Industrial water pollution tracing method based on convolutional neural network and water quality fluorescent fingerprints

Abstract

The invention belongs to the technical field of regional water environment supervision, and discloses an industrial water pollution traceability method based on a convolutional neural network and water quality fluorescence fingerprints, and the method comprises the steps: obtaining wastewater water quality fluorescence fingerprint information based on a three-dimensional fluorescence spectrum technology, and constructing a CNN pollution traceability classification model by using three-dimensional fluorescence spectrum data as input; and outputting the category of the sample, and inputting the three-dimensional fluorescence spectrum data of the unknown sample into the trained CNN pollution traceability classification model to obtain a suspected pollution source. The industrial water pollution tracing method based on the convolutional neural network and the water quality fluorescent fingerprints is developed for industrial wastewater-urban sewage compound sewage treatment plants and industrial park sewage treatment plants, has the characteristics of simplicity and convenience in operation, low cost, high efficiency and good accuracy, has a relatively high practical application value, and has a wide application prospect. The standard-exceeding emission pollution source can be quickly identified, and technical support is provided for regional water environment supervision.

Description

Technical Field

[0001] The present invention belongs to the technical field of regional water environment supervision, and relates to a method for tracing the source of industrial water pollution based on a convolutional neural network and water quality fluorescence fingerprints. Background Art

[0002] Industrial wastewater is large in volume, high in load, toxic, and difficult to treat. For industrial wastewater-urban sewage composite sewage treatment plants and industrial park sewage treatment plants, the excessive discharge of industrial wastewater is extremely likely to impact their operations, and in severe cases, will lead to the paralysis of their operations. In response to the phenomenon of excessive discharge and illegal discharge of industrial enterprises' wastewater, the traditional methods for tracing the source of water pollution adopted by regulatory authorities are usually "detecting abnormal conventional water quality parameters - collecting water samples - conducting one-by-one investigations - determining the enterprises with excessive discharge and illegal discharge". However, this method has a large workload, takes a long time, and is inefficient.

[0003] Water quality fluorescence fingerprints are fluorescence spectra representing the characteristics of water samples obtained based on the three-dimensional fluorescence spectra of water samples. Due to their morphology and intensity having a one-to-one correspondence with water samples, and high sensitivity, fast detection speed, and less sample consumption, they have good application prospects in the field of tracing the source of water pollution. They have now been widely used in the source analysis of natural water bodies such as rivers, lakes, and estuaries, but are rarely applied to the work of tracing the source of pollution at the inlet of urban sewage treatment plants or industrial park sewage treatment plants that receive industrial wastewater. Currently, the methods for extracting data characteristics from three-dimensional fluorescence spectrum data include peak seeking method, fluorescence region integration method, parallel factor analysis method, principal component analysis method, etc. These methods have many manual operation steps, take a long time, and are greatly affected by empirical values. In this context, the water quality fluorescence fingerprint pollution tracing technology based on machine learning algorithms has become a research hotspot in the field, gradually realizing the automation and intelligence of tracing the source of water pollution. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for tracing the source of industrial water pollution based on a convolutional neural network and water quality fluorescence fingerprints, which has the characteristics of simple operation, short time consumption, and high accuracy.

[0005] The technical solution of the present invention:

[0006] A method for tracing the source of industrial water pollution based on a convolutional neural network and water quality fluorescence fingerprints, comprising the following steps:

[0007] (1) Preliminary investigation: Investigate the enterprise name, affiliated industry, geographical location, main products, and main production processes of sewage discharge enterprises;

[0008] (2) Sample collection: Collect the wastewater samples at the wastewater discharge outlets of each enterprise and the mixed water samples at the total inlet of the sewage treatment plant, and number the water samples according to the sampling points;

[0009] (3) Pretreatment and Scanning of Water Samples: Before testing, filter the water samples with a 0.45μm glass fiber filter membrane; mix the wastewater samples of each enterprise with the mixed water sample at the total inlet of the sewage treatment plant in different proportions to form gradient water samples, so that the volume ratio of the enterprise's wastewater sample in the gradient water sample is 5%-60%; before performing three-dimensional fluorescence spectroscopy scanning detection, adjust the gradient water sample and the filtered water sample to UV 254 (ultraviolet-visible absorbance at a wavelength of 254nm) < 0.05cm -1 to eliminate the inner filter effect; perform on-machine scanning on the gradient water sample and the filtered water sample to obtain the three-dimensional fluorescence spectra of each water sample; use ultrapure water as a blank sample and scan its three-dimensional fluorescence spectrum;

[0010] (4) Data Pretreatment: Perform blank subtraction, Raman scattering normalization, and change negative values to 0 on the scanned three-dimensional fluorescence spectra to obtain the three-dimensional fluorescence standard spectrogram and three-dimensional fluorescence spectrum standard data of each water sample;

[0011] (5) Data Arrangement: Convert the three-dimensional fluorescence spectrum standard data of each sewage sample into a row vector corresponding to the sample, arrange them row by row in sequence, and assign the corresponding sample category label to each sample row vector, and store them in a dataset;

[0012] (6) Construction of Convolutional Neural Network Pollution Source Tracing Classification Model: Randomly divide the dataset described in step (5) into a training set and a test set according to a ratio of 8:2, introduce a neural network to construct a convolutional neural network pollution source tracing classification model, that is, a CNN classification model. This CNN classification model includes 5 parts:

[0013] Input Layer: Input the row vectors of the three-dimensional fluorescence spectrum standard data of each sewage sample into the neural network;

[0014] Convolutional Layer: Extract the features of the input data through the convolutional kernel. The more the number of convolutional layers and convolutional kernels, the more complete the data feature extraction, but it will increase the computational complexity of the model and the possibility of overfitting;

[0015] Pooling Layer: Usually located after each convolutional layer, used for data dimensionality reduction. While reducing the amount of data processing, it can also retain the original effective information, improve the training speed of the model, and to a certain extent prevent overfitting;

[0016] Fully Connected Layer: Perform non-linear weighted processing on the complex features learned previously and map them to the sample category label space vector;

[0017] Output Layer: Output the category label of the sample, usually using a support vector machine or a Softmax classifier;

[0018] (7) Classification and identification: After obtaining the three-dimensional fluorescence spectrum data of the sewage sample of an unknown enterprise, calculations are performed according to the same processing flow as above to obtain the standard data row vector of the three-dimensional fluorescence spectrum of the water sample. Subsequently, the pre-trained CNN classification model is used to predict and classify the spectral data of the sample, and finally the category label of the sample is obtained. The pollution source corresponding to this category label is the suspected pollution source enterprise.

[0019] Further, in step (3), the three-dimensional fluorescence spectrum is a fluorescence spectrogram obtained by scanning in the three-dimensional mode of a fluorescence spectrophotometer. The excitation wavelength scanning range of this fluorescence spectrogram is 220 - 450 nm, the emission wavelength scanning range is 220 - 600 nm, and the scanning interval is 5 nm for both.

[0020] Further, in step (4), the blank subtraction is to subtract the three-dimensional fluorescence raw data of the ultrapure water scanned in the corresponding batch from the three-dimensional fluorescence raw data of each sewage; the Raman scattering normalization is to calculate the ratio of the integral value of the fluorescence intensity in the range of the excitation wavelength of 350 nm and the emission wavelength of 370 - 430 nm between the three-dimensional fluorescence data of each sewage sample after blank subtraction and the three-dimensional fluorescence data of the ultrapure water; the negative value changed to 0 is to replace the negative numerical values in the three-dimensional fluorescence data of each sewage sample after Raman scattering normalization with zero values, displace the primary Rayleigh scattering data and the secondary Rayleigh scattering data, and subtract and fill the data on the primary Raman scattering line.

[0021] Further, in step (6), the training set is used to train the convolutional neural network pollution source tracing classification model; the test set is used to test the accuracy of the model; the setting principle of the sample category label is to set the gradient water samples and the original water samples of the same enterprise to the same category label, and the category labels between the wastewater samples of different enterprises are different from each other.

[0022] The construction of the CNN classification model can be appropriately adjusted according to the training situation. Multiple convolutional layers and convolutional layers with different sizes and quantities can be set. When the expected output of the training set samples is exactly the same as the actual output of the CNN classification model, it indicates that the neural network model training is successful. On this basis, when the accuracy rate of the test set reaches more than 90%, this model can be used to predict the category of unknown samples.

[0023] Advantages of the present invention:

[0024] ①The present invention develops a method for tracing industrial water pollution based on convolutional neural network and water quality fluorescence fingerprint for industrial wastewater-urban sewage composite sewage treatment plants and industrial park sewage treatment plants. All tests can be completed using a single fluorescence spectrophotometer, with few required devices, simple operation, high efficiency, and high accuracy. ②The convolutional neural network pollution tracing classification model proposed by the present invention has strong operability and low cost, and the samples participating in the training comprehensively simulate the proportion of various industrial wastewaters in the total influent of the sewage treatment plant, having great practical application value. It can quickly identify the pollution sources of excessive emissions and provide technical support for regional water environment supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the industrial water pollution tracing method provided by the present invention;

[0026] Figure 2a 、 2b 2c, 2d, and 2e are respectively the typical water quality fluorescence fingerprints of the total influent (a) of XYW Sewage Treatment Plant, the wastewater of Enterprise CD (b), Enterprise YTE (c), Enterprise ZA (d), and Enterprise BSLR (e);

[0027] Figure 3a 、 3b are respectively the effect diagrams of the training set and test set of the CNN pollution tracing classification model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The following further describes the specific embodiments of the present invention in combination with the drawings and technical solutions.

[0029] A method for tracing industrial water pollution based on convolutional neural network and water quality fluorescence fingerprint provided by the present invention includes the following steps (as Figure 1 shown):

[0030] (1) Preliminary investigation: Investigate the name, affiliated industry, geographical location, main products, and main production processes of sewage-discharging enterprises;

[0031] (2) Sample collection: Collect wastewater samples from the wastewater discharge outlets of each enterprise and mixed water samples from the total influent of the sewage treatment plant, and number the water samples according to the sampling points;

[0032] (3) Pretreatment and scanning of samples: Before testing, filter the water samples with a 0.45μm glass fiber filter membrane. Mix the wastewater of each enterprise with the water sample of the total influent of the sewage treatment plant at different ratios to form gradient water samples, so that the volume ratios of the enterprise wastewater in the gradient water samples are 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, and 60% respectively. Before performing three-dimensional fluorescence spectrum scanning detection, adjust all gradient water samples and the original water samples to UV254 (Ultraviolet-visible absorbance at a wavelength of 254 nm) < 0.05 cm -1 To eliminate the inner filter effect. All gradient water samples and the original water sample were scanned on the machine to obtain their three-dimensional fluorescence spectra. For each batch of samples scanned, ultrapure water was used as the blank sample to scan its three-dimensional fluorescence spectrum;

[0033] (4) Data preprocessing: The original data of the scanned three-dimensional fluorescence spectra was preprocessed by subtracting the blank, Raman scattering normalization, and changing negative values to 0, etc., to obtain the three-dimensional fluorescence standard spectrogram and three-dimensional fluorescence spectrum standard data of each water sample;

[0034] (5) Data arrangement: The three-dimensional fluorescence spectrum standard data of each sewage water sample was converted into a row vector corresponding to the sample, arranged in sequence by row, and the corresponding sample category label was assigned to each sample row vector, and stored in a data set;

[0035] (6) Construction of the convolutional neural network pollution source tracing classification model: The data set described in step (5) was randomly divided into a training set and a test set according to a ratio of 8:2, and a neural network was introduced to construct a convolutional neural network pollution source tracing classification model, that is, the CNN classification model. This CNN classification model includes 5 parts:

[0036] Input layer: The row vectors of the three-dimensional fluorescence spectrum standard data of each sewage water sample were passed into the neural network;

[0037] Convolutional layer: The features of the input data were extracted through the convolutional kernel. The more the number of convolutional layers and convolutional kernels, the more complete the data feature extraction, but it will increase the computational amount of the model and the possibility of overfitting;

[0038] Pooling layer: Usually located after each convolutional layer, used for data dimensionality reduction. While reducing the amount of data processing, it can also retain the original effective information, improve the training speed of the model, and to a certain extent prevent overfitting;

[0039] Fully connected layer: Non-linearly weighted process the complex features learned previously and map them to the sample category label space vector;

[0040] Output layer: Output the category label of the sample, usually using a support vector machine or a Softmax classifier;

[0041] (7) Classification and recognition: After obtaining the three-dimensional fluorescence spectrum data of the unknown enterprise sewage sample, calculations were performed according to the same processing flow as above to obtain the row vector of the three-dimensional fluorescence spectrum standard data of the water sample. Subsequently, the pre-trained CNN classification model was used to predict and classify the spectral data of the sample, and finally the category label of the sample was obtained. The pollution source corresponding to this category label is the suspected pollution source enterprise.

[0042] In step (3) of this embodiment, the three-dimensional fluorescence spectrum is a fluorescence spectrogram obtained by scanning in the three-dimensional mode of a fluorescence spectrophotometer. The excitation wavelength scanning range of this fluorescence spectrogram is 220 - 450 nm, the emission wavelength scanning range is 220 - 600 nm, and the scanning interval is 5 nm for both;

[0043] In step (4), the blank subtraction is to subtract the three-dimensional fluorescence raw data of the ultrapure water scanned in the corresponding batch from the three-dimensional fluorescence raw data of each sewage; the Raman scattering normalization is to calculate the ratio of the integral value of the fluorescence intensity in the range of the excitation wavelength of 350 nm and the emission wavelength of 370 - 430 nm between the three-dimensional fluorescence data of each sewage sample after blank subtraction and the three-dimensional fluorescence data of the ultrapure water; the negative value changed to 0 is to replace the negative numerical values in the three-dimensional fluorescence data of each sewage sample after Raman scattering normalization with zero values, displace the primary Rayleigh scattering data and the secondary Rayleigh scattering data, and subtract and fill the data on the primary Raman scattering line;

[0044] In step (6), the training set is used to train the convolutional neural network pollution source tracing classification model; the test set is used to test the accuracy of the model; the setting principle of the sample class labels is to set the gradient water samples and the original water samples of the same enterprise to the same class label, and the class labels between the wastewater samples of different enterprises are different from each other;

[0045] The construction of the CNN classification model can be appropriately adjusted according to the training situation. Multiple convolutional layers and convolutional layers with different sizes and quantities can be set. When the expected output of the training set samples is exactly the same as the actual output of the CNN classification model, it indicates that the neural network model training is successful. On this basis, when the accuracy rate of the test set reaches more than 90%, this model can be used to predict the class of unknown samples.

[0046] Example

[0047] In addition to receiving domestic sewage, XYW Sewage Treatment Plant also accepts industrial wastewater from surrounding electronic production, food production, machinery production, etc. These industrial wastewaters have the characteristics of complex pollutant components, high concentration, high toxicity, etc. Exceeding the standard discharge leads to a relatively high risk of accidents in the sewage treatment plant. Taking four enterprises, namely CD (electronic production), YTE (electronic production), ZA (food production), and BSLR (machinery production) as examples, the industrial water pollution source tracing method is described in detail.

[0048] (1) Preliminary investigation: Investigate the industries, geographical locations, main products, and main production processes of the four enterprises CD, YTE, ZA, and BSLR.

[0049] (2) Sample collection: Collect wastewater samples from the wastewater discharge outlets of the four enterprises CD, YTE, ZA, and BSLR and the mixed water sample at the total inlet of the XYW Sewage Treatment Plant, and number the water samples according to the sampling points.

[0050] (3) Pretreatment and Scanning of Samples: Before testing, filter the water samples with a 0.45-μm glass fiber filter membrane. Mix the wastewater from four enterprises, namely CD, YTE, ZA, and BSLR, with the influent water sample of XYW Wastewater Treatment Plant at different ratios to form gradient water samples, such that the volume ratios of the enterprise wastewater in the gradient water samples are 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, and 60% respectively. Before performing three-dimensional fluorescence spectroscopy scanning detection, adjust all gradient water samples and the original water sample to UV 254 (ultraviolet-visible absorbance at a wavelength of 254 nm) < 0.05 cm -1 to eliminate the inner filter effect. Perform on-machine scanning on all gradient water samples and the original water sample to obtain their three-dimensional fluorescence spectra. For each batch of samples scanned, use ultrapure water as a blank sample and scan its three-dimensional fluorescence spectrum.

[0051] The test conditions for three-dimensional fluorescence spectroscopy are as follows: the excitation wavelength scanning range is 220 - 450 nm, the emission wavelength scanning range is 220 - 600 nm, the scanning interval is 5 nm for both, the slit width is 5 nm, the scanning speed is 12,000 nm / min, and the photomultiplier tube voltage is 400 V.

[0052] (4) Data Pretreatment: Perform pretreatment on the original three-dimensional fluorescence spectroscopy data of the wastewater from XYW Wastewater Treatment Plant and the four enterprises CD, YTE, ZA, and BSLR, including blank subtraction, Raman scattering normalization, and changing negative values to 0, to obtain the three-dimensional fluorescence standard spectrograms and three-dimensional fluorescence spectroscopy standard data of each water sample. The typical three-dimensional fluorescence standard spectrograms of the wastewater from XYW Wastewater Treatment Plant and the four enterprises CD, YTE, ZA, and BSLR are as shown in Figure 2a 、 2b 、2c, 2d, and 2e.

[0053] (5) Data Arrangement: Convert the three-dimensional fluorescence spectroscopy standard data of 192 gradient water samples and the original water samples of each enterprise in 4 batches into row vectors corresponding to the samples. Arrange each sample row vector in sequence by row and store them in a dataset. Set the sample class labels corresponding to the row vectors of the CD gradient water samples and the original water sample to 1, the sample class labels corresponding to the row vectors of the YTE gradient water samples and the original water sample to 2, the sample class labels corresponding to the row vectors of the ZA gradient water samples and the original water sample to 3, and the sample class labels corresponding to the row vectors of the BSLR gradient water samples and the original water sample to 4.

[0054] (6) Construction of the convolutional neural network pollution source tracing classification model: Randomly divide the data set obtained in step (5) into a training set and a test set according to a ratio of 8:2, and introduce a neural network to construct a convolutional neural network pollution source tracing classification model, that is, a CNN classification model. Among them, the training set contains 153 sample data, and the test set contains 39 sample data. The constructed CNN classification model includes five parts:

[0055] Input layer: Input the three-dimensional fluorescence spectrum standard data row vectors of 192 water samples into the neural network;

[0056] Convolutional layer: Two convolutional layers are constructed, with a default step size of 2. The number of convolutional kernels in the first convolutional layer is 8, and the size is 2×2. The number of convolutional kernels in the second convolutional layer is 16, and the size is 2×2. To increase the non-linearity of the classification model and improve the model training speed, a Relu activation function layer is set after each convolutional layer;

[0057] Pooling layer: To reduce the amount of data processing and prevent model overfitting, a max pooling layer is set between the two convolutional layers;

[0058] Fully connected layer: Perform non-linear weighted processing on the complex features learned previously and map them to the sample class label space vector;

[0059] Output layer: Construct the output layer with a Softmax classifier, and output a class label vector containing 4 sample classes.

[0060] The effect diagrams of the training set and test set of the constructed CNN pollution source tracing classification model are as shown in Figure 3a 、 3b shown.

[0061] (7) Classification and recognition: Randomly select the wastewater of the above four enterprises for blind sample testing. After obtaining the three-dimensional fluorescence spectrum data of 10 blind samples, calculate according to the same processing flow as above to obtain the three-dimensional fluorescence spectrum standard data row vectors of the blind samples. Subsequently, use the pre-trained CNN classification model to predict and classify the spectral data of the blind samples, and finally obtain the class label of the blind samples. The pollution source corresponding to this class label is the suspected pollution source enterprise. The suspected pollution source enterprises obtained by classifying and predicting 10 blind samples are completely consistent with their corresponding actual pollution source enterprises.

[0062] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

Claims

1. A method for tracing the source of industrial water pollution based on convolutional neural network and water quality fluorescent fingerprint, characterized in that: The steps include: (1) Preliminary research: Investigate the name, industry, geographical location, main products and main production processes of wastewater discharge enterprises; (2) Sample collection: Collect wastewater samples from the wastewater outlets of each enterprise and mixed water samples from the main water inlet of the sewage treatment plant, and number the water samples according to the sampling points; (3) Pretreatment and scanning of water samples: Before testing, filter the water samples with a glass fiber filter membrane; mix the wastewater samples of each enterprise with the mixed water sample of the total water inlet of the sewage treatment plant in different proportions to form a gradient water sample, so that the volume of the enterprise's wastewater sample in the gradient water sample accounts for 5%-60%; before performing three-dimensional fluorescence spectrum scanning detection, adjust the gradient water sample and the filtered water sample to UV 254 <0.05cm -1 To eliminate the inner filter effect; the gradient water samples and filtered water samples were scanned on the machine to obtain the three-dimensional fluorescence spectrum of each water sample; ultrapure water was used as a blank sample to scan its three-dimensional fluorescence spectrum; (4) Data preprocessing: The scanned three-dimensional fluorescence spectrum is subjected to blank subtraction, Raman scattering normalization, and negative values ​​are converted to 0 to obtain the three-dimensional fluorescence standard spectrum and three-dimensional fluorescence standard data of each water sample; (5) Data collation: The three-dimensional fluorescence spectrum standard data of each sewage sample is converted into row vectors of the corresponding samples, arranged in sequence by row, and a sample category label corresponding to each sample row vector is assigned and stored in a data set; (6) Convolutional neural network pollution source tracing classification model construction: A neural network is introduced to construct a convolutional neural network pollution source tracing classification model. The data set described in step (5) is randomly divided into a training set and a test set in a ratio of 8:

2. The convolutional neural network pollution source tracing classification model is trained to finally obtain a trained convolutional neural network pollution source tracing classification model, which includes five parts: Input layer: The three-dimensional fluorescence spectrum standard data row vector of each water sample is passed into the convolutional neural network pollution tracing classification model; Convolution layer: extract the features of the three-dimensional fluorescence spectrum standard data of each water sample through the convolution kernel; Pooling layer: located after each convolutional layer, used for data dimensionality reduction; Fully connected layer: Perform nonlinear weighting on the previously learned features and map them to the sample category label space vector; Output layer: Output the category label of the sample, using support vector machine or Softmax classifier; (7) Classification and identification: After obtaining the three-dimensional fluorescence spectrum data of the wastewater sample of the unknown enterprise, the calculation is performed according to the same processing flow as above to obtain the three-dimensional fluorescence spectrum standard data row vector of the water sample; Subsequently, the pre-trained convolutional neural network pollution source tracing classification model is used to predict and classify the spectral data of the sample, and finally the category label of the sample is obtained. The pollution source corresponding to the category label is the suspected pollution source enterprise.

2. The industrial water pollution source tracing method according to claim 1, characterized in that: In step (3), the three-dimensional fluorescence spectrum is obtained by scanning in the three-dimensional mode of the fluorescence spectrophotometer, the excitation wavelength scanning range of the fluorescence spectrum is 220-450nm, the emission wavelength scanning range is 220-600nm, and the scanning interval is 5nm.

3. The industrial water pollution source tracing method according to claim 1 is characterized in that: In step (4), the blank subtraction is to subtract the three-dimensional fluorescence data of the ultrapure water scanned in the corresponding batch from the three-dimensional fluorescence data of each water sample; the Raman scattering normalization is to calculate the ratio of the three-dimensional fluorescence data of each water sample after the blank subtraction to the integral value of the fluorescence intensity in the three-dimensional fluorescence data of the ultrapure water with an excitation wavelength of 350nm and an emission wavelength of 370 to 430nm; the negative value is changed to 0 to replace the negative value in the three-dimensional fluorescence data of each water sample after Raman scattering normalization with a zero value, replace the primary Rayleigh scattering data and the secondary Rayleigh scattering data, and subtract and fill the data on the primary Raman scattering line.

4. The industrial water pollution source tracing method according to claim 1, characterized in that: In step (6), the training set is used to train the convolutional neural network pollution source tracing classification model; the test set is used to test the accuracy of the model; the principle of setting the sample category labels is to set the gradient water samples and raw water samples of the same enterprise to the same category labels, and the category labels of wastewater samples from different enterprises are different.

Images