A method for malodorous gas component identification and concentration monitoring by fusing multi-task multi-layer perceptron

Abstract

The application provides a method for malodorous gas component identification and concentration monitoring by fusing multi-task multi-layer perceptron, which has the following characteristics: step S1, constructing a sensor array to collect original response signals of malodorous gas; step S2, preprocessing the original response signals and enhancing feature engineering to construct an enhanced feature vector; step S3, constructing a multi-task multi-layer perceptron model, which adopts a shared feature layer and a task-specific head architecture to output component identification results and concentration monitoring results; and step S4, training the multi-task multi-layer perceptron model by using the enhanced feature vector and adopting a weighted fusion loss function to cooperatively optimize identification tasks and regression tasks to obtain a trained multi-task multi-layer perceptron model. The method is accurate, anti-interference and stable for malodorous gas component identification and concentration monitoring.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, specifically to a method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors. Background Technology

[0002] Odorous gases (mainly including hydrogen sulfide, methanethiol, ammonia, and multi-component mixtures) are volatile pollutants generated during the collection, crushing, fermentation, and exhaust emission of municipal solid waste. They not only have a strong, pungent odor but also cause irreversible damage to the human respiratory and nervous systems, while simultaneously polluting the atmosphere and disrupting the ecological balance. With increasingly stringent environmental regulations in my country, the current "Odor Pollutant Emission Standard" (GB14554-1993) and "Technical Specification for Environmental Monitoring of Odor Pollution" (HJ905-2017) have set forth systematic requirements for boundary concentration limits, monitoring site selection, and quality control. Accurate detection of odorous gases has become a core requirement in the field of environmental monitoring.

[0003] Currently, the industry mainly relies on the following technical means for odor gas monitoring:

[0004] I. Laboratory Testing Techniques

[0005] Gas chromatography and gas chromatography-mass spectrometry are the mainstream methods for traditional odor gas detection. While these technologies offer high detection accuracy, they have significant limitations: First, the detection cycle is long, requiring on-site sampling followed by laboratory analysis, making real-time online monitoring impossible and thus unsuitable for rapid on-site detection scenarios. Second, the equipment is bulky and cumbersome to operate, relying on specialized technicians and making rapid on-site deployment difficult. Third, due to the use of single-sensor detection technology, when dealing with multi-component odor gases, they are susceptible to cross-interference from multiple components, resulting in low component identification accuracy and large concentration monitoring errors, failing to meet the requirements for simultaneous detection of multi-component odor gases.

[0006] II. Sensor array combined with traditional algorithm monitoring technology

[0007] To address the limitations of single sensors, the industry has gradually adopted multi-channel sensor arrays to acquire multi-dimensional signals, combined with traditional algorithms for signal processing. However, this technology still faces key technical bottlenecks: First, signal processing deficiencies: the raw signals acquired by sensor arrays contain a large amount of noise, drift, and redundant information. Traditional feature extraction methods struggle to capture the nonlinear correlation between different gas components and sensor responses, resulting in insufficient component identification accuracy. Second, algorithm logic deficiencies: traditional algorithms often employ single-task models, failing to fully utilize the inherent correlation between component categories and concentrations, leading to model redundancy and low computational efficiency. Furthermore, traditional machine learning algorithms have limited fitting capabilities for high-dimensional nonlinear sensor signals, requiring manual feature extraction, which is insufficient to handle the complex response characteristics of mixed odorous gases. Ultimately, the component identification accuracy and concentration monitoring precision fail to meet the high-precision requirements of practical applications. Third, poor environmental adaptability: existing algorithm optimization research for sensor arrays focuses primarily on improving single performance indicators, lacking collaborative optimization design for the "identification-monitoring" dual tasks. It also fails to fully consider the practical needs for anti-interference and high stability in odorous gas detection, lacking reliability in complex environments.

[0008] Therefore, how to achieve accurate and simultaneous improvement in the identification and concentration monitoring of odorous gas components, while strengthening the system's anti-interference and stability, has become a key issue that urgently needs to be addressed. Summary of the Invention

[0009] This invention is made to solve the above-mentioned problems, and aims to provide a method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors.

[0010] This invention provides a method for identifying and monitoring the concentration of odorous gas components by integrating a multi-task multilayer perceptron, characterized by the following steps: Step S1, constructing a sensor array to collect the raw response signals of odorous gases; Step S2, preprocessing and enhancing the feature engineering of the raw response signals to construct enhanced feature vectors; Step S3, constructing a multi-task multilayer perceptron model, which adopts an architecture of shared feature layers and task-specific heads to output component identification results and concentration monitoring results; Step S4, training the multi-task multilayer perceptron model using the enhanced feature vectors, and using a weighted fusion loss function to collaboratively optimize the identification and regression tasks to obtain a trained multi-task multilayer perceptron model; Step S5, inputting the raw response signals of odorous gases collected in real time by the sensor array, after processing in Step S2, into the trained multi-task multilayer perceptron model, thereby outputting the component categories and concentration values ​​of each component of the odorous gas.

[0011] In the odor gas component identification and concentration monitoring method integrating multi-task multilayer sensor provided by the present invention, it may also have the following features: wherein, in step S1, the sensor array includes 5 electrochemical sensors, 4 metal oxide sensors, 1 photoionization sensor and 1 temperature and humidity sensor, forming an 11-channel sensor array, used to monitor single or mixed components in hydrogen sulfide, methanethiol and ammonia, with a sampling frequency of 1 second.

[0012] The odor gas component identification and concentration monitoring method integrating multi-task multilayer sensor provided by the present invention may also have the following feature: wherein the enhanced feature engineering in step S2 includes:

[0013] Step S2-1: Calculate the feature distance of each sample data using the dynamic warping time algorithm and remove abnormal samples;

[0014] Step S2-2: Extract the time-series mean of the 10-dimensional sensor response signal as the basic response feature;

[0015] Step S2-3: Select the first two-dimensional sensor signals that are sensitive to malodorous gases, and extract their time series standard deviation and time series maximum value respectively;

[0016] Step S2-4: Concatenate the 10-dimensional mean feature, the 2-dimensional standard deviation feature, and the 2-dimensional maximum feature in sequence to form a 14-dimensional enhanced feature vector.

[0017] The odor gas component identification and concentration monitoring method integrating multi-task multilayer sensor provided by this invention may also have the following feature: wherein, in step S3, the multi-task multilayer sensor model includes a shared feature layer and a task-specific head.

[0018] The shared feature layer consists of two fully connected layers, each followed by BatchNorm1d, ReLU activation, and Dropout, with a Dropout ratio of 0.08.

[0019] The task-specific head includes a recognition head and a concentration regression head. The recognition head consists of a two-layer fully connected network and is used to output the probability distribution of seven component categories.

[0020] The concentration regression head includes three independent branches designed for different gases, corresponding to the concentration regression of hydrogen sulfide, methanethiol, and ammonia, respectively. Furthermore, the hydrogen sulfide and methanethiol branches employ a 2-layer fully connected network, while the ammonia branch employs a 3-layer fully connected network with an additional BatchNorm1d layer.

[0021] The odor gas component identification and concentration monitoring method integrating multi-task multilayer sensor provided by the present invention may also have the following feature: the seven component categories include: hydrogen sulfide, methanethiol, ammonia, hydrogen sulfide and methanethiol, hydrogen sulfide and ammonia, methanethiol and ammonia, hydrogen sulfide and methanethiol, and ammonia.

[0022] The odor gas component identification and concentration monitoring method integrating multi-task multilayer sensor provided by this invention may also have the following feature: wherein, in step S4, the weighted fusion loss function is a linear weighted sum of the identification loss and the regression loss.

[0023] CrossEntropyLoss was used for identification loss, MSELoss for regression loss, AdamW was used for the optimizer, and StepLR was used for the scheduler.

[0024] In the odor gas component identification and concentration monitoring method fused with multi-task multilayer perceptron provided by the present invention, it may also have the following features: wherein, in step S2, the enhanced feature vector needs to be normalized before being input into the multi-task multilayer perceptron model; the training set and test set used in step S2 are randomly divided in a ratio of 8:2, the random seed is 42, and LabelEncoder is used to encode the 7 component category labels for adapting to multi-identification task training;

[0025] MinMaxScaler is used to normalize the 14-dimensional feature vector to the [0,1] interval;

[0026] The concentration labels for hydrogen sulfide and methanethiol were normalized using MinMaxScaler, while the concentration labels for ammonia were normalized using StandardScaler.

[0027] In the odor gas component identification and concentration monitoring method fused with multi-task multilayer sensor provided by the present invention, it may also have the following features: wherein, the training process in step S4 adopts the following parameters: the dataset is divided into training set and test set in a ratio of 8:2, the batch size is 32, and the total number of training rounds is 1300 rounds.

[0028] In the odor gas component identification and concentration monitoring method integrating multi-task multilayer sensor provided by the present invention, it may also have the following features: wherein, the concentration monitoring results in step S5 need to be inversely normalized: hydrogen sulfide and methanethiol are subjected to MinMaxScaler inverse transformation, and ammonia is subjected to StandardScaler inverse transformation, in order to restore the true concentration value in ppm, and the component category is determined from the 7 component probability distributions output by the multi-task multilayer sensor model through the argmax function.

[0029] The role and effect of invention

[0030] The method for identifying and monitoring the concentration of malodorous gas components by integrating multi-task multilayer sensors according to the present invention solves the technical problem that traditional laboratory detection techniques cannot collect gas signals online in real time by constructing a sensor array and collecting the original response signals of malodorous gases.

[0031] The odor gas component identification and concentration monitoring method involving the present invention solves the technical problems of traditional feature extraction methods, which are difficult to capture the nonlinear correlation between gas and sensor response and low feature recognition, by preprocessing and enhancing the feature engineering of the original response signal and constructing an enhanced feature vector. This method addresses the technical problems of the original signal collected by the sensor array containing a large amount of noise, drift and redundant information, and the fact that the original signal is difficult to capture by the sensor response. The method achieves the technical effects of eliminating abnormal signals, mining effective features and improving the distinguishability of gas component features. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of a method for identifying and monitoring the concentration of odorous gas components by integrating multi-task, multi-layer sensors. Detailed Implementation

[0033] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate a method for identifying and monitoring the concentration of odorous gas components that integrates a multi-task multilayer sensor.

[0034] Figure 1 This is a schematic diagram of a method for identifying and monitoring the concentration of odorous gas components by integrating multi-task, multi-layer sensors.

[0035] like Figure 1 As shown, the odor gas component identification and concentration monitoring method integrating multi-task multilayer sensors in this embodiment includes the following steps:

[0036] Step S1: Construct a sensor array to collect the raw response signals of the malodorous gas.

[0037] In this embodiment, the sensor array in step S1 includes 5 electrochemical sensors, 4 metal oxide sensors, 1 photoionization sensor and 1 temperature and humidity sensor, forming an 11-channel sensor array. After the 11-channel sensor array is zero-point calibrated and standard gas calibrated, it is used to monitor single or mixed components in hydrogen sulfide, methanethiol and ammonia.

[0038] Concentration gradients were set for three odorous gases: hydrogen sulfide, methanethiol, and ammonia. The concentration gradients for hydrogen sulfide and methanethiol were 0.1, 0.5, 1, and 2 ppm, while the concentration gradient for ammonia was 0.1, 0.5, 1, and 5 ppm. The sensor array sampled at a frequency of 1 second. The raw response signals of the collected odorous gases were stored in an Excel file on the terminal device. The storage file contained key information such as the sensor response values ​​of each channel of the sensor array and the acquisition time.

[0039] Simultaneously, a label file was created in Excel format. The label file contained the sample ID, component category labels, and concentration labels for the three gases corresponding to the sample. The component category labels were: hydrogen sulfide label, methanethiol label, ammonia label, hydrogen sulfide and methanethiol label, hydrogen sulfide and ammonia label, methanethiol and ammonia label, and hydrogen sulfide, methanethiol, and ammonia label. The concentration labels for the three gases were: hydrogen sulfide concentration label, methanethiol concentration label, and ammonia concentration label, all in ppm. The sample ID was used to achieve precise matching between the raw response signals of the odorous gases collected by the sensor array and the component category and concentration labels in the label file.

[0040] Step S2 involves preprocessing and enhancing the feature engineering of the original response signal to construct an enhanced feature vector.

[0041] In this embodiment, step S2 is based on the original odor gas response signal and corresponding storage file and tag file collected and completed by step S1 for tag matching. The data preprocessing and enhanced feature engineering are performed. The specific operation is as follows: Step S2-1, the feature distance of each sample data is calculated by using the dynamic regularization time algorithm. Based on the feature distance threshold, abnormal and unreasonable storage files caused by factors such as equipment vibration and environmental fluctuations are removed, thereby removing abnormal samples and ensuring the validity of the input data.

[0042] Step S2-2: Extract the time-series mean of the 10-dimensional sensor response signal and retain the basic response characteristics of each sensor to different odor gas components as the basic response features;

[0043] Step S2-3: Select the first two-dimensional sensor signals that are sensitive to malodorous gases, and extract their time series standard deviation and time series maximum value respectively to enhance the signal differentiation of different components and mixed components of gases;

[0044] Steps S2-4 involve sequentially concatenating the 10-dimensional mean feature, the 2-dimensional standard deviation feature, and the 2-dimensional maximum feature to form a 14-dimensional enhanced feature vector, thus addressing the issues of insufficient information and low recognition of traditional single features.

[0045] The 14-dimensional enhanced feature vector is normalized in the input: the training set and test set used in step S2 are randomly divided in an 8:2 ratio, with a random seed of 42. LabelEncoder is used to encode the 7 component category labels for training to adapt to multiple recognition tasks.

[0046] MinMaxScaler was used to normalize the 14-dimensional feature vector to the [0,1] interval to eliminate the influence of differences in the signal dimensions of different sensors.

[0047] The concentration labels for hydrogen sulfide and methanethiol were normalized using MinMaxScaler, and the concentration labels for ammonia were normalized using StandardScaler to improve the model's adaptability to high ammonia nitrogen scenarios and training stability.

[0048] Step S3: Construct a multi-task multilayer perceptron model. The multi-task multilayer perceptron model adopts an architecture of shared feature layer and task-specific head, which is used to output component identification results and concentration monitoring results.

[0049] In this embodiment, the multi-task multilayer perceptron model in step S3 includes a shared feature layer and a task-specific head.

[0050] The shared feature layer consists of two fully connected network layers, with each layer followed by BatchNorm1d, ReLU activation, and Dropout. The Dropout ratio is 0.08, which is used to extract general features for cross-component identification and concentration regression dual tasks, while effectively suppressing overfitting of multi-task multilayer perceptron models.

[0051] The task-specific head includes an identification head and a concentration regression head. The identification head consists of a two-layer fully connected network and is used to output the probability distribution of seven component categories. The seven component categories are: hydrogen sulfide, methanethiol, ammonia, hydrogen sulfide and methanethiol, hydrogen sulfide and ammonia, methanethiol and ammonia, hydrogen sulfide and methanethiol, and ammonia.

[0052] The concentration regression head includes three independent branches for different gases, corresponding to the concentration regression of hydrogen sulfide, methanethiol, and ammonia, respectively. The hydrogen sulfide and methanethiol branches employ a 2-layer fully connected network, while the ammonia branch uses a 3-layer fully connected network with an additional BatchNorm1d layer. This multi-task multilayer perceptron model then outputs the probability of identifying odorous gas components and the normalized concentration monitoring results.

[0053] Step S4: Train the multi-task multilayer perceptron model using enhanced feature vectors, and use a weighted fusion loss function to collaboratively optimize the recognition and regression tasks to obtain the trained multi-task multilayer perceptron model.

[0054] In this embodiment, the enhanced feature vector used in step S4 is the 14-dimensional enhanced feature vector obtained in step S2. The weighted fusion loss function in step S4 is a linear weighted sum of the recognition loss and the regression loss. The recognition loss uses CrossEntropyLoss, the regression loss uses MSELoss, the optimizer uses AdamW, and the scheduler uses StepLR.

[0055] In step S4, the training dataset is divided into training and testing sets in an 8:2 ratio, with a batch size of 32 and a total of 1300 training rounds. After training, the multi-task multilayer perceptron model with the best performance on the testing set is selected as the final trained multi-task multilayer perceptron model.

[0056] Step S5: The raw response signal of the odorous gas collected in real time by the sensor array is processed by step S2 and then input into the trained multi-task multilayer perceptron model, thereby outputting the component categories and concentration values ​​of each component of the odorous gas.

[0057] In this embodiment, the concentration monitoring results in step S5 are subjected to inverse normalization: hydrogen sulfide and methanethiol are subjected to inverse MinMaxScaler transformation, and ammonia is subjected to inverse StandardScaler transformation, to restore the true concentration value in ppm. The component category of the current odor gas is determined from the probability distribution of 7 components output by the multi-task multilayer sensor model using the argmax function. Finally, the model synchronously outputs the accurate component category of the odor gas and the true concentration value of each component, completing the integrated real-time monitoring of odor gas component category identification and concentration.

[0058] Verification of the field application effect of the odor gas component identification and concentration monitoring method integrating multi-task multilayer sensors in this embodiment:

[0059] An 11-channel sensor array based on this method was fixed at a pre-designated odor gas monitoring node within a wet waste plant workshop, 1.5m from the gas emission source. 24-hour continuous online monitoring was conducted, and the entire detection process was stable with no downtime or data anomalies. The on-site component identification accuracy at the wet waste plant workshop remained consistently above 90%. The concentration monitoring accuracy fully met the actual on-site detection and treatment requirements. Specifically, the R² values ​​for hydrogen sulfide, methanethiol, and ammonia concentration regressions were 0.932, 0.925, and 0.918, respectively, with MAE all less than 0.08 ppm and RMSE all less than 0.1 ppm. This on-site application verification demonstrates that this method can accurately and in real-time reflect the component changes and concentration fluctuations of odor gases in the workshop, providing reliable technical data support for on-site odor pollution control, ventilation system optimization, and operator safety protection.

[0060] The role and effect of the embodiments

[0061] According to the odor gas component identification and concentration monitoring method involving multi-task multilayer perceptron involved in this embodiment, the multi-task multilayer perceptron model adopts an architecture of shared feature layer and task-specific head to output component identification results and concentration monitoring results. This solves the technical problems of traditional algorithms that use single-task models, do not utilize the inherent correlation between component categories and concentrations, have model redundancy and low computational efficiency, and have limited ability to fit high-dimensional nonlinear signals.

[0062] Furthermore, by constructing an 11-channel sensor array consisting of 5 electrochemical sensors, 4 metal oxide sensors, 1 photoionization sensor, and 1 temperature and humidity sensor, and using a sampling frequency of 1 second to complete the acquisition of raw response signals of odorous gases, and by achieving accurate matching of signals and tags through sample IDs, the technical problems of poor selectivity of traditional single sensors, severe cross-interference of multi-component gases, and the inability of laboratory detection techniques to acquire signals online in real time have been solved. This has achieved multi-channel, continuous, and real-time signal acquisition of single and mixed components of hydrogen sulfide, methanethiol, and ammonia.

[0063] Furthermore, by eliminating outliers through a dynamic warping time algorithm and constructing a 14-dimensional enhanced feature vector using a feature extraction method combining 10-dimensional time-series mean, 2-dimensional time-series standard deviation, and 2-dimensional time-series maximum value, and by applying normalization processing to the feature vector and concentration label, the technical problems of low recognition rate of traditional feature extraction methods and the impact of differences in signal dimensions on model training performance were solved. This approach achieved the effects of eliminating invalid data, mining effective features of gas components, and enhancing the distinguishability of signals from different components. At the same time, it improved the model's adaptability and training stability in high ammonia nitrogen scenarios, laying a high-quality data foundation for accurate model inference.

[0064] Furthermore, by training the multi-task multilayer perceptron model using enhanced feature vectors and co-optimizing the recognition and regression tasks using a weighted fusion loss function, a well-trained multi-task multilayer perceptron model is obtained. This achieves a collaborative optimization design for the "recognition-monitoring" dual tasks, solving the technical problems of model performance imbalance caused by single-task training and difficulty in simultaneously improving the accuracy of recognition and concentration monitoring. This enhances the technical effect of improving the model's accuracy in recognizing odorous gas components and the accuracy of concentration monitoring.

[0065] Furthermore, the raw response signals acquired in real time by the sensor array are processed in step S2 and then input into the trained multi-task multilayer perceptron model. The component categories are determined by the argmax function, and differentiated inverse normalization processing is used for different gases to restore the true concentration values. This solves the technical problems of traditional detection technologies, such as the inability to achieve real-time online identification and synchronous concentration monitoring of multi-component odorous gases, cumbersome on-site detection processes, delayed result output, and low concentration restoration accuracy. It achieves integrated, real-time, and high-precision detection of odorous gas components and concentrations, thereby synchronously outputting accurate component categories and true concentration values ​​in ppm.

[0066] Furthermore, by applying this method in a real-world wet waste plant workshop and completing 24-hour continuous online monitoring verification, a component identification accuracy rate of over 90% was achieved. The R² values ​​for the regression of hydrogen sulfide, methanethiol, and ammonia concentrations reached 0.932, 0.925, and 0.918, respectively, with MAE all less than 0.08 ppm and RMSE all less than 0.1 ppm. This verified the stability and anti-interference ability of this method in complex field environments and solved the technical problems of poor environmental adaptability and insufficient detection reliability in complex scenarios of existing sensor arrays combined with traditional algorithms.

[0067] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors, characterized in that, Includes the following steps: Step S1: Construct a sensor array to collect the raw response signals of the malodorous gas; Step S2: Preprocess and enhance the feature engineering of the original response signal to construct an enhanced feature vector; Step S3: Construct a multi-task multilayer perceptron model. The multi-task multilayer perceptron model adopts an architecture of shared feature layer and task-specific head, which is used to output component identification results and concentration monitoring results. Step S4: The enhanced feature vector is used to train the multi-task multilayer perceptron model. A weighted fusion loss function is used to collaboratively optimize the recognition task and the regression task to obtain the trained multi-task multilayer perceptron model. Step S5: The raw response signal of the odorous gas collected in real time by the sensor array is processed by step S2 and then input into the trained multi-task multilayer perceptron model, thereby outputting the component categories and concentration values ​​of each component of the odorous gas.

2. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 1, characterized in that: in, The sensor array described in step S1 includes 5 electrochemical sensors, 4 metal oxide sensors, 1 photoionization sensor and 1 temperature and humidity sensor, forming an 11-channel sensor array for monitoring single or mixed components in hydrogen sulfide, methanethiol and ammonia, with a sampling frequency of 1 second.

3. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 2, Its features are: in, The enhanced feature engineering in step S2 includes: Step S2-1: Calculate the feature distance of each sample data using the dynamic warping time algorithm and remove abnormal samples; Step S2-2: Extract the time-series mean of the 10-dimensional sensor response signal as the basic response feature; Step S2-3: Select the first two-dimensional sensor signals that are sensitive to malodorous gases, and extract their time series standard deviation and time series maximum value respectively; Step S2-4: The 10-dimensional mean feature, the 2-dimensional standard deviation feature, and the 2-dimensional maximum value feature are concatenated in sequence to form a 14-dimensional enhanced feature vector.

4. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 1, characterized in that: in, The multi-task multilayer perceptron model described in step S3 includes a shared feature layer and a task-specific head. The shared feature layer consists of two fully connected layers, each followed by BatchNorm1d, ReLU activation, and Dropout, with the Dropout ratio being 0.

08. The task-specific head includes a recognition head and a concentration regression head. The recognition head consists of a two-layer fully connected network and is used to output the probability distribution of seven component categories. The concentration regression head includes three independent branches for different gases, corresponding to the concentration regression of hydrogen sulfide, methanethiol, and ammonia, respectively. The hydrogen sulfide and methanethiol branches use a 2-layer fully connected network, while the ammonia branch uses a 3-layer fully connected network with an additional BatchNorm1d layer.

5. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 4, characterized in that: in, The seven component categories include: hydrogen sulfide, methanethiol, ammonia, hydrogen sulfide and methanethiol, hydrogen sulfide and ammonia, methanethiol and ammonia, hydrogen sulfide and methanethiol, and ammonia.

6. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 5, characterized in that: in, The weighted fusion loss function mentioned in step S4 is a linear weighted sum of the identification loss and the regression loss. CrossEntropyLoss was used for identification loss, MSELoss for regression loss, AdamW was used for the optimizer, and StepLR was used for the scheduler.

7. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 6, characterized in that: in, The enhanced feature vectors described in step S2 need to be normalized before being input into the multi-task multilayer perceptron model: the training set and test set used in step S2 are randomly divided in an 8:2 ratio, with a random seed of 42. LabelEncoder is used to encode the 7 component category labels for adapting to multi-recognition task training. The 14-dimensional feature vector is normalized to the [0,1] interval using MinMaxScaler; The concentration labels for hydrogen sulfide and methanethiol were normalized using MinMaxScaler, while the concentration labels for ammonia were normalized using StandardScaler.

8. The method for identifying and monitoring the concentration of odorous gas components by integrating multi-task multilayer sensors according to claim 7, characterized in that: in, The training process described in step S4 uses the following parameters: the dataset is divided into training and test sets in an 8:2 ratio, the batch size is 32, and the total number of training rounds is 1300.

9. The method for identifying and monitoring the concentration of odorous gas components by integrating a multi-task multilayer sensor according to claim 8, characterized in that: in, The concentration monitoring results in step S5 need to be reverse normalized: hydrogen sulfide and methanethiol are subjected to MinMaxScaler inverse transformation, and ammonia is subjected to StandardScaler inverse transformation to restore the true concentration value in ppm. The component categories are determined from the probability distribution of the 7 components output by the multi-task multilayer perceptron model through the argmax function.

Images