Reverse identification method for release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm

Through the combination of deep learning and genetic algorithms, the reverse identification model is built, which solves the problem of difficult identification of multiple indoor pollution sources, realizes more accurate pollution source identification and air quality management, and improves pollutant transmission simulation and management in the indoor environment efficiency.

CN118981979APending Publication Date: 2024-11-19GUILIN UNIV OF ELECTRONIC TECH
View PDF 0 Cites 7 Cited by

Patent Information

Application Number
CN202411051636.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

The prior art is difficult to accurately identify and quantify the release intensity of multiple indoor pollution sources, limiting the efficiency and targeted nature of indoor air quality management, especially in complex environments, which are difficult to deal with unknown or hidden pollution sources.

Method used

Using the reverse identification method based on deep learning and genetic algorithms, a deep neural network model is constructed by collecting and preprocessing indoor environmental parameters and pollutant concentration data, and using genetic algorithms to optimize model parameters to achieve accurate identification and quantification of pollution source release intensity. .

Benefits of technology

It improves the identification accuracy and management efficiency of pollution source release intensity, overcomes the limitations of traditional methods, can more accurately simulate pollutant propagation and distribution, and improves the prediction performance and generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981979A_ABST
    Figure CN118981979A_ABST
Patent Text Reader

Abstract

The invention discloses a reverse identification method for release intensity of multiple indoor pollution sources based on deep learning and a genetic algorithm, and the method comprises the following steps: 1, carrying out the data collection and preprocessing of indoor environment parameters, including the comprehensive collection and preprocessing of parameters such as temperature, humidity and air velocity, so as to guarantee the data quality; 2, constructing and training a deep neural network model, and learning a relationship between pollution source release and pollutant concentration; 3, optimizing model parameters by using a genetic algorithm, and searching optimal parameter configuration through natural selection, crossover and mutation operations; 4, performing reverse analysis on the indoor pollutant concentration by applying the optimized model, and identifying the release intensity of each pollution source; and 5, verifying a reverse identification result and adjusting the model to ensure accuracy and reliability. The method combines deep learning with a genetic algorithm, improves the prediction precision and the adaptive adjustment capability of the system, and is suitable for multiple fields.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring and management, and in particular to a reverse identification method of the release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm, which is used to improve indoor air quality and guide the implementation of environmental improvement measures. Background Art

[0002] With the improvement of people's living standards and the enhancement of health awareness, indoor air quality has received widespread attention. Indoor air pollution has become one of the important factors affecting human health. Indoor air pollution sources are diverse, including but not limited to tobacco smoke, harmful substances released by building materials, volatile organic compounds in furniture and decorations, and microbial pollution generated by daily activities. Therefore, accurately identifying and quantifying the release intensity of multiple indoor pollution sources is of great significance for formulating effective air quality improvement measures and protecting residents' health.

[0003] However, existing indoor air quality monitoring and management methods mainly rely on direct monitoring of the concentration of specific pollutants, often ignoring the interactions between pollution sources and pollutant propagation characteristics in complex indoor environments. In addition, direct monitoring methods are difficult to quantitatively analyze unknown or hidden pollution sources, limiting the efficiency and pertinence of indoor air pollution control. Therefore, it is urgent to develop a new method that can accurately reversely identify the release intensity of multiple indoor pollution sources in order to more effectively evaluate and manage indoor air quality. This method should be able to comprehensively analyze indoor environmental parameters and use advanced data processing and optimization algorithms to accurately identify and quantify the release intensity of pollution sources in complex indoor environments.

[0004] The disclosure of the above background technology content is only used to assist in understanding the concept and technical solution of the present invention. It does not necessarily belong to the prior art of this patent application. In the absence of clear evidence that the above content has been disclosed on the filing date of this patent application, the above background technology should not be used to evaluate the novelty and creativity of the present application. Summary of the invention

[0005] In view of the above technical problems, the present invention provides a reverse identification method of the release intensity of multiple indoor pollution sources based on deep learning and genetic algorithms, aiming to solve the limitations of indoor air quality management in the prior art, especially the problem that the release intensity of pollution sources in complex indoor environments is difficult to accurately identify. This method comprehensively utilizes the advantages of deep learning in processing large amounts of complex data and the high efficiency of genetic algorithms in global optimization problems. Starting from indoor environmental monitoring data, it reversely calculates and accurately identifies the release intensity of each indoor pollution source, providing a new technical solution for the accurate identification and quantification of the release intensity of multiple indoor pollution sources.

[0006] To achieve the above object, the present invention adopts the following technical solution:

[0007] A reverse identification method for the release intensity of multiple indoor pollution sources based on deep learning and genetic algorithms mainly includes collecting indoor environmental parameters (such as temperature, humidity, air velocity) and pollutant concentration data of monitoring points, and using computational fluid dynamics simulation to obtain a stable thermal flow field of the indoor environment. Then a deep neural network model is constructed, the input layer receives the pre-processed environmental parameters and concentration data, and the output layer predicts the release intensity of each pollution source. Genetic algorithms are used to optimize the hyperparameters of the deep learning model to improve the prediction performance of the model. The trained model is used to predict the test set data, predict the release intensity of each pollution source, and compare it with the actual value to verify the accuracy and reliability of the model, including the following steps:

[0008] Step 1: Data collection and preprocessing of indoor environmental parameters

[0009] Comprehensive data collection of building indoor environmental parameters is carried out, and the collected data is pre-processed to ensure the data quality that can be processed by the subsequent deep learning model;

[0010] Step 2: Model construction and training based on deep learning

[0011] Based on the pre-processed indoor environmental parameters and pollutant concentration data, a deep neural network model is constructed. The model learns the complex relationship between pollution source release and indoor pollutant concentration through multi-layer nonlinear transformation. During the training process of the deep learning model, the back propagation algorithm is used to adjust the network parameters to minimize the difference between the predicted concentration and the actual concentration.

[0012] Step 3: Genetic algorithm optimization model parameters

[0013] Use genetic algorithms to optimize the structure and hyperparameters of deep learning models, and find the optimal network structure and parameter configuration by simulating the process of natural selection, including selection, crossover, and mutation operations, to improve the model's predictive performance and generalization ability;

[0014] Step 4: Reverse identification of pollution source release intensity

[0015] The trained and optimized deep learning model is applied to the reverse analysis of indoor pollutant concentrations. The genetic algorithm is combined with the actual monitored indoor pollutant concentrations to reversely calculate the release intensity of each pollution source to identify the release intensity of each pollution source. In this process, the genetic algorithm gradually approaches the actual release intensity of the pollution source through continuous iterative search.

[0016] Step 5: Result verification and model adjustment

[0017] The release intensity of the pollution source obtained by reverse identification is compared with the actual situation to verify the accuracy and reliability of the model. If there is a large deviation between the identification result and the actual value, it is necessary to return to steps 2 and 3 to adjust and optimize the parameters of the deep learning model and genetic algorithm until a satisfactory identification effect is achieved.

[0018] Furthermore, in step 1, the indoor environmental parameters are analyzed by a deep learning model to predict indoor air flow and pollution distribution patterns. This step first involves the comprehensive collection of indoor environmental parameters, including but not limited to temperature, humidity, air velocity, etc., which are crucial for understanding indoor air quality and flow characteristics. The collected data is preprocessed, including denoising, normalization, and feature extraction, to ensure the quality and validity of the input data. Next, a deep learning neural network is used to build a model to simulate and predict indoor air flow and the distribution of pollutants. The key to this step is that the deep learning model can identify and learn the complex nonlinear relationship between indoor environmental parameters and pollutant distribution, thereby providing an accurate basic model for subsequent reverse identification work. In addition, a large amount of measured data will be used in the training process of the deep learning model to ensure the accuracy and generalization ability of the model prediction. This step not only provides a scientific prediction model for reverse identification of the release intensity of pollution sources, but also lays the foundation for the application of genetic algorithms in subsequent steps.

[0019] Furthermore, in step one, comprehensive collection of building indoor environmental parameters is performed, including temperature, humidity, air velocity, etc., which are crucial for simulating the stable heat flow field released by multiple indoor pollution sources.

[0020] Furthermore, the data described in step 1 is preprocessed, including data cleaning, standardization and denoising.

[0021] Furthermore, in step 1, advanced data processing techniques, such as autoencoders, are used to perform feature extraction and noise reduction to ensure the quality of the input data.

[0022] Furthermore, in step 1, computational fluid dynamics numerical simulation technology is used to obtain the stable thermal flow field inside the building based on the collected environmental parameters. This step provides accurate basic data for subsequent pollutant propagation analysis.

[0023] Furthermore, in step 2, a genetic algorithm is used to optimize the parameters of the deep learning model established in step 1 to improve the accuracy and stability of the model in predicting indoor pollutant distribution and pollution source release intensity. Specifically, this step includes the following operations:

[0024] S1: Initialize genetic algorithm parameters: including population size, crossover probability, mutation probability, and the encoding method for defining individuals. Usually, the hyperparameters of the deep learning model are used as the gene encoding in the genetic algorithm.

[0025] S2: Population generation: In the initial stage of the genetic algorithm, a set of hyperparameter configurations of the deep learning model is randomly generated as the initial population, and each individual represents a set of possible deep learning model configurations;

[0026] S3: Fitness function evaluation: For each individual in the population, the performance of the model in the indoor pollutant distribution prediction task is evaluated according to the parameter configuration of the deep learning model it represents, and indicators such as prediction accuracy and error rate are used as fitness evaluation criteria;

[0027] S4: Selection operation: According to the fitness of individuals, roulette and tournament selection methods are used to select individuals with high fitness for subsequent crossover and mutation operations to ensure the inheritance of excellent genes;

[0028] S5: Crossover and mutation operations: New individuals are generated in the population through crossover and mutation operations. The crossover operation allows two individuals to exchange their genetic information, while the mutation operation randomly changes part of the individual's genes to introduce new genetic diversity.

[0029] S6: Generation of a new generation of population: Repeat the selection, crossover and mutation operations to generate a new generation of population. The generation of each generation of population is an optimization of the previous generation of population.

[0030] S7: Termination condition judgment: When the preset number of iterations is reached or the model performance is no longer significantly improved, the algorithm terminates, and finally the deep learning model configuration represented by the individual with the highest fitness is selected as the optimal configuration.

[0031] Step 2 not only optimizes the parameter configuration of the deep learning model, improves the prediction accuracy and generalization ability of the model, but also provides a solid foundation for reverse identification of the release intensity of multiple indoor pollution sources.

[0032] Furthermore, in step 2, on the basis of obtaining a stable thermal flow field, deep learning technology is further applied to learn and simulate the propagation and distribution patterns of pollutants in the indoor environment by training a deep neural network model.

[0033] Furthermore, the deep learning model training described in step 2 is to apply the deep learning architecture of deep neural network to fully explore data features and improve prediction accuracy.

[0034] Furthermore, the deep neural network model in step 2 can predict the pollutant concentration distribution in various indoor areas under different conditions based on the input of environmental parameters and thermal flow field.

[0035] Furthermore, the hyperparameters described in step S2 include learning rate, number of network layers, and number of neurons in each layer.

[0036] Furthermore, step three involves using the optimized deep learning model to simulate the indoor pollutant distribution based on indoor environmental parameters and measured monitoring data, and then generating a transmission matrix A between the pollution source and the monitoring point. This transmission matrix is ​​calculated based on the relationship between the pollutant concentration predicted by the model and the actual monitoring point concentration, which provides a key mathematical basis for the reverse identification of the release intensity of the pollution source, including the following steps:

[0037] S1: Indoor pollutant distribution simulation: Using the deep learning model optimized in step 2, combined with indoor environmental parameters and pollution source information, simulate indoor air flow and pollutant distribution. This simulation process takes into account the complex dynamics of the indoor environment and can accurately predict the pollutant concentration at the monitoring point under different conditions.

[0038] S2: Generation of transmission matrix A: Based on the simulation results of the deep learning model, the contribution rate of each pollution source to the pollutant concentration at each monitoring point is calculated to construct the transmission matrix A. Each element in the matrix represents the influence intensity of a specific pollution source on the concentration of the corresponding monitoring point, reflecting the propagation characteristics of pollutants in the indoor environment.

[0039] S3: Verification and correction of the transmission matrix A: Verify the accuracy of the transmission matrix A by comparing it with the actual monitoring data. If necessary, further fine-tune the deep learning model to ensure that the transmission matrix can accurately reflect the transmission relationship between the pollution source and the monitoring point.

[0040] S4: Prepare for the reverse identification process: Use the verified and corrected transfer matrix A to lay the foundation for the subsequent reverse identification work. On this basis, use the inverse algorithm to accurately infer the release intensity of each pollution source from the monitoring point concentration data.

[0041] Furthermore, the indoor environmental parameters described in step S2 include temperature, humidity, and air flow rate.

[0042] Furthermore, in step 3, in order to further improve the prediction performance and accuracy of the deep learning model, a genetic algorithm is used to globally optimize the parameters of the deep learning model, including the adjustment of parameters such as network structure and learning rate. By simulating the mechanism of natural selection and genetic variation, the optimal model parameter configuration is found to improve the accuracy and robustness of model prediction.

[0043] Furthermore, in step three, the parameters of the genetic algorithm, such as population size, crossover rate and mutation rate, are set in detail, and a suitable fitness function is selected to ensure the effectiveness and efficiency of the algorithm.

[0044] By executing step three, not only can the transmission matrix A reflecting the propagation characteristics of indoor pollutants be obtained, but it also provides the necessary conditions and basis for accurately reversely identifying the release intensity of multiple indoor pollution sources, thereby enhancing the scientificity and accuracy of the entire identification process.

[0045] Furthermore, step 4 uses the transmission matrix A that has been constructed and verified, combined with the pollutant concentration data of the actual monitoring point, to accurately calculate the release intensity of each indoor pollution source through the inverse algorithm. This step is the key to the entire method and directly determines the accuracy and reliability of the inverse identification results. It includes the following steps:

[0046] S1: Selection and implementation of inverse algorithm: Select appropriate inverse algorithm, including least squares method and regularization method, to deal with possible ill-conditioned problems and enhance the stability of the solution process, implement the inverse algorithm, and prepare to apply it to the transmission matrix A and the concentration data of the monitoring points;

[0047] S2: Inverse solution of the release intensity of pollution sources: Using the selected inverse algorithm, combined with the transmission matrix A obtained in step 3 and the actual pollutant concentration data of each monitoring point, calculation is performed to inversely infer the release intensity of each pollution source. This calculation process needs to fully consider the uncertainty of model prediction and measurement error to ensure the accuracy and reliability of the solution results;

[0048] S3: Analysis and verification of release intensity results: Conduct a detailed analysis of the release intensity of the pollution source obtained by inverse solution, including the size and distribution characteristics of the release intensity, and compare it with the known pollution source characteristics to verify the rationality and accuracy of the solution results;

[0049] S4: Parameter adjustment and model optimization: According to the solution results and verification analysis, if necessary, return to steps 1 and 2 to adjust and optimize the model parameters, in order to further improve the prediction accuracy of the model and the accuracy of inverse identification;

[0050] S5: Result output and application: Output the final pollution source release intensity results in an appropriate form to provide a scientific basis for indoor air quality management, guide the implementation of pollution prevention and control measures and the formulation of environmental improvement plans.

[0051] The execution of step four not only completed the accurate identification of the release intensity of multiple indoor pollution sources, but also provided important information for subsequent environmental management and improvement, demonstrating the important value and broad prospects of this method in practical applications.

[0052] Furthermore, in step 4, based on the optimized deep learning model and combined with the actual monitoring data, the inverse algorithm is applied to calculate and identify the release intensity of each indoor pollution source. This process uses the relationship between the pollutant concentration distribution predicted by the deep learning model and the measured data to reversely infer the pollution source intensity through mathematical modeling and calculation. The specific steps are as follows:

[0053] S1: The indoor pollutant concentration distribution predicted by the deep learning model is combined with the model parameters optimized by the genetic algorithm to construct the transmission matrix A between the release intensity of multiple pollution sources and the concentration of the monitoring point through the nonlinear network structure inside the deep neural network;

[0054] S2: On this basis, an inverse model is established, and the truncated singular value decomposition method is used to regularize the inverse model to enhance the stability of its solution;

[0055] S3: Based on the regularized inverse model and the transmission matrix A, combined with the pollutant concentration data of the actual monitoring point, the release intensity of each indoor pollution source is calculated in reverse. Through this step, the contribution of multiple pollution sources to air quality in the indoor environment can be accurately identified and quantified.

[0056] Furthermore, step five involves the verification and application process of the reverse identification results. In this step, the accuracy and reliability of the reverse identification method are verified by comparative analysis with the actual on-site monitoring data, and the model is adjusted and optimized as necessary according to the verification results to ensure its effectiveness and applicability in practical applications, including the following steps:

[0057] S1: Result verification: Collect the actual monitoring data corresponding to the reverse identification process, including the pollutant concentrations at each indoor monitoring point and the actual release of pollution sources, compare the pollution source release intensity obtained by reverse identification with the actual data, and evaluate the prediction accuracy of the model and the error range of reverse identification;

[0058] S2: Model adjustment: Based on the problems and errors found during the result verification process, analyze the possible causes, including improper model hyperparameter settings, insufficient data preprocessing, and inaccurate transfer matrix A. Based on this, adjust and optimize the key links such as the deep learning model structure, genetic algorithm parameters, and inverse algorithm;

[0059] S3: Re-verification: Re-calculate the inverse identification of the adjusted and optimized model and verify it with new actual monitoring data to ensure that the adjustments have effectively improved the prediction accuracy of the model and the accuracy of the inverse identification;

[0060] S4: Application promotion: After multiple rounds of verification and adjustment, and confirmation that the accuracy and stability of the model have reached a satisfactory level, the inverse identification method will be applied to a wider range of indoor air quality monitoring and management. The reverse identification results will be used to guide the formulation of pollution prevention and control measures and the implementation of environmental improvement strategies to effectively improve indoor air quality.

[0061] Furthermore, in step 5, the results obtained from the reverse identification are verified, and the accuracy and reliability of the model are evaluated by comparative analysis with field monitoring data. Based on the verification results, the model is adjusted and optimized as necessary to ensure that it achieves the best performance in practical applications.

[0062] The beneficial effects of the present invention compared with the prior art are as follows:

[0063] (1) The present invention uses a deep learning model to deeply analyze the complex nonlinear relationship between indoor environmental parameters and pollutant concentrations. Compared with traditional methods, it can more accurately simulate the propagation and distribution of pollutants indoors, thereby improving the accuracy of identifying the release intensity of pollution sources.

[0064] (2) The present invention uses a genetic algorithm to globally optimize the hyperparameters of the deep learning model and automatically find the optimal network structure and parameter configuration, overcoming the limitations of manual parameter adjustment, such as being time-consuming and relying on experience, so that the model can achieve better prediction performance.

[0065] (3) The present invention combines deep learning and genetic algorithms to not only improve the performance of the model on the training set, but also reduce overfitting through regularization techniques such as truncated singular value decomposition, thereby improving the model's generalization ability for unknown data.

[0066] (4) The deep learning model in the present invention can automatically extract key features from a large amount of complex data and perform effective dimensionality reduction processing, reducing the workload of manual feature selection and data preprocessing, while avoiding information loss caused by human factors.

[0067] (5) The inverse identification problem in the present invention is often pathological or ill-posed. The method proposed in the present invention combining the regularization technology of truncated singular value decomposition can effectively stabilize the solution process and improve the accuracy and stability of the solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart of the reverse identification method of the release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm of the present invention;

[0069] Figure 2 It is a graph of the measured values ​​of the release concentrations of the two pollution sources at the detection points;

[0070] Figure 3It is the reverse identification result diagram of the release intensity of two pollution sources;

[0071] Figure 4 It is the truncated singular value decomposition method to obtain the regularization parameter λ graph. DETAILED DESCRIPTION

[0072] A specific embodiment of the present invention is described in detail below in conjunction with the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited by the specific embodiment. It should be understood that the "up", "down", "left" and "right" directions mentioned in the following embodiments of the present invention are all based on the positions of the corresponding drawings. These words used to limit the direction are only for the convenience of explanation and do not represent limitations on the specific technical solutions of the present invention. Unless otherwise specified, the same numbers in the figure marks represent the same structure.

[0073] The present invention discloses a reverse identification method of the release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm, which mainly includes first starting with collecting indoor environmental parameters and pollutant concentration data, such as temperature, humidity, air velocity, etc., which are used to train a deep learning model after preprocessing. The model can learn the complex relationship between indoor environmental parameters and pollutant concentrations, and predict the distribution of pollutant concentrations under different conditions. In order to optimize the performance of the model, a genetic algorithm is used to automatically adjust and select the optimal model parameters, such as network structure, learning rate, etc. Through this process, the model can more accurately simulate the propagation process of indoor pollutants. Then, based on the optimized deep learning model, combined with actual monitoring data, an inverse model is established to reverse the release intensity of each pollution source. In order to improve the stability of the inverse model and the accuracy of the solution, a truncated singular value decomposition method is introduced to process the inverse model. Finally, the release intensity of each indoor pollution source is accurately calculated by the inverse model after regularization, and accurate reverse identification from indoor environmental monitoring data to pollution source release intensity is achieved. This method effectively combines the powerful capabilities of deep learning in data processing and pattern recognition with the high efficiency of genetic algorithms in global search and optimization, providing a new way to solve key problems in indoor air quality management. Figure 1 Workflow diagram of .

[0074] Step 1: Data collection and preprocessing. Collect monitoring data of indoor environmental parameters (such as temperature T, humidity H, air velocity V) and pollutant concentration C; preprocess the data, including standardization;

[0075] Step 2: Deep learning model construction and training. Construct a deep neural network model, with indoor environmental parameters and initial pollutant concentration as model input and predicted pollutant concentration distribution as output; train the model using a loss function minimization strategy, with the commonly used loss function being mean square error (MSE).

[0076] Step 3: Genetic algorithm optimizes model parameters. Define the parameters of the genetic algorithm, including population size, crossover rate, mutation rate, etc. During the optimization process, use the prediction accuracy of the deep learning model (such as the accuracy of reverse identification) as the fitness function, and iterate through the genetic algorithm to find the optimal model parameters;

[0077] Step 4: Implementation of the inverse identification algorithm. Establish an inverse model, which is based on the pollutant concentration distribution predicted by the deep learning model and calculates the release intensity of the pollution source through the inverse algorithm;

[0078] Step 5: Solve the emission intensity of pollution sources. Using the optimized deep learning model and the inverse model processed by the truncated singular value decomposition method, combined with the pollutant concentration data of the actual monitoring points, the precise emission intensity of each indoor pollution source is calculated through a mathematical solution method.

[0079] (1) In step 1:

[0080] The measurement personnel use professional equipment to accurately measure the temperature, humidity, air flow rate and other parameters of the indoor environment to ensure the authenticity and reliability of the data. These parameters directly affect the behavior of indoor air flow and heat transfer, and are important inputs for computational fluid dynamics (CFD) simulation. Based on the measured environmental parameters, the initial conditions and boundary conditions of the CFD simulation are set. For example, the measured indoor temperature is set as the initial temperature condition of the simulation, and the air flow rate and direction are used as boundary conditions. CFD software is used to perform numerical simulation of the indoor environment of the building. By solving fluid dynamics and heat transfer equations, the flow behavior of air in the room and the heat exchange process with the indoor structure are simulated. During the simulation process, the software will calculate the air flow rate, temperature distribution, humidity distribution, etc. at various locations in the room according to the set environmental conditions and physical models. After the CFD simulation is completed, the thermal flow field of the indoor environment of the building can be obtained, including detailed information such as temperature, humidity and air flow rate. Through the above steps, the thermal flow field of the indoor environment of the building obtained by CFD simulation provides scientific and accurate basic data for the subsequent reverse identification of the release intensity of multiple indoor pollution sources. This not only improves the accuracy of reverse identification, but also provides important technical support for the assessment and improvement of indoor air quality.

[0081] In the present invention, two pollution sources are released at the same time in the room as an example: pollution source A is set at a distance of 0.5 meters from the air inlet, and pollution source B is set at the same height of 0.5 meters on the wall opposite to the air inlet. They both release pollutants at a concentration of 100ppm. Figure 2 , showing the measured values ​​of the release concentrations of the two pollution sources at the detection points. After the release of pollution sources A and B, the peak concentrations of pollutants at outlet 2 and measuring point 1 reached about 0.22RF and exceeded 20RF, respectively, indicating that these locations were most significantly affected by pollution. The RF values ​​of outlet 1 and measuring point 2 were relatively low, at around 0.04RF and 0.20RF, respectively. The results show that the air flow characteristics significantly affect the distribution of pollutants, especially the high concentrations of pollutants at outlet 2 and measuring point 1. These data verify the effectiveness of CFD simulation and provide a key reference for indoor air quality control and pollution source management. In view of the slow characteristics of air flow in the simulated environment, two monitoring points are placed at the key positions of the air inlet and outlet, namely monitoring point 1 (0.5m, 1.6m) and monitoring point 2 (1.7m, 0.1m). The release mode of the pollution source is configured in detail by activating the user-defined function, and then a stable thermal flow field in the room is constructed in the CFD simulation environment. The concentrations of the monitoring points are shown in Table 1:

[0082] Table 1 Pollutant concentrations at two indoor monitoring points

[0083]

[0084] Considering that the pollutant concentration at each monitoring point is 0 before the pollution source begins to release, the concentration change measured at the monitoring point after the pollution source releases can be directly regarded as the actual release concentration of the pollution source monitored at that point.

[0085] (2) In step 2:

[0086] In the reverse identification of the release intensity of multiple pollution sources, deep neural networks are used to establish a nonlinear mapping relationship between indoor environmental parameters and pollutant concentrations. This process involves training a model that can predict pollutant concentrations based on input environmental parameters. The following is a detailed description of the key calculation formulas and steps of deep neural networks in this reverse identification task.

[0087] Assume that the deep neural network consists of L layers, and each layer has n l neurons, where l = 1, 2, ..., L. The input layer receives indoor environmental parameters (such as temperature, humidity, air flow rate, etc.), and the output layer predicts the corresponding pollutant concentration. For each layer l, the calculation of the neural network can be expressed as:

[0088] z (l+1) =W (l) a (l) +b(l)

[0089] a (l+1) =f(z (l+1) )

[0090] Where W (l) is the weight matrix of the lth layer, b (l) is the bias vector, a (l) is the activation output of layer l, f is the activation function (such as ReLU or Sigmoid, z (l+1) is a linear combination of the weighted input and the bias.

[0091] The loss function is used to evaluate the difference between the model's predicted pollutant concentration and the actual concentration. Commonly used loss functions include mean square error, expressed as:

[0092]

[0093] In the formula, m is the number of samples, y is i is the actual pollutant concentration of the ith sample, is the pollutant concentration predicted by the model.

[0094] The gradient of the loss function with respect to each parameter is calculated through the back-propagation algorithm, and then the network parameters are updated using an optimization algorithm (such as gradient descent):

[0095]

[0096] Where a is the learning rate, are the gradients of the loss function with respect to the weights and biases, respectively.

[0097] Through the above calculation steps, the deep neural network can learn the complex mapping relationship between indoor environmental parameters and pollutant concentrations, and then realize the accurate reverse identification of the release intensity of multiple indoor pollution sources. In this step, by training the deep neural network, the model learns the mapping relationship between indoor environmental parameters (such as temperature, humidity, air flow rate, etc.) and the pollutant concentration at the monitoring point. The input of the network is indoor environmental parameters and pollution source information, and the output is the prediction of the pollutant concentration at the corresponding monitoring point. After the deep neural network training is completed, the trained model is used to simulate the impact of each pollution source separately, and to simulate the comprehensive effect under the simultaneous action of all pollution sources. Through these simulations, the specific numerical value of the impact of each pollution source on the concentration of each monitoring point can be obtained. Once the transmission matrix A is obtained, the release intensity vector q of the pollution source can be estimated by solving the linear equation system C=A·q according to the pollutant concentration vector C actually measured at the monitoring point, where q represents the release intensity of each pollution source.

[0098] (3) In step 3:

[0099] At the beginning of the genetic algorithm, a population needs to be initialized first. The population consists of multiple individuals, each of which represents a set of parameters for a deep neural network. These parameters include network weights, biases, etc., which can be represented by real numbers, binary codes, or other encoding methods. The initialization process is usually random to ensure that the parameter space is widely explored. The fitness of each individual represents the performance of the parameter set in the deep learning task. The calculation of fitness is usually based on the performance of the network on the validation set, such as classification accuracy, prediction error, etc. A higher fitness means that the individual (parameter set) is better and has a greater chance of being retained to the next generation. According to the fitness of the individual, excellent individuals are selected from the current population as parents of the next generation. The selection process can be implemented by methods such as roulette and tournament selection, and its formula can be expressed as:

[0100]

[0101] Among them, P(x i ) is the probability that an individual is selected according to its fitness. f(x i ) is the individual x i The fitness of a species is the number of individuals in the population. A higher fitness means a higher probability of being selected. n is the total number of individuals in the current population.

[0102] Ensure that individuals with higher fitness have a higher chance of being selected. In order to prevent the algorithm from converging to the local optimal solution too early and increase the diversity of the population, mutation operations are performed on the offspring individuals. Mutation is achieved by randomly changing a part of the genes (parameter values) in the individual. This simulates the mutation mechanism in biological evolution and can explore new parameter spaces. After selection, crossover and mutation operations, a new population is generated. The crossover operation involves the exchange of randomly selected paternal genes to generate new offspring. In real number coding, a simple crossover operation is a single-point crossover, which can be mathematically expressed as:

[0103] y1=[x 1,1 ,...,x 1,k ,x 2,k+1 ,...,x 2,d ]

[0104] y2=[x 2,1 ,...,x 2,k ,x 1,k+1 ,...,x 1,d ]

[0105] Here, parents x1 and x2 are cut at random positions to produce offspring y1 and y2, where d is the length of the gene.

[0106] The mutation operation usually occurs at each gene position with a small probability μ, which can be expressed as:

[0107]

[0108] Among them, x i is the gene before mutation, x i ' is the mutated gene, It is a random number drawn from a certain distribution (such as normal distribution) and is used to simulate gene mutation.

[0109] This process is repeated continuously, and each iteration produces a new generation of individuals. The iteration continues until the termination condition is met, such as reaching the preset number of iterations or fitness convergence. Finally, the individual with the highest fitness is selected as the optimal network parameter. The process of optimizing deep neural networks with genetic algorithms is a dynamic, iterative optimization process that continuously explores and utilizes parameter space to find the optimal solution by simulating natural selection and genetic variation mechanisms. This method does not rely on gradient information and is particularly suitable for complex, non-convex optimization problems. It can effectively improve the performance of deep neural networks on specific tasks.

[0110] The prediction accuracy of deep neural networks (DNN) on the validation set or the error of reverse identification is used as the fitness function, which can be expressed as f(P i )=1 / (1+MSE), where MSE is the mean square error.

[0111] Randomly pair the parent individuals and exchange their genes (i.e. parameter values) with a certain probability to generate new offspring. m and P f , the offspring individual may be P child =αP m +(1-α)P f , where α is the crossover rate.

[0112] For each individual (i.e., a set of DNN parameters), the model output is calculated through forward propagation, and then the loss function is calculated. The performance of each individual is evaluated using the fitness function. The higher the fitness, the greater the possibility that the individual will be selected as the next generation. The selection, crossover, mutation, and fitness evaluation processes are repeated until the stopping condition is met (such as reaching the maximum number of iterations or the fitness is no longer significantly improved). The individual with the highest fitness is selected as the final DNN model parameter for the reverse identification of the release intensity of multiple indoor pollution sources. Through the above calculation process, the genetic algorithm helps to find the optimal deep neural network parameters, thereby improving the performance of the model in the reverse identification task and ensuring the accuracy and efficiency of reverse identification.

[0113] (4) In step 4:

[0114] In the inverse identification problem, especially when dealing with the calculation of the release intensity of multiple indoor pollution sources, regularization technology is the key to improving the stability and accuracy of the solution. Using the regularization matrix L and the process of calculating the regularization parameter λ, the release intensity q of the pollution source can be accurately estimated from the monitoring data.

[0115] The regularization matrix L is usually used to control the smoothness of the solution or constrain specific properties of the solution. In different application scenarios, L can be an identity matrix, a differential operator, etc., and its selection depends on the specific requirements and prior knowledge of the problem.

[0116] The regularization parameter λ greatly affects the solution result, which balances the weights of the data fitting term and the regularization term.

[0117] The specific steps are as follows:

[0118] 1) Model establishment. It is assumed that the propagation of pollutants in indoor environments can be described by a linear model, namely:

[0119] C=Aq+k

[0120] Where A is the known transmission matrix, C is the pollutant concentration obtained at the monitoring point, q is the estimated release intensity of the pollution source, and k is the observation noise.

[0121] 2) Introducing sparse priors. In the framework of the sparse Bayesian learning (SBL) model, a sparse prior is introduced for the emission intensity q of the pollution source. The common approach is to assume that each element of q is independent and obeys a zero-mean Gaussian distribution, and its variance is controlled by a hyperparameter vector γ, that is:

[0122] p(q|γ)=∏ N i=1 N(q i ∣0,γ i )

[0123] In the formula, N(·|0,γ i ) means the mean is 0 and the variance is γ i Gaussian distribution.

[0124] 3) Learning of hyperparameters. The hyperparameter γ is learned using the maximum evidence framework (also called type II maximum likelihood). This involves calculating the marginal likelihood function (also called evidence function) and maximizing it to find the optimal γ.

[0125] 4) Estimation of the release intensity of pollution sources. Given the optimal γ, the posterior distribution of q is calculated using the Bayesian theorem. Under the SBL framework, the posterior distribution of q is usually also a Gaussian distribution, and its mean vector q^ can be used as an estimate of the release intensity of the pollution source. The calculation formula is:

[0126] q^=(A T A+D γ -1 ) -1 A T C

[0127] Where D γ is a diagonal matrix whose diagonal elements are γ.

[0128] 5) Algorithm iteration. Update γ and ^q through iteration until convergence. In each iteration, update γ according to the current ^q, and then update ^q with the new γ.

[0129] Through the sparse Bayesian learning method, accurate estimation of the release intensity of multiple indoor pollution sources can be effectively achieved, especially in complex environments where the location and number of pollution sources are unknown. The SBL method can not only improve the accuracy of the estimation, but also naturally identify important pollution sources, thus playing an important role in indoor air quality assessment and management. In step 2, the transmission matrix A has been obtained using a deep neural network, and logλ=1.0352231 is calculated using a self-compiled program, and then λ=0.1871 is obtained. It is known that C=(47.08 26.28) T , by q=(A T A+λ 2 L T L) -1 ×(A T C), the calculated value of the release intensity of indoor pollution sources is q = (16.23 32.01) T . Figure 3 The reverse identification results of the release intensity of two pollution sources are shown, by carefully showing the pollutant concentrations changing over time. Among them, sub-graphs (a) and (b) show the concentration changes of two different pollution sources at different time points, highlighting their respective environmental impact characteristics. Sub-graph (c) captures a sudden emission event, showing a sharp increase in pollutant concentrations, while sub-graph (d) depicts the gradual downward trend of pollutant concentrations after an initial increase, indicating that the reverse identification algorithm can effectively simulate the entire process of a pollution event. These graphical results verify the application effect of the deep learning model in reverse identification of pollution source release intensity, and provide important data support for further model optimization and precise environmental monitoring.

[0130] It is known that in step 1, the actual release intensity of multiple indoor pollution sources is obtained as q'=(17.0833.49) T , compare q with q', the results are shown in Table 2:

[0131] Table 2 Calculated values ​​of pollution source release intensity and actual values ​​of pollution source release intensity (unit: ppm)

[0132]

[0133] From the comparison results of release intensity, it can be observed that the inverse model shows good accuracy in identifying the release intensity of multiple indoor pollution sources. Even taking into account the existing errors, the release intensity obtained by inverse identification is basically consistent with the release intensity of the actual pollution source within the allowable error range. This shows that the adopted inverse model can not only accurately capture the contribution of each pollution source to indoor air quality, but also effectively reverse derive the actual release amount of each pollution source from the monitoring data. This result provides a powerful tool for the assessment and management of indoor air quality, especially in complex indoor environments where it is necessary to accurately identify and quantify the impact of multiple pollution sources. The application of this inverse model is of great significance. By further optimizing the model and adjusting the parameters, it is expected to further improve the identification accuracy and provide more accurate data support for indoor environmental quality control and pollution control.

[0134] (5) In step 5:

[0135] Using a deep learning model optimized by a genetic algorithm and an inverse model processed by the truncated singular value decomposition (TSVD) method, the release intensity of each indoor pollution source is accurately calculated through a mathematical solution method. First, the pollutant concentration data of the actual monitoring point is input into the optimized model, and the TSVD method is used for regularization to stabilize the solution process and reduce the model's sensitivity to noise. This method can effectively deal with ill-conditioned problems and improve the accuracy and reliability of the solution.

[0136] Figure 4 The process of obtaining the regularization parameter using the truncated singular value decomposition method is shown. In the figure, sub-figure (a) shows the relationship between the logarithm of the singular value corresponding to the regularization parameter (log|λ|) and the cumulative energy (log|Δq|), where it can be observed that as the singular value increases, the cumulative energy gradually increases; sub-figure (b) depicts the relationship between the regularization parameter and the model error (log|q|), showing that the error decreases as the singular value increases. These two figures together guide the selection of the regularization parameter, ensuring that the model's sensitivity and adaptability to the data are maintained while reducing the computational error.

[0137] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.

Claims

1. A reverse identification method of the release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing of indoor environmental parameters Comprehensive data collection of building indoor environmental parameters is carried out, and the collected data is pre-processed to ensure the data quality that can be processed by the subsequent deep learning model; Step 2: Model construction and training based on deep learning Based on the pre-processed indoor environmental parameters and pollutant concentration data, a deep neural network model is constructed. The model learns the complex relationship between pollution source release and indoor pollutant concentration through multi-layer nonlinear transformation. During the training process of the deep learning model, the back propagation algorithm is used to adjust the network parameters to minimize the difference between the predicted concentration and the actual concentration. Step 3: Genetic algorithm optimization model parameters Use genetic algorithms to optimize the structure and hyperparameters of deep learning models, and find the optimal network structure and parameter configuration by simulating the process of natural selection, including selection, crossover, and mutation operations, to improve the model's predictive performance and generalization ability; Step 4: Reverse identification of pollution source release intensity The trained and optimized deep learning model is applied to the reverse analysis of indoor pollutant concentrations. The genetic algorithm is combined with the actual monitored indoor pollutant concentrations to reversely calculate the release intensity of each pollution source to identify the release intensity of each pollution source. In this process, the genetic algorithm gradually approaches the actual release intensity of the pollution source through continuous iterative search. Step 5: Result verification and model adjustment The release intensity of the pollution source obtained by reverse identification is compared with the actual situation to verify the accuracy and reliability of the model. If there is a large deviation between the identification result and the actual value, it is necessary to return to steps 2 and 3 to adjust and optimize the parameters of the deep learning model and genetic algorithm until a satisfactory identification effect is achieved.

2. According to claim 1, the inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm is characterized by: In step one, the indoor environmental parameters are analyzed through a deep learning model to predict indoor air flow and pollution distribution patterns. This step first involves the comprehensive collection of indoor environmental parameters, and the collected data is preprocessed; then, a deep learning neural network is used to build a model to simulate and predict indoor air flow and the distribution of pollutants; in addition, a large amount of measured data will be used in the training process of the deep learning model to ensure the accuracy and generalization ability of the model prediction.

3. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 1 is characterized by: The parameters described in step one include temperature, humidity, and air flow rate.

4. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 1 is characterized by: The data described in step 1 are preprocessed, including data cleaning, standardization, and denoising.

5. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 1 is characterized by: In step 2, a genetic algorithm is used to optimize the parameters of the deep learning model established in step 1 to improve the accuracy and stability of the model in predicting the distribution of indoor pollutants and the release intensity of pollution sources. Specifically, this step includes the following operations: S1: Initialize genetic algorithm parameters: including population size, crossover probability, mutation probability, and the encoding method for defining individuals. Usually, the hyperparameters of the deep learning model are used as the gene encoding in the genetic algorithm. S2: Population generation: In the initial stage of the genetic algorithm, a set of hyperparameter configurations of the deep learning model is randomly generated as the initial population, and each individual represents a set of possible deep learning model configurations; S3: Fitness function evaluation: For each individual in the population, the performance of the model in the indoor pollutant distribution prediction task is evaluated according to the parameter configuration of the deep learning model it represents, and indicators such as prediction accuracy and error rate are used as fitness evaluation criteria; S4: Selection operation: According to the fitness of individuals, roulette and tournament selection methods are used to select individuals with high fitness for subsequent crossover and mutation operations to ensure the inheritance of excellent genes; S5: Crossover and mutation operations: New individuals are generated in the population through crossover and mutation operations. The crossover operation allows two individuals to exchange their genetic information, while the mutation operation randomly changes part of the individual's genes to introduce new genetic diversity. S6: Generation of a new generation of population: Repeat the selection, crossover and mutation operations to generate a new generation of population. The generation of each generation of population is an optimization of the previous generation of population. S7: Termination condition judgment: When the preset number of iterations is reached or the model performance is no longer significantly improved, the algorithm terminates, and finally the deep learning model configuration represented by the individual with the highest fitness is selected as the optimal configuration.

6. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 5 is characterized by: The hyperparameters described in step S2 include learning rate, number of network layers, and number of neurons in each layer.

7. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 1 is characterized by: Step 3 involves using the optimized deep learning model to simulate the indoor pollutant distribution based on indoor environmental parameters and measured monitoring data, and then generate a transmission matrix A between the pollution source and the monitoring point. This transmission matrix is ​​calculated based on the relationship between the pollutant concentration predicted by the model and the actual monitoring point concentration, which provides a key mathematical basis for the reverse identification of the release intensity of the pollution source, including the following steps: S1: Indoor pollutant distribution simulation: Using the deep learning model optimized in step 2, combined with indoor environmental parameters and pollution source information, simulate indoor air flow and pollutant distribution. This simulation process takes into account the complex dynamics of the indoor environment and can accurately predict the pollutant concentration at the monitoring point under different conditions. S2: Generation of transmission matrix A: Based on the simulation results of the deep learning model, the contribution rate of each pollution source to the pollutant concentration at each monitoring point is calculated to construct the transmission matrix A. Each element in the matrix represents the influence intensity of a specific pollution source on the concentration of the corresponding monitoring point, reflecting the propagation characteristics of pollutants in the indoor environment. S3: Verification and correction of the transmission matrix A: Verify the accuracy of the transmission matrix A by comparing it with the actual monitoring data. If necessary, further fine-tune the deep learning model to ensure that the transmission matrix can accurately reflect the transmission relationship between the pollution source and the monitoring point. S4: Prepare for the reverse identification process: Use the verified and corrected transfer matrix A to lay the foundation for the subsequent reverse identification work. On this basis, use the inverse algorithm to accurately infer the release intensity of each pollution source from the monitoring point concentration data.

8. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 7 is characterized by: The indoor environmental parameters in step S2 include temperature, humidity, and air flow rate.

9. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 1 is characterized by: Step 4: Using the transmission matrix A that has been constructed and verified, combined with the pollutant concentration data of the actual monitoring point, the release intensity of each indoor pollution source is accurately calculated through the inverse algorithm, including the following steps: S1: Selection and implementation of inverse algorithm: Select appropriate inverse algorithm, including least squares method and regularization method, to deal with possible ill-conditioned problems and enhance the stability of the solution process, implement the inverse algorithm, and prepare to apply it to the transmission matrix A and the concentration data of the monitoring points; S2: Inverse solution of the release intensity of pollution sources: Using the selected inverse algorithm, combined with the transmission matrix A obtained in step 3 and the actual pollutant concentration data of each monitoring point, calculation is performed to inversely infer the release intensity of each pollution source. This calculation process needs to fully consider the uncertainty of model prediction and measurement error to ensure the accuracy and reliability of the solution results; S3: Analysis and verification of release intensity results: Conduct a detailed analysis of the release intensity of the pollution source obtained by inverse solution, including the size and distribution characteristics of the release intensity, and compare it with the known pollution source characteristics to verify the rationality and accuracy of the solution results; S4: Parameter adjustment and model optimization: According to the solution results and verification analysis, if necessary, return to steps 1 and 2 to adjust and optimize the model parameters, in order to further improve the prediction accuracy of the model and the accuracy of inverse identification; S5: Result output and application: Output the final pollution source release intensity results in an appropriate form to provide a scientific basis for indoor air quality management, guide the implementation of pollution prevention and control measures and the formulation of environmental improvement plans.

10. The inverse identification method of release intensity of multiple indoor pollution sources based on deep learning and genetic algorithm according to claim 1, characterized in that: Step 5 involves the verification and application of the reverse identification results. In this step, the accuracy and reliability of the reverse identification method are verified by comparing and analyzing the actual on-site monitoring data, and the model is adjusted and optimized as necessary based on the verification results to ensure its effectiveness and applicability in practical applications. The steps are as follows: S1: Result verification: Collect the actual monitoring data corresponding to the reverse identification process, including the pollutant concentrations at each indoor monitoring point and the actual release of pollution sources, compare the pollution source release intensity obtained by reverse identification with the actual data, and evaluate the prediction accuracy of the model and the error range of reverse identification; S2: Model adjustment: Based on the problems and errors found during the result verification process, analyze the possible causes, including improper model hyperparameter settings, insufficient data preprocessing, and inaccurate transfer matrix A. Based on this, adjust and optimize the key links such as the deep learning model structure, genetic algorithm parameters, and inverse algorithm; S3: Re-verification: Re-calculate the inverse identification of the adjusted and optimized model and verify it with new actual monitoring data to ensure that the adjustments have effectively improved the prediction accuracy of the model and the accuracy of the inverse identification; S4: Application promotion: After multiple rounds of verification and adjustment, confirming that the accuracy and stability of the model have reached a satisfactory level, the inverse identification method will be applied to a wider range of indoor air quality monitoring and management. In the process of management, the reverse identification results are used to guide the formulation of pollution prevention and control measures and the implementation of environmental improvement strategies. To effectively improve indoor air quality.

Citation Information

Cited By

  • Intelligent substation secondary equipment defect diagnosis method

    CN119337215A

  • Indoor pollutant emission concentration estimation method

    CN120833870A

  • A method of estimating indoor pollutant emission concentration

    CN120833870B

  • High and large space air quality monitoring method and system based on physical information neural network

    CN121784250A

  • Method and system for monitoring air quality in high and large space based on physical information neural network

    CN121784250B