A prediction method for the spatial distribution of heavy metals in garbage based on hyperspectral technology

Abstract

The present invention belongs to the field of heavy metal pollution treatment in waste incineration products, and provides a method for predicting the spatial distribution of heavy metals in pit-stored waste based on hyperspectral technology, including: establishing a database of heavy metal contents in waste, a hyperspectral database of waste, the correlation between heavy metals and characteristic spectra, the association between variables and heavy metals, and the information on heavy metal contents and spatial distributions in waste; using a hyperspectral image recognition system to identify waste materials, searching for the reference range of specific heavy metal contents in the heavy metal database, comparing the inversion information, and identifying abnormal heavy metal content and distribution data. The neural network and deep learning algorithm adopted by the present invention have good adaptability to nonlinear problems, can reduce the influence brought by the possible collinearity problem between characteristic bands and reduce the uncertainty brought by variables, and can detect and predict various types of waste in waste incineration plants.

Description

Technical Field

[0001] The present invention relates to the field of heavy metal pollution treatment in waste incineration products, and relates to a method for predicting the spatial distribution of heavy metals in waste, and the present invention also includes the application of the method for predicting the spatial distribution of heavy metals in waste. Background Art

[0002] During the waste incineration process, the heavy metals contained in the waste will migrate and transform with high temperature into bottom ash, fly ash or flue gas. If the subsequent treatment is insufficient, a part of the heavy metals will be discharged into the surrounding environment with the flue gas, causing serious pollution. Under the existing technical conditions, the following problems mainly exist: it is impossible to monitor the heavy metal emissions in the flue gas in real time; at present, there is a lack of stable and reliable on-line monitoring equipment for heavy metals in the flue gas, and it is difficult to timely grasp the heavy metal concentration in the flue duct. To ensure that the emissions meet the standards, it is often necessary to add a large amount of adsorbents such as activated carbon, but due to the lack of front-end prediction, the dosage of the adsorbent is difficult to accurately control, which not only increases the cost but also poses a potential risk of exceeding the emission standards. The cost of harmless treatment of fly ash is high; a large amount of heavy metals are often enriched in fly ash, and solidification or stabilization treatment is required, and the treatment cost is often high. If the types and content levels of heavy metals in fly ash can be known in advance, it will help to select the most suitable treatment strategy according to specific elements, thereby reducing the harmless treatment expenses. There is a lack of means for warning of heavy metal exceeding the standard; the types and components of the waste in the storage pit are extremely complex, and some heavy metals may seriously exceed the standard at specific positions or local accumulations. However, the existing technology has insufficient monitoring of the heavy metal distribution of the waste in the storage pit, let alone warning of exceeding the standard. Only when tracing back in the later stage of incineration or when the emission detection exceeds the standard, it is often too late. Traditional manual detection is time-consuming and laborious; in the traditional way, it is necessary to take large-scale samples of the waste in the storage pit and send them for inspection, which cannot quickly and dynamically reflect the heavy metal content status at different positions in the storage pit; at the same time, the time-consuming and manpower investment are also very large, and it is difficult to meet the current demand for real-time or near-real-time monitoring.

[0003] The development of hyperspectral technology provides a new idea for quickly predicting the heavy metal distribution in soil and solid waste. Compared with ordinary remote sensing images, hyperspectral images have more bands and higher spectral resolution, and can not only obtain spatial information, but also capture the internal component characteristics reflected by the difference in the reflection wavelengths of substances. Since the composition of waste is significantly different from that of soil, the soil heavy metal prediction model cannot be directly copied; at the same time, the composition of the waste in the storage pit is also affected by various factors such as temperature, moisture content, and industrial / municipal waste mixing ratio, and it is difficult to ensure the accuracy using traditional linear algorithms.

[0004] In this context, if the hyperspectral scanning data of the visible area can be combined with auxiliary variables (such as local temperature, moisture content of garbage, garbage source, and the ratio of industrial / municipal solid waste, etc.), and a non-linear algorithm model is used to accurately predict the spatial distribution of heavy metals within the scannable range, the monitoring of local over-standard risks can be effectively achieved, and important references can be provided for optimizing the incineration conditions and regulating heavy metal emissions. However, for the characteristics of the garbage in the storage pit, which is complex in composition and involves multi-factor coupling, there is still a lack of mature hyperspectral prediction models and analysis methods; especially in the scenario of rapid detection limited to the "scannable area of the equipment", relevant research and reports are even more limited.

[0005] Based on this, the present invention proposes a method for predicting the spatial distribution of heavy metals in the garbage in the storage pit based on hyperspectral technology. After collecting hyperspectral data of the scannable area, the present invention obtains characteristic spectral information through spectral transformation methods, and then combines non-linear algorithms with relevant auxiliary variables to establish an inversion model of heavy metal content, helping to identify and locate abnormal heavy metals on the surface or visible area of the garbage in the storage pit, and realizing the prediction of the spatial distribution of heavy metals within the scannable area. This method not only avoids blind extrapolation of the invisible area, but also can quickly, dynamically, and macroscopically grasp the heavy metal distribution of the garbage in the storage pit, providing a scientific basis for ensuring the compliance emissions during the incineration process and subsequent treatment. Summary of the Invention

[0006] The purpose of the present invention is to assist in optimizing the incineration conditions and the operating conditions of incineration pollutant control equipment by predicting the spatial distribution and content of heavy metals in the garbage in the storage pit.

[0007] The present invention provides a method for predicting the spatial distribution of heavy metals in the garbage in the storage pit based on hyperspectral technology. Aiming at the complex composition of the garbage in the storage pit, diverse types of heavy metals, and variable contents, traditional sampling and detection are labor-intensive and cannot reflect the heavy metal distribution and content of the garbage in the storage pit in real time. Hyperspectral technology is used to predict the spatial distribution and content of heavy metals in the garbage in the storage pit before incineration, realize the monitoring of the heavy metal output during the garbage incineration process, and based on the prediction results, assist in optimizing the garbage incineration conditions and the pollution control process of the incineration products, reducing the heavy metal pollution caused by garbage incineration.

[0008] To achieve the above purpose, the solution adopted by the present invention includes the following steps:

[0009] Step 1: Sample the garbage, analyze the heavy metal content of different garbage, and establish a database of heavy metal content of garbage;

[0010] Step 2: Collect hyperspectral data of the garbage samples and establish a database of hyperspectral data of the garbage;

[0011] Step 3: Use spectral transformation methods to denoise and enhance the signal of the hyperspectral data, and establish the correlation between heavy metals and characteristic spectra;

[0012] Step 4: Using a non-linear algorithm method and combining auxiliary variables, establish the correlation between variables and heavy metals, invert the heavy metal content distribution, and establish the information on the heavy metal content and spatial distribution of the garbage.

[0013] Step 5: Use a hyperspectral image recognition system to identify the garbage materials in the storage pit, search for the reference range of specific heavy metal contents in the heavy metal database, compare the inverted information, and identify abnormal heavy metal content and distribution data.

[0014] The present invention provides a method for predicting the spatial distribution of heavy metals in storage pit garbage based on hyperspectral technology. By utilizing the advantages and characteristics of hyperspectral and combining the analysis of hyperspectral images, the spatial distribution of storage pit garbage can be predicted quickly. Compared with traditional digestion methods, more accurate and rapid results can be obtained, which is of great significance for controlling heavy metal emissions from garbage incineration.

[0015] In the present invention, the applicable garbage includes various heavy metal-containing waste materials, including but not limited to domestic waste, sludge or industrial waste, and is particularly applicable to various solid-shaped garbage.

[0016] Preferably, the sampling method in Step 1 is as follows:

[0017] 1) For garbage in a non-piled state, first convert it into a piled state and then take samples.

[0018] 2) Take samples of the piled garbage according to the grid method.

[0019] Among them, the method of sampling by the grid method is:

[0020] Heap the storage pit garbage into a square with a thickness of 40 to 60 cm, vertically connect the trisection points on the opposite sides in the length and width directions, divide the square into a grid of nine small grids in a tic-tac-toe pattern.

[0021] For each small grid, take sampling points at distances greater than or equal to 50 cm from the length and width of the small grid, take out all the garbage in the vertical direction at a distance of 40 to 60 cm from the surface of the small grid, and fully mix the garbage taken out from each small grid on a clean and dry surface.

[0022] 3) Reduce the sampled garbage according to the quartering method.

[0023] Among them, in a preferred example of the present invention, the method of reducing by the quartering method is:

[0024] Lay the evenly mixed garbage sample into a circle or a square, divide the figure into a cross grid of four parts by the symmetry axes that bisect the figure perpendicularly in pairs, randomly discard two parts, mix the remaining garbage, and repeat the quartering method until the remaining garbage amount reaches the set sampling amount.

[0025] 4) Carry out steps such as drying and digestion on the landfill waste samples, and measure the heavy metal content of the waste samples using instruments such as atomic spectrometers and ICPs.

[0026] Preferably, the auxiliary variable information includes monthly average temperature, monthly maximum temperature, waste moisture content, mixing ratio of domestic waste and industrial waste, waste source, etc.

[0027] In the present invention, the ICP instrument refers to an icp-aes analyzer (atomic emission spectrometer), which is mainly used for qualitative and quantitative analysis of inorganic elements.

[0028] In the present invention, the first-order differential method, continuous spectrum removal method, envelope line removal method, reciprocal logarithm method, and second-order differential method can adopt conventional methods and means in the art.

[0029] In the present invention, hyperspectral signal denoising spectral transformation methods such as Haar wavelet, Mexican Hat wavelet, Meyer wavelet, Sym6 wavelet, Daubechies4 wavelet, and Daubechies6 wavelet can adopt conventional techniques in the art.

[0030] In the present invention, nonlinear algorithms such as support vector machines, BP neural networks, wavelet neural networks, fuzzy neural networks, genetic gene GA algorithms, and random forest algorithms can adopt conventional methods and means in the art.

[0031] Preferably, in the method for predicting the spatial distribution of heavy metals in landfill waste based on hyperspectral technology of the present invention, the spectral transformation method may include the following steps:

[0032] 1) Implement using programming languages such as Matlab, Python, C / C++, and R by calling modules such as Wavelet and Wavelet Toolbox and the functions they contain. Wavelet transform is a mathematical method for spectral analysis, and its formula is:

[0033]

[0034] 2) Perform signal enhancement processing on the denoised spectrum, and the methods used include the first-order differential method, continuous spectrum removal method, envelope line removal method, reciprocal logarithm method, second-order differential method, etc.

[0035] 3) The method for determining the characteristic spectrum of heavy metals is: the corresponding wavelength band with the largest absolute value of the correlation coefficient at a 0.05 two-sided significance confidence level for each heavy metal monitoring index.

[0036] Preferably, for the method for predicting the spatial distribution of heavy metals in storage pit garbage based on hyperspectral technology of the present invention, the establishment of the inversion model may include the following steps:

[0037] 1) Construct a non-linear method inversion model based on the auxiliary variables and the heavy metal contents of the sampling points;

[0038] 2) Import the auxiliary variables and the spectral data of the sampling points into the constructed non-linear method inversion model to obtain the predicted heavy metal contents of the sampling points;

[0039] 3) Import the auxiliary variables and the spectral data of the non-sampling points into the constructed non-linear method inversion model to obtain the predicted heavy metal contents of the non-sampling points;

[0040] 4) Export the heavy metal content data of the sampling points and the non-sampling points to form a prediction map of the spatial distribution of heavy metals in garbage.

[0041] Preferably, in the method for predicting the spatial distribution of heavy metals in storage pit garbage based on hyperspectral technology of the present invention, the step of judging abnormal inversion data may include: comparing the obtained heavy metal inversion data with the reference range of the specific heavy metal content monitoring index in the database described in step 1. If all heavy metal characteristics exceed 10% of the database limit value, it is considered that the detection result of the sampling area is accurate; if it exceeds the reference range, it is abnormal inversion data.

[0042] A possible design of the method for predicting the spatial distribution of heavy metals in storage pit garbage based on hyperspectral technology provided by the present invention is to apply the established database and prediction model method to the real-time monitoring of storage pit garbage in a waste incineration plant, and realize the working condition control of waste incineration according to the monitoring results. Its advantages are as follows:

[0043] 1. The neural network and deep learning algorithm adopted by the present invention have good adaptability to non-linear problems, and can reduce the influence brought by the possible collinearity problem between characteristic bands;

[0044] 2. The present invention can detect and predict various types of garbage in a waste incineration plant.

[0045] 3. The present invention combines auxiliary variables and non-linear algorithms, and can reduce the uncertainty brought by variables. Description of the Drawings

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0047] Figure 1 The method flow chart of the soil heavy metal content spatial prediction method according to the embodiment of the present invention is shown. Specific implementation manners

[0048] The present invention discloses a method for predicting the spatial distribution of heavy metals in landfill waste based on hyperspectral technology. First, samples of the waste are taken, the heavy metal content of different wastes is analyzed, and a heavy metal content database of the landfill waste is established. Secondly, hyperspectral images of the waste are collected, the original hyperspectral data is preprocessed, the characteristic spectra of the heavy metals in the waste are established, and the correlation between the characteristic spectra and the heavy metal content is used. Combining with auxiliary variables, a nonlinear method is used to establish an inversion model to predict the spatial distribution of heavy metals in the waste. By comparing the obtained prediction data with the reference range of the heavy metal content of specific landfill waste in the heavy metal database, abnormal data is judged.

[0049] The present invention provides a method for predicting the spatial distribution of heavy metals in landfill waste based on hyperspectral technology, including:

[0050] Taking samples of the waste, analyzing the heavy metal content of different wastes, and establishing a heavy metal content database of the landfill waste;

[0051] Performing hyperspectral scanning on the waste to establish a hyperspectral database of the waste;

[0052] Using spectral transformation methods to denoise and enhance the signal of the hyperspectral data, and establishing the correlation between the heavy metals and the characteristic spectra;

[0053] Using a nonlinear algorithm method, combining with auxiliary variables, establishing the association between the variables and the heavy metals, inversely calculating the distribution of the heavy metal content, and establishing the heavy metal content and spatial distribution information of the landfill waste;

[0054] Using a hyperspectral image recognition system to identify the waste materials, searching for the reference range of the heavy metal content of specific heavy metals in the heavy metal database, comparing the inversion information, and identifying abnormal heavy metal content and distribution data.

[0055] Among them, the heavy metals in the garbage include one or more of Cd, As, Pb, Cr, Cu, Zn, and Ni. The garbage includes one or more of domestic garbage, sludge, or industrial waste. The enhanced spectral transformation methods of the hyperspectral include, but are not limited to, the first-order differential method, the continuum removal method, the envelope removal method, the reciprocal logarithm method, and the second-order differential method. The signal denoising spectral transformation methods of the hyperspectral include, but are not limited to, Haar wavelet, Mexican Hat wavelet, Meyer wavelet, Sym6 wavelet, Daubechies4 wavelet, Daubechies6 wavelet, etc. The non-linear algorithms include, but are not limited to, support vector machine, BP neural network, wavelet neural network, fuzzy neural network, genetic algorithm GA, random forest algorithm, etc. The auxiliary variables include: monthly average temperature, monthly maximum temperature, garbage moisture content, mixing ratio of domestic garbage and industrial waste, garbage source, etc.

[0056] The determination method of the abnormal heavy metal content includes: determining the garbage type information of the spatial site according to the image information, searching for the types and contents of heavy metals contained in the garbage of this type in the database, and judging whether the measured heavy metal content at this site exceeds the corresponding threshold in the database. If the inversion data of this spatial site exceeds or is lower than the threshold, it is judged that there is a problem with the inversion data, and a third prompt message is sent to prompt that there may be an error in the inversion data of the spatial site.

[0057] Example 1 Collection of garbage samples

[0058] 1) For garbage in a non-piled state, first convert it to a piled state and then take samples.

[0059] 2) Sampling is carried out according to the grid method.

[0060] Among them, the method of sampling by the grid method is as follows:

[0061] The garbage in the storage pit is piled into a square with a thickness of 40 to 60 cm. The trisection points on the opposite sides in the length and width directions are vertically connected, and the square is divided into a grid of nine small grids.

[0062] For each small grid, sampling points with a distance greater than or equal to 50 cm from the length and width of the small grid are taken, and all the garbage in the vertical direction with a distance of 40 to 60 cm from the surface of the small grid is taken out, and the garbage taken out from each small grid is fully mixed on a clean and dry surface.

[0063] 3) Quartering is carried out on the sampled garbage.

[0064] Among them, the method of quartering is as follows:

[0065] Lay the evenly mixed garbage samples into a circular or square shape, divide the shape into a cross grid with the symmetry axes that bisect the figure perpendicularly in pairs, a total of 4 parts, randomly discard 2 parts, mix the remaining garbage, and repeat the quartering method until the remaining garbage amount reaches the set sampling amount.

[0066] Example 2: Determination of heavy metal content in garbage

[0067] 1) Crush the large-sized items in the garbage samples brought back to the laboratory to 100 - 200 mm, then place various components in separate dry containers, put them in an electrothermal blast drying oven, and dry them at (105 ± 5) °C for 4 - 8 h. After cooling for 0.5 h, weigh them. Repeat drying for 1 - 2 h, cool for 0.5 h, and then weigh again. Repeat until the difference between the two weighings is less than one percent of the sample amount.

[0068] 2) Grind the dried samples to less than 0.5 mm with a grinder, perform microwave digestion, conduct 5 parallel experiments for each sample, dilute and make up the volume, and then use an inductively coupled plasma mass spectrometer (ICP-MS) to determine the heavy metal content. Record the obtained data.

[0069] Example 3: Establish characteristic spectra

[0070] 1) Sample collection and hyperspectral imaging

[0071] · From the garbage samples obtained in Example 1 and Example 2, select 100 representative samples: 70 of them as the training sample set, 15 as the test sample set, and 15 as the validation sample set.

[0072] · Use a hyperspectral imaging system to collect hyperspectral images of the above samples in the near-infrared spectral range (350 - 2500 nm).

[0073] · To ensure image quality, the collection should be carried out in a light-shielded or semi-light-shielded environment, and ensure that the light source is stable, the sample surface is relatively dry and flat.

[0074] 2) Hyperspectral data preprocessing

[0075] 1. Black and white correction:

[0076] Before collecting the images, perform whiteboard and blackboard corrections on the system to eliminate the influence of the sensor in aspects such as dark current and uneven illumination. The obtained original hyperspectral reflectance data is denoted as R(λ), where λ is the wavelength.

[0077] 2. Denoising processing:

[0078] Denoise the original spectrum using wavelet transform. In the example, wavelet functions such as Haar wavelet, Daubechies4 or Sym6 can be selected, and the high-frequency noise can be removed through decomposition and threshold processing; on this basis, moving smoothing or Savitzky-Golay smoothing filter can also be superimposed to suppress local noise points.

[0079] 3. Spectrum cutting and splicing:

[0080] If the device images different wavelength bands separately, the image data of multiple bands can be spliced into a complete spectral data cube; moderately truncate the edge bands with low signal-to-noise ratio (such as <380nm or >2500nm), and retain the effective spectral range, such as 400 - 2450nm.

[0081] 3) Signal enhancement and primary selection of characteristic bands

[0082] 1. Low-order differential processing:

[0083] Perform first-order and second-order differentials on the denoised reflectance spectrum. Through differential processing, part of the continuous background can be removed, and the spectral absorption characteristic peaks / valleys can be enhanced, thereby improving the recognition ability.

[0084]

[0085] 2. Primary selection of spectral characteristics:

[0086] Calculate the correlation coefficient (or other discriminant indicators) between the differential spectral data and the measured heavy metal content for each wavelength band one by one. Set the significance level, such as p<0.05 or 0.01, and screen out the wavelength bands that have a significant correlation with the target heavy metal content, which is recorded as the candidate characteristic band set.

[0087] 4) Feature selection and verification

[0088] 1. Optimization of feature subsets:

[0089] In the candidate characteristic band set, there may be multicollinearity or redundancy between the wavelength bands. Through methods such as stepwise regression, principal component analysis (PCA) or random forest importance assessment, further screen out the optimal combination of a small number of wavelength bands (such as 3 - 6 wavelength bands) to balance accuracy and simplicity.

[0090] 2. Establish regression relationship:

[0091] Based on this feature subset, perform regression analysis on the training sample set (70 samples) and the target heavy metal content. Test various models (linear, polynomial, non-linear algorithms, etc.), calculate their fitting degrees for the training samples and the prediction accuracies for the test samples (15 samples). Record the determination coefficient R of each model on the test samples. 2, indicators such as mean square error (MSE) and mean absolute error (MAE).

[0092] 3. Finally determine the characteristic bands:

[0093] After the above evaluation, the optimal band combination that meets the accuracy requirements of the present invention is obtained, such as 450nm, 970nm, 1338nm, 1938nm. A significant correlation has been established between the FDR or SDR value using this combined band and the heavy metal content. After verification of the test samples, the highest R 2 can reach about 0.9857, indicating that this combination has good accuracy in discriminating heavy metal content.

[0094] 5) Connection between characteristic spectra and subsequent inversion models

[0095] The above-selected characteristic bands (or corresponding derived indicators, such as the numerical values of FDR and SDR) constitute the main source of the "hyperspectral characteristic information" mentioned in Example 5 of the present invention. When subsequent modeling (see Example 5) is carried out, these characteristic spectral parameters and auxiliary variables (temperature, moisture content, etc.) will be used as inputs to the non-linear algorithm model together, so as to achieve more accurate inversion of heavy metal content.

[0096]

[0097] Example 4 Collect auxiliary variable information

[0098] 1) In order to study the relationship between heavy metal content and auxiliary variable information, it is also necessary to collect auxiliary variable information.

[0099] 2) Auxiliary variable information includes monthly average temperature, monthly maximum temperature, garbage moisture content, mixing ratio of domestic waste and industrial waste, garbage source, etc.

[0100] Example 5 Establish an inversion model

[0101] 1) Data preparation and construction of the input layer

[0102] 1. The heavy metal content data obtained from Example 1 and Example 2 is used as the target value of the training data.

[0103] 2. The characteristic spectral information obtained from Example 3, that is, some key bands or spectral curve parameters (such as peaks and valleys after first-order and second-order differentiation) after noise reduction and spectral enhancement processing, is used as a part of the input variables of the model.

[0104] 3. The auxiliary variable information obtained from Example 4, including monthly average temperature, monthly maximum temperature, garbage moisture content, mixing ratio of domestic waste and industrial waste, garbage source, etc., one or more indicators constitute another part of the input variables.

[0105] 4. Combine the above characteristic spectral information with the auxiliary variables to form the input feature vectors required for the training model. In actual operation, each training sample corresponds to a hyperspectral data (characteristic spectrum), a set of auxiliary variable values, and a measured heavy metal content.

[0106] 2) Selection and structure of the non - linear algorithm model

[0107] · The present invention preferably adopts a non - linear algorithm (such as one or more of BP neural network, random forest, support vector machine SVM, wavelet neural network, fuzzy neural network, genetic algorithm GA) to deal with the complex non - linear relationship in the prediction of heavy metal content in garbage.

[0108] ● Take the BP neural network as an example for illustration as follows:

[0109] 1. Input layer: It includes several input nodes for receiving hyperspectral feature information and auxiliary variable information. For example, when 4 main characteristic spectral bands + 5 auxiliary variables are selected, the number of input layer nodes is 9.

[0110] 2. Hidden layer: 1 - 3 hidden layers can be set according to the sample size and complexity. Each hidden layer contains several neurons (the number of neurons can be selected through cross - validation or empirical formula), and activation functions such as ReLU,

[0111] Sigmoid or TanH are used to map the non - linear relationship.

[0112] 3. Output layer: It corresponds to the predicted values of heavy metal content, such as Cd, As, Pb, etc. If it is necessary to predict the heavy metal content of multiple elements simultaneously, multiple neurons can be set in the output layer corresponding to each metal element, or multiple models can be trained separately for each element.

[0113] 3) Training process

[0114] 1. Take the N samples (such as 70 training samples) obtained in Example 1 and Example 2. Each sample contains:

[0115] a) Input features: Xi = [xi1, xi2, …, xim] (including hyperspectral features and auxiliary variables),

[0116] b) Target output: Yi (the measured heavy metal content of a certain element).

[0117] 2. Use algorithms such as backpropagation or stochastic gradient descent to continuously adjust the weights and biases of the hidden layer, so that the error (usually mean square error MSE or root mean square error RMSE) between the predicted value Yi of the neural network output and the measured value Yi gradually converges.

[0118] 3. Validate the trained model on 15 test samples and evaluate the model performance according to the test results (such as indicators like R 2 , RMSE, etc.). If the expected effect is not achieved, the number of neurons in the hidden layer can be appropriately increased or decreased, or hyperparameters such as the learning rate and regularization can be adjusted until a better result is obtained.

[0119] 4) Model application and inversion

[0120] 1. After training is completed, the "heavy metal content inversion model" of the present invention is obtained - this model can input "hyperspectral feature information + auxiliary variable information" and output the corresponding predicted heavy metal content values.

[0121] 2. In the application stage:

[0122] 3. The hyperspectral data at each point (or each small block) within the scannable area is first preprocessed in the same way as in step 3, and the corresponding auxiliary variables are obtained.

[0123] 4. Feed these input features into the trained non - linear algorithm model to obtain the predicted heavy metal content at that location.

[0124] 5. Map or visualize the above results on the visible surface within the storage pit range to obtain the spatial distribution information of the heavy metals in the garbage.

[0125] 4) Comparison with polynomial model or other methods

[0126] 5) To prove the advantages of the non - linear algorithm in the present invention, linear or quasi - linear methods such as the least partial squares method, linear regression model, logarithmic model, exponential model, polynomial regression model, etc. can also be used for fitting comparison. In the present invention, the non - linear algorithm can generally better handle scenarios with high dimensions, multiple interferences, and complex coupling relationships. The test results show that its R 2 value can reach above 0.98 (see the description of Example 3), while the multiple stepwise regression or simple polynomial model may only reach a fitting degree of 0.49 - 0.56.

[0127] Example 6 Judgment of abnormal areas

[0128] The obtained heavy metal inversion data is compared with the reference range of the monitoring indicators of specific heavy metal contents in the database described in Step 1. If all heavy metal characteristics exceed the database limit value by 10%, the test results of the sampling area are considered accurate; if they exceed the reference range, they are abnormal inversion data. If the heavy metal content obtained according to the regression model exceeds 10% of the historical maximum value in the database or is less than 10% of the historical minimum value in the database, it is abnormal inversion data. If abnormal inversion data appears, the typical heavy metal content of the garbage sample should be measured to verify whether there are high or low outliers.

[0129] After preliminary research, in the processing step of hyperspectral, the method for obtaining the enhanced spectrum of the hyperspectral of the present invention is transformed into a logarithmic second-order form, and the optimal R of the regression model obtained on the training set 2 May be only 0.560. In the step of establishing the inversion model, the inversion method of the present invention is transformed into a multiple stepwise regression method, that is, fitting is performed according to the following regression equation:

[0130] Y = a + b1X1 + b2X2 + b3X3 +... + b n X n

[0131] The optimal R of its inversion model 2 May be only 0.49, lower than 0.9857 of the polynomial model described in the present invention.

Claims

1. A method for predicting the heavy metal distribution of landfill waste based on hyperspectral technology, characterized in that, It includes the following steps: Sampling the landfill waste, measuring the heavy metal content in different types of waste, and establishing a database of heavy metal content in waste; Using a hyperspectral spectrometer to collect hyperspectral data from the surface area of the landfill waste that can be directly scanned, and establishing a corresponding hyperspectral database; Using spectral transformation methods to denoise and enhance the signal of the hyperspectral data to obtain characteristic spectral information of heavy metals; Obtaining auxiliary variable information related to the waste, where the auxiliary variables include one or more of monthly average temperature, monthly maximum temperature, waste moisture content, mixing ratio of domestic waste and industrial waste, and waste source; Adopting a non-linear algorithm, combining the characteristic spectral information and the auxiliary variable information to establish an inversion model for heavy metal content, specifically including: Taking the characteristic spectral information and the auxiliary variable information as the input layer of the model; Taking the heavy metal content in the waste as the output layer; Determining the hidden layer and model parameters by fitting the training set data obtained from sampling; Inputting the hyperspectral data obtained in step 2) and the corresponding auxiliary variables into the inversion model to obtain the distribution of heavy metal content at each point in the scannable area; According to the distribution result and the reference range in the heavy metal content database, identifying any spatial position that exceeds or is lower than the set threshold range, and determining it as an abnormal heavy metal content and sending a prompt message.

2. The method for predicting the heavy metal distribution of landfill waste based on hyperspectral technology according to claim 1, characterized in that, The heavy metals in the waste include one or more of Cd, As, Pb, Cr, Cu, Zn, and Ni.

3. The method for predicting the heavy metal distribution of landfill waste based on hyperspectral technology according to claim 1, wherein The waste includes one or more of domestic waste, sludge, or industrial waste.

4. The method for predicting the heavy metal distribution of landfill waste based on hyperspectral technology according to claim 1, wherein, The enhanced spectral transformation methods for the hyperspectrum include, but are not limited to, one or more of the first derivative method, continuum removal method, envelope removal method, reciprocal logarithm method, and second derivative method.

5. The method for predicting the heavy metal distribution of landfill waste based on hyperspectral technology according to claim 1, wherein The signal denoising spectral transformation methods for the hyperspectrum include, but are not limited to, one or more of Haar wavelet, Mexican Hat wavelet, Meyer wavelet, Sym6 wavelet, Daubechies4 wavelet, and Daubechies6 wavelet.

6. The method for predicting the heavy metal distribution in landfill waste based on hyperspectral technology according to claim 1, wherein, The non-linear algorithm-based methods include, but are not limited to, one or more of support vector machine, BP neural network, wavelet neural network, fuzzy neural network, genetic algorithm GA, and random forest algorithm.

7. The method for predicting the heavy metal distribution in landfill waste based on hyperspectral technology according to claim 1, where the auxiliary variables include: one or more of monthly average temperature, monthly maximum temperature, waste moisture content, mixing ratio of domestic waste and industrial waste, and waste source.

8. The method for predicting the heavy metal distribution in landfill waste based on hyperspectral technology according to claim 1, where the determination method for abnormal heavy metal content includes the following steps: Determining the waste type information of the spatial site according to the image information; Searching for the types and contents of heavy metals contained in this type of waste in the database; Judging whether the measured heavy metal content at this site exceeds the corresponding threshold in the database. If the inversion data at this spatial site exceeds or is lower than the threshold, it is judged that there is a problem with the inversion data; Sending a third prompt message to prompt that there may be an error in the inversion data of the spatial site.

9. Application of the prediction method for heavy metal distribution in landfill waste based on hyperspectral technology according to claim 1, characterized in that, The method is applied to obtain the visible area distribution of heavy metals in the storage pit garbage, so as to assist in optimizing the operation of the garbage incineration working conditions and the incineration pollutant control equipment.

10. The application according to claim 9, wherein Combined with the real-time or regular scanning and analysis of the hyperspectral image, the heavy metal distribution of the storage pit garbage in the scannable area is predicted.

Images