A volatile organic compound real-time tracing method and system for an industrial park

Abstract

The application discloses a volatile organic compound real-time tracing method and system for an industrial park, which comprises the following steps: acquiring VOCs component spectrum and three-dimensional meteorological field data in real time by using proton transfer time-of-flight mass spectrometry and Doppler laser radar, and aligning and fusing the VOCs component spectrum and three-dimensional meteorological field data with enterprise process data through a space-time alignment algorithm; constructing a time convolution network and embedding a chemical mechanism module for constraint to obtain a TCN-MCM hybrid model, and training data loss and mechanism loss jointly through the TCN-MCM hybrid model; deploying an INT8 quantization model on a Jetson AGX Xavier platform, and performing efficient inference and uncertainty quantization through dynamic scheduling of computing power and a Monte Carlo method; establishing a transfer learning parameter library, aligning source domain and target domain feature distributions based on a maximum mean difference loss, and establishing a parameter sharing mechanism to realize rapid deployment across parks; and the scheme can significantly improve tracing accuracy and real-time performance, reduce cross-scene deployment costs, and is suitable for precise pollution control in complex industrial environments.

Description

Technical Field

[0001] This invention relates to the fields of environmental monitoring and industrial Internet of Things (IoT) technology, specifically to a method and system for real-time tracing of volatile organic compounds (VOCs) in industrial parks. Background Technology

[0002] Volatile organic compound (VOC) emissions from industrial parks are characterized by complex composition, strong spatiotemporal dynamics, and multiple diffusion pathways. Traditional monitoring and source tracing technologies have the following limitations:

[0003] 1. Limited data dimensions and insufficient fusion capabilities: Existing technologies mostly rely on single sensors (such as electrochemical sensors) or offline analysis (GC-MS), lacking multi-source data collaboration; existing fixed monitoring networks only collect concentration data and do not integrate three-dimensional meteorological fields and enterprise process parameters, resulting in source tracing models having errors exceeding 40% under complex meteorological conditions; research shows that models that do not consider the vertical profile of wind speed have a positioning deviation of more than 50 meters for high-altitude emission sources.

[0004] 2. Poor physical consistency of models: Mainstream data-driven models (such as LSTM and CNN) lack chemical mechanism constraints, which easily leads to physically unreliable solutions; existing pure data-driven models have a false alarm rate as high as 18% in scenarios involving a mixture of benzene series compounds and alkanes. In addition, traditional models are unable to explain the causes of pollution, which restricts regulatory decisions.

[0005] 3. Low real-time performance of edge computing: Existing edge devices have limited computing power, and the inference latency of complex models exceeds 10 minutes; the existing cloud-based centralized computing architecture has a full-process response time of 8-12 minutes due to data transmission bandwidth limitations, which cannot meet the minute-level response requirements for sudden leakage events.

[0006] 4. High cost of cross-scenario migration: The process differences between different parks require repeated training of the model, and the deployment cycle is as long as 2-3 months. Existing transfer learning methods require ≥1000 hours of labeled data in the target domain, and the adaptation accuracy is less than 70%, making them difficult to use.

[0007] Therefore, existing technologies have the following significant drawbacks: Data level: lack of spatiotemporal alignment of multi-source heterogeneous data, and coarse feature fusion; Model level: pure data-driven approach ignores physical laws and has poor interpretability; Computation level: inefficient scheduling of edge hardware resources and insufficient real-time performance; Application level: cross-campus migration relies on a large amount of labeled data, which is costly. Summary of the Invention

[0008] To address the aforementioned shortcomings of existing technologies, this invention provides a high-precision, high-real-time, high-portability, and low-energy-consumption method and system for real-time volatile organic compound (VOC) source tracing in industrial parks, thereby solving the technical problems mentioned in the background section. It is applicable to the precise location, diffusion simulation, and rapid cross-park deployment of pollution sources in industries such as chemical, petrochemical, and pharmaceutical.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] Firstly, a real-time source tracing method for volatile organic compounds (VOCs) in industrial parks is provided, comprising the following steps: S1: Real-time acquisition of VOCs component spectra and three-dimensional meteorological field data using proton transfer time-of-flight mass spectrometry (PTF-MS) and Doppler lidar, and alignment and fusion with enterprise process data through a spatiotemporal alignment algorithm; S2: Construction of a temporal convolutional network and embedding a chemical mechanism module for constraint, resulting in a TCN-MCM hybrid model, and joint training of data loss and mechanism loss through the TCN-MCM hybrid model; S3: Deployment of an INT8 quantization model on the Jetson AGX Xavier platform, and efficient inference and uncertainty quantification through dynamic scheduling of computing power and Monte Carlo methods; S4: Establishment of a transfer learning parameter library, alignment of source and target domain feature distributions based on maximum mean difference loss, and establishment of a parameter sharing mechanism to achieve rapid deployment across industrial parks.

[0011] Furthermore, the three-dimensional meteorological field data in step S1 includes wind speed, wind direction, and temperature vertical profiles, while the enterprise process data includes reactor temperature, valve status, and raw material flow process parameters in the industrial park. By aligning the sensor and process data time series through dynamic time warping, the discrete meteorological data is mapped into a three-dimensional continuous field using Kriging interpolation, and a spatiotemporally aligned fusion data cube is constructed.

[0012] Furthermore, in step S2, the temporal convolutional network adopts a 4-layer dilated causal convolutional structure with dilation coefficients of 1, 2, 4, and 8, a kernel size of 3, 64 output channels per layer, ReLU activation function, and a receptive field covering a 60-minute temporal window.

[0013] Furthermore, the chemical mechanism module embeds 30 pre-defined VOCs reaction pathways and outputs the model through a stoichiometric matrix constraint. The expression for its mechanism loss function is as follows:

[0014]

[0015] Among them, S i Let i be the stoichiometric matrix of the reaction group. These are the model's predicted values. This is the theoretical value based on the mechanism. for L 2-norm.

[0016] Furthermore, during joint training, the total loss function of the TCN-MCM hybrid model is a weighted sum of data loss and mechanism loss, with a weight ratio of 4:1.

[0017] Furthermore, the output of the temporal convolutional network is added to the mechanism constraint loss of the chemical mechanism module via skip connections to prevent gradient vanishing; the pre-defined chemical mechanisms in the chemical mechanism module are expressed in matrix form; when the mechanism loss L... mech When the value is greater than 0.1, an abnormal warning is triggered and a manual review process is initiated.

[0018] Furthermore, in step S3, the FP32 model is converted to INT8 quantization format through TensorRT optimization, and a layer-by-layer calibration method is used to reduce accuracy loss; the computing power is dynamically scheduled to automatically switch between CPU / GPU computing based on the GPU utilization threshold.

[0019] Furthermore, the method for uncertainty quantification in step S3 using the Monte Carlo method is as follows: Monte Carlo is enabled during the inference phase, 100 random forward propagations are performed, and a 95% confidence interval is output.

[0020] The output of uncertainty quantification is:

[0021] , ;

[0022] Where N=100, the confidence interval is calculated as follows: ±1.96 / .

[0023] Furthermore, the transfer learning parameter library in step S4 covers three types of industrial parks: petrochemical, pharmaceutical, and coating. The expression for the maximum mean difference loss is as follows:

[0024]

[0025] in, The eigenmap corresponding to the Gaussian kernel function, bandwidth =1.0, , The number of samples in the source and target domains. Let i be the i-th sample in the source domain. For the j-th sample in the target domain, Let Hilbert be the norm in the regenerated kernel space;

[0026] The parameter sharing mechanism includes freezing the parameters of the first three convolutional layers of the time-bound convolutional network and using a course learning strategy to fine-tune the target domain data.

[0027] Secondly, a system for real-time source tracing of volatile organic compounds (VOCs) in industrial parks is provided, comprising: a multi-source data acquisition module for acquiring VOCs component spectra, three-dimensional meteorological field data, and enterprise process data; a spatiotemporal alignment and fusion module for aligning and fusing multi-source data; a TCN-MCM hybrid model for coupling temporal feature extraction and chemical mechanism constraints; an edge computing optimization module for efficient inference and uncertainty quantification through dynamic scheduling of computing power and Monte Carlo methods; a cross-park transfer learning module for rapid deployment across parks; and a visualization platform for rendering three-dimensional pollution diffusion paths and real-time alarm push notifications.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. This solution uses proton transfer time-of-flight mass spectrometry to accurately analyze VOCs components, constructs a three-dimensional meteorological field using lidar, and integrates enterprise process logs. Through spatiotemporal alignment algorithms, data errors and positioning errors can be controlled within 1 second and 10 meters, respectively, thereby improving the accuracy of source tracing.

[0030] 2. The TCN-MCM hybrid model in this scheme serves as an organically integrated framework, coupling temporal feature extraction with chemical mechanism constraints. TCN is responsible for learning complex spatiotemporal patterns from data, while MCM is responsible for verifying and constraining the brain's thinking results using chemical rules. This achieves a balance between accuracy, physical reliability, real-time performance, and transferability, and significantly reduces the false alarm rate compared to traditional pure data-driven methods.

[0031] 3. This solution uses edge-end INT8 quantization and dynamic scheduling strategies to ensure inference latency of ≤1 second and end-to-end response time of <3 minutes, meeting the minute-level response requirements for sudden leakage events and enabling real-time rapid analysis.

[0032] 4. The transfer learning mechanism based on maximum mean difference loss significantly reduces the target domain annotation data requirement (≤100 hours), shortens the deployment cycle of new parks to 7 days, and achieves an adaptation accuracy of ≥85%;

[0033] 5. This solution reduces the power consumption of edge devices by 40%, supports 72 hours of uninterrupted operation, and has good engineering applicability and promotion value. Attached Figure Description

[0034] Figure 1 This is a flowchart of a method for real-time tracing of volatile organic compounds (VOCs) in industrial parks. Detailed Implementation

[0035] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0036] like Figure 1 As shown, one method for real-time tracing of volatile organic compounds (VOCs) in industrial parks according to this scheme includes the following steps:

[0037] S1: Real-time acquisition of VOCs component spectra and three-dimensional meteorological field data is achieved using proton transfer time-of-flight mass spectrometry and Doppler lidar, and then aligned and fused with enterprise process data through a spatiotemporal alignment algorithm;

[0038] The three-dimensional meteorological field data in step S1 includes wind speed, wind direction, and vertical temperature profiles. The enterprise process data includes reactor temperature, valve status, and raw material flow rate process parameters in the industrial park. By dynamically warping the sensor and process data time series, the discrete meteorological data is mapped into a three-dimensional continuous field using Kriging interpolation, and a spatiotemporally aligned fusion data cube is constructed. The sampling frequency of the VOCs component spectrum is 1 Hz, the sampling flow rate is 5 L / min, its mass range is 20-500 amu, and the resolution is ≥10000 m / Δm. The Doppler lidar wavelength is 1550 nm, the spatial resolution is 5 m × 5 m × 10 m, the vertical detection height is 500 m, and the horizontal scanning angle is 360°.

[0039] In step S1, baseline correction (moving average window of 10 seconds) and mass spectrum peak alignment (mass deviation tolerance ±0.01 amu) are performed on the proton transfer time-of-flight mass spectrometry data; during dynamic time warping (DTW), the alignment window width is 30 seconds; the lidar data is used to invert the three-dimensional wind field through the velocity-azimuth display (VAD) algorithm, and anisotropic kriging interpolation is used to generate a spatial continuous field with a horizontal to vertical range ratio of 3:1 and a field radius of 50m; when acquiring enterprise process data, an industrial IoT gateway is used to access the enterprise DCS system through the OPC UA protocol, and the timestamp alignment error is ≤1 second.

[0040] S2: Construct a temporal convolutional network and embed a chemical mechanism module for constraint to obtain a TCN-MCM hybrid model, and jointly train the data loss and mechanism loss through the TCN-MCM hybrid model;

[0041] In step S2, the temporal convolutional network adopts a 4-layer dilated causal convolutional structure with dilation coefficients of 1, 2, 4, and 8, a kernel size of 3, 64 output channels per layer, ReLU activation function, and a receptive field covering a 60-minute temporal window.

[0042] The chemical mechanism module embeds 30 preset VOCs reaction pathways, including benzene photolysis and alkane -OH oxidation, etc., and outputs the model through a stoichiometric matrix constraint. The expression for its mechanism loss function is as follows:

[0043]

[0044] Among them, S i Let i be the stoichiometric matrix of the reaction group. These are the model's predicted values. This is the theoretical value based on the mechanism. for L 2-norm (Euclidean distance).

[0045] During joint training, the TCN-MCM hybrid model uses a total loss function that is a weighted sum of data loss and mechanism loss with a weight ratio of 4:1. It employs the Adam optimizer with a learning rate of 1e-4 and a batch size of 32.

[0046] The output of the temporal convolutional network is added to the mechanism constraint loss of the chemical mechanism module via skip connections to prevent gradient vanishing. The pre-defined chemical mechanisms in the chemical mechanism module are expressed using a matrix representation; for example, the photolysis reaction path of benzene is encoded as: from C6H6+... hv C6H5·+H•, corresponding to the stoichiometric matrix S=[-1,+1,+1], the rate constant k1 is calculated by the Arrhenius equation; when the mechanistic loss L mech When the value is greater than 0.1, an abnormal warning is triggered and a manual review process is initiated.

[0047] S3: Deploy the INT8 quantization model on the Jetson AGX Xavier platform and perform efficient inference and uncertainty quantization through dynamic scheduling of computing power and Monte Carlo methods;

[0048] In step S3, the model is deployed on the NVIDIA Jetson AGX Xavier hardware platform with a GPU computing power of 32 TOPS. The FP32 model is converted to INT8 quantization format using TensorRT 8.4. Layer-by-layer calibration is used to reduce accuracy loss, with a maximum error of <1%. A dynamic resource scheduling strategy is adopted, setting the task priority queue to real-time tasks > offline tasks. Based on the GPU utilization threshold (>90%), the system automatically switches to CPU parallel computing (8 OpenMP threads).

[0049] Dynamic resource scheduling logic: real-time tasks (such as sudden leak events) are given priority in GPU computing resources (CUDA stream priority 0); offline tasks (such as model retraining) are limited to GPU memory usage ≤50%, and when this is exceeded, they are switched to CPU asynchronous computing.

[0050] TensorRT optimization specifically includes two steps: layer fusion and INT8 quantization calibration. The key point is to merge Conv1D+BN+ReLU in TCN into a single computing node and use the entropy minimization method to select the dynamic range. The calibration set contains 1000 sets of representative input data.

[0051] The method for uncertainty quantification in step S3 using the Monte Carlo method is as follows: enable Monte Carlo during the inference phase, perform 100 random forward propagations, and output a 95% confidence interval;

[0052] The output of uncertainty quantification is:

[0053] , ;

[0054] Where N=100, the confidence interval is calculated as follows: ±1.96 / .

[0055] S4: Establish a transfer learning parameter library, align the feature distributions of the source and target domains based on the maximum mean difference loss, and establish a parameter sharing mechanism to achieve rapid deployment across campuses.

[0056] The transfer learning parameter library in step S4 covers three types of industrial parks: petrochemical, pharmaceutical, and coating. The amount of source model training data is ≥100,000 hours. The expression for the maximum mean difference loss is:

[0057]

[0058] in, The eigenmap corresponding to the Gaussian kernel function, bandwidth =1.0, , The number of samples in the source and target domains. Let i be the i-th sample in the source domain. For the j-th sample in the target domain, Let Hilbert be the norm in the regenerated kernel space;

[0059] The parameter sharing mechanism includes freezing the parameters of the first three convolutional layers of the time-limited convolutional network. The first layer is a dilated convolution (dilation coefficient 1, kernel weights are not updated), and the second layer is a dilated convolution (dilation coefficient 2, only the bias term can be fine-tuned). A course learning strategy is used to fine-tune the target domain data, with an initial learning rate of 5e-5, a decay coefficient of 0.5 every 50 rounds, and a total number of training rounds ≤200.

[0060] Feature alignment between the source and target domains is achieved through a deep adaptation network (DAN). Multi-kernel maximum mean difference (MK-MMD) loss is introduced in three hidden layers (with the core adaptation layer having a dimension of 512) to achieve multi-layer feature alignment between the source and target domains.

[0061] This solution also provides a system for real-time source tracing of volatile organic compounds (VOCs) in industrial parks, comprising: a multi-source data acquisition module for acquiring VOCs component spectra, three-dimensional meteorological field data, and enterprise process data; a spatiotemporal alignment and fusion module for aligning and fusing multi-source data; a TCN-MCM hybrid model for coupling temporal feature extraction and chemical mechanism constraints; an edge computing optimization module for efficient inference and uncertainty quantification through dynamic scheduling of computing power and Monte Carlo methods; a cross-park transfer learning module for rapid deployment across parks; and a visualization platform for rendering three-dimensional pollution diffusion paths and real-time alarm push notifications.

[0062] The following is a deployment test of a petrochemical industrial park, which includes:

[0063] Hardware configuration: 3 PTR-TOFMS units (covering 10km²) + 1 millimeter-wave weather radar; Model training: 2000 sets of historical data were used, with a training period of 50 epochs, and the loss function converged to 0.12; Performance verification: Source tracing response time: 3 minutes and 28 seconds (target enterprise ID: E-1023); Accuracy: 93.6% (compared with manual verification results); Computational resource consumption: GPU utilization of edge devices remained stable at 63±5%.

[0064] Cross-campus migration verification includes:

[0065] Source domain: An electronics park in the Yangtze River Delta (1500 data sets); Target domain: A paint park in the Pearl River Delta (200 data sets); Transfer effect: Initial accuracy: 68.4%; After 72 hours of fine-tuning: 89.2%; Model training energy consumption: 77% lower than training from scratch.

[0066] In summary, the proposed solution for real-time volatile organic compound (VOC) tracing in industrial parks can significantly improve tracing accuracy and real-time performance, reduce cross-scenario deployment costs, and is suitable for precise pollution control in complex industrial environments.

[0067] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for real-time source tracing of volatile organic compounds (VOCs) in industrial parks, characterized in that, Includes the following steps: S1: Real-time acquisition of VOCs component spectra and three-dimensional meteorological field data is achieved using proton transfer time-of-flight mass spectrometry and Doppler lidar, and then aligned and fused with enterprise process data through a spatiotemporal alignment algorithm; S2: Construct a temporal convolutional network and embed a chemical mechanism module for constraint to obtain a TCN-MCM hybrid model, and jointly train the data loss and mechanism loss through the TCN-MCM hybrid model; In step S2, the temporal convolutional network adopts a 4-layer dilated causal convolutional structure with dilation coefficients of 1, 2, 4, and 8, a kernel size of 3, 64 output channels per layer, ReLU activation function, and a receptive field covering a 60-minute temporal window. The chemical mechanism module embeds 30 pre-set VOCs reaction pathways and outputs the model through a stoichiometric matrix constraint. The expression for its mechanism loss function is as follows: Among them, S i Let i be the stoichiometric matrix of the reaction group. These are the model's predicted values. This is the theoretical value based on the mechanism. for L 2-norm; S3: Deploy the INT8 quantization model on the Jetson AGX Xavier platform and perform efficient inference and uncertainty quantization through dynamic scheduling of computing power and Monte Carlo methods; The FP32 model is converted to INT8 quantization format through TensorRT optimization, and a layer-by-layer calibration method is used to reduce accuracy loss; the computing power is dynamically scheduled to automatically switch between CPU / GPU computing based on the GPU utilization threshold; S4: a transfer learning parameter library is established, the feature distributions of the source domain and the target domain are aligned based on the maximum mean difference loss, and a parameter sharing mechanism is established to achieve rapid deployment across campuses.

2. The method for real-time tracing of volatile organic compounds in industrial parks according to claim 1, characterized in that, The three-dimensional meteorological field data in step S1 includes wind speed, wind direction and temperature vertical profiles, and the enterprise process data includes reactor temperature, valve status and raw material flow process parameters in the industrial park. By dynamically warping and aligning the time series of sensor and process data, we can map discrete meteorological data into a three-dimensional continuous field using Kriging interpolation and construct a spatiotemporally aligned fused data cube.

3. The method for real-time tracing of volatile organic compounds in industrial parks according to claim 1, characterized in that, During joint training, the total loss function of the TCN-MCM hybrid model is a weighted sum of data loss and mechanism loss, with a weight ratio of 4:

1.

4. The method for real-time tracing of volatile organic compounds in industrial parks according to claim 3, characterized in that, The output of the temporal convolutional network is added to the mechanism constraint loss of the chemical mechanism module via skip connections to prevent gradient vanishing; the preset chemical mechanisms in the chemical mechanism module are expressed in matrix form; when the mechanism loss L... mech When the value is greater than 0.1, an abnormal warning is triggered and a manual review process is initiated.

5. The method for real-time tracing of volatile organic compounds in industrial parks according to claim 1, characterized in that, The method for uncertainty quantification in step S3 using the Monte Carlo method is as follows: enable Monte Carlo during the inference phase, perform 100 random forward propagations, and output a 95% confidence interval; The output of uncertainty quantification is: ， ； Where N=100, the confidence interval is calculated as follows: ±1.96 / .

6. The method for real-time tracing of volatile organic compounds in industrial parks according to claim 1, characterized in that, The transfer learning parameter library in step S4 covers three types of industrial parks: petrochemical, pharmaceutical, and coating. The expression for the maximum mean difference loss is: in, The eigenmap corresponding to the Gaussian kernel function, bandwidth =1.0, , The number of samples in the source and target domains. Let i be the i-th sample in the source domain. For the j-th sample in the target domain, Let Hilbert be the norm in the regenerated kernel space; The parameter sharing mechanism includes freezing the parameters of the first three convolutional layers of the time-bound convolutional network and using a course learning strategy to fine-tune the target domain data.

7. A system employing the real-time source tracing method for volatile organic compounds in industrial parks as described in any one of claims 1-6, characterized in that, include: The multi-source data acquisition module is used to acquire VOCs component spectra, three-dimensional meteorological field data, and enterprise process data; The spatiotemporal alignment and fusion module is used for the alignment and fusion of multi-source data. The TCN-MCM hybrid model is used to couple temporal feature extraction with chemical mechanism constraints. The edge computing optimization module performs efficient inference and uncertainty quantification through dynamic scheduling of computing power and Monte Carlo methods. The cross-campus migration learning module is used for rapid deployment across campuses. A visualization platform for rendering 3D pollution diffusion paths and sending real-time alarms.

Images