A comprehensive toxicant monitoring method for microbial electrochemical sensors based on machine learning models

Abstract

The present invention discloses a method for realizing simultaneous monitoring of concentrations of multiple poisons by combining a machine learning model and a microbial electrochemical system. In view of the problem that traditional microbial electrochemical sensors are difficult to identify different types of poisons, this system is composed of a microbial electrochemical sensor for detecting different types of comprehensive poisons in water and a data acquisition system. Data analysis is performed based on a machine learning regression model, and an anode electroactive biofilm sensor element for a microbial electrolytic cell configuration is constructed. The operation of the reactor and data acquisition are successfully realized through the data acquisition system. By using different algorithms and characteristic value combinations in a targeted manner, the concentrations of different types of poisons such as heavy metals, nitrites and antibiotics are simultaneously quantified. The present invention provides a solution for the wide application of microbial electrochemical sensors in water quality monitoring.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental detection, and in particular relates to a complex poison detection method based on machine learning for data analysis, a microbial electrochemical sensor for implementing the method, and a method of using the same. Background Art

[0002] With the acceleration of urbanization and industrialization, water pollution has become increasingly serious. In order to cope with the complex types of pollutants, the development of high-precision rapid water quality detection methods has become an urgent problem to be solved. Existing water quality detection methods, including physical and chemical methods, have high costs, long detection time, and the need for non-in-situ analysis in the laboratory. Therefore, on the basis of achieving high-precision and high-accuracy comprehensive toxic water quality detection, it is particularly important to develop low-cost, in-situ real-time detection technology that can be online. In recent years, microbial electrochemical toxicity sensors with microbial electrochemical systems (MES) as the core have been widely studied because they can achieve low-cost real-time monitoring of broad-spectrum toxicants. MES sensors generally use anode electroactive biofilms as sensing units. When impacted by toxicants, the metabolism of electroactive microorganisms is affected, the electron transfer rate is reduced, and the electrical signals generated by the system change accordingly. However, since electroactive biofilms respond to all toxic shocks as changes in electrical signals, it is difficult for researchers to directly obtain information about each toxicant by analyzing a single electrical signal when dealing with actual water bodies with complex pollutants. This hinders the further application of MES sensors for early warning of toxicity and water quality monitoring in actual water bodies.

[0003] Machine learning is a process of mining deep rules through data. In recent years, it has been widely used in water treatment and environmental monitoring. In the framework of machine learning, by selecting appropriate algorithms and parameters, a part of the data set is used as sample data (training set) to build a reasonable statistical model, and another part of the data set is used as a validation set to test the accuracy of the model, and finally it can make accurate predictions and decisions on the newly input data. Through regression modeling, machine learning can achieve in-depth analysis of the relationship between different types of poisons and response electrical signals, and finally realize the quantification of multiple poisons by MES sensors. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for a comprehensive poison monitoring system of a microbial electrochemical sensor based on machine learning for data analysis. A microbial electrolytic cell (MEC) sensor is constructed with an anode electroactive biofilm as a sensing element, and different machine learning algorithms and electrical signal characteristic data are combined to achieve rapid and accurate monitoring of multiple poisons in the same batch of MECs.

[0005] The technical solution of the present invention is a method of a comprehensive poison monitoring system of a microbial electrochemical sensor based on machine learning for data analysis, which specifically includes the following steps:

[0006] 1. A comprehensive poison monitoring method of a microbial electrochemical sensor based on a machine learning model, characterized by being achieved by the following steps:

[0007] 1) Processing the collected electrical signal data of the microbial electrochemical sensor before and after the toxic shock, and finally obtaining the electrical signal data of each microbial electrochemical sensor arranged in time sequence;

[0008] 2) For each reactor, by processing the experimental data and the collected electrical signal data, the characteristic values ​​of the descriptive electrical signal and the characteristic values ​​of the response to the poison impact are obtained. The characteristic values ​​of the descriptive electrical signal are the maximum current (I_max) generated during the operation of the MEC reactor, the steady-state current (I_stable) when the biofilm in the reactor is domesticated and mature, the system reaches a stable state, the time required to reach a steady state (t_stable), and the time point of injecting the poison (t_inject);

[0009] 3) Training the machine learning model, selecting a specific algorithm according to the type of poison, selecting specific descriptive electrical signal feature values ​​and response poison shock feature values ​​as input, and using the concentration of the poison as output for regression model training, using part of the processed electrical signal data set as the training set and the other part as the validation set, and using the model to predict the poison concentration of the validation set samples;

[0010] 4) Machine Learning Model Evaluation

[0011] The accuracy of the machine learning model is evaluated. When the model is accurate enough, the predicted value can be used as its actual toxicant concentration, thereby realizing the detection of comprehensive toxicant concentration.

[0012] 2. Optionally, the characteristic value of the response to the poison shock is the current and the current drop rate every 6 hours after the poison shock, that is, the current (I_6h) and the current drop rate (DropRatio_6h) for 6 hours, the current (I_12h) and the current drop rate (DropRatio_12h) for 12 hours, the current (I_18h) and the current drop rate (DropRatio_18h) for 18 hours, and the current (I_24h) ​​and the current drop rate (DropRatio_24h) ​​for 24 hours.

[0013] 3. Optionally, the maximum current (I_max) generated during the operation of the reactor is the third highest current value in the electrical signal data of the reactor to exclude possible measurement deviations and outliers caused by accidental disconnections during operation; the reactor current reaches 90% of I_max to enter a stable state, and this time point is the time required to reach a steady state (t_stable), and the average current from entering a steady state to being impacted by the poison is the steady-state current (I_stable); the reactor is impacted by the poison starting from the time of poison injection; the current drop rate (DropRatio) formula is:

[0014]

[0015] Among them, DropRatio is DropRatio_6h, DropRatio_12h, DropRatio_18h, DropRatio_24h, and correspondingly, I_drop is I_6h, I_12h, I_18h, I_24h.

[0016] 4. Optionally, the feature group used to train the manganese chloride quantitative model is I_max, I_stable, t_stable, and DropRatio_12h;

[0017] The feature group used to train the quantitative model of sodium nitrite is I_max, I_stable, t_stable, and DropRatio_24h; the feature group used to train the quantitative model of tetracycline hydrochloride is I_max, I_stable, t_stable, DropRatio_6h, and DropRatio_12h.

[0018] 5. Optionally, 90% of the data is used for the training set and 20% of the data is used for the validation set.

[0019] 6. Optionally, the machine learning models used to train manganese chloride, sodium nitrite, and tetracycline hydrochloride are partial least squares (PLS), K-nearest neighbor (KNN), and neural network (NNET), respectively.

[0020] Beneficial Effects

[0021] Compared with the prior art, the present invention combines the front-end microbial electrochemical water quality monitoring with the back-end machine learning data analysis. Through the operation of the same batch of reactors and the impact of poisons, it overcomes the shortcomings of traditional microbial electrochemical sensors that cannot simultaneously identify comprehensive poisons, and realizes the simultaneous quantification of the concentrations of multiple poisons, providing a new technology for the further application of microbial electrochemical sensors in water quality detection and toxicity warning of complex pollutant water bodies. The machine learning model is assembled from models that quantify each specific poison. These models use different algorithms and feature sets for each poison, which provides the possibility of customizing models for specific water bodies and establishing a broad-spectrum poison analysis database. The use of wireless data acquisition systems provides broad application prospects for the intelligentization and Internet of Things of microbial electrochemical online water quality monitoring and early warning systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of water quality detection using a microbial electrochemical sensor that performs data analysis based on machine learning in the present invention.

[0023] Figure 2 This is the time-current curve of the MEC reactor with a manganese chloride concentration of 9 mg / L, a sodium nitrite concentration of 9 mg / L, and a tetracycline hydrochloride concentration of 6 mg / L.

[0024] Figure 3 Schematic diagram for defining characteristic values ​​of electrical signal data

[0025] Figure 4 This is a graph showing the machine learning model's prediction results for the concentrations of three poisons in the validation set reactor. DETAILED DESCRIPTION

[0026] Example 1: Microbial electrochemical sensor based on machine learning for data analysis to simultaneously detect manganese chloride, sodium nitrite, and tetracycline hydrochloride

[0027] 1. Construction of a microbial electrolytic cell reactor with anode electroactive biofilm as the sensing element and comprehensive toxicant monitoring

[0028] 1) Construction of a microbial electrolytic cell reactor using anode electroactive biofilm as a sensing element

[0029] Electroactive biofilms are enriched by a two-electrode MEC system. The main body of the reactor is a 100 mL blue-capped reagent bottle (not limited to this size); 1.5×1.5 cm 2 The stainless steel mesh was perpendicular to the bottom of the reagent bottle as the cathode; a graphite rod with a diameter of 1 cm and a height of 1.5 cm was used as the anode, one of the bottom surfaces of which was parallel to the stainless steel mesh, and the other surfaces were covered with silicone rubber to ensure that the bottom surface was the only surface for enriching biofilm. An external voltage of 0.7 V was applied to the reactor to enrich the anode biofilm, and the biofilm status was monitored by current.

[0030] 2) Microbial electrolytic cell sensor with anode electroactive biofilm as sensing element for toxicity monitoring

[0031] like Figure 2 As shown, after all reactors reach their stable current, poisons are added to the reactors. Manganese chloride, sodium nitrite, and tetracycline hydrochloride are added to simulate the effects of comprehensive poisons including heavy metals, nitrites, and antibiotics on electroactive biofilms. In order to ensure that the electrical signal data as sample data for the machine learning model is sufficiently representative and universal, the concentration range of the three poisons is 1-10 mg / L, and the concentration of each poison in each reactor is randomly generated and not exactly the same. The three poisons are configured with 50 mmol / L phosphate buffer solution (PBS) to form a total of 1 mL of a poison mixed solution and injected into the reactor. Under the condition that the reactor pH and substrate concentration are not affected, the changes in the reactor current after being impacted by the poison are continued to be monitored. Finally, data of 23 MEC reactors and their electrical signals after being impacted by poisons of different concentrations are obtained. 2. Data processing and machine learning modeling of the collected sample data

[0032] 1) Perform data cleaning on the collected electrical signal data

[0033] The electrical signal data saved by the computer carries labels such as the sampling timestamp, the IP address of the data acquisition system, and the equipment number. These data are processed into an electrical signal data pattern for each reactor arranged in chronological order to facilitate subsequent feature value extraction and machine learning modeling.

[0034] 2) Extract feature groups and select feature values ​​for electrical signal data

[0035] The electrical signal data of 23 reactors after data sorting are used as the database. Figure 3For each MEC reactor, a series of characteristic values ​​are obtained by processing the experimental data and the collected electrical signal data: characteristic values ​​of descriptive electrical signals, including the maximum current (I_max) generated during the operation of the MEC reactor, the steady-state current (I_stable) when the biofilm in the reactor is acclimated and mature, the time required to reach the steady state (t_stable), and the time point of poison injection (t_inject); characteristic values ​​of response to poison shock, including the current 6 hours after the poison shock (Drop_6h) and the current drop rate (DropRatio_6h), the current 12 hours after the poison shock (Drop_12h) and the current drop rate (DropRatio_12h), the current 18 hours after the poison shock (Drop_18h) and the current drop rate (DropRatio_18h), the current 24 hours after the poison shock (Drop_24h) ​​and the current drop rate (DropRatio_24h), the definition and calculation method of the characteristic values ​​are the same as above.

[0036] All eigenvalue data are normalized and each eigenvalue data set is mapped to the (0, 1) interval. The normalization formula is as follows:

[0037]

[0038] Data sets of 20 reactors were randomly selected as training sets to train the machine learning model, and the other 3 were used as training sets to verify that the model can achieve concentration prediction and evaluate the prediction accuracy of the model.

[0039] The feature group used to train the manganese chloride concentration model is I_max, I_stable, t_stable, and DropRatio_12h; the feature group used to train the sodium nitrite concentration model is I_max, I_stable, t_stable, and DropRatio_24h; the feature group used to train the tetracycline hydrochloride concentration model is I_max, I_stable, t_stable, DropRatio_6h, and DropRatio_12h.

[0040] 3) Use different algorithms for machine learning modeling for different poisons

[0041] The machine learning models used to train manganese chloride, sodium nitrite, and tetracycline hydrochloride were partial least squares (PLS), K nearest neighbor (KNN), and neural network (NNET) respectively. Since the concentration of the poison is a continuous variable, a regression model was established based on the characteristic value data of the electrical signal in the training set for supervised machine learning. The characteristic value data of the electrical signal in the verification set was input into the model to obtain the concentration of each poison for predicting the continuous variable.

[0042] Partial least squares regression analysis is a regression analysis method for multiple eigenvalues ​​and multiple output values. It realizes a multivariate regression method that combines canonical correlation analysis for variable correlation analysis and principal component analysis for data simplification. By integrating and screening the information of the electrical signal adjustment data, the PLS algorithm can extract new comprehensive variables with the best explanatory power for the concentration of toxicants from the entire feature group for regression modeling. The K nearest neighbor method is to find a predetermined number (k) of points closest to the new point in distance (standard Euclidean distance) from the training electrical signal data sample, and then make predictions based on these points. Based on the feature group of electrical signal data with continuous labels, KNN can realize the training of regression models. In the regression model, the training of neural networks relies on multi-layer perceptrons (MLP). The multi-layer perceptron under the NNET framework includes an input layer, an output layer, and a nonlinear hidden layer. The first input layer consists of a group of neurons {x_i|x_1, x_2, ..., x_m} representing the input features, that is, a series of electrical signal eigenvalues ​​used for modeling; the neurons in the hidden layer perform weighted linear summation transformation w on the values ​​of the previous layer. 1 x 1 +w 2 x 2 +...+w m x m And the nonlinear activation function transformation g(·): R→R (identity transformation in regression model); finally, the output layer receives the continuous variable transformed by the hidden layer. 4) Evaluate the prediction accuracy of the machine learning model

[0043] After training the regression models for manganese chloride, sodium nitrite and tetracycline hydrochloride through the corresponding algorithms and feature groups, the feature value group of each reactor in the validation set was input into the model to obtain the predicted concentrations of different poisons, such as Figure 4 As shown in the figure, it can be seen that the concentrations predicted by the model are very different from the concentrations of the poisons actually injected, which preliminarily shows that the quantitative models of the three poisons are accurate enough.

[0044] The accuracy of the model was further evaluated by calculating the root mean square error (RMSE) and mean absolute error (MAE) of the model.

[0045] MAE measures the average of all errors in the prediction process and is calculated as follows:

[0046]

[0047] in, is the predicted value output when the feature group of system sample i is given, y i is the actual value.

[0048] RMSE measures the standard deviation of the prediction error during the prediction process and is calculated as follows:

[0049]

[0050] in, is the predicted value output when the feature group of system sample i is given, y i is the actual value.

[0051] After verification of the validation set, the MAE of the concentration quantification model of manganese chloride, sodium nitrite, and tetracycline hydrochloride can be reduced to 0.20, 0.18, and 0.26, and the RMSE can be reduced to 0.21, 0.20, and 0.23.

[0052] It is demonstrated that by using a large amount of electrical signal data for reaction toxicity monitoring and constructing different machine learning algorithms, it is possible to achieve quantitative detection of multiple poisons simultaneously on the same batch of microbial electrochemical sensors, laying the foundation for the further promotion of microbial electrochemistry.

[0053] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A comprehensive poison monitoring method for microbial electrochemical sensors based on machine learning models, characterized in that: This is accomplished by following these steps: 1) Processing the collected electrical signal data of the microbial electrochemical sensor before and after the toxic shock, and finally obtaining the electrical signal data of each microbial electrochemical sensor arranged in time sequence; 2) For each reactor, by processing the experimental data and the collected electrical signal data, the characteristic values ​​of the descriptive electrical signal and the characteristic values ​​of the response to the poison impact are obtained; the characteristic values ​​of the descriptive electrical signal are the maximum current (I_max) generated during the operation of the MEC reactor, the steady-state current (I_stable) when the biofilm in the reactor is domesticated and mature, the system reaches a stable state, the time required to reach the steady state (t_stable), and the time point of injecting the poison (t_inject); 3) Training the machine learning model, selecting a specific algorithm according to the type of poison, selecting specific descriptive electrical signal feature values ​​and response poison shock feature values ​​as input, and using the concentration of the poison as output for regression model training, using part of the processed electrical signal data set as the training set and the other part as the validation set, and using the model to predict the poison concentration of the validation set samples; 4) Machine Learning Model Evaluation The accuracy of the machine learning model is evaluated. When the model is accurate enough, the predicted value can be used as its actual toxicant concentration, thereby realizing the detection of comprehensive toxicant concentration.

2. The comprehensive poison monitoring method of a microbial electrochemical sensor based on a machine learning model according to claim 1 is characterized in that: The characteristic value of the response to the poison shock is the current and the current drop rate every 6 hours after the poison shock, that is, the current (I_6h) and the current drop rate (DropRatio_6h) for 6 hours, the current (I_12h) and the current drop rate (DropRatio_12h) for 12 hours, the current (I_18h) and the current drop rate (DropRatio_18h) for 18 hours, and the current (I_24h) ​​and the current drop rate (DropRatio_24h) ​​for 24 hours.

3. The comprehensive poison monitoring method of a microbial electrochemical sensor based on a machine learning model according to claim 1 is characterized in that: The maximum current (I_max) generated during the operation of the reactor is the third highest current value in the electrical signal data of the reactor, in order to exclude the possible measurement deviation and outliers caused by accidental disconnection during operation; the reactor current reaches 90% of I_max to enter the stable state, and this time point is the time required to reach the steady state (t_stable), and the average current from entering the steady state to being impacted by the poison is the steady-state current (I_stable); the reactor is impacted by the poison starting from the time of poison injection; the current drop rate (DropRatio) formula is: Among them, DropRatio is DropRatio_6h, DropRatio_12h, DropRatio_18h, DropRatio_24h, and correspondingly, I_drop is I_6h, I_12h, I_18h, I_24h.

4. The comprehensive poison monitoring method of a microbial electrochemical sensor based on a machine learning model according to claim 1 is characterized in that: The feature group used to train the quantitative model of manganese chloride is I_max, I_stable, t_stable, and DropRatio_12h; the feature group used to train the quantitative model of sodium nitrite is I_max, I_stable, t_stable, and DropRatio_24h; the feature group used to train the quantitative model of tetracycline hydrochloride is I_max, I_stable, t_stable, DropRatio_6h, and DropRatio_12h.

5. The comprehensive poison monitoring method of a microbial electrochemical sensor based on a machine learning model according to claim 1 is characterized in that: The machine learning models used to train manganese chloride, sodium nitrite, and tetracycline hydrochloride are Partial Least Squares (PLS), K-Nearest Neighbor (KNN), and Neural Network (NNET), respectively.

Images