Intelligent detection method and system for multiple pollutants in urban solid waste incineration process

Abstract

The invention provides an intelligent detection method and system for multiple pollutants in the urban solid waste incineration process, and belongs to the technical field of artificial intelligence. Inputting the input variables into the multi-module echo state network model to obtain predicted concentrations corresponding to the to-be-detected pollutants output by the multi-module echo state network model; wherein the multi-module echo state network model comprises two sub-modules, each sub-module is constructed for one pollutant, and each sub-module is provided with an independent input layer, a reserve pool and an output layer. According to the method, the multi-pollutant collaborative detection model based on the multi-module echo state network is constructed, and the concentration of various pollutants in the urban solid waste incineration process is accurately predicted through state sharing among the sub-modules.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for intelligently detecting multiple pollutants in a municipal solid waste incineration process. Background Art

[0002] With the continuous acceleration of urbanization, the amount of urban solid waste generated is increasing day by day. Urban solid waste incineration has become an important means of effectively disposing of urban solid waste due to its technical characteristics of harmlessness, reduction and resource utilization. However, in the process of urban solid waste incineration, nitrogen oxides (NO X ) and carbon monoxide (CO), which pose potential hazards to the environment.

[0003] For pollutant emission monitoring, the current detection method based on independent modeling of single pollutants is commonly used, that is, a corresponding prediction model is constructed for each pollutant separately. However, this independent modeling method ignores the coupled correlation characteristics of pollutant emissions during the incineration of municipal solid waste and fails to take into account the mutual influence between different pollutants, resulting in limited model accuracy and insufficient generalization performance. At the same time, the existing technology lacks effective data fusion strategies and model interaction mechanisms, making it difficult to accurately capture the complex dynamic characteristics and time delay characteristics of pollutant emissions. Ultimately, the prediction results of pollutant concentrations are highly biased and cannot meet the high-precision and high-stability requirements of environmental protection departments for the coordinated detection of multiple pollutants.

[0004] Therefore, how to improve the prediction accuracy of pollutants to be detected has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The present invention provides a method, system, electronic equipment and storage medium for intelligent detection of multiple pollutants in a municipal solid waste incineration process, which are used to address the defects in the prior art and achieve accurate prediction of the concentration of pollutants to be detected.

[0006] The present invention provides a method for intelligently detecting multiple pollutants in a municipal solid waste incineration process, comprising the following steps: Obtain relevant input variables from the municipal solid waste incineration process; Inputting the input variables into a multi-module echo state network model to obtain predicted concentrations corresponding to each pollutant to be detected output by the multi-module echo state network model; The multi-module echo state network model includes two sub-modules, each of which is constructed for a pollutant, and each of which has an independent input layer, a storage pool, and an output layer; The input layer is used to receive the input variables; The reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data, and the previous state of the reserve pools of other submodules introduced through the shared weight matrix; The output layer is used to output the predicted concentration of the pollutant to be detected according to the output weight matrix and the current state of each storage tank.

[0007] According to the present invention, a method for intelligently detecting multiple pollutants in a municipal solid waste incineration process further includes a method for training the multi-module echo state network model: Obtain input variable samples and use real pollutant concentration data as the target output vector for model training; Initialize the number of input layer nodes, the number of reserve pool nodes, and the number of output layer nodes of each submodule respectively, and randomly initialize the input weight matrix of each submodule according to uniform distribution; For each of the submodules, a diagonal matrix is randomly generated according to a preset singular value distribution, and a random orthogonal matrix is simultaneously generated. A reserve pool weight matrix and a shared weight matrix are further obtained by matrix singular value decomposition, and the reserve pool weight matrix and the shared weight matrix are made to satisfy the maximum singular value constraint condition; Initializing the initial state of each submodule reservoir randomly in a uniformly distributed manner, and updating the current state of each reservoir based on the input variable samples input to the corresponding submodule input layer, the previous state of its own reservoir, and the previous state of the reservoirs of other submodules introduced through shared weight matrix fusion; The current state of each sub-module reserve pool is collected and a state matrix is formed. The robust loss function with a penalty term is used as the optimization target. The target output vector is combined, and the output weight matrix of each sub-module is iteratively updated to obtain the multi-module echo state network model.

[0008] According to the intelligent detection method for multiple pollutants in a municipal solid waste incineration process provided by the present invention, the constraint condition of the maximum singular value is expressed by the following formula: ; Where, is the weight matrix of the i-th reserve pool, is the shared weight matrix, is the maximum singular value of the weight matrix of the i-th reserve pool, is the maximum singular value of the shared weight matrix.

[0009] According to a method for intelligent detection of multiple pollutants in a municipal solid waste incineration process provided by the present invention, the current state of each storage pool is updated based on the input variable samples input to the input layer of the corresponding submodule, the previous state of its own storage pool, and the previous state of the storage pools of other submodules introduced by shared weight matrix fusion, specifically including: The current status of each reserve pool is updated using the following formula: ; In the formula, t is the current time, t-1 is the previous time, is the current state of the storage pool of the i-th submodule, is the state of the reserve pool of the i-th submodule at the previous moment, is the previous state of the reserve pool of the jth other submodule, is the input weight matrix of the i-th submodule, is the input variable sample of the i-th submodule, is the reserve pool weight matrix of the i-th submodule, is the shared weight matrix, is a non-linear activation function.

[0010] According to a method for intelligent detection of multiple pollutants in a municipal solid waste incineration process provided by the present invention, the current state of each submodule storage pool is collected to form a state matrix, a robust loss function with a penalty term is used as the optimization target, the target output vector is combined, and the output weight matrix of each submodule is iteratively updated to obtain the multi-module echo state network model, including: Initialize the sample weight matrix of each submodule to the identity matrix; Calculating the updated output weight matrix of each submodule according to the state matrix, sample weight matrix, target output vector and penalty term for constraining the weight size of each submodule; Calculating the network error of each submodule according to the updated output weight matrix, state matrix and target output vector of each submodule; Calculating a scale estimate of each of the submodules according to the median of the absolute value of the network error of each of the submodules; Calculate the updated sample weight matrix using the Welsch robust loss function based on the network error and scale estimation value of each submodule; The updating step of the sample weight matrix is repeated until the maximum number of training iterations is reached or the network error meets the preset convergence condition, and the optimized output weight matrix is obtained, thereby completing the construction of the multi-module echo state network model.

[0011] According to the present invention, a method for intelligently detecting multiple pollutants in a municipal solid waste incineration process is provided, wherein the method obtains relevant input variables from the municipal solid waste incineration process, including: Collect raw data from the municipal solid waste incineration process; Using a Hampel filtering algorithm to replace outliers on the original data to obtain data after outlier replacement, wherein the optimal sliding window size of the Hampel filtering algorithm is determined by maximizing the mean absolute error of the data before and after filtering, and within the determined sliding window, identifying outliers by calculating the local median and local standard deviation of the data, and replacing the outliers with the local median; Normalizing the data after replacing the outliers using a Min-Max normalization algorithm to obtain normalized data; According to the pollutant generation mechanism of the municipal solid waste incineration process and the normalized mutual information algorithm, a process variable having a high correlation with the concentration of each pollutant to be detected is determined from the normalized data; According to the process variable, a maximum average cross-correlation function algorithm is used to analyze the time delay relationship between the process variable and the concentration of the pollutant to be detected, determine the optimal time delay parameter of the process variable, and perform time delay processing on the process variable based on the determined optimal time delay parameter to obtain the input variable.

[0012] The present invention also provides an intelligent detection system for multiple pollutants in the municipal solid waste incineration process, comprising the following modules: The first processing module is used to obtain relevant input variables from the municipal solid waste incineration process; The second processing module is used to input the input variables into the multi-module echo state network model to obtain the predicted concentrations corresponding to the pollutants to be detected output by the multi-module echo state network model; wherein the multi-module echo state network model includes two sub-modules, each of the sub-modules is constructed for a pollutant, and each of the sub-modules has an independent input layer, a reserve pool and an output layer; wherein the input layer is used to receive the input variables; the reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data and the previous state of the reserve pools of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

[0013] The present invention also provides an electronic device comprising a memory, a processor and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements any of the above-described intelligent detection methods for multiple pollutants in the urban solid waste incineration process.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for intelligent detection of multiple pollutants in a municipal solid waste incineration process as described above is implemented.

[0015] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described intelligent detection methods for multiple pollutants in a municipal solid waste incineration process.

[0016] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: By acquiring relevant input variables from the municipal solid waste incineration process and feeding them into a multi-module echo state network model, the collaborative and precise detection of multiple pollutant concentrations during the process can be achieved efficiently. The designed multi-module echo state network model consists of two submodules, each independently constructed for a specific pollutant and possessing its own input layer, reservoir, and output layer. This modular design effectively accounts for the characteristic differences between different pollutants, ensuring the accuracy of single pollutant detection while improving overall detection efficiency and system scalability. Each submodule has its own input layer to receive input variables that have undergone preprocessing and time-delay analysis, enabling accurate capture of key process variables for different pollutants and effectively reducing the negative impact of data redundancy on model predictions. Furthermore, each submodule's reservoir receives the corresponding input layer data and simultaneously fuses the previous state of its own reservoir with the previous state of the reservoirs of other submodules, introduced via a shared weight matrix. This allows each submodule to independently learn the characteristics of a single pollutant while further capturing the potential interactions between multiple pollutants, significantly improving the generalization performance and collaborative detection capabilities of the network model. At the same time, the output layer accurately outputs the concentration of the pollutants to be detected based on the output weight matrix and the current state of each reservoir. This echo state network-based structural design, through the state memory and evolution capabilities of the reservoirs, can more efficiently capture and dynamically track the complex nonlinear characteristics and temporal changes of pollutants in the municipal solid waste incineration process, thereby accurately predicting the concentration of the pollutants to be detected. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 This is one of the flow charts of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention.

[0019] Figure 2 This is the second flow chart of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention.

[0020] Figure 3 This is the third flow chart of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention.

[0021] Figure 4 This is the fourth flow chart of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention.

[0022] Figure 5 The present invention provides a NO X Test result graph.

[0023] Figure 6 The present invention provides a NO X Test error graph.

[0024] Figure 7 This is a CO test result diagram provided by the present invention.

[0025] Figure 8 This is a CO test error diagram provided by the present invention.

[0026] Figure 9 It is a structural schematic diagram of the multi-pollutant intelligent detection system for the municipal solid waste incineration process provided by the present invention.

[0027] Figure 10 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0028] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field without making creative efforts based on the embodiments of the present invention are within the scope of protection of the present invention.

[0029] It should be noted that, in the description of the present invention, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. The orientation or positional relationship indicated by the terms "upper", "lower", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the system or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0030] The terms "first," "second," and so forth, used herein are used to distinguish similar objects, not to describe a specific order or precedence. It should be understood that such terms are interchangeable where appropriate, allowing embodiments of the present invention to be implemented in an order other than that illustrated or described herein. Furthermore, the terms "first," "second," and so forth generally distinguish objects of a single type, and do not limit the number of objects. For example, the first object may be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the connected objects.

[0031] The following combination Figures 1-10 The present invention describes the intelligent detection method, system, electronic equipment and storage medium for multiple pollutants in the municipal solid waste incineration process.

[0032] Figure 1 This is one of the flow charts of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention, such as Figure 1 As shown, including but not limited to the following steps: Step 101: Obtain relevant input variables from the municipal solid waste incineration process.

[0033] In this embodiment, relevant input variables are first obtained from the municipal solid waste incineration process to provide accurate and effective support for subsequent intelligent detection of pollutant concentrations. Specifically, to accurately predict the concentrations of multiple pollutants during the municipal solid waste incineration process, it is necessary to first collect raw data generated during the municipal solid waste incineration process. This data can be obtained through the distributed control system deployed in the municipal solid waste incineration plant to ensure that the obtained data accurately reflects the actual municipal solid waste incineration process and pollutant emissions.

[0034] In one possible implementation, Figure 2 This is the second flow chart of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention, such as Figure 2 As shown, step 101 specifically includes steps 201-205: Step 201: Collecting raw data from the municipal solid waste incineration process.

[0035] In this embodiment, in order to achieve accurate and coordinated detection of multiple pollutant concentrations in the municipal solid waste incineration process, it is first necessary to accurately and reliably obtain the original data of the municipal solid waste incineration process. This is because the pollutants generated in the municipal solid waste incineration process, such as nitrogen oxides (NO X ) and carbon monoxide (CO), both of which pose varying degrees of harm to the ecological environment and human health, making real-time monitoring of their emission concentrations crucial for pollution control. Failure to accurately collect raw data from the municipal solid waste incineration process makes it difficult to effectively establish subsequent pollutant concentration prediction models and achieve ideal prediction accuracy.

[0036] Specifically, this embodiment uses a distributed control system (DCS) installed in an MSW incineration plant to collect real-time raw data from the MSW incineration process, such as flue gas oxygen content, upper and lower furnace temperatures, flue gas temperature at the waste heat boiler economizer outlet, deacidification tower inlet temperature, and primary fan inlet flow rate. This data is periodically collected at a fixed sampling interval (10 seconds in this embodiment) and stored in the DCS, ensuring that the data continuously and dynamically reflects the actual changes in the MSW incineration process and pollutant emissions.

[0037] Step 202: Use the Hampel filtering algorithm to replace outliers in the original data to obtain data after outlier replacement, wherein the optimal sliding window size of the Hampel filtering algorithm is determined by maximizing the mean absolute error of the data before and after filtering, and under the determined sliding window, the local median and local standard deviation of the data are calculated to identify outliers and replace the outliers with the local median.

[0038] In this embodiment, the raw data from the municipal solid waste incineration process is acquired in real time through a distributed control system. This data is inevitably affected by environmental interference, sensor failures, or random measurement errors, and often contains some abnormal data. Directly using this abnormal data to construct a pollutant concentration prediction model would severely interfere with the model's ability to capture the changing patterns of the real data, reducing the accuracy and reliability of the subsequent multi-module echo state network model. Therefore, it is necessary to perform outlier processing on the raw data to improve data quality and provide a highly reliable data foundation for subsequent modeling.

[0039] Specifically, in order to effectively identify and replace abnormal data, it is first necessary to determine the optimal sliding window size for the Hampel filter algorithm. In this embodiment, the optimal sliding window size is determined by maximizing the mean absolute error (MAE) of the data before and after filtering. Specifically, the mean absolute error (MAE) can be expressed by the following formula: ; in, represents the i-th original data point, represents the i-th data point after filtering, and n is the number of samples. By adjusting the sliding window size, MAE is maximized to obtain the optimal sliding window size.

[0040] After determining the optimal sliding window size, the local median and local standard deviation are further calculated for the data in each window. and local standard deviation The calculation formulas are as follows: ; ; in, is the normalization factor, which takes a value of 1.4826. It is the median of the absolute value of the difference between the data and the median, and is calculated as follows: ; Then, we use three times the local standard deviation as a threshold to identify outliers. If the absolute deviation between a data point and the local median exceeds three times the local standard deviation, then: ; The data point is judged as abnormal data, and the local median corresponding to the window where the point is located is used Replacement is performed to achieve effective identification and replacement of abnormal data.

[0041] Through the outlier processing of the above-mentioned Hampel filtering algorithm, abnormal interference in the collected data can be removed efficiently and accurately, significantly improving data quality, thereby providing reliable data guarantee for the subsequent multi-module echo state network model to accurately predict pollutant concentrations, and ultimately effectively improving the accuracy and stability of intelligent detection of pollutants in the municipal solid waste incineration process.

[0042] Step 203: The data after replacing the outliers is normalized using the Min-Max normalization algorithm to obtain normalized data.

[0043] In this example, while the raw data processed by the Hampel filter algorithm effectively replaces outliers, significant differences in dimension and scale still exist between the variables. These differences lead to inconsistent numerical ranges and scales between the variables, which can easily lead to larger-scale variables masking the characteristic information of smaller-scale variables during the subsequent training of the multi-module echo state network model. This in turn reduces the model's ability to capture the inherent patterns of the data, limiting the accuracy and generalization performance of the prediction model.

[0044] Therefore, in order to eliminate the adverse effects of this dimensional difference on model training, this embodiment further uses the Min-Max normalization algorithm to perform unified normalization processing on the data after replacing the outliers, so as to ensure that all variables are within a unified scale range.

[0045] Specifically, the Min-Max normalization algorithm is implemented by mapping all data of each variable to the interval (0, 1), thereby unifying the numerical scale of the variable. The Min-Max normalization algorithm can be implemented using the following formula: ; in, and Represent the maximum and minimum values of one-dimensional data respectively, Indicates the value after normalization of data.

[0046] Through the above-mentioned normalization processing method, the raw data of each urban solid waste incineration process are strictly mapped to the same numerical scale range, making the weight contribution between variables more balanced, avoiding the phenomenon that larger-scale variables dominate the model prediction while smaller-scale variables are ignored, thereby helping the subsequent multi-module echo state network model to capture the true change characteristics of the data more objectively and impartially.

[0047] Step 204: According to the pollutant generation mechanism of the municipal solid waste incineration process and the normalized mutual information algorithm, process variables having a high correlation with the concentration of each pollutant to be detected are determined from the normalized data.

[0048] In this example, while the raw data obtained through Min-Max normalization is uniformly scaled, the resulting data typically contains a large number of variables, not all of which are directly and significantly correlated with the concentration of the pollutant being tested. Using all of these variables directly as input for subsequent modeling would not only increase the data processing burden but also reduce the accuracy and generalization performance of the model due to interference from redundant or irrelevant variables.

[0049] Therefore, in order to effectively improve the prediction accuracy and training efficiency of the multi-module echo state network model, this embodiment further combines the intrinsic mechanism of pollutant generation with the normalized mutual information algorithm to select process variables for the normalized data, thereby determining the key process variables that are highly correlated with the concentration of the pollutant to be detected.

[0050] Specifically, based on the pollutant generation mechanism during the incineration of municipal solid waste, we preliminarily clarified the factors affecting pollutants (such as nitrogen oxides NO X and carbon monoxide (CO) are key factors in the generation and emission of NO X In terms of generation pathways, the main ones include thermal, fuel, and rapid. Emission concentrations are influenced by factors such as the incinerator's furnace temperature, excess air coefficient, solid waste composition, and denitrification agent dosage. CO generation, on the other hand, is primarily related to inadequate combustion conditions, inappropriate furnace temperature, uneven air supply, and the nature of the solid waste. Therefore, based on the pollutant generation mechanism, a preliminary screening of relevant variables was conducted.

[0051] Based on the preliminary screening of the mechanism analysis, this embodiment further uses the normalized mutual information algorithm to quantitatively analyze the primary screening variables to accurately determine the correlation strength between each variable and the concentration of the pollutant to be detected. Specifically, the normalized mutual information algorithm can quantitatively evaluate the nonlinear correlation between each process variable and the concentration of the pollutant to be detected, thereby efficiently identifying the key variables that are most closely related to the change in pollutant concentration. The calculation process of normalized mutual information (NMI) is as follows: First, let the random variable 、 The marginal probability distributions of 、 , the joint probability distribution is , then the mutual information between variables The calculation formula is: ; Then, calculate the marginal entropy of the two random variables separately and : ; ; Finally, the normalized mutual information (NMI) is defined as follows: ; By calculating the above normalized mutual information, the degree of correlation between each variable and the pollutant concentration can be intuitively quantified, thereby accurately determining the process variables that are highly relevant to the prediction of pollutant concentration and avoiding the introduction of redundant or irrelevant variables.

[0052] Step 205: Based on the process variables, a maximum average cross-correlation function algorithm is used to analyze the time delay relationship between the process variables and the concentration of the pollutant to be detected, and the optimal time delay parameters of the process variables are determined. The process variables are then subjected to time delay processing based on the determined optimal time delay parameters to obtain input variables.

[0053] In this embodiment, although key process variables with a high correlation with the concentration of the pollutants to be detected have been obtained through the aforementioned mechanism analysis and screening of the normalized mutual information algorithm, the impact of each process variable on the pollutant concentration during the municipal solid waste incineration process is usually not immediate, but there is a significant time delay. For example, the impact of changes in process variables such as the incinerator furnace temperature or air flow on the pollutant concentration often takes a certain amount of time to gradually manifest. If these time delay effects are not effectively analyzed and processed, the true relationship between each variable and the pollutant concentration cannot be accurately established during the model training process, thereby seriously reducing the prediction accuracy and practical applicability of the multi-module echo state network model.

[0054] Therefore, in order to more accurately capture the true correlation between each process variable and the pollutant concentration, this embodiment further uses the maximum average cross-correlation function algorithm to analyze the time delay relationship between each process variable and the pollutant concentration to be detected, and to clarify the optimal time delay parameters between the process variable and the pollutant concentration, so as to effectively reflect the delay characteristics in the actual physical process.

[0055] Specifically, the maximum average cross-correlation function algorithm determines the optimal time delay between process variables and pollutant concentration by calculating the cross-correlation function value between process variables and pollutant concentration variables. Assume that the variable sequences are and , the time delay between the two variable sequences is d, then the cross-correlation function between the variables is The calculation formula is as follows: ; in, and is the average of the two series, and is the standard deviation of the two series.

[0056] The delay relationship between each process variable and pollutant concentration is calculated according to the cross-correlation function value, and the time delay corresponding to the maximum absolute value of the cross-correlation function is selected as the optimal time delay parameter.

[0057] After determining the optimal delay parameters for each process variable, this embodiment further performs delay processing on each process variable based on these delay parameters, that is, determining the delay time between the process variable and the gas variable. : ; According to the determined time delay, the process variables are time-shifted accordingly, thereby effectively restoring the true impact relationship of the process variables on the pollutant concentration and forming the final input variables suitable for subsequent model construction and prediction.

[0058] Step 102: Input the input variables to the multi-module echo state network model to obtain the predicted concentrations corresponding to the pollutants to be detected output by the multi-module echo state network model; wherein the multi-module echo state network model includes two sub-modules, each sub-module is constructed for a pollutant, and each sub-module has an independent input layer, a reserve pool and an output layer; wherein the input layer is used to receive input variables; the reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data and the previous state of the reserve pool of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

[0059] In this example, after completing the collection of raw data from the municipal solid waste incineration process, outlier replacement, normalization, key variable screening, and time-delay analysis, high-quality input variables were obtained. To achieve coordinated and accurate detection of multiple pollutants, these input variables were fed into a multi-module echo state network model. Through parallel processing across multiple submodules, the concentration of each pollutant to be detected was predicted.

[0060] In the specific implementation process, the multi-module echo state network model consists of two sub-modules, each of which is independently constructed for a pollutant, and each sub-module contains an input layer, a storage pool, and an output layer. Since the pollutants emitted during the incineration of municipal solid waste may have different generation mechanisms and time delay characteristics (such as NO X The multi-module design can be used to carry out targeted modeling and prediction for different pollutants. At the same time, by introducing a shared weight matrix between each sub-module, information interaction and sharing can be achieved, thereby enhancing the overall learning efficiency and generalization ability of the network.

[0061] When input variables are fed into the multi-module echo state network model, they first pass through the independent input layers of each submodule. At this point, the submodule uses the input weight matrix to map the input variables to the reservoir space. Subsequently, the reservoir updates its current state by combining its own reservoir state at the previous moment, the data from the submodule's input layer, and the previous state of the reservoirs of other submodules introduced by the shared weight matrix. This update process fully integrates the temporal and feature information from different pollutant tasks, ensuring that as the reservoir state evolves, it not only focuses on the characteristics of the pollutants in this module but also promptly obtains feature information from other modules, thereby capturing the complex cross-relationships that affect pollutant emissions.

[0062] After completing the update of the reserve pool state, the output layer of each submodule performs a linear or nonlinear combination of the current state of the reserve pool according to the output weight matrix to obtain the predicted concentration of the corresponding pollutant.

[0063] Specifically, the formula for calculating the predicted concentration is: Where, is the predicted concentration output by the i-th submodule, is the output weight matrix of the i-th submodule, is the current state of the reserve pool of the i-th submodule.

[0064] Because the multi-module echo state network introduces the idea of multi-task learning in its network structure design, it can make full use of the similar or related dynamic characteristics that may exist between different pollutants. It not only improves the detection accuracy of a single pollutant, but also shortens the training time of the network model, greatly improving the generalization performance and stability of the overall model.

[0065] In the above manner, the input variables are input into the multi-module echo state network model in step 102, thereby obtaining the predicted concentration of each pollutant to be detected, which can significantly improve the NO concentration in the process of municipal solid waste incineration. X , CO and other key pollutant concentration prediction accuracy.

[0066] In one possible implementation, the method further includes a training method for a multi-module echo state network model. Figure 3 This is the third flow chart of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention, such as Figure 3 As shown, it specifically includes steps 301-305: Step 301: Obtain input variable samples and use the actual pollutant concentration data as the target output vector for model training.

[0067] In this example, effective supervised training of the multi-module echo state network model requires both input variable samples and their corresponding real-world pollutant concentration data. The input variable samples provide the model with key information about pollutant emissions during the municipal solid waste incineration process, while the real-world pollutant concentration data serves as the target output vector, helping the model continuously adjust network parameters during training to ensure that predictions align with actual emissions and gradually converge to the desired level.

[0068] Specifically, first, based on the data collection and processing process described in steps 101 and 102 above, high-quality input variables can be obtained after replacing outliers, normalization, variable screening, and time delay analysis. Then, in order for the multi-module echo state network model to accurately learn the emission characteristics of each pollutant during training, it is necessary to simultaneously obtain the actual pollutant concentration information corresponding to these input variables. This can usually be obtained from the actual measurement values recorded by the incineration plant's distributed control system or continuous flue gas emission monitoring system, ensuring that the measurement values correspond to the same time series interval, so that they are accurately paired with the input variables.

[0069] By simultaneously collecting input variable samples and their corresponding real pollutant concentration data, and inputting the real concentration data into the training process as the target output vector, the model can compare the predicted output with the target output in each round of training, and continuously approach the real value through error feedback and corresponding network weight adjustment. In this process, the model not only learns the temporal evolution of different pollutants in the process of municipal solid waste incineration, but also captures the characteristics of process variables that are strongly correlated with pollutant emissions. With this supervised learning method oriented to actual concentration data, the training process of the multi-module echo state network model is more targeted and effective, and can ultimately significantly improve the detection of NO X The prediction accuracy and stability of multiple key pollutants such as CO and CO provide solid data support and model foundation for the subsequent realization of multi-pollutant intelligent detection and real-time regulation.

[0070] Step 302: Initialize the number of input layer nodes, the number of reserve pool nodes, and the number of output layer nodes of each submodule respectively, and randomly initialize the input weight matrix of each submodule according to uniform distribution.

[0071] In this embodiment, to ensure that the multi-module echo state network model possesses effective nonlinear mapping capabilities and sufficient state capacity, it is necessary to first appropriately initialize the number of input layer nodes, the number of reservoir nodes, and the number of output layer nodes of each submodule. The input weight matrix of each submodule is also randomly initialized according to a uniform distribution. The purpose of initializing these network structure parameters and weights is to provide the network with diverse and unbiased initial conditions before model training begins, thereby ensuring that the network can fully learn the complex dynamic relationships between different pollutant emission characteristics during subsequent training phases.

[0072] Specifically, when determining the number of input layer nodes, it is necessary to consider the number of key process variables in the municipal solid waste incineration process to ensure that the network can receive and represent all process data that have a significant impact on pollutant emissions. For the selection of the number of reserve pool nodes, it is necessary to balance the network capacity and computational complexity: if the number of reserve pool nodes is too small, the network's ability to express time series features is insufficient; if the number of reserve pool nodes is too large, the computational burden will be increased, increasing the risk of overfitting. In the actual deployment of this embodiment, a priori experience and multiple experimental adjustments can be used to obtain a relatively appropriate reserve pool size. The number of output layer nodes is closely related to the number of pollutant types that need to be detected; since this embodiment is aimed at the collaborative detection of multiple pollutants, the output layer of each submodule usually only corresponds to the concentration prediction value of one pollutant to be detected, so the number of output layer nodes of the submodule is 1 in this embodiment.

[0073] After configuring the number of nodes in the network structure, this embodiment further randomly initializes the input weight matrix of each submodule according to a uniform distribution. This means that the connection weights between the input layer and the reservoir are randomly sampled within a preset interval (for example, [-0.5, 0.5]). This random initialization method allows the model to maintain a diverse input mapping in its initial state, helping to avoid problems such as premature parameter concentration or gradient vanishing. It also provides a richer set of initial conditions for subsequent training, allowing the network to fully learn the various state changes characteristic of the municipal solid waste incineration process.

[0074] Step 303: For each submodule, a diagonal matrix is randomly generated according to a preset singular value distribution, and a random orthogonal matrix is generated at the same time. The reserve pool weight matrix and the shared weight matrix are further obtained by matrix singular value decomposition calculation, and the reserve pool weight matrix and the shared weight matrix are made to meet the constraint condition of the maximum singular value.

[0075] In this embodiment, in order to ensure that the multi-module echo state network model can simultaneously meet the network stability and echo state properties, especially to maintain strong learning and generalization capabilities under the collaborative detection task of multiple pollutants in the municipal solid waste incineration process, it is necessary to specially design the generation process of the reserve pool weight matrix and the shared weight matrix.

[0076] Specifically, this embodiment randomly generates a diagonal matrix according to a preset singular value distribution , and simultaneously generate multiple random orthogonal matrices ,in, , N is the number of nodes in the reserve pool. Diagonal matrix The generation formula is as follows: , where the singular values Need to meet .

[0077] Then, the weight matrix of each reserve pool is obtained by matrix singular value decomposition and the shared weight matrix ,in, ; The weight matrix generated by the singular value decomposition has the same singular value distribution as the corresponding diagonal matrix. Therefore, the preset singular values can ensure that the weight matrix meets the constraint of the maximum singular value.

[0078] The reason for such matrix generation and constraints is mainly to ensure that the network will not experience unbounded growth or numerical overflow during the state evolution process, thereby satisfying the echo state properties required by the echo state network and ensuring that the current state of the reserve pool can be uniquely determined by historical inputs.

[0079] In one possible implementation, the constraint condition for the maximum singular value is expressed by the following formula: ; Where, is the weight matrix of the i-th reserve pool, is the shared weight matrix, is the maximum singular value of the weight matrix of the i-th reservoir, is the maximum singular value of the shared weight matrix.

[0080] By strictly constraining singular values in the generation stage, the possibility of numerical oscillation or divergence in the network during training can be reduced, thereby effectively improving the stability and robustness of the network in the task of dynamic pollutant prediction.

[0081] After ultimately selecting the reservoir weight matrix and shared weight matrix that meet the aforementioned singular value distribution requirements, each submodule's reservoir can achieve both independent and mutually integrated information processing during state updates: on the one hand, the reservoir retains the ability to independently learn the characteristics of a single pollutant; on the other hand, the shared weight matrix incorporates the previous state of the reservoirs of other submodules to fully capture the complex dynamic characteristics of the interconnectedness between multiple pollutants. This allows the network to not only structurally implement multi-task learning but also maintain the core properties of the echo state network at the algorithmic level, ultimately achieving higher accuracy, faster convergence, and better generalization performance in predicting multi-pollutant concentrations in municipal solid waste incineration processes.

[0082] Step 304: Randomly initialize the initial state of each sub-module reservoir in a uniformly distributed manner, and update the current state of each reservoir based on the input variable samples input to the corresponding sub-module input layer, the previous state of its own reservoir, and the previous state of the reservoirs of other sub-modules introduced through the shared weight matrix fusion.

[0083] In this embodiment, after initializing and constraining the singular values of the reserve pool weight matrix and the shared weight matrix, each submodule reserve pool is randomly initialized and its current state is updated in real time using the actual input variable samples, the previous state of its own reserve pool, and the previous state of other submodules introduced via the shared weight matrix. This is necessary because the multi-module echo state network needs to take into account the independent learning ability of each submodule for different pollutant characteristics in the collaborative prediction of multiple pollutants, while effectively integrating the time series information from other submodules to capture the potential mutual influence mechanisms during the municipal solid waste incineration process.

[0084] In its implementation, this embodiment first randomly initializes the initial state of each submodule's reservoir, for example, with a uniform distribution within the interval [0, 1] to ensure sufficient diversity in the initial state. Subsequently, the reservoir state of each submodule at time t is updated using the following formula.

[0085] In one possible implementation, the current state of each storage pool is updated using the following formula: ; In the formula, t is the current time, t-1 is the previous time, is the current state of the storage pool of the i-th submodule, is the state of the reserve pool of the i-th submodule at the previous moment, is the previous state of the reserve pool of the jth other submodule, is the input weight matrix of the i-th submodule, is the input variable sample of the i-th submodule, is the reserve pool weight matrix of the i-th submodule, is the shared weight matrix, is a non-linear activation function.

[0086] By introducing the previous state of other sub-modules into the update equation in the form of a shared weight matrix, each sub-module can retain its exclusive learning ability for its own pollutants within the reserve pool, while also fully sharing the temporal dynamic characteristics of other pollutants.

[0087] The effectiveness of this state update mechanism lies in the ability of the multi-module echo state network to capture the unique driving factors behind different pollutant emissions while also not neglecting the linkages and mutual influences between them during the actual process of municipal solid waste incineration. By retaining the independent input weights and reservoir weight matrix of each sub-module, the model can finely model the characteristics of individual pollutants. The cross-information introduced by the shared weight matrix allows for a comprehensive understanding of the potential coupling relationships between multiple pollutants. Ultimately, this significantly improves the accuracy and robustness of the multi-module echo state network in the coordinated prediction of multiple pollutants, providing a flexible and efficient state representation for subsequent output layer weight training and actual multi-pollutant concentration detection.

[0088] Step 305: Collect the current state of each sub-module reserve pool and form a state matrix. Take the robust loss function with a penalty term as the optimization target, combine it with the target output vector, and iteratively update the output weight matrix of each sub-module to obtain a multi-module echo state network model.

[0089] In this example, to further improve the accuracy and robustness of the multi-module echo state network model in predicting various pollutant concentrations, it is necessary to iteratively optimize the output weight matrices of each submodule after updating the reservoir state. To this end, this example utilizes a robust loss function with a penalty term and combines it with the target output vector to iteratively update the output weight matrices of each submodule, ultimately achieving a stable and highly accurate multi-module echo state network model.

[0090] In one possible implementation, Figure 4 This is the fourth flow chart of the intelligent detection method for multiple pollutants in the municipal solid waste incineration process provided by the present invention, such as Figure 4 As shown, step 305 specifically includes steps 401-406: Step 401: Initialize the sample weight matrix of each submodule to the identity matrix.

[0091] In this embodiment, in order to accurately measure the role played by each training sample in the model training process, it is necessary to set corresponding initial weight matrices for all training samples while introducing a robust loss function. This allows for targeted adjustment of the degree of influence of each sample on the update of the output weight matrix during subsequent iterative optimization. Specifically, step 401 first initializes the sample weight matrices of each submodule to the identity matrix. This ensures that all samples receive consistent attention during the initial training phase, while also retaining sufficient flexibility for subsequent iterative processes.

[0092] In actual operation, each submodule maintains a diagonal matrix corresponding to the number of samples, and the diagonal elements are used to represent the relative importance of each sample in the iterative calculation. This diagonal matrix is denoted as , initialization time Here, I is the identity matrix, meaning that in the initial stage, the weight of each training example is set to 1. When the model is not familiar with the training data, this indiscriminate treatment can fully reflect the overall distribution characteristics of the training set, providing a benchmark for the subsequent robust loss function to identify and process abnormal samples.

[0093] It is worth emphasizing that this initialization strategy lays an important foundation for the subsequent Welsch robust loss function to work: during the iterative update process, if certain sample points have large errors, the model will gradually reduce their influence on the training process, preventing large-error samples from causing excessive perturbations when updating the output weight matrix, thereby significantly improving the network's robustness to abnormal data. At the same time, for samples with smaller errors and close to the main distribution area, higher weights will still be retained, allowing the network to focus more on the main characteristic patterns and pollutant emission patterns during the municipal solid waste incineration process, ultimately achieving more stable and accurate prediction capabilities for multiple pollutant concentrations.

[0094] Step 402: Calculate the updated output weight matrix of each submodule based on the state matrix, sample weight matrix, target output vector, and penalty term for constraining the weight size of each submodule.

[0095] In this embodiment, after collecting the reservoir state of each submodule, it is necessary to calculate the updated output weight matrix of each submodule based on the state matrix, sample weight matrix, target output vector, and penalty term used to constrain the weight size. The purpose of this is to ensure that the network can fully utilize the internal state of the reservoir that has been collected to characterize the key characteristics of different pollutants in the municipal solid waste incineration process, and on the other hand, The penalty term effectively suppresses the unconstrained growth of network weights during training, thereby maintaining the generalization ability of the model and reducing the risk of overfitting.

[0096] In the specific implementation, each submodule first records its reservoir state matrix as , the target output vector is recorded as , and the sample weight matrix that has been initialized or iteratively updated is recorded as The updated output weight matrix can be obtained using the following formula: : ; in, For the i-th submodule The above formula shows that while minimizing the prediction error, the weight constraint (i.e., the penalty term) is also incorporated into the overall optimization objective to prevent the output weight from growing uncontrollably due to a small number of abnormal samples or training errors.

[0097] By in the state matrix and the target output vector Introduce the sample weight matrix , this embodiment can not only highlight the focus on key training samples, but also dynamically adjust the contribution of each sample to network training in subsequent iterations of the Welsch robust loss function. In this way, when encountering samples that are obviously inconsistent with the pollutant emission characteristics or have large measurement errors, the model will gradually reduce its impact on the weight update process, thereby further enhancing the network's robustness to abnormal data. Finally, the output weight matrix update strategy with a penalty term can enable the multi-module echo state network model to have a high sensitivity to the overall data distribution when learning the multi-pollutant emission patterns of the municipal solid waste incineration process, while also maintaining appropriate constraints to avoid overfitting, ultimately achieving a balance between prediction accuracy and stability.

[0098] Step 403: Calculate the network error of each submodule based on the updated output weight matrix, state matrix and target output vector of each submodule.

[0099] In this embodiment, in order to enable the multi-module echo state network model to more accurately perceive the deviations and deficiencies in the training process, it is necessary to recalculate the network error of each submodule after updating the output weight matrix, so as to evaluate the gap between the network prediction result and the actual pollutant concentration. Specifically, when the output weight matrix of each submodule is After the update is completed, the status matrix of the submodule can be used And the corresponding target output vector , the network error is calculated by the following formula : ; in, is the network error vector of the i-th submodule; represents the state matrix of the storage pool collected by the i-th submodule, which can fully describe the dynamic characteristics of the pollutant emissions during the incineration of municipal solid waste; It is the actual pollutant concentration data corresponding to the prediction target of this sub-module.

[0100] The reason for the above network error calculation is to be able to combine the robust loss function mechanism in subsequent iterations to perform weighted correction on abnormal samples or data with large noise. The model can clearly identify which sample points differ significantly from the predicted results and dynamically adjust the weights of these samples in the next iteration, thereby reducing their interference in updating the output weight matrix. This ensures that the network remains highly sensitive to the majority of reasonable samples while effectively suppressing training fluctuations caused by extreme outliers. Ultimately, through continuous iteration, the prediction accuracy for multiple pollutants and the overall robustness of the model are significantly improved.

[0101] Step 404: Calculate the scale estimation value of each submodule according to the median of the absolute value of the network error of each submodule.

[0102] In this embodiment, in order to fully measure the overall distribution characteristics of the network error after each output weight matrix update, it is necessary to further estimate the error scale of each submodule after calculating the network error. This provides a basis for determining the degree of abnormality when subsequently updating the sample weight matrix. Specifically, step 404 first performs absolute value processing on the network error vector of each submodule at the current iteration time and takes its median as the basis for a robust estimate of the overall error distribution of the submodule. To make this estimate closer to the standard deviation under a normal distribution, in actual operation, this median is also divided by 0.6745 to form the so-called "scale estimate."

[0103] In the specific implementation, the scale estimation value of the i-th submodule can be expressed by the following formula The calculation process: ; in, Represents the network error vector obtained by this submodule after this iterative update, It is the absolute value processing of the error corresponding to each sample. is the median operation. It can avoid the distortion of the error mean value caused by extreme values or a small number of abnormal samples to a certain extent, and better characterize the overall error distribution characteristics of the submodule at the current moment.

[0104] The purpose of estimating the scale of the network error is to provide a dynamic threshold for the error level of each sample in the subsequent calculation of the Welsch robust loss function. If certain samples appear too extreme relative to the overall scale estimate, their weights can be reduced in the next update; otherwise, their weights can be maintained or moderately increased. Through this mechanism, the multi-module echo state network model can better balance the overall data distribution and local anomalies in the prediction of multi-pollutant concentrations in municipal solid waste incineration, thereby achieving a precise and robust training process, ultimately effectively improving the model's accuracy and generalization performance.

[0105] Step 405: Calculate the updated sample weight matrix using the Welsch robust loss function based on the network error and scale estimation value of each submodule.

[0106] In this embodiment, to further enhance the multi-module echo state network model's ability to suppress abnormal samples or extreme errors, it is necessary to update the sample weight matrix using the Welsch robust loss function after obtaining the network error and scale estimates for each submodule. This reduces the training process's sensitivity to large errors and maintains a higher focus on samples with normal errors. The Welsch robust loss function was chosen primarily because it can smoothly process large deviations while maintaining relatively stable discrimination for small deviations, helping to improve the overall robustness and prediction accuracy of the network.

[0107] In the specific implementation, this embodiment calculates the network error of each submodule obtained in step 404 The corresponding scale estimate Combined with the weight function in the Welsch function, the sample weight matrix The diagonal elements of are adjusted one by one. Specifically, if represents the error of the i-th submodule on the n-th sample, and the corresponding sample weight can be updated by the following formula: ; ; in, is the scale factor, is the scale factor adjustment parameter, is the default scale factor constant, which is 2.985 in this embodiment; is the scale estimation value calculated in step 404. According to this formula, when Relative to When it is too big, It will be significantly reduced, which means that the model will gradually weaken the influence of such samples on the update of the output weight matrix in subsequent iterations; for samples with small errors or relatively reasonable, the weight level will be maintained at a higher level, so that the model will continue to focus on the main pollutant emission characteristics.

[0108] The above-mentioned process of updating the sample weight matrix is equivalent to setting a dynamic "trust" coefficient for each training sample. This not only suppresses the disturbance of abnormal data during training, but also comprehensively improves the network's ability to characterize the patterns of real pollutant emissions while ensuring sufficient learning of normal data. Finally, by adjusting for large error samples using the Welsch robust loss function, this embodiment enables the multi-module echo state network model to better balance model accuracy and robustness when dealing with complex municipal solid waste incineration emission data, fully exploring the potential coupling relationships between different pollutants and providing a more stable and efficient modeling foundation for subsequent predictions.

[0109] Step 406: Repeat the sample weight matrix update step until the maximum number of training iterations is reached or the network error meets the preset convergence condition, and obtain the optimized output weight matrix, thereby completing the construction of the multi-module echo state network model.

[0110] In this embodiment, to achieve stable convergence of the multi-module echo state network model during training, the aforementioned sample weight matrix update step is repeated until the maximum number of training iterations is reached or the network error meets the preset convergence condition. The reason for this is that the output weight matrix update and sample weight adjustment in the initial training phase are often only a rough correction and are insufficient to fully explore the multi-pollutant emission patterns during the municipal solid waste incineration process. Continuous iteration allows the model to continuously correct the weight allocation of noise samples and conduct in-depth learning based on the common patterns of most samples, thereby obtaining more accurate and robust model results as it gradually approaches the optimal solution.

[0111] In specific implementation, this embodiment will execute the process from step 401 to step 405 in each round of training, that is, after initializing or adjusting the sample weight matrix, the output weight matrix is calculated, and then the network error is evaluated, and the scale estimate is used to guide the further correction of the sample weight. If at the end of a certain round of iteration, the network error is less than the pre-set threshold, it is considered that the network has learned the data to an ideal level, and the convergence gain of subsequent iterations is limited, and the training can be terminated; if multiple consecutive rounds of iterations are still insufficient to reduce the error below the threshold, but the maximum number of training times has been reached, the training is stopped based on the current optimal weight output, and the final output weight matrix is saved. Through this cyclical iterative process, the network can not only balance the influence of the main features and outliers in different training samples, but also significantly reduce the cumulative error of multi-pollutant prediction in an overall sense. Ultimately, as the network error gradually converges, this embodiment successfully obtains a multi-module echo state network model with high generalization ability and prediction accuracy, which can quickly and accurately detect NO in the actual scenario of municipal solid waste incineration. X The emission concentrations of various key pollutants such as CO and CO provide reliable support for scientific decision-making on subsequent environmental management and control measures.

[0112] Furthermore, this embodiment uses historical data stored in the distributed control system of a municipal solid waste incineration plant in Beijing to verify and analyze the proposed intelligent detection method for multiple pollutants in the municipal solid waste incineration process. The details are as follows: First, 1,700 sets of measured data were selected from a distributed control system, with a sampling interval of 10 seconds. The first 1,000 sets (approximately 60% of the total data) were used as training samples, and the last 700 sets (approximately 40% of the total data) were used as test samples. This data segmentation helps fully evaluate the generalization ability and prediction accuracy of the multi-module echo state network in detecting different pollutant concentrations.

[0113] In the data preprocessing stage, in order to deal with the sensor anomalies or noise interference that inevitably occur during the original data collection process, this embodiment uses the Hampel filtering algorithm to detect outliers and replaces the data points that are judged to be abnormal with their local medians, ensuring that the data used for subsequent modeling is relatively clean and reliable. Based on the pollutant generation mechanism in the process of municipal solid waste incineration and the normalized mutual information algorithm, the pollutants related to NO are further screened out. X , CO and other key pollutant concentrations have a high correlation with the process variables, and the NO X concentrations with the variables shown in Table 1 as NO X The input of the submodule is the CO concentration at the previous moment and the variables shown in Table 2, which greatly reduces the potential interference of redundant variables on the prediction accuracy of the model.

[0114] Table 1 is the variable description table of submodule 1 (NO X ):

[0115] Table 2 is the variable description table of submodule 2 (CO):

[0116] To further capture the time delay effect of each variable on the pollutant concentration, this embodiment uses the maximum average cross-correlation function algorithm to perform time delay analysis on the above process variables, and obtains the time delay analysis results shown in Table 3. According to the time delay analysis, it can be seen that the lead time of the impact of some variables on the pollutant concentration can reach 120 seconds, while the time delay of other variables is relatively short (such as 60 seconds, 80 seconds, 100 seconds, 110 seconds, etc.). This fully demonstrates that in the process of municipal solid waste incineration, each process variable affects NO X The reaction speeds of CO concentration are different, and corresponding time delay analysis and processing are required to ensure the rationality and accuracy of the model input.

[0117] Table 3 is the delay analysis result table:

[0118] Based on the above data, this example uses the first 1000 samples (about 60%) to construct NO X The remaining 700 samples (about 40%) are used to test the generalization ability and prediction accuracy of the network in actual application scenarios. To ensure that the multi-module echo state network conforms to the echo state attribute, this embodiment configures the weight matrix of each submodule reserve pool when configuring the network. 、 The maximum singular value interval is set between 0.33 and 0.77, and the shared weight matrix The maximum singular value of is controlled within the range of 0.11 to 0.22. This can effectively constrain the stability of the network internal state update. In addition, in order to enhance the robustness of the network to noisy data and suppress overfitting, this embodiment introduces the output weight matrix learning Penalty 、 , both are set to 0.01; and the scale factor adjustment parameter 、 It is set to 0.67 and combined with the Welsch robust loss function to further improve the suppression effect of abnormal samples.

[0119] After completing the construction of the multi-module echo state network, this embodiment uses the network to XThe results show that the network has high prediction accuracy and good generalization ability on both training and test samples. Figure 5 and Figure 6 , Figure 5 The present invention provides a NO X Test result graph; Figure 6 The present invention provides a NO X Test error graph. Figure 5 and Figure 6 Shown are NO X The prediction results and error distribution in the test phase, the horizontal axis is NO X Number of test samples, vertical axis is NO X Concentration (unit: mg / m³), it can be seen that the predicted curve is relatively close to the actual measured value, and the error is maintained in a reasonable range at most times. Figure 7 and Figure 8 , Figure 7 This is a CO test result diagram provided by the present invention; Figure 8 This is a CO test error diagram provided by the present invention. Figure 7 and Figure 8 The figure shows the prediction results and error distribution of CO in the testing phase. The horizontal axis is the number of CO test samples, and the vertical axis is the CO concentration (unit: mg / m³), which also shows good prediction consistency.

[0120] To quantitatively evaluate the prediction accuracy, this example uses three indicators: root mean square error (RMSE), normalized root mean square error (NRMSE), and coefficient of determination (R²) to evaluate model performance. The calculation formula is as follows: ; ; ; in, The network is The output value at time The network is The expected output value at time t, is the average of all sample expected output values, is the number of test samples. According to the calculation results, the multi-module echo state network model proposed in this embodiment is X As shown in Table 4, the network has achieved excellent performance in terms of modeling efficiency, network structure and detection accuracy from the comprehensive evaluation of indicators such as RMSE, NRMSE and R², and can detect NO in the process of municipal solid waste incineration. X Accurate and rapid detection of CO concentration.

[0121] Table 4 is the test results:

[0122] In summary, through the comprehensive application of a series of methods such as data preprocessing, key process variable screening, delay analysis and multi-module echo state network structure design in this embodiment, the network is X It demonstrates high efficiency, high precision, and stability in the tasks of predicting CO2 concentrations, and can support the coordinated detection and real-time monitoring of the concentrations of multiple pollutants in the process of municipal solid waste incineration, and has good application potential and promotion value.

[0123] Reference Figure 9 , Figure 9 This is a schematic diagram of the structure of the intelligent detection system for multiple pollutants in the municipal solid waste incineration process provided by the present invention. The system includes: The first processing module is used to obtain relevant input variables from the municipal solid waste incineration process; The second processing module is used to input the input variables into the multi-module echo state network model to obtain the predicted concentration corresponding to each pollutant to be detected output by the multi-module echo state network model; The multi-module echo state network model includes two sub-modules, each of which is constructed for a pollutant, and each sub-module has an independent input layer, storage pool and output layer; The input layer is used to receive input variables; The reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data, and the previous state of the reserve pools of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

[0124] In one possible embodiment, the system further includes a training module for: obtaining input variable samples and using actual pollutant concentration data as target output vectors for model training; Initialize the number of input layer nodes, reserve pool nodes, and output layer nodes of each submodule separately, and randomly initialize the input weight matrix of each submodule according to uniform distribution; For each submodule, a diagonal matrix is randomly generated according to the preset singular value distribution, and a random orthogonal matrix is also generated. The reserve pool weight matrix and the shared weight matrix are further obtained through matrix singular value decomposition, and the reserve pool weight matrix and the shared weight matrix are made to meet the maximum singular value constraint. Initialize the initial state of each submodule's reservoir randomly in a uniformly distributed manner, and update the current state of each reservoir based on the input variable samples input to the corresponding submodule's input layer, the previous state of its own reservoir, and the previous state of the reservoirs of other submodules introduced through shared weight matrix fusion; The current state of each sub-module reserve pool is collected to form a state matrix. The robust loss function with a penalty term is used as the optimization objective. Combined with the target output vector, the output weight matrix of each sub-module is iteratively updated to obtain a multi-module echo state network model.

[0125] In a possible implementation, the training module is further configured to update the current state of each storage pool using the following formula: ; In the formula, t is the current time, t-1 is the previous time, is the current state of the storage pool of the i-th submodule, is the state of the reserve pool of the i-th submodule at the previous moment, is the previous state of the reserve pool of the jth other submodule, is the input weight matrix of the i-th submodule, is the input variable sample of the i-th submodule, is the reserve pool weight matrix of the i-th submodule, is the shared weight matrix, is a non-linear activation function.

[0126] In a possible implementation, the training module is further configured to: Initialize the sample weight matrix of each submodule to the identity matrix; Calculate the updated output weight matrix of each submodule based on the state matrix, sample weight matrix, target output vector, and penalty term used to constrain the weight size of each submodule; Calculate the network error of each submodule based on the updated output weight matrix, state matrix and target output vector of each submodule; Calculate the scale estimate of each submodule based on the median of the absolute value of the network error of each submodule; According to the network error and scale estimation value of each submodule, the updated sample weight matrix is calculated through the Welsch robust loss function; Repeat the sample weight matrix update step until the maximum number of training iterations is reached or the network error meets the preset convergence condition, and obtain the optimized output weight matrix, thereby completing the construction of the multi-module echo state network model.

[0127] In a possible implementation, the first processing module is further configured to: Collect raw data from the municipal solid waste incineration process; The Hampel filter algorithm is used to replace outliers in the original data to obtain data after outlier replacement. The optimal sliding window size of the Hampel filter algorithm is determined by maximizing the mean absolute error of the data before and after filtering. Under the determined sliding window, the local median and local standard deviation of the data are calculated to identify outliers and replace them with the local median. The data after replacing outliers are normalized using the Min-Max normalization algorithm to obtain normalized data; Based on the pollutant generation mechanism of the municipal solid waste incineration process and the normalized mutual information algorithm, process variables with a high correlation with the concentration of each pollutant to be detected are determined from the normalized data; According to the process variables, the maximum average cross-correlation function algorithm is used to analyze the time delay relationship between the process variables and the concentration of the pollutants to be detected, and the optimal time delay parameters of the process variables are determined. The process variables are then subjected to time delay processing based on the determined optimal time delay parameters to obtain the input variables.

[0128] It should be noted that the intelligent detection system for multiple pollutants in the municipal solid waste incineration process provided by the present invention can execute the intelligent detection method for multiple pollutants in the municipal solid waste incineration process of any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0129] Figure 10 Schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 10As shown, the electronic device may include: a processor 1010 (processor), a communication interface 1020 (Communications Interface), a memory 1030 (memory) and a communication bus 1040, wherein the processor 1010, the communication interface 1020, and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 can call the logic instructions in the memory 1030 to execute the intelligent detection method for multiple pollutants in the municipal solid waste incineration process, which method includes: obtaining relevant input variables from the municipal solid waste incineration process; inputting the input variables into the multi-module echo state network model to obtain the predicted concentrations corresponding to each pollutant to be detected output by the multi-module echo state network model; wherein the multi-module echo state network model includes two sub-modules, each sub-module is constructed for a pollutant, and each sub-module has an independent input layer, a reserve pool and an output layer; wherein the input layer is used to receive input variables; the reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data and the previous state of the reserve pool of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

[0130] Furthermore, the logic instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0131] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the intelligent detection method for multiple pollutants in the urban solid waste incineration process provided by the above-mentioned embodiments, the method including: obtaining relevant input variables from the urban solid waste incineration process; inputting the input variables into a multi-module echo state network model to obtain the predicted concentration corresponding to each pollutant to be detected output by the multi-module echo state network model; wherein the multi-module echo state network model includes two sub-modules, each sub-module is constructed for a pollutant, and each sub-module has an independent input layer, a reserve pool and an output layer; wherein the input layer is used to receive input variables; the reserve pool is used to receive corresponding input layer data through an input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data and the previous state of the reserve pools of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

[0132] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the intelligent detection method for multiple pollutants in the urban solid waste incineration process provided by the above-mentioned embodiments, the method comprising: obtaining relevant input variables from the urban solid waste incineration process; inputting the input variables into a multi-module echo state network model to obtain the predicted concentrations corresponding to each pollutant to be detected output by the multi-module echo state network model; wherein the multi-module echo state network model includes two sub-modules, each sub-module is constructed for a pollutant, and each sub-module has an independent input layer, a reserve pool and an output layer; wherein the input layer is used to receive input variables; the reserve pool is used to receive corresponding input layer data through an input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data and the previous state of the reserve pools of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

[0133] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. That is, they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0134] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods of each embodiment or certain portions of the embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for intelligent detection of multiple pollutants in the process of municipal solid waste incineration, characterized in that: include: Obtain relevant input variables from the municipal solid waste incineration process; Inputting the input variables into a multi-module echo state network model to obtain predicted concentrations corresponding to each pollutant to be detected output by the multi-module echo state network model; The multi-module echo state network model includes two sub-modules, each of which is constructed for a pollutant, and each of which has an independent input layer, a storage pool, and an output layer; The input layer is used to receive the input variables; The reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data, and the previous state of the reserve pools of other submodules introduced through the shared weight matrix; The output layer is used to output the predicted concentration of the pollutant to be detected according to the output weight matrix and the current state of each storage tank.

2. The intelligent detection method for multiple pollutants in the municipal solid waste incineration process according to claim 1 is characterized in that: Also included is a training method for the multi-module echo state network model: Obtain input variable samples and use real pollutant concentration data as the target output vector for model training; Initialize the number of input layer nodes, the number of reserve pool nodes, and the number of output layer nodes of each submodule respectively, and randomly initialize the input weight matrix of each submodule according to uniform distribution; For each of the submodules, a diagonal matrix is randomly generated according to a preset singular value distribution, and a random orthogonal matrix is simultaneously generated. A reserve pool weight matrix and a shared weight matrix are further obtained by matrix singular value decomposition, and the reserve pool weight matrix and the shared weight matrix are made to satisfy the maximum singular value constraint condition; Initializing the initial state of each submodule reservoir randomly in a uniformly distributed manner, and updating the current state of each reservoir based on the input variable samples input to the corresponding submodule input layer, the previous state of its own reservoir, and the previous state of the reservoirs of other submodules introduced through shared weight matrix fusion; The current state of each sub-module reserve pool is collected and a state matrix is formed. The robust loss function with a penalty term is used as the optimization target. The target output vector is combined, and the output weight matrix of each sub-module is iteratively updated to obtain the multi-module echo state network model.

3. The intelligent detection method for multiple pollutants in the municipal solid waste incineration process according to claim 2 is characterized in that: The constraint condition of the maximum singular value is expressed by the following formula: ； Where, is the weight matrix of the i-th reserve pool, is the shared weight matrix, is the maximum singular value of the weight matrix of the i-th reserve pool, is the maximum singular value of the shared weight matrix.

4. The intelligent detection method for multiple pollutants in the municipal solid waste incineration process according to claim 2 is characterized in that: The updating of the current state of each reserve pool according to the input variable samples input to the input layer of the corresponding submodule, the previous state of its own reserve pool, and the previous state of the reserve pools of other submodules introduced by the shared weight matrix fusion specifically includes: The current status of each reserve pool is updated using the following formula: ； In the formula, t is the current time, t-1 is the previous time, is the current state of the storage pool of the i-th submodule, is the state of the reserve pool of the i-th submodule at the previous moment, is the previous state of the reserve pool of the jth other submodule, is the input weight matrix of the i-th submodule, is the input variable sample of the i-th submodule, is the reserve pool weight matrix of the i-th submodule, is the shared weight matrix, is a non-linear activation function.

5. The intelligent detection method for multiple pollutants in the municipal solid waste incineration process according to claim 2 is characterized in that: The current state of each submodule reserve pool is collected to form a state matrix, a robust loss function with a penalty term is used as the optimization target, the target output vector is combined, and the output weight matrix of each submodule is iteratively updated to obtain the multi-module echo state network model, including: Initialize the sample weight matrix of each submodule to the identity matrix; Calculating the updated output weight matrix of each submodule according to the state matrix, sample weight matrix, target output vector and penalty term for constraining the weight size of each submodule; Calculating the network error of each submodule according to the updated output weight matrix, state matrix and target output vector of each submodule; Calculating a scale estimate of each of the submodules according to the median of the absolute value of the network error of each of the submodules; Calculate the updated sample weight matrix using the Welsch robust loss function based on the network error and scale estimation value of each submodule; The updating step of the sample weight matrix is repeated until the maximum number of training iterations is reached or the network error meets the preset convergence condition, and the optimized output weight matrix is obtained, thereby completing the construction of the multi-module echo state network model.

6. The intelligent detection method for multiple pollutants in the municipal solid waste incineration process according to claim 1 is characterized in that: The relevant input variables obtained from the municipal solid waste incineration process include: Collect raw data from the municipal solid waste incineration process; Using a Hampel filtering algorithm to replace outliers on the original data to obtain data after outlier replacement, wherein the optimal sliding window size of the Hampel filtering algorithm is determined by maximizing the mean absolute error of the data before and after filtering, and within the determined sliding window, identifying outliers by calculating the local median and local standard deviation of the data, and replacing the outliers with the local median; Normalizing the data after replacing the outliers using a Min-Max normalization algorithm to obtain normalized data; According to the pollutant generation mechanism of the municipal solid waste incineration process and the normalized mutual information algorithm, a process variable having a high correlation with the concentration of each pollutant to be detected is determined from the normalized data; According to the process variable, a maximum average cross-correlation function algorithm is used to analyze the time delay relationship between the process variable and the concentration of the pollutant to be detected, determine the optimal time delay parameter of the process variable, and perform time delay processing on the process variable based on the determined optimal time delay parameter to obtain the input variable.

7. An intelligent detection system for multiple pollutants in the process of municipal solid waste incineration, characterized by: include: The first processing module is used to obtain relevant input variables from the municipal solid waste incineration process; The second processing module is used to input the input variables into the multi-module echo state network model to obtain the predicted concentrations corresponding to the pollutants to be detected output by the multi-module echo state network model; wherein the multi-module echo state network model includes two sub-modules, each of the sub-modules is constructed for a pollutant, and each of the sub-modules has an independent input layer, a reserve pool and an output layer; wherein the input layer is used to receive the input variables; the reserve pool is used to receive the corresponding input layer data through the input weight matrix, and update the current state of each reserve pool according to the previous state of its own reserve pool, the corresponding input layer data and the previous state of the reserve pools of other sub-modules introduced through the shared weight matrix; the output layer is used to output the predicted concentration of the corresponding pollutant to be detected according to the output weight matrix and the current state of each reserve pool.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method for intelligent detection of multiple pollutants in the urban solid waste incineration process as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for intelligent detection of multiple pollutants in the municipal solid waste incineration process as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for intelligent detection of multiple pollutants in the municipal solid waste incineration process as described in any one of claims 1 to 6 is implemented.

Images