A method and system for monitoring and early warning of volatile organic compound emissions in a composite production line

Abstract

This invention discloses a method and system for monitoring and early warning of volatile organic compound (VOC) emissions from a composite production line, belonging to the field of VOC emission monitoring technology. The method acquires time-series data of VOC concentration, equipment status parameters, and environmental ventilation volume synchronously collected by array-type sensor groups in each section of the composite production line. It performs adaptive signal decomposition processing based on multi-scale sample entropy on the concentration time-series data of each section, separating the intrinsic fluctuation mode components contributing to local emissions and the extrinsic fluctuation mode components causing transmission interference from neighboring sections. Using the intrinsic fluctuation mode components and equipment status parameter time-series data as input, a dynamic causal graph network is constructed through a time-convolution gated cyclic unit to generate an emission transmission delay matrix. Based on this matrix and ventilation volume, a dynamic pollution contribution deviation index is calculated, triggering an early warning and locating the abnormal section. This invention can eliminate transmission interference and achieve accurate source tracing of abnormal emissions.

Description

Technical Field

[0001] This invention relates to the field of volatile organic compound (VOC) emission monitoring technology, specifically to a method and system for monitoring and early warning of VOC emissions from a composite production line. Background Technology

[0002] Composite production lines typically consist of multiple continuously arranged process sections, each emitting volatile organic compounds (VOCs) during processes such as coating, printing, and drying. Because these sections are spatially adjacent and share a ventilation system, VOCs emitted from any section can easily diffuse and be transmitted to adjacent sections by airflow. This results in the concentration time-series signals collected by sensors in each section being a mixture of local emission contributions and transmission interference from neighboring sections. Existing monitoring methods mostly directly set fixed thresholds for alarms on the raw concentration signals or use simple correlation analysis to estimate the inter-section impact. These methods fail to effectively decouple local emissions from transmission interference, leading to frequent instances where abnormal warnings point to the wrong section, and the accuracy of source tracing is insufficient to meet actual control requirements.

[0003] Existing signal processing methods often perform overall empirical mode decomposition or wavelet denoising on concentration time-series data, failing to distinguish the differences in how different modal components characterize local emissions and transmission interference. The extracted feature sequences are therefore unable to accurately reflect the emission dynamics of the specific work section. In analyzing causal interactions between work sections, conventional methods typically assume a fixed time lag or use linear metrics such as the Pearson correlation coefficient to assess the correlation, neglecting the dynamic impact of ventilation fluctuations and changes in operating conditions on emission transmission delays. This makes it impossible to obtain real-time and accurate emission transmission lag relationships between work sections. When abnormal emissions occur in a work section, due to the lack of precise transmission delay information, existing technologies struggle to trace the scope of impact and the source of responsibility, often misclassifying neighboring affected work sections as the source of the abnormality.

[0004] To address the above issues, it is necessary to solve how to accurately separate the local emission contribution of each section from the mixed concentration signal, eliminate the transmission interference from adjacent sections, and dynamically capture the causal relationship and time delay of volatile organic compound emission transmission between sections, so that the pollution contribution calculation can adapt to changes in ventilation conditions and accurately locate the abnormal source section. Summary of the Invention

[0005] This invention provides a method and system for monitoring and early warning of volatile organic compound (VOC) emissions from a composite production line. It addresses the problems of existing monitoring methods being unable to accurately decouple local emissions from transmission interference in mixed concentration signals and being unable to dynamically acquire emission transmission delays between production sections. This enables accurate early warning of abnormal VOC emissions from composite production lines and the location of the source production section.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] This invention provides a method for monitoring and early warning of volatile organic compound emissions from a composite production line, comprising:

[0008] Acquire time-series data of volatile organic compound concentration, equipment status parameters, and environmental ventilation volume from array-type sensor groups configured in each section of the composite production line.

[0009] Adaptive signal decomposition processing based on multi-scale sample entropy is performed on the time series data of volatile organic compound concentration in each section to separate the endogenous fluctuation mode components that are contributed by local emissions and the exogenous fluctuation mode components that are transmitted interference from neighboring sections. The mode complexity is used to distinguish between locally generated fluctuations and transmission interference, and to eliminate the interference of crosstalk from neighboring sections on the judgment of local emissions.

[0010] Using the intrinsic fluctuation mode components and the time series data of the equipment state parameters as input, a dynamic causal graph network based on time convolution gated recurrent units is constructed to generate a volatile organic compound emission transmission delay matrix between each section, thereby realizing a deep correlation between the equipment operating status and local emission characteristics and accurately restoring the emission transmission path and time lag relationship across sections.

[0011] Based on the emission propagation delay matrix and the environmental ventilation volume time series data, the dynamic pollution contribution deviation index of each section in the current sampling window is calculated. The emission propagation characteristics are integrated with the ventilation diffusion conditions to quantify the degree of deviation of the pollution contribution of each section relative to the normal operating conditions in real time.

[0012] Based on the comparison results between the dynamic pollution contribution deviation index and the preset multi-level early warning threshold range, the corresponding level of emission early warning signal is triggered and the abnormal section is located, so as to realize accurate source tracing and graded response to the abnormal emission section.

[0013] In a preferred embodiment of the present invention, when separating the intrinsic and extrinsic wave mode components, empirical mode decomposition is performed on the time-series data of volatile organic compound concentration for each process segment to obtain multiple intrinsic mode function components and one residual component. The sample entropy value of each intrinsic mode function component is calculated and sorted in ascending order. The intrinsic mode function components whose entropy values ​​fall within the first preset percentage interval after sorting are accumulated and reconstructed to generate the intrinsic wave mode component. The intrinsic mode function components whose entropy values ​​fall within the last preset percentage interval after sorting are accumulated and reconstructed to generate the extrinsic wave mode component. This decomposition method adaptively distinguishes between local ordered fluctuations and transmitted random interference based on signal complexity, improving separation accuracy.

[0014] Preferably, when generating the volatile organic compound (VOC) emission transmission delay matrix between different work sections, the intrinsic fluctuation mode components of each work section are concatenated with the time-series data of the equipment state parameters of that work section along the time axis to obtain the enhanced emission feature sequence of each work section. The enhanced emission feature sequences of all work sections are aligned by time steps and then input into the temporal convolutional gated recurrent unit. This unit extracts the local temporal patterns of each work section through temporal convolutional layers and captures long-term dependencies across work sections through gated recurrent unit layers. The causal correlation strength time series between every two different work sections is extracted from the output state of the gated recurrent unit layer. Peak detection processing is performed on the causal correlation strength time series, and the lag time step corresponding to each peak is determined as the emission transmission delay value between the two work sections. All emission transmission delay values ​​constitute the emission transmission delay matrix. During this process, the expansion rate of the temporal convolutional layer in the temporal convolutional gated recurrent unit is dynamically adjusted according to the maximum delay value of the emission transmission delay matrix to adapt to changes in emission propagation speed under different operating conditions.

[0015] Furthermore, when extracting the causal association strength time series, the historical hidden state sequence of the first work segment and the current hidden state vector of the second work segment are obtained from the output state of the gated recurrent unit layer; each historical hidden state in the historical hidden state sequence of the first work segment is multiplied by the current hidden state vector of the second work segment to generate an initial association scoring sequence; attention weight normalization processing based on learnable query vectors is performed on the initial association scoring sequence to generate the causal association strength time series. The attention mechanism automatically focuses on key influencing points in historical moments, improving the accuracy of time delay estimation and the ability to capture complex temporal dependencies.

[0016] In one technical solution of the present invention, when calculating the dynamic pollution contribution deviation index of each work section in the current sampling window, for the target work section, all transmission delay values ​​of the work section as the upstream source are extracted from the emission transmission delay matrix to form the pollution propagation feature vector of the work section; a sliding window segmentation process is performed on the environmental ventilation volume time series data, and the coefficient of variation of the environmental ventilation volume in the current sampling window is calculated as the diffusion condition feature value of the work section; the pollution propagation feature vector and the diffusion condition feature value are concatenated and input into a pre-trained deviation index regressor, which outputs the dynamic pollution contribution deviation index. This index integrates emission delay characteristics and real-time ventilation fluctuations, and can sensitively reflect the dynamic changes in the emission contribution of the work section.

[0017] Preferably, during the calculation of the dynamic pollution contribution deviation index, when the coefficient of variation of the environmental ventilation volume time series data exceeds the preset ventilation fluctuation threshold, a ventilation volume weighted correction factor is introduced to dynamically attenuate and compensate the deviation index, thereby avoiding false alarms caused by instantaneous changes in ventilation.

[0018] Furthermore, before inputting the pollution propagation feature vector and diffusion condition feature value into the deviation index regressor, outlier pruning based on quantiles is performed on the pollution propagation feature vector to obtain a truncated propagation feature vector. The truncated propagation feature vector and the diffusion condition feature value are then dimensionally aligned and concatenated to generate a fused feature vector. The fused feature vector is then input into the deviation index regressor, which is composed of multiple fully connected layers stacked together. Through layer-by-layer nonlinear mapping, a dynamic pollution contribution deviation index is output, improving the robustness of the deviation index to abnormal time delay values ​​and noise.

[0019] As another preferred embodiment of the present invention, when triggering an emission early warning signal and locating an abnormal work section, the dynamic pollution contribution deviation index is compared with a first early warning threshold and a second early warning threshold, wherein the second early warning threshold is greater than the first early warning threshold. When the dynamic pollution contribution deviation index is greater than or equal to the first early warning threshold and less than the second early warning threshold, a first-level emission early warning signal is triggered, and the abnormal work section is recorded in the pending confirmation list. When the dynamic pollution contribution deviation index is greater than or equal to the second early warning threshold, a second-level emission early warning signal is triggered, and the abnormal work section is directly marked as the source responsibility work section. The two-level early warning mechanism balances the timeliness of early warning with the accuracy of responsibility identification.

[0020] When a Level 1 warning signal is generated, the intrinsic wave mode components of the upstream and downstream adjacent sections corresponding to the abnormal section identifier are extracted within the current sampling window. The dynamic time curvature distances between the intrinsic wave mode components of the abnormal section identifier and the intrinsic wave mode components of the upstream and downstream adjacent sections are calculated. When the dynamic time curvature distance shows a higher similarity to the upstream adjacent section than to the downstream adjacent section, an upstream tracing identifier is added to the abnormal section identifier in the pending confirmation list; when the similarity to the downstream adjacent section is even higher, a downstream tracing identifier is added. This step can indicate the possible source direction of pollution transmission in the initial warning stage, assisting in rapid investigation.

[0021] After triggering an early warning and locating the abnormal work section, the system can retrieve the historical dynamic pollution contribution deviation index sequence of the abnormal work section within the same equipment status parameter range from the historical early warning database based on the identified abnormal work section identifier. The system then applies an autoregressive integral moving average model to the historical dynamic pollution contribution deviation index sequence to generate a deviation index prediction trend line for the abnormal work section. The system compares the current dynamic pollution contribution deviation index with the predicted value at the corresponding time on the deviation index prediction trend line to calculate the deviation rate. When the deviation rate exceeds a preset drift threshold, an equipment performance degradation early warning is generated and associated with the time-series data of the equipment status parameters of the abnormal work section. This allows for the detection of equipment performance degradation trends before emissions exceed limits, supporting preventative maintenance.

[0022] Furthermore, after comparing the dynamic pollution contribution deviation index with the multi-level warning threshold interval, the comparison results, the triggered emission warning signal level, and the identified abnormal work section are packaged to generate a warning event record. This warning event record is used as a feedback tag and, together with the time-series data of volatile organic compound concentration, equipment status parameters, and environmental ventilation volume within a preset time window before the warning is triggered, forms a training sample pair. The connection weights of the time-convolutionally gated recurrent units in the dynamic causal graph network are incrementally fine-tuned using the training samples to obtain the updated dynamic causal graph network. Through feedback learning from real-time warning events, the model continuously adapts to production line condition drift, maintaining the long-term accuracy of causal time delay inference and deviation index regression.

[0023] This invention also includes a volatile organic compound (VOC) emission monitoring and early warning system for a composite production line. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the aforementioned VOC emission monitoring and early warning method for a composite production line. This system can be integrated into a production line monitoring platform to continuously monitor, isolate interferences, trace causes, and provide tiered early warnings for VOC emissions from multiple sections of the composite production line, improving the timeliness of emission anomaly detection and the accuracy of responsibility determination.

[0024] The beneficial effects of this invention are:

[0025] By employing adaptive signal decomposition based on multi-scale sample entropy, empirical mode decomposition is performed on the time-series data of volatile organic compound (VOC) concentrations for each work section to obtain multiple intrinsic mode function (EMF) components. The sample entropy of each component is calculated and sorted according to its value. Low sample entropy components are reconstructed as internal source fluctuation mode components, and high sample entropy components are reconstructed as external source fluctuation mode components. Low sample entropy represents strong regularity and self-similarity within the signal, mainly reflecting the local emission dynamics caused by the operation of equipment within the work section itself; high sample entropy corresponds to random interference introduced by airflow transmission and the influence of neighboring work sections. This separation method can remove transmission interference from neighboring work sections from the mixed signal, ensuring that the signal components input to subsequent causal analysis steps are mainly composed of local emission contributions, eliminating the confusion effect of external interference on inter-work section causal inference, thereby improving the accuracy of the emission transmission delay matrix and the reliability of the evaluation of abnormal pollution contribution.

[0026] Using endogenous fluctuation mode components and equipment state parameter time-series data as input, a dynamic causal graph network is constructed using temporal convolutional gated recurrent units to generate emission transmission delay matrices between different work sections. The temporal convolutional layer extracts local temporal patterns from the enhanced emission feature sequences of each work section through dilated convolution, while the gated recurrent unit layer captures long-range dependencies across work sections. From the hidden states of the gated recurrent unit layer, an attention mechanism is used to extract the time series of causal association strength between any two different work sections, and peak detection determines the lag time corresponding to each peak as the emission transmission delay value. The resulting delay matrix adaptively follows changes in environmental ventilation volume and work section operating status, reflecting the dynamic lag relationship of emission propagation between work sections in real time. Based on this delay matrix and combined with the environmental ventilation volume variation characteristics, a dynamic pollution contribution deviation index is calculated, enabling accurate determination of the actual contribution of each work section to overall pollution even under fluctuating ventilation conditions. When an early warning is triggered, the deviation index and upstream / downstream propagation relationships are used to locate abnormal work section identifiers, effectively distinguishing between the true source work section and adjacent work sections affected by transmission. Attached Figure Description

[0027] The invention will now be further described with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart of a method for monitoring and early warning of volatile organic compound emissions from composite production lines;

[0029] Figure 2 This is a flowchart for calculating the dynamic pollution contribution deviation index;

[0030] Figure 3 This is a flowchart of a method for monitoring and early warning of volatile organic compound emissions from composite production lines. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] See Figure 1 This invention provides a method for monitoring and early warning of volatile organic compound (VOC) emissions from a composite production line. This method is implemented by a processor in a computer system executing a computer program stored in memory. During operation, the method acquires time-series data on VOC concentration, equipment status parameters, and environmental ventilation volume synchronously collected by pre-configured array sensor groups in each section of the composite production line. These three types of data are strictly aligned on the time axis. After acquiring the data, adaptive signal decomposition processing based on multi-scale sample entropy is performed on the VOC concentration time-series data of each section. This processing coarsely granulates the concentration signal across multiple scales, calculates the sample entropy at each scale, and adaptively selects the intrinsic mode function (IMF) components based on the distribution characteristics of the sample entropy. Components with low sample entropy characteristics are classified as internal wave mode components, reflecting the local emission contribution of the section; components with high sample entropy characteristics are classified as external wave mode components, reflecting transmission interference from neighboring sections. After signal separation, a dynamic causal graph network based on time-convolutional gated recurrent units (GRUs) is constructed using the intrinsic fluctuation mode components and equipment state parameter time-series data as input. This network extracts local time-series patterns from the input sequences of each work section through time-convolutional layers, captures long-term dependencies across work sections through GRU layers, extracts the causal correlation strength between work sections from the hidden states of the GRU layers, determines the lag time step through peak detection, and generates a volatile organic compound (VOC) emission transmission delay matrix between work sections. After obtaining the emission transmission delay matrix, it is used together with the environmental ventilation volume time-series data to calculate the dynamic pollution contribution deviation index of each work section within the current sampling window. This index reflects the degree of abnormality in the pollution contribution of the work section under the current ventilation conditions and emission transmission relationships. The calculated dynamic pollution contribution deviation index is compared with a preset multi-level warning threshold range. Based on the comparison results, the corresponding level of emission warning signal is triggered, and the abnormal work section is identified.

[0033] In practice, empirical mode decomposition (EMD) is performed on the time-series data of volatile organic compound (VOC) concentration for each section of the composite production line. For the VOC concentration time-series data of a section, EMD decomposes the original time-series signal into multiple intrinsic mode function (IMF) components with frequencies ranging from high to low, and a residual component. Each IMF component reflects the oscillation mode at different time scales in the original signal, while the residual component represents the overall trend of the signal.

[0034] For each intrinsic mode function component obtained from empirical mode decomposition, the sample entropy value of each intrinsic mode function component is calculated. The calculation process for the sample entropy value is as follows: [The text then abruptly shifts to a different topic:] ...a sequence of length... The sequence of intrinsic mode function components is denoted as Set the embedding dimension Similarity tolerance Build according to the sequence number. indivual A dimensional vector, where the dimensional vector is... The vectors are represented as , The range of values ​​is Calculate any two dimensional vector and The distance between two vectors is defined as the maximum absolute difference between corresponding elements. For each... Value, statistically satisfied And the distance is less than the similarity tolerance. of dimensional vector Count the number of , and calculate the sum of that number and . The ratio of , denoted as For all of Calculate the average value to obtain Increase the embedding dimension to Repeat the above process to build dimensional vector Statistically, those that meet the conditions Count, calculate ratio And calculate the average value to get The sample entropy values ​​of the intrinsic mode function components are calculated according to the following formula:

[0035]

[0036] in, For the embedded dimension, set as ; For similarity tolerance, it is set to the standard deviation of the original intrinsic mode function component sequence. The standard deviation of the original intrinsic mode function component sequence is times that of This multiple setting can effectively distinguish eigenmode function components of different complexities; The length of the intrinsic mode function component sequence; For the embedding dimension is The average probability of matching template pairs; For the embedding dimension is The average probability of matching template pairs; This represents the natural logarithm. The larger the sample entropy value, the higher the sequence complexity and the stronger the randomness of the intrinsic mode function components.

[0037] After calculating the sample entropy values ​​of all intrinsic mode function (IMF) components, all IMF components are sorted in ascending order of their respective sample entropy values, resulting in a sorted IMF component sequence. IMF components with smaller sample entropy values ​​exhibit stronger fluctuation regularity and tend to reflect deterministic emission characteristics generated by local equipment operation and production processes. Conversely, IMF components with larger sample entropy values ​​exhibit stronger fluctuation randomness and tend to reflect nondeterministic disturbances caused by external transmission and diffusion.

[0038] In one implementation, the preset percentage interval is determined based on the contribution distribution of emission characteristics of each production segment. All intrinsic mode function (IMF) components whose entropy values ​​fall within the top 30% of the sequence after sorting are selected. These selected IMF components are then cumulatively reconstructed point-by-point on the time axis. This cumulative reconstruction involves directly adding the amplitudes of each selected IMF component at the corresponding time point to generate the intrinsic fluctuation mode component of the current production segment. In a composite production line environment, changes in volatile organic compound (VOC) concentration caused by emissions from the production segment itself often manifest as continuous, highly periodic local fluctuations in the signal. The IMF components corresponding to these fluctuations have low sample entropy values, and the top 30% interval is sufficient to stably capture this emission contribution.

[0039] Simultaneously, all intrinsic mode function (IMF) components located within the last 30% of the sequence after sorting are selected. These selected IMF components are then reconstructed point-by-point to generate the extrinsic fluctuation mode components for the current work section. The IMF components within the last 30% interval correspond to signal components with high sample entropy and high randomness. These components mainly originate from transmission interference caused by the diffusion of emissions from neighboring work sections with airflow and crosstalk. Through the above separation, the intrinsic fluctuation mode components and extrinsic fluctuation mode components independently describe the different sources of emissions in the work section. The intrinsic fluctuation mode components no longer contain transmission interference from neighboring work sections, and the extrinsic fluctuation mode components no longer contain local deterministic emission information.

[0040] In practical implementation, for each section of the composite production line, the intrinsic fluctuation mode components obtained by empirical mode decomposition and sample entropy sorting reconstruction of the volatile organic compound (VOC) concentration time-series data for that section, along with the equipment status parameter time-series data collected synchronously for that section, are subjected to feature stitching processing on a time-by-time axis. The feature stitching process is as follows: at each sampling time, the value of the intrinsic fluctuation mode component at that time is concatenated with one or more parameter values ​​from the equipment status parameter time-series data at that time in a predetermined order to form a multi-dimensional vector. This multi-dimensional vector constitutes the stitched feature vector for that sampling time. The stitched feature vectors from all sampling times are arranged in chronological order to form the enhanced emission feature sequence for that section. The enhanced emission feature sequence for each section simultaneously contains the intrinsic fluctuation characteristics of local emissions within the section and the dynamic changes in equipment operating status.

[0041] After constructing the enhanced emission feature sequences for all work sections, the enhanced emission feature sequences for all work sections are aligned according to the same sampling time index to form a multivariate time series input matrix. Each row of this multivariate time series input matrix corresponds to a sampling time, and each column corresponds to a feature dimension.

[0042] The multivariate temporal input matrix is ​​fed into a temporal convolutional gated recurrent unit (TRN). The TRN consists of cascaded temporal convolutional layers and gated recurrent unit layers. Each temporal convolutional layer contains multiple parallel dilated convolutional kernels, each with a kernel size of 3. The dilation rate is dynamically adjusted based on the maximum delay value of the emission propagation delay matrix. During implementation, when the emission propagation delay matrix has not yet been generated, the initial dilation rate is set to 1. In subsequent iterations, after obtaining the latest emission propagation delay matrix, the maximum value of all emission propagation delay values ​​in the matrix is ​​read and used as the new dilation rate, which is then assigned to the dilation parameter of the temporal convolutional layer to adapt to the causal correlation capture requirements across different time spans. The temporal convolutional layer performs independent dilated convolution operations along the time dimension on the enhanced emission feature sequence for each work segment, extracting the temporal pattern features of each segment within a local time window and generating the convolutional feature sequence for each segment.

[0043] The output of the temporal convolutional layer is fed into the gated recurrent unit (ROU) layer. The ROU layer contains multiple gated recurrent unit nodes, each corresponding to a work segment and responsible for recurrently processing the convolutional feature sequence of that segment. The hidden states of all gated recurrent unit nodes interact at each time step, capturing long-term dependencies across work segments. The hidden state sequence output by the gated recurrent unit layer stores the context-aware features of each work segment at each time step.

[0044] Extract the time series of causal association strength between every two different work segments from the output states of the gated cyclic unit layer. Select the first and second work segments, and obtain the historical hidden state sequence of the first work segment at all historical time steps and the current hidden state vector of the second work segment at the current time step from the output states of the gated cyclic unit layer. Perform a dot product operation between each historical hidden state vector in the historical hidden state sequence of the first work segment and the current hidden state vector of the second work segment. The result of the dot product operation is a scalar representing the initial association strength of the first work segment to the second work segment at the current time step. Traverse all historical time steps to obtain an initial association score sequence.

[0045] Attention weight normalization based on a learnable query vector is performed on the initial association score sequence. A learnable query vector with the same dimension as the hidden state vector is set, and this learnable query vector is updated through backpropagation during model training. The dot product of each historical hidden state vector in the historical hidden state sequence of the first work segment and this learnable query vector is performed to obtain the corresponding attention weight scalar. The initial association scores at corresponding time steps in the initial association score sequence are weighted using the attention weight scalar. The weighting method is to multiply the attention weight scalar by the initial association score, sum the weighted scores over all historical time steps, and then normalize them. The specific form of the normalization process is to input the weighted score sequence into a softmax function to obtain a normalized causal association strength time series. The value at each time step in the causal association strength time series represents the causal association strength score between the first and second work segments at that time step. The calculation process of the causal association strength time series is expressed by the following formula:

[0046]

[0047] in, Indicates the first section For the second section At time step The time series of causal association strength; Indicates the first section Steps in history The hidden state vector; Indicates the second work section At the current time step The hidden state vector; This represents the learnable query vector, with the same dimension as the hidden state vector, and the initial values ​​are randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.01. Represents the dot product operation of vectors; This represents the softmax function, which exponentially normalizes the input vector. The overall parameters of the temporal convolutional gated recurrent unit include the kernel weights of the temporal convolutional layer, the gating parameters of the gated recurrent unit layer, and the learnable query vector. The dataset is obtained through end-to-end supervised training using a pre-constructed training dataset. The input to the training dataset consists of enhanced emission feature sequence samples from all work sections within a historical time period, the supervision labels are known emission propagation delay labels, and the loss function is the mean absolute error.

[0048] After obtaining the time series of the causal correlation strength between the first and second work sections, peak detection processing is performed on the causal correlation strength time series. The peak detection processing uses a threshold cross-validation method, setting the peak threshold for causal correlation strength to 1.5 times the mean of the causal correlation strength time series. The values ​​of each historical time step in the causal correlation strength time series are iterated. When the value of a historical time step is greater than the peak threshold and greater than the values ​​of its two adjacent time steps, that historical time step is identified as a peak point. The lag time step corresponding to the identified peak point, i.e., the time difference between the time step where the peak point is located and the current time step, is determined as the emission transmission delay value between the first and second work sections. All work section pairs are iterated, and the emission transmission delay values ​​of all work section pairs are used to form an emission transmission delay matrix. The first... Line number The elements of the column represent the first The first section, as the upstream source, is related to the... The emission propagation delay value caused by each work section. The emission propagation delay matrix is ​​incrementally updated after each new causal correlation intensity time series calculation is completed, and is used for dynamic adjustment of the expansion rate of the next round of temporal convolutional layers.

[0049] In specific implementation, please refer to Figure 2 For the target work section as the calculation object, all emission transmission delay values ​​of the target work section as the upstream source are extracted from the emission transmission delay matrix. The emission transmission delay matrix is ​​a two-dimensional matrix, in which the first... Line number The elements of the column represent the first... The first section, as the upstream, is related to the second... The emission propagation delay values ​​affecting each work section are extracted. The operation iterates through the rows corresponding to the target work section in the emission propagation delay matrix, arranging all non-empty emission propagation delay values ​​in that row in column order to form the pollution propagation feature vector of the target work section. Each dimension of the pollution propagation feature vector corresponds to a downstream work section affected by the emissions from the target work section, with the dimension value being the emission propagation delay value from the target work section to that downstream work section. The pollution propagation feature vector describes the time delay distribution characteristics of the propagation of emissions from the target work section to other work sections in the composite production line.

[0050] In one implementation, a sliding window segmentation process is performed on the environmental ventilation volume time-series data. The window length of the current sampling window is set to be consistent with the sampling window length of the volatile organic compound concentration time-series data, which is set to 60 sampling times. The sampling frequency is 1 Hz, the same as the sampling frequency of the array sensor group. Ventilation volume values ​​at the current sampling time and the previous 59 sampling times are extracted from the environmental ventilation volume time-series data, totaling 60 ventilation volume values, forming the ventilation volume data sequence within the current sampling window. The coefficient of variation of the environmental ventilation volume is calculated for the ventilation volume data sequence within the current sampling window. The calculation process of the coefficient of variation of the environmental ventilation volume is as follows: calculate the mean of the ventilation volume data sequence within the current sampling window, calculate the standard deviation of the ventilation volume data sequence within the current sampling window, and divide the standard deviation by the mean. The ratio obtained is the coefficient of variation of the environmental ventilation volume. The coefficient of variation of the environmental ventilation volume serves as a characteristic value of the diffusion conditions of the target section, reflecting the degree of fluctuation of ventilation conditions within the current sampling window.

[0051] After obtaining the pollution propagation feature vector, quantile-based outlier pruning is performed. The specific steps of quantile-based outlier pruning are as follows: The distribution of all emission propagation delay values ​​in the pollution propagation feature vector is statistically analyzed, and the 5th and 95th quantiles of the emission propagation delay values ​​are calculated. The 5th quantile corresponds to the value at the 5th percentile after sorting all emission propagation delay values ​​from smallest to largest, and the 95th quantile corresponds to the value at the 95th percentile after sorting all emission propagation delay values ​​from smallest to largest. All emission propagation delay values ​​in the pollution propagation feature vector below the 5th quantile are replaced with the 5th quantile value, and all emission propagation delay values ​​above the 95th quantile are replaced with the 95th quantile value, resulting in the truncated propagation feature vector. Quantile-based outlier pruning eliminates the interference of extremely deviated emission propagation delay values ​​in the pollution propagation feature vector on subsequent calculations.

[0052] After outlier pruning, the truncated propagation feature vector and diffusion condition feature values ​​are concatenated after dimensional alignment. The dimensional alignment process is as follows: the truncated propagation feature vector is a multi-dimensional vector with the number of dimensions equal to the number of downstream sections affected by the target section; the diffusion condition feature values ​​are scalars. The diffusion condition feature values ​​are added as a new dimension and appended to the end of the truncated propagation feature vector, forming a fused feature vector with one more dimension than the original vector.

[0053] The fused feature vector is input into a pre-trained deviation exponential regressor. The core architecture of the deviation exponential regressor consists of multiple stacked fully connected layers. In one implementation, the deviation exponential regressor contains three fully connected layers: the first fully connected layer has the same input dimension as the fused feature vector and an output dimension of 64; the second fully connected layer has an input dimension of 64 and an output dimension of 32; and the third fully connected layer has an input dimension of 32 and an output dimension of 1. A modified linear unit activation function is connected after the first and second fully connected layers, while no activation function is used after the third fully connected layer. The deviation exponential regressor maps the fused feature vector to a scalar output value through layer-by-layer nonlinear mapping. This scalar output value is the dynamic pollution contribution deviation index of the target section in the current sampling window.

[0054] The pre-training process of the deviation exponential regressor employs supervised learning. The training dataset consists of samples from multiple sampling windows within historical time periods. The input to each training sample is the fused feature vector of the corresponding sampling window, and the supervision label is a manually labeled reference value for the dynamic pollution contribution deviation index. The reference value for the dynamic pollution contribution deviation index is determined based on the percentage deviation between the actual volatile organic compound (VOC) emission contribution of the target section within the corresponding sampling window and the baseline contribution under normal operating conditions; the larger the percentage deviation, the higher the reference value. During training, mean squared error is used as the loss function, and the Adam optimizer is used for parameter updates. The learning rate is set to 0.001, the training epochs are set to 200, and the batch size is set to 32. After training, the weight parameters of the fully connected layers of the deviation exponential regressor are fixed and saved.

[0055] During the calculation of the dynamic pollution contribution deviation index, the coefficient of variation of the environmental ventilation volume time series data is continuously monitored. A preset ventilation fluctuation threshold of 0.3 is set. This setting is based on the fact that when the coefficient of variation reaches 0.3, the ventilation volume fluctuation amplitude has reached 30% of the average ventilation volume, indicating significant airflow instability and a non-negligible impact on the pollutant diffusion path. When the coefficient of variation of the environmental ventilation volume exceeds the preset ventilation fluctuation threshold of 0.3, a ventilation volume weighted correction factor is introduced to dynamically attenuate and compensate for the deviation index value output by the deviation index regressor. The ventilation volume weighted correction factor is denoted as... Ventilation volume weighted correction factor The value of is determined by the following formula:

[0056]

[0057] in, The coefficient of variation of environmental ventilation volume is the ratio of the standard deviation to the mean of the ventilation volume data sequence within the current sampling window. This indicates the preset ventilation fluctuation threshold, with a value of 0.3. Represented by natural constant An exponential function with base 0.5. The coefficient of variation of the ambient ventilation volume. Less than or equal to the preset ventilation fluctuation threshold At that time, the ventilation volume weighted correction factor The value is 1, and no attenuation compensation is performed; when the coefficient of variation of the ambient ventilation volume is 1. Greater than the preset ventilation fluctuation threshold At that time, the ventilation volume weighted correction factor The value of decreases exponentially with the increase of the coefficient of variation, and the range is . Ventilation volume weighted correction factor Multiplying the original dynamic pollution contribution deviation index output by the deviation index regressor yields the final dynamic pollution contribution deviation index after dynamic attenuation compensation. This attenuation compensation mechanism appropriately reduces the absolute value of the deviation index when ventilation fluctuates drastically, reducing the risk of misjudging normal concentration fluctuations as abnormal emission contributions due to random changes in ventilation conditions.

[0058] In practice, after obtaining the dynamic pollution contribution deviation index of each section in the composite production line within the current sampling window, the dynamic pollution contribution deviation index of each section is compared with preset first and second warning thresholds. The preset first warning threshold is set to 0.45, and the preset second warning threshold is set to 0.75. The first warning threshold of 0.45 is based on the fact that, in historical normal operating data, the upper quartile of the long-term statistical distribution of the dynamic pollution contribution deviation index of each section is 0.42. Setting it to 0.45 provides approximately 7% buffer margin while covering the normal fluctuation range. The second warning threshold of 0.75 is based on the fact that, in historical abnormal operating data, the lowest value of the dynamic pollution contribution deviation index of the corresponding section when an abnormal emission event is confirmed is 0.78. Setting it to 0.75 ensures that all known abnormal events are captured while allowing for a certain lead time.

[0059] For any section in the composite production line, when the dynamic pollution contribution deviation index of that section is greater than or equal to the first warning threshold of 0.45 and less than the second warning threshold of 0.75, a Level 1 emission warning signal is triggered. The Level 1 emission warning signal includes a warning level identifier, a timestamp of the triggering time, and an abnormal section identifier. Simultaneously, the abnormal section identifier is recorded in a pending confirmation list, which is a string list structure stored in system memory to temporarily store abnormal section identifiers for which responsibility has not yet been definitively determined.

[0060] Simultaneously with triggering the generation of a Level 1 emission warning signal, a source tracing analysis based on dynamic time curvature distance is performed. The system extracts the intrinsic fluctuation mode components (IMMs) of the section corresponding to the abnormal section identifier within the current sampling window. Simultaneously, it extracts the IMMs of the upstream and downstream adjacent sections of the corresponding abnormal section within the current sampling window. The upstream and downstream adjacent sections are predefined in the system configuration file according to the process flow sequence of the composite production line.

[0061] After obtaining three sets of intrinsic wave mode components, the dynamic time curvature distance between the intrinsic wave mode components of the section corresponding to the abnormal section identifier and the intrinsic wave mode components of the upstream adjacent section is calculated. Simultaneously, the dynamic time curvature distance between the intrinsic wave mode components of the section corresponding to the abnormal section identifier and the intrinsic wave mode components of the downstream adjacent section is also calculated. The calculation process for the dynamic time curvature distance is as follows: Let the sequence of intrinsic wave mode components of the section corresponding to the abnormal section identifier be a vector. , length is Let the sequence of intrinsic wave mode components of the upstream adjacent section be a vector. , length is Build a OK The cumulative distance matrix of the column, the cumulative distance matrix of the column Line number Column elements Indicates sequence prefix With sequence prefix The minimum cumulative distance between them. The cumulative distance matrix is ​​calculated recursively as follows: Initialize to For boundary points and , For boundary points and , For the remaining positions and , .in, This indicates the absolute value operation. Ultimately... The value of is the dynamic time bending distance between the intrinsic wave mode component of the section corresponding to the abnormal section identifier and the intrinsic wave mode component of the upstream adjacent section. Using the same calculation process, the intrinsic wave mode component sequence of the upstream adjacent section is replaced with the intrinsic wave mode component sequence of the downstream adjacent section to calculate the dynamic time bending distance between the intrinsic wave mode component of the section corresponding to the abnormal section identifier and the intrinsic wave mode component of the downstream adjacent section.

[0062] When comparing the magnitudes of two dynamic time curvature distances, if the dynamic time curvature distance between the section corresponding to the abnormal section identifier and its upstream neighboring section is less than the dynamic time curvature distance between the section corresponding to the abnormal section identifier and its downstream neighboring section, it indicates that the intrinsic wave mode component of the abnormal section is more similar to the intrinsic wave mode component of the upstream neighboring section in terms of time curvature. In this case, an upstream traceability identifier string is appended to the string of the abnormal section identifier in the pending confirmation list. Conversely, if the dynamic time curvature distance between the section corresponding to the abnormal section identifier and its upstream neighboring section is greater than the dynamic time curvature distance between the section corresponding to the abnormal section identifier and its downstream neighboring section, it indicates that the intrinsic wave mode component of the abnormal section is more similar to the intrinsic wave mode component of the downstream neighboring section in terms of time curvature. In this case, a downstream traceability identifier string is appended to the string of the abnormal section identifier in the pending confirmation list.

[0063] For any section in the composite production line, when the dynamic pollution contribution deviation index of that section is greater than or equal to the second warning threshold of 0.75, a secondary emission warning signal is triggered. The secondary emission warning signal includes a warning level identifier, a timestamp of the triggering time, and an abnormal section identifier. Simultaneously, this abnormal section identifier is directly marked as the source responsibility section, and the source responsibility section identifier, its corresponding dynamic pollution contribution deviation index value, and the triggering time are written into the confirmation record table of the system warning log. The confirmation record table is a structured data table containing columns for section identifier, deviation index, timestamp, and responsibility determination.

[0064] After triggering the corresponding level of emission warning signal and locating the abnormal work section, the equipment performance degradation warning process is executed. Based on the located abnormal work section identifier, the historical dynamic pollution contribution deviation index sequence within the same equipment status parameter range for that abnormal work section identifier is retrieved from the historical warning database. The historical warning database is a time-series database maintained by the system, indexed by work section identifier and equipment status parameter range, storing the historical dynamic pollution contribution deviation index values ​​recorded for each warning event. The equipment status parameter range is divided into several segments according to the value range of the equipment status parameters. During retrieval, the current equipment status parameter is matched to the corresponding segment, and all historical dynamic pollution contribution deviation index values ​​stored in that segment are extracted and arranged chronologically to form a historical dynamic pollution contribution deviation index sequence.

[0065] An autoregressive integral moving average (ARM) model was used to fit the historical dynamic pollution contribution deviation index sequence. The specific steps of the ARM model fitting process were as follows: First-order differencing was performed on the historical dynamic pollution contribution deviation index sequence to eliminate the trend term; after differencing, the... The elements are ,in Indicates the first position in the original sequence The historical dynamic pollution contribution deviation index value was analyzed. An autoregressive moving average model was fitted to the differencing sequence. The order of the autoregressive moving average model was determined using the Akaike Information Criterion. Alternative models were selected with regression orders ranging from 0 to 3 and moving average orders ranging from 0 to 3. All combinations were iterated, and the combination that minimized the Akaike Information Criterion value was chosen as the final model order. The parameters of the differencing sequence were estimated using the determined-order autoregressive moving average model, employing the maximum likelihood estimation method. After parameter estimation, the fitted autoregressive integral moving average model was used to recursively predict future times. The predicted values ​​for each time point obtained from the recursive prediction were then inversely differencing to generate the deviation index prediction trend line for this abnormal section.

[0066] After obtaining the deviation index prediction trend line, the deviation index of the dynamic pollution contribution at the current moment is compared with the predicted value at the corresponding moment on the deviation index prediction trend line to calculate the deviation rate. The formula for calculating the deviation rate is:

[0067]

[0068] in, Indicates the deviation rate; This indicates the deviation of the dynamic pollution contribution from the index value at the current moment; This represents the predicted value at the current moment that deviates from the index prediction trend line; This indicates absolute value calculation. The preset drift threshold is set to 0.25. This value is set based on the fact that during normal equipment performance degradation, the deviation rate of continuously monitored data should not exceed 0.20. Setting it to 0.25 allows for a certain tolerance margin. When the deviation rate... When the drift threshold of 0.25 is exceeded, an equipment performance degradation warning is generated. The equipment performance degradation warning includes the abnormal section identifier, the current dynamic pollution contribution deviation index value, the predicted value value, the deviation rate value, and the generation timestamp. The equipment performance degradation warning is also associated with the time series data of the equipment status parameters of the abnormal section in the current sampling window through the association field.

[0069] Optionally, after comparing the dynamic pollution contribution deviation index with the preset first and second warning thresholds, an incremental fine-tuning update of the dynamic causal graph network is performed. The comparison results, including the relationship between the dynamic pollution contribution deviation index and the warning threshold interval, the triggered emission warning signal level, and the identified abnormal work section, are packaged into a warning event record. The warning event record uses a structured data format, including comparison result fields, warning level fields, and abnormal work section identification fields. The warning event record is used as a feedback tag, and combined with the time-series data of volatile organic compound concentration, equipment status parameters, and environmental ventilation volume within a preset time window before the warning is triggered, to form a training sample pair. The preset time window is set to 120 sampling times, with a sampling frequency of 1 Hz, meaning that the three types of time-series data are extracted within 120 seconds before the warning is triggered. Using this training sample pair, the connection weights of the graph convolutional gated recurrent units in the dynamic causal graph network are incrementally fine-tuned. The specific steps of incremental fine-tuning are as follows: The time-series data from the training sample pairs are input into the current dynamic causal graph network, and the emission transmission delay matrix output by the network is calculated via forward propagation. A monitoring signal is constructed based on the actual emission source relationship corresponding to the abnormal work section identifier in the early warning event record. The monitoring signal is a binary label indicating whether the emission transmission delay between target work sections is present or absent. The error between the network's predicted delay and the monitoring signal is calculated using the cross-entropy loss function. The gradient of the connection weights is calculated using a stochastic gradient descent algorithm with a learning rate set to 0.0001. The gradient is backpropagated to update the gate parameters and kernel weights in the graph convolutional gated recurrent unit. The learning rate is set to 0.0001 because a smaller learning rate is needed in the incremental fine-tuning scenario to avoid catastrophic forgetting of learned parameters; 0.0001 is one-tenth of the initial training learning rate. After completing the incremental fine-tuning update, the updated dynamic causal graph network is obtained, and the updated dynamic causal graph network is called in the calculation of the next sampling window.

[0070] In specific implementation, please refer to Figure 3 A volatile organic compound (VOC) emission monitoring and early warning system for a composite production line includes a memory, a processor, and a computer program stored in the memory and running on the processor. The memory uses a non-volatile storage medium, specifically a solid-state drive (SSD), with a storage capacity of at least 512 gigabytes, for persistently storing the computer program, intermediate data generated during operation, and a historical early warning database. The processor is a central processing unit (CPU) with a clock speed of at least 2.5 GHz and at least eight processing cores, used to execute the instruction sequences in the computer program. The memory also includes a dynamic random access memory (DRAM) area with a capacity of at least 32 gigabytes, used to temporarily load timing data and intermediate matrices during calculations when the processor executes the computer program.

[0071] The memory stores time-series data on volatile organic compound (VOC) concentration, equipment status parameters, and environmental ventilation volume, collected by array-type sensor groups configured in each section of the composite production line. The VOC concentration time-series data is collected by photoionization detector arrays installed above and at the boundaries of each section, with a sampling frequency set to 1 Hz, generating a concentration value at each sampling moment. The equipment status parameter time-series data is output in real time by programmable logic controllers (PLCs) installed on the production equipment in each section, including at least one parameter among equipment operating speed, heating temperature, and feeding rate, with a sampling frequency aligned to 1 Hz with the VOC concentration time-series data. The environmental ventilation volume time-series data is collected by an array of wind speed sensors deployed within the composite production line workshop, with wind speed sensors installed at the top air inlets and side exhaust outlets, also with a sampling frequency set to 1 Hz. All three types of time-series data are timestamped upon entering the memory to ensure strict alignment of the timeline in all subsequent processing steps.

[0072] The processor executes a computer program, loading and running program modules that implement all steps of the monitoring and early warning method. The computer program is written with a modular architecture, including a data acquisition module, a signal decomposition module, a causal network construction module, a deviation index calculation module, an early warning determination module, and a model update module. The data acquisition module is responsible for reading three types of time-series data from memory and performing timestamp alignment verification. The signal decomposition module performs adaptive signal decomposition processing based on multi-scale sample entropy on the volatile organic compound concentration time-series data of each work section to separate endogenous and exogenous fluctuation mode components. The causal network construction module constructs a dynamic causal graph network based on time-convolutionally gated cyclic units using endogenous fluctuation mode components and equipment state parameter time-series data as input to generate an emission propagation delay matrix. The deviation index calculation module calculates the dynamic pollution contribution deviation index for each work section within the current sampling window based on the emission propagation delay matrix and environmental ventilation volume time-series data. The early warning determination module triggers emission early warning signals and identifies abnormal work sections based on the comparison result between the dynamic pollution contribution deviation index and preset multi-level early warning threshold intervals. The model update module is responsible for incrementally fine-tuning the connection weights of the dynamic causal graph network using newly generated training samples after an early warning event occurs.

[0073] During the execution of the computer program by the processor, the data acquisition module reads time-series data on volatile organic compound concentrations, equipment status parameters, and environmental ventilation volumes for all sections of the composite production line within the current sampling window from the time-series data storage area of ​​the memory. The data window length is 60 sampling times. The data acquisition module verifies whether the timestamps of the three types of data are consistent. For missing time steps, linear interpolation is used to fill in the gaps. The linear interpolation operation is as follows: for a missing time step, the values ​​of the two valid times before and after the missing time step are taken, and the interpolation result is calculated according to the time distance ratio. The three types of data after filling in the gaps are organized into a matrix form of the number of sections multiplied by the number of time steps, and then passed to the signal decomposition module.

[0074] After receiving the time-series data of volatile organic compound (VOC) concentration, the signal decomposition module independently performs empirical mode decomposition (EMD) processing on the concentration time-series data for each process segment, obtaining multiple intrinsic mode function (IMF) components and one residual component. For each IMF component, a sample entropy value is calculated. The embedding dimension of the sample entropy calculation parameter is set to 2, and the similarity tolerance is set to 0.15 times the standard deviation of the corresponding IMF component. After sorting the sample entropy values ​​in ascending order, the top 30% of IMF components are accumulated and reconstructed into intrinsic fluctuation mode components, and the bottom 30% of IMF components are accumulated and reconstructed into extrinsic fluctuation mode components. The intrinsic and extrinsic fluctuation mode components are stored separately in the intermediate result area of ​​the memory.

[0075] The causal network construction module reads the intrinsic fluctuation mode components of each work segment from the intermediate result area of ​​the memory and concatenates them with the time-series data of equipment state parameters step by step to generate an enhanced emission feature sequence. The enhanced emission feature sequences of all work segments are aligned by time step and then fed into a dynamic causal graph network based on a time-convolutionally gated recurrent unit. The expansion rate of the time-convolutional layer of the dynamic causal graph network is initially set to 1 and dynamically adjusted in subsequent iterations based on the maximum delay value of the generated emission propagation delay matrix. The hidden states output by the gated recurrent unit layer are used to calculate the time series of causal association strength between work segments. A normalized causal association strength score is generated through attention normalization processing weighted by learnable query vectors. After peak detection, the emission propagation delay matrix is ​​obtained and stored in memory.

[0076] The deviation index calculation module extracts the pollution propagation feature vector of the target section from the emission propagation delay matrix and performs outlier pruning based on the 5th and 95th quantiles. Simultaneously, it extracts the ventilation volume sequence of the current sampling window from the environmental ventilation volume time series data and calculates the coefficient of variation as a diffusion condition feature value. The truncated propagation feature vector and diffusion condition feature value are concatenated and input into a deviation index regressor composed of three stacked fully connected layers, outputting a dynamic pollution contribution deviation index. When the coefficient of variation of the environmental ventilation volume exceeds a preset ventilation fluctuation threshold of 0.3, an error is introduced using the formula... A determined ventilation volume weighted correction factor is used for dynamic attenuation compensation. The dimensions of the three fully connected layers deviating from the exponential regressor are, respectively, the input dimension to 64, 64 to 32, and 32 to 1. The first two layers are followed by a modified linear unit activation function.

[0077] The early warning judgment module reads the dynamic pollution contribution deviation index of each work section and compares it one by one with the preset first early warning threshold of 0.45 and the second early warning threshold of 0.75. For work sections with a dynamic pollution contribution deviation index greater than or equal to 0.45 and less than 0.75, a first-level emission early warning signal is triggered, the corresponding work section identifier is recorded in the pending confirmation list, and upstream and downstream source tracing analysis based on dynamic time curvature distance is initiated, adding the upstream or downstream source identifier to the pending confirmation list. For work sections with a dynamic pollution contribution deviation index greater than or equal to 0.75, a second-level emission early warning signal is triggered, and the corresponding work section identifier is directly marked as the source responsibility work section. The early warning signal and the abnormal work section identifier are written to the system early warning log and pushed to the monitoring terminal display interface. The early warning determination module also retrieves the historical dynamic pollution contribution deviation index sequence within the same equipment status parameter range from the historical early warning database based on the abnormal section identifier, performs autoregressive integral moving average model fitting, calculates the deviation rate between the current deviation index and the predicted trend line, and generates an equipment performance degradation early warning prompt when the deviation rate exceeds the preset drift threshold of 0.25 and associates it with the corresponding equipment status parameter time series data.

[0078] After each warning determination, the model update module packages the comparison results, warning level, and abnormal section identifier of this warning into a warning event record. This record is then combined with three types of time-series data from the 120 sampling times prior to the warning to form a training sample pair. The module performs an incremental fine-tuning update on the connection weights of the temporal convolutional gated recurrent units in the dynamic causal graph network. The updated network parameters overwrite the original network parameters stored in the memory.

[0079] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for monitoring and early warning of volatile organic compound emissions from a composite production line, characterized in that, The method includes: Acquire time-series data of volatile organic compound concentration, equipment status parameters, and environmental ventilation volume from array-type sensor groups configured in each section of the composite production line. Adaptive signal decomposition processing based on multi-scale sample entropy was performed on the time series data of volatile organic compound concentration in each section to separate the endogenous wave mode components that belong to the local emission contribution and the exogenous wave mode components that belong to the transmission interference of neighboring sections. Using the intrinsic fluctuation mode components and the time series data of the equipment state parameters as input, a dynamic causal graph network based on time convolution gated recurrent units is constructed to generate a volatile organic compound emission transmission delay matrix between each section. Based on the emission transmission delay matrix and the environmental ventilation time series data, calculate the dynamic pollution contribution deviation index of each section in the current sampling window; Based on the comparison results between the dynamic pollution contribution deviation index and the preset multi-level early warning threshold range, the corresponding level of emission early warning signal is triggered and the abnormal section is identified.

2. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 1, characterized in that, The intrinsic wave mode components attributable to local emissions and the extrinsic wave mode components attributable to transmission interference from neighboring work sections were separated, including: Empirical mode decomposition was performed on the time series data of volatile organic compound concentration for each section to obtain multiple intrinsic mode function components and one residual component. Calculate the sample entropy value of each intrinsic mode function component, and sort all intrinsic mode function components in ascending order of sample entropy value; The intrinsic mode function components whose entropy values ​​are within the first preset percentage range after sorting are accumulated and reconstructed to generate the intrinsic fluctuation mode components. The intrinsic mode function components whose entropy values ​​are within the preset percentage range after sorting are accumulated and reconstructed to generate the exogenous wave mode components.

3. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 1, characterized in that, Generate the volatile organic compound emission transfer delay matrix between each work section, including: The intrinsic fluctuation mode components of each section are combined with the time series data of the equipment status parameters of that section along the time axis to generate the enhanced emission characteristic sequence of each section. The enhanced emission feature sequences of all work sections are aligned by time step and then input into the temporal convolutional gated recurrent unit. This unit extracts the local temporal pattern of each work section through the temporal convolutional layer and then captures the long-term dependency relationship across work sections through the gated recurrent unit layer. Extract the time series of causal correlation strength between every two different work sections from the output state of the gated loop unit layer; Peak detection processing is performed on the time series of causal correlation strength, and the lag time step corresponding to each peak is determined as the emission transmission delay value between the two sections. All emission transmission delay values ​​constitute the emission transmission delay matrix.

4. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 3, characterized in that, The expansion rate of the temporal convolutional layer in the temporal convolutional gated recurrent unit is dynamically adjusted according to the maximum delay value of the emission propagation delay matrix.

5. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 3, characterized in that, Extract the time series of causal correlation strength between every two different work segments from the output state of the gated loop unit layer, including: Obtain the historical hidden state sequence of the first section and the current hidden state vector of the second section from the output state of the gated loop unit layer; Perform a dot product operation between each historical hidden state in the historical hidden state sequence of the first work section and the current hidden state vector of the second work section to generate an initial associated scoring sequence. The initial association score sequence is subjected to attention weight normalization based on learnable query vectors to generate the causal association strength time series.

6. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 1, characterized in that, Based on the emission transmission delay matrix and the environmental ventilation time series data, calculate the dynamic pollution contribution deviation index for each work section within the current sampling window, including: For the target section, all transmission delay values ​​of the section as the upstream source are extracted from the emission transmission delay matrix to form the pollution propagation feature vector of the section. The environmental ventilation volume time series data is processed by sliding window segmentation, and the coefficient of variation of the environmental ventilation volume within the current sampling window is calculated as the diffusion condition characteristic value of the section. The pollution propagation feature vector and the diffusion condition feature value are concatenated and then input into a pre-trained deviation index regressor, which outputs the dynamic pollution contribution deviation index.

7. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 6, characterized in that, In the process of calculating the dynamic pollution contribution deviation index, when the coefficient of variation of the environmental ventilation volume time series data exceeds the preset ventilation fluctuation threshold, a ventilation volume weighted correction factor is introduced to dynamically attenuate and compensate the deviation index.

8. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 6, characterized in that, The concatenation of the pollution propagation feature vector and the diffusion condition feature value is then input into a pre-trained deviation exponential regressor, including: The pollution propagation feature vector is subjected to outlier pruning based on quantiles to obtain a truncated propagation feature vector; The truncated propagation feature vector is dimensionally aligned with the diffusion condition feature value and then concatenated to generate a fused feature vector. The fused feature vector is input into the deviation index regressor, which is composed of multiple fully connected layers stacked together. The regressor outputs the dynamic pollution contribution deviation index through layer-by-layer nonlinear mapping.

9. The method for monitoring and early warning of volatile organic compound emissions from a composite production line according to claim 1, characterized in that, Based on the comparison results between the dynamic pollution contribution deviation index and the preset multi-level early warning threshold range, the corresponding level of emission early warning signal is triggered and the abnormal section is identified, including: The dynamic pollution contribution deviation index is compared with the first warning threshold and the second warning threshold, respectively, and the second warning threshold is greater than the first warning threshold. When the dynamic pollution contribution deviation index is greater than or equal to the first warning threshold and less than the second warning threshold, a first-level emission warning signal is generated, and the abnormal section identifier is recorded in the list to be confirmed. When the dynamic pollution contribution deviation index is greater than or equal to the second warning threshold, a secondary emission warning signal is triggered, and the abnormal section is directly marked as the source responsibility section.

10. A monitoring and early warning system for volatile organic compound emissions from a composite production line, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for monitoring and early warning of volatile organic compound emissions from a composite production line as described in any one of claims 1 to 9.

Images