A traceability method and system based on multi-dimensional electrochemical fingerprint spectrum

Abstract

The present application relates to the technical field of electrochemical tracing technology, and provides a tracing method and system based on multi-dimensional electrochemical fingerprint spectrum, which comprises the following steps: collecting multiple types of electrochemical signals on the same sample jointly and fusing modeling to realize accurate identification of liquid sources. Two cycles of cyclic voltammetry curves are continuously collected and difference is calculated, dynamic changes of the electrode interface caused by irreversible active substances are converted into dynamic background fingerprints; differential pulse voltammetry signals are synchronously collected to capture weak redox response; Fourier transform alternating current voltammetry signals are collected, and the capacitive background is naturally inhibited by means of high-order harmonic, and the time domain information is orthogonal and complementary. In combination with local baseline correction in the partition, non-linear background drift is adapted, peak characteristics are accurately extracted, and finally time-frequency characteristics are fused to build a library and select a classification model. The present application can overcome high background interference of a complex system, improve recognition accuracy and robustness, and can complete detection by using a single device, which is convenient for rapid deployment on site.

Description

Technical Field

[0001] This invention relates to the field of electrochemical traceability technology, and more specifically, to a traceability method and system based on multi-dimensional electrochemical fingerprinting. Background Technology

[0002] The content in this section only provides background information related to this invention and may not constitute prior art.

[0003] With the rapid advancement of urbanization and industrialization, water pollution incidents such as illegal discharge of industrial wastewater into urban drainage networks, mixed connections between stormwater and sewage pipe networks, and sudden chemical leaks occur frequently. These incidents are characterized by their hidden sources, complex composition, and significant temporal and spatial fluctuations, posing a significant challenge to urban water environment supervision and pollution emergency response. Electrochemical detection technology, due to its portable equipment, rapid response, and ease of on-site deployment, has been gradually applied in the fields of water quality monitoring and pollution source tracing.

[0004] However, existing electrochemical-based wastewater source tracing schemes still have obvious technical defects: First, existing voltammetric source tracing methods mostly use single-cycle cyclic voltammetric curves or differential pulse voltammetric curves as the basis for identification. The differences in signals between cycles generated by continuous scanning are generally attributed to test repeatability errors and ignored. This fails to explore the dynamic background changes caused by the adsorption, passivation, polymerization and deposition of irreversible electrochemical active species in wastewater on the electrode surface, and loses independent characteristic dimensions that can characterize the source of wastewater.

[0005] Secondly, existing solutions mostly rely on time-domain volt-ampere signals, which have limited ability to suppress capacitive double-layer currents. Wastewater is a typical non-ideal electrochemical system with strong background interference and complex composition. The capacitive background current can easily overwhelm the weak Faraday response, resulting in low accuracy of effective feature extraction. Furthermore, the single time-domain information dimension is insufficient to distinguish wastewater with similar components, making it difficult to ensure the source tracing accuracy of complex wastewater systems.

[0006] Third, some existing source tracing solutions require the combined use of multiple methods such as spectral detection, multi-parameter sensor arrays, or bioelectrodes, which makes the detection process cumbersome and the hardware deployment cost high, making it difficult to meet the needs of rapid on-site early warning and source tracing; solutions that only use a single electrochemical method generally suffer from insufficient information dimensions and weak background interference resistance, and the robustness and resolution of identification in high background sewage scenarios are difficult to meet actual regulatory requirements. Summary of the Invention

[0007] The purpose of this invention is to provide a source tracing method and system based on multi-dimensional electrochemical fingerprinting to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows: Firstly, this application provides a source tracing method based on multi-dimensional electrochemical fingerprinting, including: S101: Collect a liquid sample after coarse filtration to remove large particulate matter. When the conductivity or ohmic drop of the liquid sample is within the first preset range, dilute the liquid sample according to a set ratio and add an inert supporting electrolyte so that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. Divide the liquid sample into unoxidized samples and oxidized samples of equal volume. The oxidized samples are subjected to controlled oxidation electrolysis through multiple three-electrode systems corresponding to different materials, while the unoxidized samples are not electrolyzed. S102, for unoxidized and oxidized samples, the following electrochemical signals are collected sequentially: under the same scanning conditions, the first and second cyclic voltammetry curves are continuously collected; a differential pulse voltammetry curve is collected; a single-frequency Fourier transform AC voltammetry signal is collected, and a single-frequency preset amplitude AC disturbance is superimposed on the DC linear ramp of the Fourier transform AC voltammetry signal, and serial acquisition is performed independently according to multiple different frequencies to obtain current-time data at each frequency; S103, calculate the difference curve between the first and second cyclic voltammetry curves; perform local baseline correction on the first, second, and differential pulse voltammetry curves according to the potential intervals, and extract characteristic parameters of each net redox peak. The characteristic parameters include one or more of the following: peak potential, peak current, half-maximum width, peak area, peak symmetry factor, and spacing between adjacent peaks; based on current-time data, extract oxidation features including oxidation integral charge, initial average current, steady-state average current, and time constant; use the difference curve, characteristic parameters, and oxidation features as time-domain features; based on current-time data, segment, window, and perform fast Fourier transform processing on each frequency track of the Fourier transform AC voltammetry signal according to a dynamic window, and extract the amplitude and phase of the fundamental wave and preset higher harmonics as frequency-domain harmonic features; S104 integrates time-domain features with frequency-domain harmonic features corresponding to Fourier transform AC volt-ampere signals to construct a multi-dimensional electrochemical fingerprint library with source labels. S105 takes features from a multi-dimensional electrochemical fingerprint database as input and liquid source category as output label. It trains from a variety of pre-set candidate classification models and selects the model with the highest classification accuracy on the blind test set as the final source tracing model. S106, execute S101 to S104 sequentially on the liquid sample to be tested to obtain the current multi-dimensional electrochemical fingerprint database. Input the multi-dimensional electrochemical fingerprint database into the final traceability model and output the liquid source category to which the liquid sample to be tested belongs.

[0008] Furthermore, in multiple three-electrode systems corresponding to different materials, the working electrode in any three-electrode system adopts one of glassy carbon, platinum, boron-doped diamond, or titanium-based inert electrodes.

[0009] Furthermore, liquid samples are collected at pipeline nodes or discharge points in the target area.

[0010] Furthermore, serial acquisition is performed independently at multiple different frequencies, specifically including: A frequency set is preset, which contains multiple different frequency values ​​within a preset frequency range; For the same liquid sample, multiple complete DC scans are performed independently in sequence. During each scan, only a single frequency AC disturbance is superimposed on the DC ramp. The acquisition order of each frequency is randomized to eliminate system drift introduced by the measurement order.

[0011] Furthermore, before performing independent serial acquisition based on multiple different frequencies, the process also includes: In a blank solution of pure supporting electrolyte corresponding to the liquid sample, the solution resistance and double layer capacitance were measured by electrochemical impedance spectroscopy. Based on the solution resistance and double-layer capacitance, the available bandwidth at each frequency is calculated, and frequencies exceeding the available bandwidth are removed from the frequency set.

[0012] Furthermore, local baseline correction is performed separately for each potential interval, specifically including: For the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve, the asymmetric least squares algorithm was used for baseline fitting. The baseline fitting results obtained by the asymmetric least squares algorithm were cross-validated by combining the sliding window polynomial fitting method. The corresponding baseline was subtracted from the original curve to obtain the correction result.

[0013] Furthermore, the amplitude and phase of the fundamental wave and the preset higher harmonics are extracted, specifically including: When acquiring the phase of each higher harmonic, a dual-channel encoding method consisting of sine and cosine components is used to represent the phase value to avoid phase jumps at positive and negative boundaries; and the phase of each harmonic measured at the same frequency in a pure supporting electrolyte blank solution is used as a reference standard to perform absolute correction on the phase of each harmonic of the liquid sample.

[0014] Furthermore, a multi-dimensional electrochemical fingerprint library with source labels is constructed, specifically including: The data of the Fourier transform AC volt-ampere signal on each frequency track are interpolated and aligned to a common DC potential grid. Using DC potential as the first dimension, harmonic order as the second dimension, and AC frequency as the third dimension, the harmonic amplitude and phase corresponding to each coordinate point are stacked to form a three-dimensional fingerprint tensor. The time-domain features corresponding to the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, the differential curve, and the differential pulse voltammetry curve are used as additional data channels and combined with the three-dimensional fingerprint tensor to form a multi-dimensional electrochemical fingerprint of the liquid sample.

[0015] Furthermore, candidate classification models include: support vector machine model, random forest model, gradient boosting tree model, partial least squares discriminant analysis model, K-nearest neighbor model, one-dimensional convolutional neural network model, and deep learning model based on residual network combined with self-attention mechanism.

[0016] Secondly, this application also provides a traceability system based on multi-dimensional electrochemical fingerprinting, including: The sample processing module is used to collect liquid samples after coarse filtration to remove large particulate matter. When the conductivity or ohmic drop of the liquid sample is within a first preset range, the liquid sample is diluted according to a set ratio and an inert supporting electrolyte is added so that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. The liquid sample is divided into unoxidized samples and oxidized samples of equal volume. The oxidized samples are subjected to controlled oxidation electrolysis through multiple three-electrode systems corresponding to different materials, while the unoxidized samples are not electrolyzed. The signal acquisition module is used to sequentially acquire the following electrochemical signals for unoxidized and oxidized samples: continuously acquire the first and second cyclic voltammetry curves under the same scanning conditions; acquire a differential pulse voltammetry curve; acquire a single-frequency Fourier transform AC voltammetry signal, which is superimposed with a preset amplitude AC disturbance of a single frequency on a DC linear ramp, and is independently acquired serially according to multiple different frequencies to obtain current-time data at each frequency; The feature extraction module is used to calculate the difference curve between the first and second cyclic voltammetry curves; perform local baseline correction on the first, second, and differential pulse voltammetry curves according to the potential range, and extract the characteristic parameters of each net redox peak. The characteristic parameters include one or more of the following: peak potential, peak current, half-maximum width, peak area, peak symmetry factor, and spacing between adjacent peaks; based on current-time data, extract oxidation features including oxidation integral charge, initial average current, steady-state average current, and time constant; use the difference curve, characteristic parameters, and oxidation features as time-domain features; based on current-time data, segment, window, and perform fast Fourier transform processing on each frequency track of the Fourier transform AC voltammetry signal according to a dynamic window, and extract the amplitude and phase of the fundamental wave and preset higher harmonics as frequency-domain harmonic features. The fingerprint database construction module is used to fuse time-domain features with frequency-domain harmonic features corresponding to Fourier transform AC volt-ampere signals to construct a multi-dimensional electrochemical fingerprint database with source tags. The model selection module is used to train from a variety of preset candidate classification models with features in the multi-dimensional electrochemical fingerprint database as input and liquid source category as output label, and select the model with the highest classification accuracy on the blind test set as the final source tracing model. The source tracing prediction module is used to sequentially execute S101 to S104 on the liquid sample to be tested to obtain the current multi-dimensional electrochemical fingerprint database. The multi-dimensional electrochemical fingerprint database is then input into the final source tracing model, which outputs the liquid source category to which the liquid sample to be tested belongs.

[0017] The beneficial effects of this invention are as follows: This invention first performs coarse filtration to remove impurities from the collected liquid samples, then dilutes samples with insufficient conductivity and adds an inert supporting electrolyte to ensure detection stability. Subsequently, by jointly collecting multiple types of electrochemical signals on the same sample and fusing them into a model, the accurate identification of the liquid source is achieved.

[0018] This invention continuously acquires two cycles of cyclic voltammetry curves for the same sample under consistent conditions and calculates the differential results. Utilizing the dynamic changes in irreversible electrochemically active substances on the electrode surface, such as adsorption, polymerization, or interfacial passivation, the inter-cycle differences, originally considered as testing errors, are transformed into dynamic background fingerprints characterizing sample properties, supplementing the time-domain feature information. Simultaneously, differential pulse voltammetry curves are acquired to capture weak redox responses in high-background regions; then, single-frequency Fourier transform AC voltammetry signals are acquired. Their higher harmonics naturally suppress background interference from double-layer capacitance, obtaining a frequency-domain nonlinear fingerprint orthogonal to the time-domain method, complementing the time-domain information and alleviating the problem of weak Faraday signals being submerged by background in complex systems. Furthermore, interval-based local baseline correction is applied to each time-domain feature to adapt to the nonlinear background drift of complex liquids, accurately extracting interpretable redox peak features.

[0019] This invention integrates time-domain and frequency-domain features to construct a multi-dimensional fingerprint database, selects the best suitable model from multiple candidate classification models, and relies on multi-dimensional complementary fingerprint information and suitable classification algorithms to effectively overcome high background interference in complex liquid systems, improve the accuracy and robustness of source identification, and only requires a single electrochemical detection device to complete the entire process, which is convenient for rapid on-site deployment and application. Attached Figure Description

[0020] Figure 1 A flowchart of a source tracing method based on multi-dimensional electrochemical fingerprinting provided by the present invention; Figure 2 This is a schematic diagram of the three-electrode electrochemical measurement device in this invention; Figure 3a This is a schematic diagram of principal component analysis (PCA) of the unoxidized three-electrode characteristics in this invention; Figure 3bThis is a schematic diagram of principal component analysis (PCA) of the oxidation-enhanced three-electrode characteristics in this invention; Figure 4 This is a schematic diagram illustrating the blind sample test performance of different classification models in this invention; Figure 5 This is a comparison chart of the accuracy rates of different data types and dimension combinations in this invention; Figure 6 This is a schematic diagram of the confusion matrix for blind sample testing of five types of wastewater in this invention; Figure 7 This is a schematic diagram of a traceability system based on multi-dimensional electrochemical fingerprinting provided by the present invention.

[0021] In the diagram: 201, Sample Processing Module; 202, Signal Acquisition Module; 203, Feature Extraction Module; 204, Fingerprint Database Construction Module; 205, Model Optimization Module; 206, Source Tracing and Prediction Module. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] like Figure 1 As shown in the embodiment of the present invention, a source tracing method based on multi-dimensional electrochemical fingerprinting includes: S101, collect a liquid sample after coarse filtration to remove large particulate matter. When the conductivity or ohmic drop of the liquid sample is within a first preset range, dilute the liquid sample according to a set ratio and add an inert supporting electrolyte so that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. Divide the liquid sample into unoxidized samples and oxidized samples of equal volume. The oxidized samples are subjected to controlled oxidation electrolysis through multiple three-electrode systems corresponding to different materials, while the unoxidized samples are not electrolyzed.

[0024] Specifically, liquid samples are collected at pipe network nodes or discharge outlets in the target area. The target area can cover urban drainage pipe network systems, industrial park sewage systems, and areas suspected of water pollution diffusion. Specific sampling points include the junctions of drainage pipe trunk lines and branches, industrial wastewater discharge outlets, discharge outlets in residential and commercial service areas, stormwater pipe network outlets, and key monitoring points suspected of pollution. The principle behind selecting these sampling points is that these locations are key interfaces for the inflow, transmission, and discharge of wastewater from different sources. The collected samples can completely retain the dissolved components and electrochemical activity characteristics of the corresponding discharge source, ensuring that the subsequently constructed electrochemical fingerprint has a direct correspondence with the wastewater source. This avoids dilution of source characteristics due to long-distance transportation or excessive mixing of multiple sources, effectively improving the directionality and accuracy of source tracing.

[0025] The collected liquid samples are first subjected to coarse filtration to remove large suspended particles. Coarse filtration can be performed using quantitative membrane filtration, sand core filtration devices, or dedicated online coarse filtration units to retain large-sized solid phases and floating impurities such as silt, fiber debris, floating oil, and colloidal aggregates. The principle behind this treatment is that large suspended particles are easily adsorbed and deposited on the working electrode surface of the three-electrode system during subsequent electrochemical measurements. This physically blocks the active sites of the electrodes and disrupts the uniform structure of the double layer at the electrode interface, leading to baseline drift, abnormal attenuation of peak current, and decreased batch-to-batch reproducibility of fingerprints. By removing large suspended particles through coarse filtration, the physical interference of solid impurities on the electrode interface can be eliminated without changing the composition and content of soluble electrochemical active species in the sample, ensuring the stability of subsequent multi-dimensional electrochemical signal acquisition and the reliability of feature extraction.

[0026] In measurement scenarios requiring the acquisition of single-frequency Fourier transform AC volt-ampere signals, the conductivity or ohmic drop level of the liquid sample must first be evaluated. When the conductivity or ohmic drop of the liquid sample falls within a first preset range—that is, when the sample conductivity is too low and the ohmic drop is too large, insufficient to support high-precision AC volt-ampere potential control—the liquid sample is diluted according to a set ratio, and an inert supporting electrolyte is added to ensure that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. The first preset range corresponds to the interval in electrochemical measurements where the ohmic drop significantly affects the accuracy of potential control. Its setting principle is as follows: Fourier transform AC volt-ampere measurements rely on the superposition of a precise DC potential slope and an AC disturbance signal. Insufficient solution conductivity leads to excessive solution resistance, resulting in a significant ohmic voltage drop. This causes the actual potential applied to the electrode interface to deviate from the set value, inducing amplitude attenuation and phase shift of AC harmonics. In particular, it significantly reduces the signal-to-noise ratio of higher harmonics, compromising the stability and discriminability of the frequency domain fingerprint. Adding an inert supporting electrolyte with no electrochemical activity can significantly improve the ionic strength of the solution, reduce the solution resistance, and control the ohmic voltage drop within an acceptable range, ensuring the accuracy of the applied potential and the quality of the harmonic signal. The dilution-salting treatment method avoids the salting-out effect and localized concentration unevenness caused by directly adding a solid supporting electrolyte, ensuring uniform dissolution of the supporting electrolyte. It also moderately reduces the passivation deposition rate of high-concentration pollutants on the electrode surface, extending the electrode activity stability period during multi-frequency serial acquisition. The preset threshold is typically set to 0.1 mol / L. This concentration provides sufficient ionic strength support for AC voltammetry measurements without altering the existing active species and redox behavior in the wastewater due to excessively high salt concentrations, ensuring that the treated sample still accurately reflects the source characteristics of the original wastewater. For cyclic voltammetry, differential pulse voltammetry, and controlled oxidation electrolysis, in order to preserve the non-ideal electrolyte system properties of liquid samples to the greatest extent and to realistically reproduce the interfacial adsorption, charge transfer, and mass transfer behavior of active species in complex liquids, no additional inert supporting electrolyte is added; only in the Fourier transform AC voltammetry measurement scenario, the electrolyte concentration is adjusted as needed according to the above rules, so as to balance the authenticity of the fingerprint of the non-ideal system and the acquisition and reproducibility of AC harmonic signals.

[0027] In one specific embodiment, a liquid sample was collected from a rainwater drainage outlet in an industrial park. Due to the significant dilution effect of rainwater, the sample had low conductivity. Testing showed that its conductivity and ohmic drop were within a first preset range, failing to meet the basic requirements for Fourier transform AC voltammetry measurement. In this case, the sample was diluted at a volume ratio of 1:2, and potassium chloride was added as an inert supporting electrolyte. After thorough mixing, the final concentration of the supporting electrolyte was measured to be 0.15 mol / L, reaching and exceeding the preset threshold of 0.1 mol / L, thus qualifying it for subsequent electrochemical fingerprinting. However, for samples from electroplating wastewater discharge outlets within the same industrial park, which have high salt content and sufficient conductivity, with ohmic drops outside the first preset range, dilution and the addition of supporting electrolyte are unnecessary. After coarse filtration to remove large suspended particles, a full range of electrochemical signals was directly acquired to maximize the preservation of the original wastewater's electrochemically active components and source fingerprint characteristics.

[0028] Furthermore, liquid samples were divided into unoxidized and oxidized samples of equal volume. Oxidized samples underwent controlled oxidation electrolysis using multiple three-electrode systems with different materials, while unoxidized samples were not electrolyzed. In other words, after sample grouping, the two samples maintained identical parameters in terms of temperature, settling conditions, and processing time, except for the application of oxidation potential, with volume deviation controlled within 2% to ensure comparability of data before and after oxidation. Unoxidized samples were not subjected to any electrolysis potential and were simply settling under identical conditions for the corresponding duration to preserve their original electrochemical fingerprint. Oxidized samples underwent constant-potential controlled oxidation electrolysis using a three-electrode system. The working electrode could be glassy carbon, platinum, boron-doped diamond, or titanium-based inert electrodes, the counter electrode was platinum or titanium, and the reference electrode was uniformly Ag / AgCl to ensure a consistent potential reference.

[0029] This invention only incorporates a forward oxidation electrolysis process and does not include reductive electrolysis. The core principle is to use electrolysis as an active excitation method to induce interfacial oxidation, adsorption state changes, or directional transformation of electrochemically active substances in the sample with different valence states, complex states, adsorbed states, or kinetically inert states, rather than simply removing contaminants. The oxidation electrolysis is performed with a constant forward oxidation potential of +0.6 to +1.2 V applied under an Ag / AgCl reference electrode. The electrolysis time is controlled between 30 and 180 seconds, the electrolysis temperature is maintained at 20 to 30°C, and the stirring speed is set to 200 to 800 r / min. This potential range promotes controlled transformation of oxidizable components in the sample while avoiding excessively high potentials that could trigger violent gas evolution and non-target side reactions. During electrolysis, current-time data are continuously recorded at a sampling frequency of at least 10 Hz. After electrolysis, the applied potential is stopped, and the oxidized sample is allowed to stand for 5 to 30 seconds to eliminate interference from stirring, concentration polarization, and instantaneous charging current on subsequent detection.

[0030] Under specific implementation conditions, 10.00 mL of the same wastewater sample was taken as the unoxidized sample and the oxidized sample, respectively. KCl was added to achieve a final concentration of 0.10 mol / L, and the samples were treated at 25℃ and 400 r / min. The unoxidized sample was kept for the same duration without applying a potential, while the oxidized sample was given a constant potential of +1.0V for 60s. After electrolysis, stirring was stopped and the sample was allowed to stand for 10s. If continuous bubbles or current exceeding the instrument range occurred during the detection process, the oxidation potential was appropriately reduced or the electrolysis time was shortened, and the actual treatment parameters and abnormal conditions were recorded simultaneously.

[0031] S102, for the unoxidized sample and the oxidized sample, the following electrochemical signals are collected sequentially: under the same scanning conditions, the first cycle cyclic voltammetry curve and the second cycle cyclic voltammetry curve are collected continuously; a differential pulse voltammetry curve is collected; a single-frequency Fourier transform AC voltammetry signal is collected, wherein the Fourier transform AC voltammetry signal is superimposed with a single-frequency preset amplitude AC disturbance on a DC linear ramp, and is independently serially collected according to multiple different frequencies to obtain current-time data at each frequency.

[0032] Specifically, in this step, the three-electrode system uses a glassy carbon electrode as the working electrode, a platinum wire electrode as the counter electrode, and Ag / AgCl (saturated KCl) as the reference electrode. Figure 2 As shown, this system can precisely control the potential difference between the working electrode and the reference electrode, eliminate potential deviation caused by electrode polarization, and ensure the consistency of the potential reference for multiple signal acquisitions.

[0033] Specifically, this step employs at least three independent three-electrode detection units to acquire signals for both unoxidized and oxidized samples. All detection units share an Ag / AgCl (saturated KCl) reference electrode and a platinum wire counter electrode. The working electrodes are a glassy carbon electrode (2–5 mm in diameter), a gold disk electrode, and a boron-doped diamond electrode, respectively. Figure 2 As shown, different electrode materials possess differentiated surface chemistry, electron transfer capabilities, adsorption characteristics, and potential windows, providing complementary interfacial responses and forming a multi-interface, multi-window joint fingerprint, avoiding information loss caused by the insensitivity of a single electrode to certain substances. Among them, glassy carbon electrodes have a wide response range and a stable carbon surface, making them suitable for capturing comprehensive information on aromatic structures, general redox substances, and interfacial adsorption processes in complex liquids, serving as a basic detection electrode. Gold disk electrodes possess the catalytic properties of noble metal surfaces, and their responses to certain metal ions and easily oxidized substances differ from those of carbon electrodes, supplementing the resolution blind spots of glassy carbon electrodes. Boron-doped diamond electrodes have a wider potential stability window and lower background current, allowing for the observation of electrochemical responses over a broader potential range and improving the discriminative power of weak Faraday signals.

[0034] During the detection process, the scanning range was set according to the stable potential window of different electrode materials: the cyclic voltammetry and differential pulse voltammetry potential range for the glassy carbon electrode was -1.3 to +1.4 V, for the gold disk electrode it was -0.6 to +1.1 V, and for the boron-doped diamond electrode it was -1.5 to +1.8 V. The acquisition process for all electrodes remained consistent, and the detection order of unoxidized samples, oxidized samples, and the three working electrodes was randomly arranged within the sample to reduce systematic errors caused by measurement sequence, electrode state changes, and instrument drift. The three-electrode system can precisely control the potential difference between the working electrode and the reference electrode, eliminate potential deviations caused by electrode polarization, and ensure the consistency of the potential reference for multiple signal acquisitions.

[0035] For each working electrode, under the same scanning conditions, the first and second cycles of cyclic voltammetry were continuously acquired. The scanning window for cyclic voltammetry was -1.3 to +1.4 V (relative to the Ag / AgCl reference electrode), and the scan rate was 20 to 100 mV·s. -1 Between two acquisition cycles, there is no standing time, no solution change, and no electrode re-polishing. Furthermore, the initial potential, scanning direction, and scanning rate of the second cycle are strictly consistent with the first cycle. Maintaining consistent scanning conditions between cycles ensures that the current difference between cycles originates solely from irreversible processes induced at the electrode interface by the liquid sample's own components, including surfactant adsorption, deposition of easily passivated heavy metal oxide films, and film formation by electropolymerizable organic matter. This avoids introducing additional errors through manipulated variables. Continuous, uninterrupted acquisition maximizes the preservation of the first cycle's modifying effect on the electrode interface, enabling subsequent difference curves to accurately characterize the dynamic background changes of irreversible electrochemically active species in the sample. This transforms the inter-cycle differences, originally considered testing errors, into an independent fingerprint dimension, enriching the feature hierarchy of the time-domain information.

[0036] After completing two cycles of cyclic voltammetry acquisition on each working electrode, a differential pulse voltammetry curve was simultaneously acquired. The differential pulse voltammetry scan window was -1.3 to +1.4 V, the pulse amplitude was 25 to 50 mV, the pulse width was 40 to 60 ms, the step potential was 4 to 10 mV, and the equivalent scan rate was 10 to 30 mV·s. -1 Differential pulse voltammetry, by sampling the current difference before and after the pulse, can effectively deduct part of the double-layer charging current. In high-background wastewater systems, its ability to detect weak redox peaks is superior to that of conventional cyclic voltammetry. It can supplement the subtle distinguishing features that two cycles of cyclic voltammetry cannot fully resolve, complementing the time-domain differential information and further improving the resolution of samples from similar sources.

[0037] Subsequently, a single-frequency Fourier transform AC volt-ampere signal was synchronously acquired on each working electrode. The Fourier transform AC volt-ampere signal was then superimposed with a single-frequency AC disturbance of a preset amplitude on a DC linear ramp, with the bottom DC ramp scan rate ranging from 1 to 5 mV·s. -1 The peak-to-peak amplitude of the superimposed AC disturbance is 20–80 mV; and multiple independent serial acquisitions are performed at different frequencies to obtain current-time data streams at each frequency. Superimposing a large-amplitude AC disturbance on the DC ramp can excite the nonlinear response of the electrochemical reaction at the electrode interface. The high-order harmonics obtained after subsequent Fourier transform can naturally suppress the contribution of double-layer capacitance at the acquisition level, obtaining a nonlinear fingerprint with low capacitance background. This fingerprint complements the time-domain volt-ampere information, solving the problem of weak Faraday signals being submerged in high-background sewage using time-domain methods.

[0038] Before conducting multi-frequency serial acquisition, a frequency set is pre-defined, containing multiple different frequency values ​​within a preset frequency range. For example, this set may include two or more frequencies from 5Hz, 15Hz, 50Hz, 150Hz, and 400Hz, covering approximately 1.5 orders of magnitude. Selecting multiple frequency points covering 1.5 orders of magnitude allows for the simultaneous detection of electrode processes at different timescales. Low frequencies correspond to slow reactions and interfacial adsorption processes, while high frequencies correspond to rapid electron transfer and double-layer relaxation processes. The combination of multiple frequencies can comprehensively capture electrochemically active components with different kinetic characteristics, expanding the information capacity of the frequency domain fingerprint. In actual detection, it is preferable to retain three characteristic frequencies: 50Hz, 150Hz, and 400Hz. These frequencies are used to extract the second harmonic at 50Hz, the third harmonic at 150Hz, and the fourth harmonic at 400Hz, while retaining the phase information of the third harmonic at 150Hz. This maximizes the discriminative power of the frequency domain fingerprint while controlling the data dimensions.

[0039] Furthermore, before independently acquiring data serially at multiple different frequencies, the solution resistance and double-layer capacitance were measured using electrochemical impedance spectroscopy in a blank solution of pure supporting electrolyte corresponding to the liquid sample. Based on these values, the available bandwidth at each frequency was calculated, and frequencies exceeding the available bandwidth were removed from the frequency set, with a bandwidth margin of 2–3 times. By measuring the solution resistance and double-layer capacitance using electrochemical impedance spectroscopy, the RC cutoff frequency of the system can be accurately calculated. After margin screening, potential control distortion and harmonic signal distortion caused by ohmic drop in the solution at high frequencies can be avoided, ensuring that the harmonic data of each selected frequency has effective electrochemical significance and preventing invalid frequency domain data from interfering with the discrimination accuracy of subsequent classification models.

[0040] For the same liquid sample, multiple complete DC scans are performed independently and sequentially. During each scan, only a single frequency AC disturbance is superimposed onto the DC ramp. The acquisition order of each frequency is randomized to eliminate system drift introduced by the measurement sequence. Specifically, single-frequency independent serial acquisition, rather than simultaneous superposition of multiple frequencies, avoids intermodulation distortion between large-amplitude AC signals, ensuring the purity and quantitative reliability of each harmonic. Randomizing the acquisition order can offset the non-uniform effects of system drift, such as the slow passivation of the electrode surface over time and minor evaporation of the solution, on different frequency tracks, ensuring the comparability and reproducibility of fingerprints at each frequency and improving the stability of the subsequent fingerprint database.

[0041] In one specific embodiment, a three-electrode system was constructed using a 3mm diameter glassy carbon electrode as the working electrode, a platinum wire electrode as the counter electrode, and a saturated KClAg / AgCl electrode as the reference electrode. Signal acquisition was performed sequentially on an electroplating wastewater sample from an industrial park. The cyclic voltammetry scan window was set to -1.3 to +1.4 V, and the scan rate was 50 mV·s. -1 The initial potential was -1.3V. The first cycle of cyclic voltammetry was completed by scanning forward to +1.4V and then scanning back to the initial potential in the negative direction. No settling, electrolyte changes, or electrode polishing were performed between the two cycles. The second cycle of cyclic voltammetry was then acquired immediately with identical parameters. Subsequently, differential pulse voltammetry was acquired, with a pulse amplitude of 50mV, a pulse width of 50ms, a step potential of 8mV, and an equivalent scan rate of approximately 20mV·s. -1 Before formally acquiring the Fourier transform AC voltammetric signal, electrochemical impedance spectroscopy was first tested in a 0.15 mol / L potassium chloride blank supporting electrolyte. The solution resistance and double-layer capacitance were measured, and the RC cutoff frequency was calculated to confirm that frequencies up to 400 Hz all met the 2-fold bandwidth margin. The frequency sets of 5 Hz, 15 Hz, 50 Hz, 150 Hz, and 400 Hz were retained. For the five frequencies, a randomly generated acquisition sequence was used, and a complete DC scan was run independently in sequence at a DC scan rate of 3 mV·s. -1 The peak-to-peak AC amplitude is 40mV. Each scan only superimposes a single frequency AC disturbance, and current-time data streams corresponding to five frequencies are collected respectively.

[0042] S103, calculate the difference curve between the first-cycle cyclic voltammetry curve and the second-cycle cyclic voltammetry curve; perform local baseline correction on the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve according to the potential interval, and extract the characteristic parameters of each net redox peak. The characteristic parameters include one or more of the following: peak potential, peak current, half-maximum width, peak area, peak symmetry factor, and adjacent peak spacing; based on the current-time data, extract oxidation features including oxidation integral charge, initial average current, steady-state average current, and time constant; use the difference curve, the characteristic parameters, and the oxidation features as time-domain features; based on the current-time data, perform segmentation, windowing, and fast Fourier transform processing on each frequency track of the Fourier transform AC voltammetry signal according to a dynamic window, and extract the amplitude and phase of the fundamental wave and the preset higher harmonics as frequency-domain harmonic features.

[0043] Specifically, the difference curve between the first and second cyclic voltammetry curves is first calculated. Since the two curves were continuously acquired under completely consistent scanning parameters, with no settling, solution replacement, or electrode re-polishing between cycles, the two curves were first aligned point-by-point according to the same potential sampling points, and then subtracted accordingly to obtain the difference curve ΔCV = CV1 - CV2. The principle of this difference calculation is that the strictly consistent scanning conditions eliminate the interference of operational errors and differences in the initial state of the electrodes, allowing the difference result to reflect only the irreversible processes initiated at the electrode interface by the wastewater sample's own components during the first scan, including surfactant adsorption, deposition of easily passivated heavy metal oxide films, and film formation of electropolymerizable organic matter. This transforms the inter-cycle current difference, usually considered a test repeatability error, into an independent dynamic background fingerprint dimension, effectively characterizing the types and content of irreversible electrochemically active species in wastewater, thus overcoming the information limitation of a single-cycle cyclic voltammetry cycle, which can only reflect the quasi-steady-state response.

[0044] The aforementioned inter-circle difference calculation is performed independently on each working electrode, without averaging the original curves of different electrodes, to fully preserve the interface response specificity of each electrode material. Based on this, for the same working electrode, the corresponding curves of oxidized and unoxidized samples are aligned to a common potential grid, and the oxidation difference curve is calculated. That is, the potential-current value corresponding to the oxidation difference curve is the current difference between the oxidized and unoxidized samples at the same potential. Furthermore, the root mean square, positive peak value, negative peak value, absolute area, and zero-crossing position of the oxidation difference curve are extracted. Simultaneously, the changes in peak current, peak area, and peak potential of each curve before and after oxidation are statistically analyzed. This transforms the differences in the content of oxidizable components, oxidation ease, mass transfer characteristics, and interface passivation rate of the samples into quantifiable feature dimensions.

[0045] After completing the differential calculation, local baseline correction was performed on the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve, respectively, according to their potential intervals. Wastewater is a typical non-ideal electrochemical system, and its background current exhibits complex nonlinear drift with potential changes. Global baseline subtraction is prone to correction deviations in different potential intervals. Therefore, a segmented local baseline correction strategy was adopted to adapt to the background characteristics of the wastewater system. Specifically, baseline fitting was performed on the original curves of the three curves using an asymmetric least squares algorithm, with the smoothing factor λ ranging from 10. 4 ~10 7 The asymmetric parameter p ranges from 0.001 to 0.01. This algorithm, through an asymmetric weighting mechanism, can fit a slowly changing background baseline while preserving the sharp redox peak shape, avoiding the misclassification of the true Faraday response within the baseline range. Simultaneously, a sliding window polynomial fitting method is used to cross-validate the baseline results of the asymmetric least squares algorithm. The sliding window width is set to 0.2V, and the polynomial order is 2. By comparing the results of the two algorithms, local fitting biases are corrected, further improving the accuracy of baseline localization. After subtracting the corresponding baseline from the original curve, the net response curve (i.e., the corrected result) after removing background interference is obtained, providing a reliable data foundation for subsequent peak feature extraction.

[0046] After baseline correction, characteristic parameters of the net redox peaks in each curve are extracted. Within a scanning window of -1.3 to +1.4 V, molecular intervals are cut in steps of 0.01 to 0.05 V. Peak detection is performed within each sub-interval, with a peak detection signal-to-noise ratio threshold (S / N) ≥ 3 and a minimum half-width (HWHM) of no less than 0.03 V to filter out spurious peak interference caused by random noise. For each valid net redox peak, one or more of the following are extracted as characteristic parameters: peak potential, peak current, HWHM, peak area, peak symmetry factor, and inter-peak spacing. Peak potential corresponds to the inherent redox characteristics of the active species and can support qualitative differentiation of components. Peak current and peak area are directly related to the concentration level of the active species. HWHM and peak symmetry factor reflect the reversibility and kinetic speed of the electrochemical reaction. Inter-peak spacing characterizes the potential distribution of multiple electron transfer processes or multiple active components.

[0047] For the current-time curves collected during the oxidation electrolysis process, oxidation kinetic characteristics are further extracted: First, the oxidation integral charge, calculated by integrating the change of electrolysis current over time, and divided by the effective electrode area when necessary to obtain the charge per unit area, used to characterize the overall oxidizable capacity of the sample. Second, the initial average current, taken as the average current within 0-5 seconds after the start of electrolysis, reflects the instantaneous response intensity of easily oxidized components. Third, the steady-state average current, taken as the average current in the last 5 seconds before the end of electrolysis, reflects the response level of the continuous oxidation process or mass transfer control stage. Fourth, the time constant, obtained by fitting the current decay curve using a single exponential decay model, characterizes the speed of oxidation reaction, electrode interface transformation, and passivation processes. The above electrolysis kinetic characteristics and oxidation differential characteristics are incorporated into the time-domain feature system. By combining the aforementioned differential curves, oxidation characteristics, and the extracted peak characteristic parameters as time-domain features, a rich and complementary time-domain fingerprint system is formed, encompassing both the dynamic background information of irreversible processes and physically interpretable Faraday response characteristics.

[0048] For Fourier transform AC volt-ampere signals, based on the current-time data stream corresponding to each frequency, each frequency track is segmented, windowed, and processed using a fast Fourier transform (FFT). Specifically, due to the significant differences in period length between different AC frequencies, a fixed window would result in insufficient period count and frequency resolution in the low-frequency band, and period redundancy and low computational efficiency in the high-frequency band. Therefore, a dynamic window segmentation strategy is adopted: each Fourier transform window contains no less than 10 complete AC cycles, i.e., a long window corresponds to the low-frequency track, and a short window corresponds to the high-frequency track, ensuring that the Fourier transform at each frequency has sufficient frequency resolution and period statistical representativeness. The DC mean is first removed from the data within each window, and then a Hanning window is added to suppress spectral leakage. Subsequently, a fast Fourier transform is performed to obtain the corresponding frequency domain spectrum, from which the amplitude and phase of the fundamental wave and the 2nd to 4th harmonics are extracted as frequency domain harmonic features. The fundamental frequency mainly reflects matrix information such as the overall conductivity and double-layer capacitance of the solution, and can be used as a reference reporting position of the system state. The second to fourth higher harmonics originate from the nonlinear response of the electrode reaction. Their double-layer capacitance contribution decreases significantly with the increase of the harmonic order. They can naturally suppress capacitive background interference at the acquisition level, obtain nonlinear frequency domain fingerprints with low capacitive pollution, and form an information orthogonal and complementary relationship with the time domain characteristics. This effectively solves the problem of weak Faraday signals being submerged by capacitive current in high background sewage systems.

[0049] When acquiring the phase information of each harmonic, a dual-channel encoding method consisting of sine and cosine components is used to represent the phase value. That is, the values ​​of the in-phase component and the quadrature component are stored simultaneously, rather than directly storing a single phase angle. This avoids the phase jump problem that occurs at the ±π boundary and prevents numerical discontinuities during subsequent feature fusion and model training. At the same time, the phase of each harmonic measured at the same frequency in a pure support electrolyte blank solution is used as a reference benchmark. Absolute correction is performed on the phase of each harmonic of the liquid sample. That is, the harmonic phase of the sample is subtracted from the corresponding blank background phase. This eliminates common interferences such as phase offset of the instrument system and inherent phase difference caused by solution resistance. The corrected phase reflects only the electrochemical kinetic characteristics of the active species in the wastewater sample, improving the comparability and reproducibility of frequency domain fingerprints between different batches and different locations. The frequency domain features of each frequency track are also extracted separately according to the electrode material, without cross-electrode averaging, thus fully preserving the nonlinear response differences of different electrode interfaces.

[0050] Taking the treatment process of electroplating wastewater samples from an industrial park as an example: First, the two continuously collected cyclic voltammetry curves are aligned point by point according to the potential sampling points within the range of -1.3 to +1.4V, and the corresponding subtraction is used to obtain the ΔCV difference curve; the CV1, CV2 and DPV curves are respectively compared with λ=10 5 Asymmetric least squares algorithm with p=0.005 was used to iterate 10 times to fit the local baseline, supplemented by second-order sliding window polynomial fitting with a window width of 0.2V for cross-validation. After subtracting the baseline, the molecular interval was cut with a step size of 0.02V. Multiple net redox peaks were detected with S / N≥3 as the threshold. The peak potential, peak current, and half-peak width of each peak were extracted and used together with the ΔCV curve to form the time-domain feature set of the sample. For the 5Hz FTACV data stream, the window length was determined to include 10 complete cycles. After DC removal, a Hanning window was added and a fast Fourier transform was performed to extract the amplitude of the fundamental wave and the 2nd, 3rd, and 4th harmonics, as well as the sine and cosine dual-channel phase components. Then, the corresponding harmonic phase measured at the same frequency in a 0.15mol / L potassium chloride blank solution was subtracted to complete the absolute phase correction. The data of each frequency track at 15Hz, 50Hz, 150Hz, and 400Hz were processed in the same way to finally obtain the complete frequency domain harmonic features of the sample.

[0051] S104 integrates time-domain features with frequency-domain harmonic features corresponding to Fourier transform AC volt-ampere signals to construct a multi-dimensional electrochemical fingerprint library with source labels.

[0052] Specifically, the Fourier transform AC volt-ampere signal data on each frequency track is first interpolated and aligned to a common DC potential grid. Because the dynamic window lengths differ for different AC frequencies, the distribution density of the original sampling points on the DC potential axis is inconsistent across frequency tracks, making direct dimensional stacking impossible. Therefore, a cubic spline interpolation method is used to uniformly map the harmonic amplitude and phase data at each frequency to a common potential coordinate system with consistent spacing. The potential grid step size is set to 0.02 V, and the potential range is consistent with the FTACV scanning window, ranging from -0.2 V to +1.0 V. This processing eliminates sampling misalignment between different frequency tracks by unifying the potential coordinates, ensuring a one-to-one correspondence between harmonic features at the same potential location for each frequency. This provides a unified dimensional benchmark for subsequent 3D tensor construction, avoiding feature matching deviations caused by differences in sampling point density and ensuring the spatial consistency of the frequency domain fingerprint.

[0053] Based on this, a three-dimensional fingerprint tensor is constructed by stacking the harmonic amplitudes and phases corresponding to each coordinate point, with DC potential as the first dimension, harmonic order as the second dimension, and AC frequency as the third dimension. The first dimension, DC potential, covers the entire scanning range, carrying the voltage-current law of harmonic changes with potential. The second dimension includes the fundamental frequency and the 2nd, 3rd, and 4th higher harmonics. At each order, the harmonic amplitude and absolutely corrected sine and cosine dual-channel phase components are stored simultaneously, fully preserving the amplitude and dynamic information of the nonlinear response. The third dimension corresponds to the frequency set after bandwidth filtering, including 5 Hz, 15 Hz, 50 Hz, 150 Hz, and 400 Hz, covering approximately 1.5 orders of magnitude of time scale. This three-dimensional tensor, through its regularized three-dimensional layout, fully preserves the intrinsic correlation of frequency domain features in the three dimensions of potential, order, and frequency, giving full play to the natural suppression of double-layer capacitance by higher harmonics at the acquisition level; it can integrate scattered multi-frequency harmonic data into a structured frequency domain fingerprint carrier, which is not only convenient for subsequent models to extract multi-scale features, but also fully presents the advantages of nonlinear fingerprints in low capacitance backgrounds, forming a positive interactive complement with time domain information.

[0054] Subsequently, the time-domain features corresponding to the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, the differential curve, and the differential pulse voltammetry curve were used as additional data channels and combined with the three-dimensional fingerprint tensor to form a multi-dimensional electrochemical fingerprint of the liquid sample.

[0055] After extracting the time-domain and frequency-domain features of a single electrode, the features of the glassy carbon electrode, gold disk electrode, and boron-doped diamond electrode are retained separately according to their electrode materials. These features are then fused at the feature level to form a multi-electrode fused feature. During the fusion process, the material-specific identifiers of each electrode are preserved, allowing the model to trace the electrode interface type corresponding to the features and enhancing the interpretability of the fingerprint. Based on this, the original fingerprint of the unoxidized sample, the post-oxidized fingerprint of the oxidized sample, the oxidation difference features of the two, and the electrolysis kinetic features are dimensionally concatenated, integrating information from the oxidation state dimension and the reaction kinetic dimension to further enrich the information hierarchy of the fingerprint. Finally, the time-domain features integrating multi-electrode response, oxidation state, and kinetic information are combined with the three-dimensional frequency-domain fingerprint tensor to form a multi-dimensional electrochemical fingerprint of the liquid sample.

[0056] All the aforementioned time-domain features have undergone local baseline correction. The original potential range is -1.3 to +1.4 V. After potential range trimming and interpolation adaptation, they are spatially aligned with the common potential grid of the three-dimensional tensor. The differential curves carry the dynamic background information of irreversible active species in wastewater, while the differential pulse voltammetry curves supplement the fine features of weak redox peaks under high background. This fusion method, under a unified potential benchmark, connects interpretable peak features, dynamic background information, and low-capacitance nonlinear fingerprints in the frequency domain in a multi-channel parallel manner. This preserves the independence of features in each dimension while establishing a corresponding correlation in the potential dimension. It can fully integrate the complementary advantages of two-loop CV differential, DPV, and FTACV signals, breaking through the information dimension limitations of single electrochemical methods, significantly improving the fingerprint's ability to distinguish wastewater with similar compositions, and the unified multi-channel structure can simultaneously adapt to the input requirements of traditional machine learning and deep learning models.

[0057] Finally, a source tag is associated with the multi-dimensional electrochemical fingerprint of each sample. The tag can be set as the wastewater source category or a specific discharge outlet number, and each tag is stored in the database to form a multi-dimensional electrochemical fingerprint library of wastewater with source tags. Taking an electroplating wastewater sample from an industrial park as an example, the FTACV data of the five frequency orbitals are first interpolated and aligned to a common potential grid with a step size of 0.02 V, and stacked to form a three-dimensional frequency domain tensor of the corresponding size. Then, the time-domain features of the four curves CV1, CV2, ΔCV, and DPV after baseline correction are combined with it as additional channels to form the complete multi-dimensional fingerprint of the sample. Finally, the source tag "industrial electroplating wastewater" is labeled and stored in the fingerprint library.

[0058] S105 takes features from a multi-dimensional electrochemical fingerprint database as input and liquid source category as output label. It is trained from a variety of pre-set candidate classification models and selects the model with the highest classification accuracy on the blind test set as the final source tracing model.

[0059] Specifically, the core logic of this step is to select the best candidate classification model from those known in the field, rather than improving the algorithm structure of the model itself. The principle is that different classification models have different fitting abilities, generalization performance and computational efficiency for high-dimensional electrochemical fingerprints. By comparing the performance under a unified benchmark, the optimal model is selected, which can be adapted to traceability scenarios with different fingerprint databases and different numbers of categories, ensuring the discrimination accuracy and robustness of the final traceability model.

[0060] When performing this step, the samples in the multi-dimensional electrochemical fingerprint database are first divided into sample-level data. The training set and validation set are divided using the sample-level grouped K-fold cross-validation (GroupKFold) strategy, and completely independent samples are reserved to form a blind test set. This division method uses complete samples as the grouping unit, which can avoid the feature fragments of the same sample from entering the training and validation stages at the same time, thus avoiding data leakage and ensuring the objectivity and credibility of the model performance evaluation.

[0061] For traditional machine learning models such as Support Vector Machine (SVM), Random Forest, Gradient Boosting Tree (XGBoost), Partial Least Squares Discriminant Analysis (PLS-DA), and K-Nearest Neighbors (KNN), the input is a flattened feature vector compressed by Principal Component Analysis (PCA). For one-dimensional convolutional neural network (1D-CNN) models and deep learning models based on residual networks combined with self-attention mechanisms (ResNet-Att-1D), the input is a multi-dimensional electrochemical fingerprint in the form of a multi-channel signal tensor. This approach matches the input format requirements of different models and fully leverages the feature extraction advantages of various models.

[0062] All candidate models were independently trained and parameter-tuned using identical training, validation, and blind test sets. The classification accuracy and macro-average F1 score on the blind test set were used as the core evaluation metrics. The blind test set was used as the final evaluation criterion because its samples were not involved in the training and parameter tuning process, and its performance best reflects the model's actual source tracing ability in the face of unknown samples. The macro-average F1 score can balance the recognition accuracy of different source categories, offsetting the interference of uneven sample distribution on the evaluation results. After training and testing all candidate models, the model with the highest classification accuracy on the blind test set was selected as the final source tracing model. This selection mechanism does not require pre-binding a specific classification algorithm, allowing the core value of the method to focus on the collection and construction of multi-dimensional electrochemical fingerprints. The solution is highly portable and can be flexibly adapted to the sample size and hardware computing power of actual scenarios.

[0063] S106, execute S101 to S104 sequentially on the liquid sample to be tested to obtain the current multi-dimensional electrochemical fingerprint database. Input the multi-dimensional electrochemical fingerprint database into the final traceability model and output the liquid source category to which the liquid sample to be tested belongs.

[0064] Specifically, the sample to be tested is sequentially processed through steps S101 to S104: sample collection and processing, multi-dimensional electrochemical fingerprint acquisition, data processing and feature extraction, and multi-dimensional electrochemical fingerprint construction. This process yields the multi-dimensional electrochemical fingerprint corresponding to the sample to be tested. The fingerprint is then input into the final traceability model determined by S105, and the liquid source category or the most similar emission outlet number to which the sample to be tested belongs is output.

[0065] This step relies on a multi-dimensional fingerprint system with time-frequency positive complementation, using a pre-trained optimal classification model as the discrimination carrier. It eliminates the need for manual comparison of peak positions and spectral features, automatically matching unknown samples with various source features in the fingerprint database. Because the fingerprint integrates dynamic background information from double-loop CV difference, fine peak features from DPV, and low-capacitance nonlinear fingerprints from the 2nd to 4th harmonics of FTACV, the model can stably extract source-specific features even in high-background wastewater systems, avoiding misjudgments caused by insufficient information in a single dimension and ensuring the accuracy and robustness of the source tracing results.

[0066] Based on the above scheme, the experimental results of this embodiment are as follows: Table 1

[0067] Table 2

[0068] In Table 2, “CV1” represents the first-cycle cyclic voltammetry curve; “CV2” represents the second-cycle cyclic voltammetry curve; “ΔCV” represents the differential curve; “DPV” represents the differential pulse voltammetry curve; and “FTACV” represents the Fourier transform AC voltammetry signal. “Three-electrode: CV / DPV / FTACV fusion” refers to the feature-level fusion of the CV1, CV2, inter-cycle differential, DPV, and FTACV harmonic characteristics of three different working electrodes: glassy carbon electrode, gold disk electrode, and boron-doped diamond electrode. “Three-electrode + FTACV + Oxidation Differential” refers to further introducing the oxidation differential characteristics of the oxidized sample relative to the unoxidized sample, based on the full signal fusion of the three electrodes. This includes the difference in current response before and after controlled oxidation of the same sample, as well as the corresponding changes in peak potential, peak current, peak area, harmonic amplitude, and phase. This transforms the differences in the content and reactivity of oxidizable components in the sample into quantifiable characteristics, amplifying the potential differences between different sources. The “three-electrode + oxidation differential + electrolysis kinetics” approach incorporates the kinetic characteristics of the controlled oxidation electrolysis process, including the oxidation integral charge, initial average current, steady-state average current, and time constant. These correspond to the overall oxidizable capacity of the sample, the instantaneous response of easily oxidizable components, the mass transfer control stage response, and the interface conversion rate, respectively, thus supplementing the explainable physical dimensions from the perspective of reaction kinetics.

[0069] From Table 1, Table 2 and Figure 4 , Figure 5 and Figure 6 The results show that: (1) Under the multidimensional fingerprint constructed in this invention, the deep learning model based on residual network combined with self-attention mechanism (ResNet-Att-1D) has the best performance among the candidate models and was selected as the final source tracing model. However, even using traditional SVM or random forest can achieve an accuracy of more than 85%, indicating that the value of this invention lies in the construction of multidimensional fingerprints rather than a specific model; (2) Ablation experiments show that the time domain ΔCV and DPV, and the frequency domain FTACV high harmonics contribute significantly and are complementary. Introducing FTACV high harmonics on the basis of time domain CV+DPV can further improve the accuracy. Introducing only the FTACV fundamental wave (which still contains a lot of capacitance) has a limited improvement, while introducing 2 to 4 high harmonics has a significant improvement, verifying the independent value of the nonlinear fingerprint information carried by the natural decapacitance of high harmonics; (3) The PCA principal component scatter plot shows that the five types of sewage aggregate in the principal component space and the inter-class separation is obvious, such as Figure 3a and 3b As shown, the confusion matrix of the blind test set shows that the correct classification percentage for each class is no less than 90%, such as... Figure 6 As shown.

[0070] It is important to note that the source tracing method based on multi-dimensional electrochemical fingerprinting provided by this invention is not limited to wastewater source tracing scenarios. It can also be extended to various liquid source tracing fields such as drinking water source pollution tracing, identification of the origin and authenticity of liquid food, verification of the source of chemical production liquids, and groundwater pollution tracing. The core technology of this method is to extract the composition, content level, and interfacial behavior characteristics of dissolved electrochemically active species in liquids through multiple sets of orthogonal electrochemical signals in the time and frequency domains, and construct a unique fingerprint. Its principle does not depend on specific pollutant components in the wastewater system. As long as the liquid to be traced contains source-specific electrochemically active substances and a stable voltammetric response signal can be collected through a three-electrode system, the same sample pretreatment, multi-dimensional signal acquisition, feature fusion, and model discrimination process can be used to complete the source tracing determination, demonstrating good scenario versatility and method portability.

[0071] To illustrate the applicability of this invention in actual river pollution source tracing scenarios, identification cases were constructed using wastewater from different sources diluted with river water. Electroplating wastewater, dyeing and printing wastewater, pharmaceutical wastewater, food processing wastewater, and domestic sewage were selected as pollution sources to be identified, and diluted with river water by approximately 10, 20, and 50 times, respectively, to simulate the actual situation where pollutant concentrations decrease and electrochemical responses are disturbed by the river water background after wastewater from different sources enters the river.

[0072] This case study employs the highest-dimensional feature combination in this invention for source identification. This feature combination includes the first-cycle cyclic voltammetry feature CV1, the second-cycle cyclic voltammetry feature CV2, the inter-cycle differential feature ΔCV, the differential pulse voltammetry feature DPV, and the 2nd to 4th harmonic Fourier transform AC voltammetry features FTACV-H obtained from three different working electrodes. It also includes the initial sample S0 before electrolytic oxidation, the sample S1 after electrolytic oxidation, and the oxidation difference feature ΔS1 between them, as well as the electrolytic oxidation charge Q1, the initial electrolytic current I0, the steady-state electrolytic current Iss, and the current decay time constant τ. By integrating information from different electrochemical methods, different working electrodes, responses before and after electrolytic oxidation, and electrolytic oxidation kinetics, a high-dimensional, multi-dimensional electrochemical fingerprint for tracing the source of diluted river wastewater is formed.

[0073] Regarding model selection, the highest-dimensional feature combination mentioned above was used as input. Support Vector Machines, Random Forests, Gradient Boosting Trees, Partial Least Squares Discriminant Analysis, K-Nearest Neighbors, One-Dimensional Convolutional Neural Networks, and deep learning models based on residual networks combined with self-attention mechanisms were compared. The ResNet-Att-1D model, which achieved the highest accuracy on the blind test set, was chosen as the final source model for this case. Therefore, the accuracies listed in Table 3 are all recognition results obtained under the condition of "highest-dimensional feature combination + optimal classification model".

[0074] Table 3

[0075] like Figure 7 As shown, based on the same inventive concept, this embodiment provides a traceability system based on multi-dimensional electrochemical fingerprinting, including: The sample processing module 201 is used to collect liquid samples after coarse filtration to remove large particulate matter. When the conductivity or ohmic drop of the liquid sample is within a first preset range, the liquid sample is diluted according to a set ratio and an inert supporting electrolyte is added so that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. The liquid sample is divided into unoxidized samples and oxidized samples of equal volume. The oxidized samples are subjected to controlled oxidation electrolysis through multiple three-electrode systems corresponding to different materials, while the unoxidized samples are not electrolyzed. The signal acquisition module 202 is used to sequentially acquire the following electrochemical signals for unoxidized and oxidized samples: continuously acquire the first and second cyclic voltammetry curves under the same scanning conditions; acquire a differential pulse voltammetry curve; acquire a single-frequency Fourier transform AC voltammetry signal, which is superimposed with a preset amplitude AC disturbance of a single frequency on a DC linear ramp, and is independently serially acquired according to multiple different frequencies to obtain current-time data at each frequency; The feature extraction module 203 is used to calculate the difference curve between the first-cycle cyclic voltammetry curve and the second-cycle cyclic voltammetry curve; to perform local baseline correction on the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve according to the potential interval, and to extract the feature parameters of each net redox peak. The feature parameters include one or more of the following: peak potential, peak current, half-maximum width, peak area, peak symmetry factor, and spacing between adjacent peaks; based on current-time data, to extract oxidation features including oxidation integral charge, initial average current, steady-state average current, and time constant; to use the difference curve, feature parameters, and oxidation features as time-domain features; and based on current-time data, to perform segmentation, windowing, and fast Fourier transform processing on each frequency track of the Fourier transform AC voltammetry signal according to a dynamic window, and to extract the amplitude and phase of the fundamental wave and the preset higher harmonics as frequency-domain harmonic features. The fingerprint database construction module 204 is used to fuse time-domain features with frequency-domain harmonic features corresponding to Fourier transform AC volt-ampere signals to construct a multi-dimensional electrochemical fingerprint database with source tags. The model selection module 205 is used to train from a variety of preset candidate classification models with features in the multi-dimensional electrochemical fingerprint database as input and liquid source category as output label, and select the model with the highest classification accuracy on the blind test set as the final source tracing model. The source tracing prediction module 206 is used to sequentially execute S101 to S104 on the liquid sample to be tested to obtain the current multi-dimensional electrochemical fingerprint database, input the multi-dimensional electrochemical fingerprint database into the final source tracing model, and output the liquid source category to which the liquid sample to be tested belongs.

[0076] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0077] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A source tracing method based on multi-dimensional electrochemical fingerprinting, characterized in that, include: S101: Collect a liquid sample after coarse filtration to remove large particulate matter. When the conductivity or ohmic drop of the liquid sample is within a first preset range, dilute the liquid sample according to a set ratio and add an inert supporting electrolyte so that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. Divide the liquid sample into unoxidized and oxidized samples of equal volume. The oxidized sample is subjected to controlled oxidation electrolysis through multiple three-electrode systems corresponding to different materials, while the unoxidized sample is not electrolyzed. S102, for the unoxidized sample and the oxidized sample, the following electrochemical signals are collected sequentially: under the same scanning conditions, the first cycle cyclic voltammetry curve and the second cycle cyclic voltammetry curve are collected continuously; a differential pulse voltammetry curve is collected; a single-frequency Fourier transform AC voltammetry signal is collected, wherein the Fourier transform AC voltammetry signal is superimposed with a single-frequency preset amplitude AC disturbance on a DC linear ramp, and is independently serially collected according to multiple different frequencies to obtain current-time data at each frequency; S103, calculate the difference curve between the first-cycle cyclic voltammetry curve and the second-cycle cyclic voltammetry curve; perform local baseline correction on the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve according to the potential range, and extract the characteristic parameters of each net redox peak. The characteristic parameters include one or more of the following: peak potential, peak current, half-maximum width, peak area, peak symmetry factor, and spacing between adjacent peaks; based on the current-time data, extract oxidation features including oxidation integral charge, initial average current, steady-state average current, and time constant; use the difference curve, the characteristic parameters, and the oxidation features as time-domain features. Based on the current-time data, each frequency track of the Fourier transform AC volt-ampere signal is segmented, windowed, and processed by fast Fourier transform according to a dynamic window to extract the amplitude and phase of the fundamental wave and the preset higher harmonics as frequency domain harmonic features. S104, the time-domain features are fused with the frequency-domain harmonic features corresponding to the Fourier transform AC volt-ampere signal to construct a multi-dimensional electrochemical fingerprint database with source tags; S105, using the features in the multi-dimensional electrochemical fingerprint database as input and the liquid source category as output label, train from a variety of preset candidate classification models, and select the model with the highest classification accuracy on the blind test set as the final source tracing model. S106, execute S101 to S104 sequentially on the liquid sample to be tested to obtain the current multi-dimensional electrochemical fingerprint database, input the multi-dimensional electrochemical fingerprint database into the final traceability model, and output the liquid source category to which the liquid sample to be tested belongs.

2. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 1, characterized in that, The liquid samples were collected at pipeline nodes or discharge outlets in the target area.

3. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 1, characterized in that, In the multiple three-electrode systems corresponding to different materials, the working electrode in any of the three-electrode systems adopts one of glassy carbon, platinum, boron-doped diamond, or titanium-based inert electrodes.

4. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 1, characterized in that, The step of independently acquiring data serially at multiple different frequencies specifically includes: A frequency set is preset, which includes multiple different frequency values ​​within a preset frequency range; For the same liquid sample, multiple complete DC scans are performed independently in sequence. During each scan, only a single frequency AC disturbance from the frequency set is superimposed on the DC ramp. The acquisition order of each frequency is randomized to eliminate system drift introduced by the measurement order.

5. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 4, characterized in that, Before performing independent serial acquisition based on multiple different frequencies, the method further includes: In the pure supporting electrolyte blank solution corresponding to the liquid sample, the solution resistance and double layer capacitance were measured by electrochemical impedance spectroscopy. Based on the solution resistance and double-layer capacitance values, the available bandwidth at each frequency is calculated, and frequencies exceeding the available bandwidth are removed from the frequency set.

6. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 1, characterized in that, The local baseline correction performed according to the potential range specifically includes: For the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve, the baseline of the original curve is fitted using an asymmetric least squares algorithm. The baseline fitting results obtained by the asymmetric least squares algorithm are then cross-validated using a sliding window polynomial fitting method. The corresponding baseline is subtracted from the original curve to obtain the correction result.

7. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 1, characterized in that, The extraction of the amplitude and phase of the fundamental wave and the preset higher harmonics specifically includes: When acquiring the phase of each higher harmonic, a dual-channel encoding method consisting of sine and cosine components is used to represent the phase value to avoid phase jumps at positive and negative boundaries; and the phase of each harmonic measured at the same frequency in a pure support electrolyte blank solution is used as a reference benchmark to perform absolute correction on the phase of each harmonic of the liquid sample.

8. The source tracing method based on multi-dimensional electrochemical fingerprinting according to claim 1, characterized in that, The candidate classification models include: support vector machine model, random forest model, gradient boosting tree model, partial least squares discriminant analysis model, K-nearest neighbor model, one-dimensional convolutional neural network model, and deep learning model based on residual network combined with self-attention mechanism.

9. A source tracing system based on multi-dimensional electrochemical fingerprinting, using the source tracing method based on multi-dimensional electrochemical fingerprinting as described in claim 1, characterized in that, include: The sample processing module is used to collect liquid samples after coarse filtration to remove large particulate matter. When the conductivity or ohmic drop of the liquid sample is within a first preset range, the liquid sample is diluted according to a set ratio and an inert supporting electrolyte is added so that the final concentration of the inert supporting electrolyte is not lower than a preset threshold. The liquid sample is divided into unoxidized samples and oxidized samples of equal volume. The oxidized samples are subjected to controlled oxidation electrolysis through multiple three-electrode systems corresponding to different materials, while the unoxidized samples are not electrolyzed. The signal acquisition module is used to sequentially acquire the following electrochemical signals for the unoxidized and oxidized samples: continuously acquire the first and second cyclic voltammetry curves under the same scanning conditions; acquire a differential pulse voltammetry curve; acquire a single-frequency Fourier transform AC voltammetry signal, wherein the Fourier transform AC voltammetry signal is superimposed with a preset amplitude AC disturbance of a single frequency on a DC linear ramp, and is independently serially acquired according to multiple different frequencies to obtain current-time data at each frequency; The feature extraction module is used to calculate the difference curve between the first-cycle cyclic voltammetry curve and the second-cycle cyclic voltammetry curve; to perform local baseline correction on the first-cycle cyclic voltammetry curve, the second-cycle cyclic voltammetry curve, and the differential pulse voltammetry curve according to the potential range, and to extract the feature parameters of each net redox peak. The feature parameters include one or more of the following: peak potential, peak current, half-maximum width, peak area, peak symmetry factor, and spacing between adjacent peaks; based on the current-time data, to extract oxidation features including oxidation integral charge, initial average current, steady-state average current, and time constant; and to use the difference curve, the feature parameters, and the oxidation features as time-domain features. Based on the current-time data, each frequency track of the Fourier transform AC volt-ampere signal is segmented, windowed, and processed by fast Fourier transform according to a dynamic window to extract the amplitude and phase of the fundamental wave and the preset higher harmonics as frequency domain harmonic features. The fingerprint database construction module is used to fuse the time-domain features with the frequency-domain harmonic features corresponding to the Fourier transform AC volt-ampere signal to construct a multi-dimensional electrochemical fingerprint database with source tags. The model selection module is used to train from a variety of preset candidate classification models with the features in the multi-dimensional electrochemical fingerprint database as input and the liquid source category as output label, and select the model with the highest classification accuracy on the blind test set as the final source tracing model. The source tracing prediction module is used to sequentially execute S101 to S104 on the liquid sample to be tested to obtain the current multi-dimensional electrochemical fingerprint database, input the multi-dimensional electrochemical fingerprint database into the final source tracing model, and output the liquid source category to which the liquid sample to be tested belongs.

Images