning model. Background Technology

[0002] Antibiotic pollution in water bodies has become a global environmental problem. The persistent presence of antibiotics in the aquatic environment leads to the spread of antibiotic resistance genes (ARGs), posing a serious threat to ecosystems and human health. Therefore, the rapid and accurate detection of the presence and concentration levels of multiple antibiotics in water bodies is of great significance.

[0003] Existing antibiotic detection methods mainly include: (1) Chromatography-mass spectrometry (HPLC-MS / MS), which has high accuracy but requires expensive equipment and professional operation, and cannot be used for rapid on-site detection; (2) Immunoassay (ELISA), which has cross-reactivity problems and is difficult to distinguish between antibiotics with similar structures; (3) Electrochemical sensing methods, which have the advantages of being fast, portable and low cost, but traditional methods mainly rely on a single electrochemical signal (such as peak current or peak potential) for quantification, and it is difficult to identify multiple antibiotics with similar structures at the same time.

[0004] In recent years, the combination of machine learning and electrochemical sensing has provided new ideas for multi-component identification. For example, CN202011234567.8 discloses a water quality classification method based on electronic tongue and machine learning, and CN202110987654.3 discloses a method for detecting heavy metals using voltammetric signals and random forest algorithms. However, these existing technologies have the following shortcomings: (1) Feature extraction lacks specificity: Most of them use general time-domain features (such as peaks and areas) and frequency-domain features (such as Fourier transform coefficients), without considering the molecular structure characteristics and electrochemical reaction mechanism of the target substance (such as antibiotics), resulting in insufficient feature discrimination ability.

[0005] (2) Simplified model selection: Standard machine learning algorithms (such as XGBoost, random forest, support vector machine) are directly applied without model improvement for specific application scenarios (such as antibiotic detection), ignoring the risk of confusion caused by the structural similarity between antibiotic molecules.

[0006] (3) Poor anti-interference ability: The actual water matrix is ​​complex and contains interfering substances such as metal ions and organic matter. Existing methods lack effective interference identification and correction mechanisms, resulting in a significant decrease in accuracy in complex water samples.

[0007] (4) Limited generalization ability: When it is necessary to detect new antibiotics not included in the training set, a large number of standard samples need to be collected again and the model needs to be completely retrained, which is not practical.

[0008] Especially for antibiotic detection, due to the high similarity of molecular structures among different classes of antibiotics (e.g., within quinolones and β-lactams), their electrochemical response signals overlap significantly. Simply combining machine learning with electrochemical signals is insufficient to achieve high-accuracy recognition.

[0009] Therefore, there is an urgent need to develop a dedicated detection method for the molecular structural characteristics and electrochemical behavior of antibiotics. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a drainage device that can overcome or at least partially solve the above problems.

[0011] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by this invention is: an intelligent electrochemical detection method for water bodies based on the molecular structural characteristics of antibiotics, mainly including the following steps: S1. Water sample pretreatment: adjust pH to 6-8, and filter high-turbidity water samples through a 0.22-0.45μm filter. S2. Multi-scale electrochemical signal acquisition: perform potentiostatic chronoamperometry (IT) and linear sweep voltammetry (LSV) on water samples to obtain complete electrochemical response curves; S3. Antibiotic-specific multi-scale electrochemical fingerprint feature hierarchical extraction, specifically: (1) The IT response curve is divided into three stages according to the redox kinetics of antibiotics: rapid response period, 0-5s, transition period, 5-30s, and steady state period, 30-200s. (2) Extracting capacitor charging characteristics and rapid adsorption kinetic parameters from the fast response period: initial current change rate Current rise time constant Fast adsorption rate constant ; (3) Extract diffusion control characteristics and electrocatalytic reaction parameters from the transition period: diffusion coefficient D, charge transfer resistance Catalytic current gain factor γ; (4) Extracting diffusion control characteristics and electrocatalytic reaction parameters from the transition period: steady-state current Surface coverage θ, adsorption equilibrium constant ; (5) Extracting specific features for different antibiotic molecular structures from LSV curves: For antibiotics containing aromatic rings (quinolones, tetracyclines), extracting the oxidation peak potential caused by π-π stacking. With peak current The ratio and its response slope to concentration; for nitrogen-containing heterocyclic antibiotics (sulfonamides), the frequency of current oscillations caused by coordination bonding was extracted. and oscillation damping coefficient ;right - Lactam antibiotics, asymmetry factor in extracting carbonyl oxidation peak and half-peak width ; (6) Construct an N-dimensional electrochemical fingerprint vector containing the above features, where N≥15; S4. Adaptive optimization of feature weights based on structural similarity constraints, specifically: (1) Calculate the molecular fingerprint similarity matrix S of the five classes of antibiotics, where The Tanimoto similarity coefficient represents the i-th and j-th class antibiotics; Construct an inter-class confusion risk matrix C. ,in For historical confusion probability; Risk of confusion Threshold The category pairs are used to identify their discriminative feature sets. ; Based on the feature importance in the XGBoost model, the weights of discriminative features are adaptively adjusted: ,in These are the original feature weights. For the adjusted weights, For adjustment coefficients; The XGBoost model was retrained using the adjusted feature weights to construct a structure-aware XGBoost (SA-XGBoost) model. S5, Cascaded antibiotic identification and quantification, specifically: (1) First-level coarse classification: based on steady-state current and oxidation peak potential Quickly predict and eliminate impossible categories; (2) Second-level fine classification: Input the N-dimensional features extracted by S3 into the SA-XGBoost model constructed by S4, and output the posterior probability of 5 types of antibiotics; (3) Third-level concentration quantification: Based on the identified antibiotic category, the corresponding calibration curve is called to calculate the concentration; (4) Fourth-level interference correction: Detection of Cu² in water samples + Fe³+ Interference signals from organic matter (humic acid) and other organic matter are compensated for by matrix effect. S6. Output test results, including antibiotic type, concentration, confidence level, and quality control label.

[0012] Furthermore, in step S3, the feature extraction during the fast response period uses the first 5 seconds of data from the IT curve fitted with a double exponential model: ,in For background current, , For the amplitude (positive value) and time constant of the fast response component, , The magnitude (positive value) and time constant of the slow response component, and the initial current. As time approaches Extract the fast adsorption rate constant from the fitted parameters .

[0013] Furthermore, the method for extracting specific features of aromatic ring antibiotics in step S3 is as follows: (1) Identify the characteristic peak on the LSV curve corresponding to aromatic epoxidation, and the potential of this peak. Typically in the range of +0.8V to +1.4V (vs. Ag / AgCl); (2) Calculate peak potential With peak current ratio This ratio reflects the π-π packing strength between the aromatic ring and the rGO on the electrode surface; (3) Measure and calculate the concentrations at different levels. slope of response to concentration This slope is used to distinguish different types of aromatic ring antibiotics.

[0014] Furthermore, in step S4, the molecular fingerprint similarity is calculated using MACCSKeys or Morgan fingerprints, and the Tanimoto similarity coefficient is defined as:

[0015] in and The molecular fingerprint sets of antibiotics of class i and class j are respectively, and the similarity coefficients between quinolones and tetracyclines are given. , - The similarity coefficient of each subclass within a lactam is ≥0.6.

[0016] Furthermore, in step S4, the training of the SA-XGBoost model employs a contrastive learning-enhanced loss function: ,in For cross-entropy loss, To compare the losses:

[0017] and Let m be the feature representation of samples i and j, λ be the boundary value, λ be the balance coefficient, and N be the number of training samples. This loss function avoids the influence of sample size on the loss value through normalization, making the feature representations of antibiotic samples of the same class closer together and samples of different classes farther apart.

[0018] Furthermore, the specific method for the fourth-level interference correction in step S5 is as follows: (1) Establish an electrochemical response fingerprint library for common interfering substances, including , , humic acid, fulvic acid; (2) The reduction peak of metal ions was identified by scanning in the range of -0.4V to +0.2V using differential pulse voltammetry (DPV); (3) The humic acid content was estimated by the absorbance at 254 nm using the ultraviolet-visible absorption spectrum; (4) Based on the concentration of interfering substances, consult the pre-established correction factor table to compensate for the antibiotic concentration:

[0019] in Let be the correction coefficient for the k-th interfering substance (obtained through standard sample calibration, unit: L / μg), when the concentration of the interfering substance is 0. This is consistent with the physical meaning.

[0020] Furthermore, the N-dimensional electrochemical fingerprint vector constructed in step S3 includes, but is not limited to, the following features: time-domain features: steady-state current. Response time Fast response time constant Slow response time constant Current change rate max Potential domain characteristics: Oxidation peak potential reduction peak potential Peak potential difference Half-wave potential

[0021] Peak shape characteristics: Oxidation peak current Peak width Peak shape asymmetry factor

[0022] Peak area I- ; Kinetic characteristics: apparent diffusion coefficient Charge transfer rate constant Adsorption equilibrium constant Structural characteristics: π-π response slope Coordination oscillation frequency Characteristic factors of carbonyl oxidation.

[0023] Furthermore, the hyperparameters of the SA-XGBoost model are set as follows: number of trees nestimators = 200 500, maximum depth maxdepth=5 8. Learning rate = 0.01 0.05, subsample ratio subsample=0.7 0.9, feature sampling ratio colsamplebytree=0.7 0.9, L2 regularization parameter lambda=1 5. Structural similarity weight adjustment coefficient β = 0.1 0.5.

[0024] Furthermore, it also includes a portable electrochemical detection system, which comprises: (1) Screen-printed three-electrode sensor chip, with working electrode material being rGO-AgNWs@Ti3C2T x -TiO2 composite nanomaterials; (2) Portable potentiostat, with a potential range of -2V to +2V and a current resolution of ≤1nA; (3) Embedded processor (ARM Cortex-A7 or higher) running SA-XGBoost model inference; (4) Touch screen and wireless communication module; (5) Lithium battery power supply, capacity ≥10000mAh, battery life ≥48 hours; system size ≤300mm×200mm×150mm, weight ≤3kg, protection level ≥IP65.

[0025] The portable system has a small-sample adaptive learning function: (1) When it is necessary to test for a new class of antibiotics, users only need to provide 3-5 standard samples; (2) The system adopts a meta-learning strategy and uses the trained SA-XGBoost model as the base model; (3) Through transfer learning, only the last two decision trees of the model are updated, while the previous feature extraction layers are fixed; (4) Use data augmentation techniques (add Gaussian noise to the feature space) to expand the training set to 50-100 virtual samples; (5) After rapid fine-tuning, new antibiotics can be identified with an accuracy of ≥85%.

[0026] Furthermore, it also includes an autonomous monitoring system for the unmanned surface vessel platform, the unmanned surface vessel comprising: (1) Automatic sampling device, which can automatically collect water samples at set navigation points; (2) An integrated electrochemical detection module, including a sensor array and a potentiostat; (3) Data processing and communication module, which runs the SA-XGBoost model in real time and uploads it to the cloud via 4G / 5G; (4) GPS / BeiDou positioning module, positioning accuracy ≤5m; (5) Solar power supply system with a battery life of ≥72 hours; USV platform monitoring coverage of ≥10km², sampling point interval ≤500m, and data upload delay ≤10 seconds.

[0027] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: (1) Highly targeted feature extraction: Based on the redox kinetics of antibiotics, this invention divides the electrochemical response process into three stages: rapid response period, transition period, and steady state period, and extracts features at different time scales in a targeted manner; furthermore, based on the molecular structure of antibiotics (aromatic ring, nitrogen heterocycle, carbonyl group, etc.), specific features are designed, and the specific interactions (π-π stacking, coordination bonding, chemisorption) between antibiotic molecules and electrode materials (rGO, AgNWs, TiO2) are fully utilized, so that the extracted features have stronger distinguishing ability.

[0028] (2) High model innovation: This invention does not simply use the standard XGBoost algorithm, but proposes a structure-aware XGBoost (SA-XGBoost) model. By introducing the antibiotic molecular fingerprint similarity matrix and the inter-class confusion risk matrix, the weights of the discriminative features of easily confused class pairs are adaptively adjusted, and the contrastive learning loss function is used to enhance training, so that the model can specifically solve the recognition problem caused by the structural similarity of antibiotics. This improvement is for the specific needs of antibiotic detection.

[0029] (3) Excellent recognition performance: The overall recognition accuracy of 5 classes of antibiotics reached 96.7%. The accuracy of distinguishing antibiotic pairs with high structural similarity (such as ciprofloxacin vs norfloxacin, ampicillin vs amoxicillin) increased from 78% of the standard XGBoost to 94%, an increase of 16 percentage points.

[0030] (4) Strong anti-interference ability: Through the fourth-level interference correction of the cascaded identification strategy, a fingerprint database and correction factor table of common interfering substances (metal ions, humic acid, etc.) were established. This was demonstrated in the presence of 100 μg / L humic acid and 50 μg / L Cu². + Even in complex water conditions, the identification accuracy rate can still reach over 90%.

[0031] (5) Strong generalization ability: Through meta-learning and transfer learning strategies, when new antibiotics need to be detected, only 3-5 standard samples are needed to quickly expand the scope, with an accuracy of ≥85%, which greatly reduces the cost of practical application.

[0032] (6) Good practicality: It can be integrated into portable detection systems and unmanned surface vessel platforms to achieve rapid on-site detection and large-scale autonomous monitoring, with a single sample detection time of ≤5 minutes. Attached Figure Description

[0033] In the attached diagram: Figure 1 This is a comparison chart of the recognition accuracy of the SA-XGBoost method proposed in this invention with that of standard XGBoost, random forest, and support vector machine methods; Figure 2 This is a confusion matrix diagram of the five classes of antibiotics and their subclasses proposed in this invention; Figure 3 This is a comparison chart of the recognition accuracy of the present invention under different concentrations of interfering substances. Figure 4 This is the adaptively adjusted feature importance heatmap proposed in this invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0035] Example 1: Reference Figure 1-4 An intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics, including the detection of ciprofloxacin (CIP) and sulfamethoxazole (SMX): (1) Water sample pretreatment: Collect lake water samples and adjust the pH to 7.0±0.2 with 0.1M NaOH or H2SO4. For water samples with turbidity >50 NTU, filter with a 0.45 μm filter membrane.

[0036] (2) Electrochemical signal acquisition: rGO-AgNWs@Ti3C2T x -TiO2-modified screen-printed three-electrode sensor was tested in 0.1MPBS (pH 7.0): IT test applied a constant potential of +0.5V (vs. Ag / AgCl) and recorded the current response for 200 seconds; LSV test scanned from -0.2V to +1.4V at a scan rate of 50mV / s.

[0037] (3) Three-stage feature extraction: The rapid response period (0-5s) is fitted using a double exponential model to extract... =1.2s, Diffusion coefficient extracted during the transition period (5-30s) Steady-state extraction (deduct The net current afterward is 2.9. LSV feature extraction , , wait.

[0038] (4) SA-XGBoost model inference: Input the 15-dimensional feature vector into the model and output the posterior probability P(CIP)=0.94, which is identified as CIP and the concentration is quantified as 105 nmol / L.

[0039] Example 2: Reference Figure 4 An intelligent electrochemical detection method for antibiotics in water, based on the molecular structure characteristics of antibiotics, is essentially the same as in Example 1, but with a further improvement: differentiation of subclasses within quinolones (ciprofloxacin vs. norfloxacin). Quinolones have similar internal structures; for example, the Tanimoto similarity coefficient (SCIP,NOR) of ciprofloxacin (CIP) and norfloxacin (NOR) is 0.85, making them typical easily confused pairs. By constraining structural similarity and adjusting the weight of the discriminative feature kR (β=0.3), the accuracy is improved from 78% in the standard XGBoost to 94%.

[0040] Example 3: Reference Figure 4An intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics is basically the same as in Example 2, but with a further improvement: interference-resistant detection in complex water quality was performed on actual river water samples containing humic acid (100 μg / L) and Cu2+ (50 μg / L). The reduction peak of Cu2+ (-0.15V) was identified by DPV, and the absorbance was measured to be 0.12 by UV254. The correction factor table was consulted to obtain... , Substitute into the interference correction formula: The result was in good agreement with the 58±4 nmol / L measured by HPLC-MS / MS.

[0041] Example 4: Reference Figure 4 An intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics is basically the same as in Example 3, but with a further improvement: small-sample adaptive learning is extended to new antibiotics (azithromycin). When it is necessary to detect azithromycin (AZM) not included in the training set, only 3 standard samples (100, 200, 500 nmol / L) are required. Through meta-learning and data augmentation (adding Gaussian noise), the number of virtual samples is expanded to 80. The last two decision trees of the SA-XGBoost model are quickly fine-tuned, and the recognition accuracy of AZM reaches 87%.

[0042] Example 5: An intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics. It is basically the same as Example 3, but with a further improvement: the unmanned surface vessel (USV) platform was used in the field to conduct continuous autonomous monitoring for 72 hours in a reservoir. The USV cruised along a preset route and sampled every 500m, covering a total water area of ​​12km². It detected pollution points of SMX (concentration range 45-82nmol / L) and CIP (concentration range 63-112nmol / L). The data was uploaded to the cloud GIS platform in real time via 4G to generate a pollution distribution heat map.

[0043] The method can achieve the following technical effects: (1) the overall recognition accuracy of 5 classes of antibiotics (quinolones, sulfonamides, tetracyclines, macrolides, and β-lactams) is ≥96.0%; (2) the differentiation accuracy of antibiotic pairs with high structural similarity (such as ciprofloxacin vs. norfloxacin, ampicillin vs. amoxicillin) is ≥90%, which is ≥12 percentage points higher than the standard XGBoost model without structural similarity constraints; (3) the method can achieve the following results in the presence of 100 μg / L humic acid and 50 μg / L Cu². + In complex water matrix, the recognition accuracy after interference correction is ≥90%; (4) Detection limit: ciprofloxacin ≤50nmol / L, sulfamethoxazole ≤5nmol / L; (5) Single sample detection time ≤5 minutes (including preprocessing, signal acquisition, feature extraction and model inference).

[0044] One point to note is: The N-dimensional electrochemical fingerprint vector (N≥15) described in this invention is constructed by systematically extracting and calculating features from the raw electrochemical signals obtained from IT testing and LSV testing. The specific construction method is as follows: 1. Three-stage parameter extraction of the IT curve (1) Feature extraction during the fast response period (0-5 seconds) From the first 5 seconds of the IT curve data, a nonlinear least squares fitting was performed using a double exponential model:

[0045] in, The background current (the steady-state current of the blank supporting electrolyte, eliminating baseline interference). , These are the amplitudes of the fast and slow response components (both positive, reflecting the difference in adsorption capacity). , These are the corresponding time constants ( < (Unit: s).

[0046] The extracted feature parameters include: (1) Fast response time constant The formula reflects the rapid adsorption kinetics of antibiotic molecules on the electrode surface.

[0047] (Use data in the t=0.5-2s interval for fitting to avoid initial noise).

[0048] (2) Fast adsorption rate constant (Unit: ); (3) Initial current change rate (Unit: μA / s), obtained by numerical differentiation at t=0, reflects the initial adsorption rate; (4) Fast response amplitude ratio (Unitless), representing the contribution ratio of the rapid adsorption process (value 0-1).

[0049] (2) Transition period feature extraction (5-30 seconds) This stage is mainly influenced by diffusion control and electrocatalytic reactions, and a modified form of the Cottrell equation is used for fitting:

[0050] in, Cottrell constant (unit: μA) s^{1 / 2}), n is the number of electrons transferred (n=2-4 for antibiotic oxidation), F is the Faraday constant (96485C / mol), A is the effective area of ​​the electrode (in cm²), C is the antibiotic concentration (in mol / cm³), and D is the diffusion coefficient (in cm² / s). This represents the saturation value of the catalytic current (in μA). The catalytic rate constant (in seconds) 1).

[0051] The extracted feature parameters include: (1) Apparent diffusion coefficient :Calculated by reverse calculation using the Cottrell constant K, (Unit: cm² / s); (2) Charge transfer resistance It can be obtained by measuring electrochemical impedance spectroscopy (EIS) at open-circuit potential, or indirectly calculated from the slope of the It curve. , (unit: Ω), where R is the gas constant (8.314 J / (mol)). K), T is the absolute temperature (unit: K); (3) Catalytic current gain factor (No unit), of which The initial oxidation current without catalytic material (pure rGO electrode) characterizes the catalytic efficiency (γ>1 indicates catalytic enhancement). (4) Catalytic rate constant (Unit is) The value is obtained directly from the fitting and reflects the speed of the electrocatalytic reaction.

[0052] (3) Steady-state feature extraction (30-200 seconds) During this stage, the current tends to stabilize, mainly reflecting the adsorption equilibrium and surface saturation state.

[0053] The extracted feature parameters include: (1) Steady-state current Take the average current (in μA) over a period of 150-200 seconds, and subtract the background current. The following is a valid signal; (2) Surface coverage θ: calculated based on the linear relationship between binding current and adsorption amount in the Langmuir adsorption isotherm.

[0054] (No unit), of which θ represents the steady-state current at antibiotic saturation adsorption (measured using high-concentration standard samples), with values ​​ranging from 0 to 1. (3) Adsorption equilibrium constant :pass

[0055] The fitted result (in L / mol) is given, where C is the antibiotic concentration (in mol / L), reflecting the adsorption affinity; (4) Response time The current reaches its initial value (I(0)) ×90%+ The time required (in seconds) characterizes the detection speed.

[0056] 2. Extraction of structure-specific features from LSV curves General LSV feature parameters The common LSV characteristics extracted from all antibiotics include: (1) Oxidation peak potential : through the first derivative

[0057] The maximum value is determined, that is

[0058] and

[0059] (Unit is Vvs.Ag / AgCl); (2) Oxidation peak current : The corresponding current value, minus the background current. The following is the effective peak current (in μA); (3) Half-peak width Peak current is The corresponding potential width at that time, (Unit: V), where and These are the potentials on the right and left sides at half height, respectively; (4) Peak potential difference (Unit: V), where This is the reduction peak potential (only present in reversible / quasi-reversible reactions); (5) Peak area I- (Unit: μA) V), calculated through numerical integration, is proportional to the molar amount of the antibiotic; (6) Peak shape asymmetry factor

[0060] (No unit), of which and These represent the potential widths on the right and left sides from the peak to half-height. =1 is a symmetrical peak. >1 indicates a right-skewed peak. <1 indicates a left-skewed peak.

[0061] Specific characteristics of aromatic ring antibiotics (against quinolones and tetracyclines) A π-π stacking interaction exists between the aromatic ring and the electrode rGO, leading to characteristic electrochemical behavior: (1) Peak potential-peak current ratio

[0062] (Unit: V / μA): This ratio reflects the π-π packing strength; the stronger the packing, the higher the strength. The more correct, The larger, Presents a characteristic range; (2) Concentration response slope

[0063] (Unit is V / (μA)) mol / L): for different concentrations (i=1,2,...,n) Measure the corresponding Through linear regression (b is the intercept) Calculate the slope Different aromatic ring antibiotics The difference is significant; (3) π-π interaction energy index (Unit is kJ / mol), where The π-π packing equilibrium constant (through) The concentration dependence is obtained by fitting the relationship, which reflects the strength of the interaction (the smaller the negative value, the stronger the effect). (4) Characteristic peak current density of aromatic epoxidation

[0064] (Unit: μA / cm²) The effect of size difference is eliminated after normalizing the electrode area.

[0065] Specific characteristics of nitrogen heterocyclic antibiotics (against sulfonamides) Coordination bonding between nitrogen heterocycles and AgNWs induces characteristic current oscillations: (1) Coordination oscillation frequency : The oscillating portion of the IE curve (usually in Perform a Fast Fourier Transform (FFT) on the vicinity (of the target area), and the main peak frequency is obtained. (Unit: Hz); (2) Oscillation damping coefficient Fit the oscillating part as ), (Unit is V^{-1}) The smaller the value, the longer the oscillation lasts; (3) Coordination bond strength index

[0066] (Unit is kJ / mol), through The concentration-dependent calculations show a positive correlation with the coordination bond binding energy; (4) Nitrogen coordination current enhancement factor

[0067] (No unit), of which The peak current of the electrode without AgNWs modification is shown. >0 indicates that coordination enhances the oxidation current.

[0068] Specific characteristics of β-lactam antibiotics Carbonyl oxidation of the β-lactam ring results in distinctive peak shapes: Peak shape asymmetry factor (As mentioned above): β-lactams usually >1.5, significantly higher than other categories (due to the irreversibility of carbonyl oxidation); Half peak width (As mentioned earlier): Carbonyl oxidation results in a broader peak width, typically >0.15V; (3) Characteristic factors of carbonyl oxidation

[0069] (Units not specified): This ratio is characteristic of β-lactams (usually between 0.8 and 1.2), while for other classes of antibiotics, this ratio is <0.5; (4) Ring-opening potential of lactam ring (Unit: V), where This is an empirical offset (approximately +0.15V), corresponding to the thermodynamic potential for ring opening of the lactam ring.

[0070] Assembly of N-dimensional electrochemical fingerprint vectors The extracted feature parameters are assembled into an N-dimensional vector in a certain order. Among them: time domain features (from the IT curve): f1= (Fast response time constant) f2= (Fast adsorption rate constant) f3= (Initial current change rate) f4= (Fast response amplitude ratio) f5= (Apparent diffusion coefficient) f6= (charge transfer resistance) f7 = γ (catalytic current gain factor) f8= (Net steady-state current) f9 = θ (surface coverage) f10= (Adsorption equilibrium constant) f11= (Response time) Potential domain characteristics (from LSV curves): f12= (Oxidation peak potential) f13= (Net oxidation peak current) f14= (Half-peak width) f15= (Peak area) f16= (Peak shape asymmetry factor) Structural specific features (added dynamically according to antibiotic class): For aromatic ring antibiotics: f17= (Concentration response slope) f18= (π-π action energy index) for nitrogen heterocyclic antibiotics: f19= (Coordination oscillation frequency) f20= (Oscillation damping coefficient) For β-lactams: f21=ζC=O (carbonyl oxidation characteristic factor). The final feature vector dimension N depends on the specific type of antibiotic, and is usually N = 15 (basic features) + 2~6 (structure-specific features) = 17~21 dimensions.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics, characterized in that, The main steps include: S1. Water sample pretreatment: adjust pH to 6-8, and filter high-turbidity water samples through a 0.22-0.45μm filter. S2. Multi-scale electrochemical signal acquisition: perform potentiostatic chronoamperometry and linear scanning voltammetry on water samples to obtain complete electrochemical response curves; S3. Antibiotic-specific multi-scale electrochemical fingerprint feature hierarchical extraction, specifically: (1) The IT response curve is divided into three stages according to the characteristics of antibiotic redox kinetics: rapid response period, transition period, and steady state period; (2) Extracting capacitor charging characteristics and rapid adsorption kinetic parameters from the fast response period: initial current change rate Current rise time constant Fast adsorption rate constant ; (3) Extract diffusion control characteristics and electrocatalytic reaction parameters from the transition period: diffusion coefficient D, charge transfer resistance Catalytic current gain factor γ; (4) Extracting diffusion control characteristics and electrocatalytic reaction parameters from the transition period: steady-state current Surface coverage θ, adsorption equilibrium constant ; (5) Extracting specific features for different antibiotic molecular structures from LSV curves: For antibiotics containing aromatic rings, extracting the oxidation peak potential caused by π-π stacking. With peak current The ratio and its response slope to concentration; for nitrogen-containing heterocyclic antibiotics, the current oscillation frequency caused by coordination bonding was extracted. and oscillation damping coefficient ;right - Lactam antibiotics, asymmetry factor in extracting carbonyl oxidation peak and half-peak width ; (6) Construct an N-dimensional electrochemical fingerprint vector containing the above features, where N≥15; S4. Adaptive optimization of feature weights based on structural similarity constraints, specifically: (1) Calculate the molecular fingerprint similarity matrix S of the five classes of antibiotics, where The Tanimoto similarity coefficient represents the i-th and j-th class antibiotics; Construct an inter-class confusion risk matrix C. ,in For historical confusion probability; Risk of confusion Threshold The category pairs are used to identify their discriminative feature sets. ; Based on the feature importance in the XGBoost model, the weights of discriminative features are adaptively adjusted: ,in These are the original feature weights. For the adjusted weights, For adjustment coefficients; The XGBoost model was retrained using the adjusted feature weights to construct a structure-aware XGBoost (SA-XGBoost) model. S5, Cascaded antibiotic identification and quantification, specifically: (1) First-level coarse classification: based on steady-state current and oxidation peak potential Quickly predict and eliminate impossible categories; (2) Second-level fine classification: Input the N-dimensional features extracted by S3 into the SA-XGBoost model constructed by S4, and output the posterior probability of 5 types of antibiotics; (3) Third-level concentration quantification: Based on the identified antibiotic category, the corresponding calibration curve is called to calculate the concentration; (4) Fourth-level interference correction: detect the interference signals of metal ions and organic matter in the water sample and perform matrix effect compensation; S6. Output test results, including antibiotic type, concentration, confidence level, and quality control label.

2. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 1, characterized in that, In step S3, feature extraction during the fast response period uses the first 5 seconds of data to fit the IT curve using a double exponential model. ,in For background current, , For the amplitude (positive value) and time constant of the fast response component, , The magnitude and time constant of the slow response component, and the initial current. As time approaches Extract the fast adsorption rate constant from the fitted parameters .

3. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 2, characterized in that, The method for extracting specific features of aromatic ring antibiotics in step S3 is as follows: (1) Identify the characteristic peak on the LSV curve corresponding to aromatic epoxidation, and the potential of this peak. Typically, it is in the range of +0.8V to +1.4V; (2) Calculate peak potential With peak current ratio This ratio reflects the π-π packing strength between the aromatic ring and the rGO on the electrode surface; (3) Measure and calculate the concentrations at different levels. slope of response to concentration This slope is used to distinguish different types of aromatic ring antibiotics.

4. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 3, characterized in that, In step S4, the molecular fingerprint similarity is calculated using MACCSKeys or Morgan fingerprints, and the Tanimoto similarity coefficient is defined as follows: in and The molecular fingerprint sets of antibiotics of class i and class j are respectively, and the similarity coefficients between quinolones and tetracyclines are given. , - The similarity coefficient of each subclass within a lactam is ≥0.

6.

5. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 4, characterized in that, In step S4, the SA-XGBoost model is trained using a contrastive learning-enhanced loss function. ,in For cross-entropy loss, To compare the losses: and Let m be the feature representation of samples i and j, λ be the boundary value, λ be the balance coefficient, and N be the number of training samples. This loss function avoids the influence of sample size on the loss value through normalization, making the feature representations of antibiotic samples of the same class closer together and samples of different classes farther apart.

6. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 1, characterized in that, The specific method for the fourth-level interference correction in step S5 is as follows: (1) Establish an electrochemical response fingerprint library for common interfering substances, including , , humic acid, fulvic acid; (2) The reduction peak of metal ions was identified by scanning in the range of -0.4V to +0.2V using differential pulse voltammetry; (3) The humic acid content was estimated by the absorbance at 254 nm using the ultraviolet-visible absorption spectrum; (4) Based on the concentration of interfering substances, consult the pre-established correction factor table to compensate for the antibiotic concentration: in This is the correction coefficient for the k-th interfering substance. When the concentration of the interfering substance is 0, This is consistent with the physical meaning.

7. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 1, characterized in that, The N-dimensional electrochemical fingerprint vector constructed in step S3 includes, but is not limited to, the following features: time-domain features: steady-state current. Response time Fast response time constant Slow response time constant Current change rate max Potential domain characteristics: Oxidation peak potential reduction peak potential Peak potential difference Half-wave potential Peak shape characteristics: Oxidation peak current Peak width Peak shape asymmetry factor Peak area I- ; Kinetic characteristics: apparent diffusion coefficient Charge transfer rate constant Adsorption equilibrium constant Structural characteristics: π-π response slope Coordination oscillation frequency Characteristic factors of carbonyl oxidation.

8. The intelligent electrochemical detection method for water bodies based on the molecular structure characteristics of antibiotics according to claim 1, characterized in that, The hyperparameters of the SA-XGBoost model are set as follows: number of trees nestimators = 200 500, maximum depth maxdepth=5 8. Learning rate = 0.01 0.05, subsample ratio subsample=0.7 0.9, feature sampling ratio colsamplebytree=0.7 0.9, L2 regularization parameter lambda=1 5. Structural similarity weight adjustment coefficient β = 0.1 0.

5.

9. The intelligent electrochemical detection method for water based on the molecular structure characteristics of antibiotics according to any one of claims 1-8, characterized in that, It also includes a portable electrochemical detection system, which comprises: (1) Screen-printed three-electrode sensor chip, with working electrode material being rGO-AgNWs@Ti3C2T x -TiO2 composite nanomaterials; (2) Portable potentiostat, with a potential range of -2V to +2V and a current resolution of ≤1nA; (3) Embedded processor, running SA-XGBoost model inference; (4) Touch screen and wireless communication module; (5) Lithium battery power supply, capacity ≥10000mAh, battery life ≥48 hours; system size ≤300mm×200mm×150mm, weight ≤3kg, protection level ≥IP65.

10. A smart electrochemical detection method for water based on the molecular structure characteristics of antibiotics according to any one of claims 1-8, characterized in that, It also includes an autonomous monitoring system for the unmanned surface vessel platform, the unmanned surface vessel comprising: (1) Automatic sampling device, which can automatically collect water samples at set navigation points; (2) An integrated electrochemical detection module, including a sensor array and a potentiostat; (3) Data processing and communication module, which runs the SA-XGBoost model in real time and uploads it to the cloud via 4G / 5G; (4) GPS / BeiDou positioning module, positioning accuracy ≤5m; (5) Solar power supply system with a battery life of ≥72 hours; USV platform monitoring coverage of ≥10km², sampling point interval ≤500m, and data upload delay ≤10 seconds.

Images