Early warning method for abnormal emission of pollution facilities based on big data analysis

Abstract

The invention discloses a big data analysis-based early warning method for abnormal emission of pollution facilities. The method comprises the following steps of S1, collecting data and constructing a data set; s2, storing and preprocessing a data set; s3, constructing a space-time propagation network of the external events by using the graph neural network model; s4, modeling and predicting the time sequence features by using a mixed depth generation model; s5, analyzing the causal relationship between the external event and the facility emission behavior through a causal inference model; s6, adopting a Bayesian optimization algorithm based on uncertainty estimation of a strategy to dynamically adjust an emission early warning threshold value; s7, generating a new abnormal emission mode label by using a self-supervised learning model and a few-sample learning technology; and S8, adopting an interpretable artificial intelligence technology to generate an abnormal source tracing report, and providing coping suggestions and action schemes. According to the method, the early warning of the abnormal emission of the pollution facilities is realized by utilizing a cross-time-space event inference and self-supervised learning method.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring technology, and in particular to an early warning method for abnormal emissions from pollution facilities based on big data analysis. Background Art

[0002] As environmental pollution becomes increasingly serious, the demand for monitoring and management of pollution facilities has increased significantly worldwide. Abnormal emissions from pollution facilities are one of the main factors leading to environmental pollution, especially waste gas and wastewater emitted during industrial production, which may cause irreversible impacts on the atmosphere and water bodies. Therefore, how to effectively monitor the emission behavior of pollution facilities, identify abnormal emissions and provide early warnings is an important research direction in the current field of environmental monitoring.

[0003] Existing pollution monitoring systems mainly rely on regularly collected emission data and manual analysis methods. These systems usually use traditional sensor networks to monitor the concentration of pollutants, and the data is transmitted to the data center for analysis via wired or wireless networks. Under this architecture, monitoring systems often face the following challenges: First, existing monitoring systems mainly rely on statically set thresholds to determine whether emissions are abnormal. These thresholds are usually set based on historical data or regulations. Once set, they will not change for a long time and cannot reflect the operating status of emission facilities and changes in the external environment in a timely manner. When environmental conditions such as temperature, humidity, wind speed, etc. change significantly, the use of static thresholds is prone to false alarms or missed alarms, and cannot dynamically adapt to complex actual conditions.

[0004] Secondly, most existing emission monitoring systems are limited to monitoring the emission behavior of a single facility, while ignoring the potential impact of external events on facility emissions. External events such as climate change, industrial activities and natural disasters may have a significant impact on the emission and diffusion of pollutants. Due to the lack of cross-temporal and spatial analysis of these external factors, it is difficult for existing systems to predict the impact of these external events on emission behavior in advance, resulting in delayed warnings and failure to provide sufficient response time for relevant departments, thereby increasing the risk of pollution accidents.

[0005] In the existing technology, the detection of abnormal emission behavior relies on a large amount of labeled data. By summarizing historical abnormal events, the system identifies possible abnormal situations in the future. However, the acquisition cost of labeled data is high, and it is difficult to cover all possible abnormal scenarios. Due to the complexity of industrial production and the diversity of emission behaviors, many potential abnormal situations cannot be identified in advance, and the anomaly detection capabilities of existing systems are therefore limited. In addition, anomaly detection systems often use black box models for prediction. Although these models have improved detection accuracy to a certain extent, they lack explainability. When faced with abnormal warnings, managers cannot understand the reasons and basis behind the warnings, which not only reduces the transparency of the system, but also makes it difficult for managers to make correct responses.

[0006] Furthermore, most existing anomaly detection technologies rely on supervised learning methods, that is, training models based on labeled data. However, in real-world environments, labeled data is often insufficient, and the scarcity of abnormal data samples further exacerbates this problem. In the case of scarce data, the detection accuracy of the model drops significantly, and it is difficult for the system to perform effective anomaly detection in the absence of labeled or few sample data. In addition, traditional anomaly detection methods lack adaptability and cannot flexibly respond to complex environmental changes and operating conditions during facility operation, resulting in a decrease in the reliability of the early warning system.

[0007] Based on the above problems, the existing pollution facility emission monitoring system still has major defects in anomaly detection and early warning. First, the setting of static thresholds limits the system's adaptability in complex dynamic environments, making it difficult to flexibly respond to changes in facilities and environmental conditions. Secondly, due to the lack of in-depth analysis of the cross-temporal and spatial correlations between external events and emission behaviors, the existing system cannot predict the potential impact of external events on emission behaviors in advance, resulting in insufficient timeliness of early warnings. In addition, the existing system's reliance on labeled data limits its anomaly detection capabilities, making it difficult to effectively respond to situations where unlabeled data or abnormal samples are scarce. Finally, due to the lack of the application of explainable artificial intelligence, the warning results of the existing system often cannot give a clear explanation and traceability path, reducing decision makers' trust and reliance on the system.

[0008] Therefore, how to provide an early warning method for abnormal emissions from pollution facilities based on big data analysis is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose an early warning method for abnormal emissions from pollution facilities based on big data analysis. The present invention makes full use of advanced technologies such as cross-temporal and spatial event inference, self-supervised learning, few-sample learning, and Bayesian optimization algorithms. By comprehensively analyzing facility emission data, external events, and environmental changes, it achieves real-time monitoring and dynamic early warning of abnormal emission behaviors. The present invention not only has strong foresight and can predict the potential impact of external events on facility emissions in advance, but also has high-precision detection capabilities and a flexible threshold adjustment mechanism, which significantly improves the accuracy and adaptability of abnormal emission warnings and reduces false alarms and missed reports.

[0010] The method for early warning of abnormal emissions from pollution facilities based on big data analysis according to an embodiment of the present invention comprises the following steps:

[0011] S1. Collect emission data, environmental data and external event data of pollution facilities through a sensor network and construct a data set. The sensor network transmits data through the Internet of Things technology;

[0012] S2, storing the data set in a big data storage platform, and using the Hadoop big data processing framework to store and preprocess the data set, wherein the preprocessing includes data cleaning, missing value filling, noise removal, and time series feature extraction;

[0013] S3. Use graph neural network models to model the complex relationship between external events and the emission behavior of pollution facilities, build a spatiotemporal propagation network of external events, and analyze the potential impact of external events on the emissions of pollution facilities;

[0014] S4. Use a hybrid deep generative model, combined with attention mechanism and variational inference, to model and predict time series features and predict external events that may affect emissions from pollution facilities in the future;

[0015] S5. Analyze the causal relationship between external events and facility emissions through causal inference models, determine the potential impact of external events on facility emissions, and generate a forecast report;

[0016] S6. Adopting the Bayesian optimization algorithm based on strategy uncertainty estimation, introducing uncertainty trade-offs in the decision-making process, and dynamically adjusting the emission warning threshold;

[0017] S7. Use a self-supervised learning model to learn normal emission patterns from unlabeled data, identify abnormal emission behaviors that deviate from the normal pattern, and quickly learn from limited abnormal samples through few-shot learning technology to generate new abnormal emission pattern labels;

[0018] S8. Use explainable artificial intelligence technology to explain the warning results and generate anomaly source tracing reports, showing the key data features that cause the anomalies, the basis for model decision-making and the tracing path, and provide response suggestions and action plans.

[0019] Optionally, the S3 specifically includes:

[0020] S31. Construct a heterogeneous graph structure based on the emission data of pollution facilities, environmental data and external event data. The nodes include pollution facilities, environmental variables and external events. The edges represent the associations between different nodes. The weight of each edge is calculated based on the spatiotemporal association:

[0021]

[0022] Among them, W ij represents the association weight between node i and node j, exp represents the exponential function, x i represents the feature vector of node i, x j represents the feature vector of node j, p i represents the spatial coordinates of node i, p j represents the spatial coordinate of node j, y i represents the timestamp of node i, y j represents the timestamp of node j, e i represents the external event characteristics of node i, e j represents the external event characteristics of node j, σ, α, γ, β and η represent the adjustment factors;

[0023] S32. Use graph neural network to process heterogeneous graphs, and update the features of nodes through the message passing mechanism of neighbor nodes:

[0024]

[0025] in, represents the feature representation of node i in the k+1 layer, σ represents the activation function, represents the feature representation of node i at layer k, represents the feature representation of node m at layer k, represents the feature representation of node j at the kth layer, N(i) represents the neighbor set of node i, λ represents the regularization coefficient, represents the bias term of node i in the kth layer, λ represents the regularization coefficient, and W im represents the association weight between node i and node m;

[0026] S33. Through the message passing mechanism of the spatiotemporal graph neural network, combined with historical external event data and environmental data, the association weights between nodes in the graph are dynamically adjusted. The state of the node at each moment is updated according to the historical propagation path, and the node features are updated synchronously in the spatial and temporal dimensions.

[0027] S34, based on the spatiotemporal model constructed by the spatiotemporal propagation network, further adjust the feature representation of each node and update it through multiple iteration cycles;

[0028] S35. Generate the final high-dimensional representation of each node through the iterative propagation mechanism of the multi-layer spatiotemporal graph neural network, use the dimensionality reduction algorithm to extract the main influencing factors, construct the spatiotemporal propagation path of external events, and predict the external event behavior that may affect the pollution facilities in the future;

[0029] S36. Based on the analysis of space-time propagation networks, combined with historical data and external environmental changes, the propagation paths of future external events are predicted and the potential impact on emissions from pollution facilities is assessed.

[0030] Optionally, the S4 specifically includes:

[0031] S41. Use a hybrid deep generative model to model and predict time series features. The generative model is jointly trained with a variational autoencoder and a generative adversarial network. In the hybrid deep generative model, the generator generates simulated emission data by sampling from a high-dimensional latent space. The discriminator is used to distinguish between real emission data and generated data, and optimizes the representation of the latent space through variational inference. The loss function of the generator is defined as:

[0032]

[0033] Among them, L VAE-GAN represents the loss function of the generator, E represents the mathematical expectation, q(z|x) represents the approximate posterior, x represents the real emission data, z represents the random variable sampled from the latent space, p(x|z) represents the conditional probability of the emission data generated by the generator according to the random variable z, represents the simulated emission data generated by the generator, D(G(z)) represents the discriminant output of the discriminator on the generated data, ξ represents the adjustment factor of the time dimension, μ represents the mean in the latent space, KL(q(z|x)||p(z)) represents the KL divergence, t i and t j A timestamp representing the time step;

[0034] S42. Introduce a multi-dimensional attention mechanism to capture the contribution of each time step in the time series to other time steps and calculate the relevance weight of each time step:

[0035]

[0036] Among them, a t,t′ represents the attention weight between time step t and time step t′, T represents the total length of the time step, and W q represents the query matrix, W k represents the key matrix, h t represents the hidden state vector at time step t, h t′ represents the hidden state vector at time step t′, σ1 represents the distance adjustment factor, and p t represents the spatial coordinates of time step t, p t′ represents the spatial coordinates of time step t′, β1 represents the spatial dependence adjustment factor, and ζ represents the regularization parameter;

[0037] S43, combining the relevance weights of each time step generated by the multi-dimensional attention mechanism, by modeling the dependency of the generation of future emission data on the historical time series features;

[0038] S44, based on the hybrid deep generative model and attention mechanism, the performance of the generator and discriminator is continuously optimized through iterative training;

[0039] S45. The future time series features generated by the generator and the feedback of the discriminator are optimized through multiple iterations, and the generated time series features are finally output to predict external events that may affect the emissions of pollution facilities in the future.

[0040] Optionally, the S5 specifically includes:

[0041] S51. Model the causal relationship between external events and facility emission behaviors through a causal inference model, define external event feature vectors and facility emission feature vectors, and use a causal graph model to capture the causal relationship between nodes, where nodes represent external events and emission behaviors;

[0042] S52. Use hybrid causal inference models to combine short-term and long-term causal effects and capture the effects of external events on different time scales;

[0043] S53. The causal path analysis between external events and facility emissions is modeled in a complex manner through a multidimensional weight mechanism:

[0044]

[0045] Among them, C global (e i ) represents an external event e i Emission behavior j The global causal effect, P(p j ∣e i ,C ij ,G ij) represents a given external event e i , Causal Path C ij and dependency path G ij The conditional probability of emission behavior under ij ,G ij ,t) represents the causal path C at time t ij and dependency path G ij causal function, α2 represents the path sensitive adjustment factor, σ2 represents the smoothing parameter, exp represents the exponential function, T i and T j Represents external event e i and emission behavior p j In the time dimension, the timestamp, β2 represents the time adjustment factor, and γ2 represents the space adjustment factor;

[0046] S54. Based on causal counterfactual inference, the predicted value of facility emission behavior when the external event does not occur is:

[0047]

[0048] in, represents the counterfactual probability of facility emissions when the external event does not occur, E obs represents the actual observed external event characteristics, E cf represents the characteristics of external events under the counterfactual hypothesis, P obs represents the observed emission behavior, P cf represents the emission behavior under the counterfactual scenario, β3, σ3 and γ3 represent adjustment factors, and η3 and λ3 represent regularization parameters;

[0049] S55. Based on the results of the hybrid causal inference model and causal counterfactual inference, a forecast report is generated to analyze in detail the potential causal paths of external events, counterfactual impacts, and future warnings on facility emission behavior.

[0050] Optionally, the S6 specifically includes:

[0051] S61. Dynamically adjust the emission warning threshold using a Bayesian optimization algorithm based on strategy-based uncertainty estimation;

[0052] S62, Bayesian optimization algorithm dynamically adjusts the warning threshold through uncertainty estimation:

[0053]

[0054] Among them, V(t) represents the dynamic adjustment threshold of the current time step t, v low represents the initial low threshold, v mid represents the initial middle threshold, v highrepresents the initial high threshold, δ1, δ2 and δ3 represent adjustment factors, exp represents the exponential function, P(t) represents the emission observation data at the current time step t, E(t) represents the external event characteristics at the current time step t, V prev (t) represents the impact of the historical threshold on the current threshold, ρ1, ρ2, ρ3, ρ4 and ρ5 represent smoothing parameters, θ1, θ2 and θ3 represent nonlinear adjustment factors of different parts in the threshold adjustment process, represents the sensitivity of the nonlinear distance metric in the process of adjusting the threshold change, ξ1, ξ2 and ξ3 represent the adjustment factors, and p threshold represents the set threshold for pollution emission warning, P(t) <p threshold Indicates that the emission is within the normal range; p critical represents the emission critical point, P(t)>p critical Indicates that the emission behavior seriously exceeds the standard;

[0055] S63, updating the probability distribution of the current threshold in real time through Gaussian process regression, adjusting the threshold using the changes in P(t) and E(t), and generating a new threshold based on historical and current data;

[0056] S64, introduce uncertainty factors, adjust the threshold strategy by estimating the risk under the current threshold strategy;

[0057] S65. The system monitors emission behavior and external event characteristics in real time, and makes adjustments based on the predicted emission behavior and current thresholds, triggering warning signals when the thresholds are exceeded.

[0058] Optionally, the S7 specifically includes:

[0059] S71. Learn normal emission patterns from unlabeled emission data through self-supervised learning models. The self-supervised learning model extracts the potential pattern of normal emissions from unlabeled data by designing “pre-tasks” to predict the next time step or reconstruct some of the missing data;

[0060] S72. In the process of self-supervised learning, the self-supervised learning model embeds the features of normal emission behavior, maps high-dimensional emission data into low-dimensional latent space, and gradually establishes the time dependency of normal emission behavior and the distribution of emission features by comparing data at different time steps;

[0061] S73. Using the generated normal emission pattern, the self-supervised learning model detects behaviors that deviate from the normal emission pattern. When new emission data flows in, the current emission behavior is compared with the learned normal emission pattern. If a significant difference is detected between the current data and the normal emission pattern, it is marked as a potential abnormal emission behavior.

[0062] S74. After initially detecting abnormal emission behavior, use a few-sample learning technique to quickly learn from a limited number of abnormal emission behavior samples. The few-sample learning technique quickly identifies the characteristics of abnormal emission behavior by comparing a small number of abnormal emission behaviors with normal emission patterns, and constructs a new abnormal emission behavior feature vector.

[0063] S75. During the few-shot learning process, the self-supervised learning model automatically generates new abnormal emission pattern labels by analyzing the differences between abnormal emission behaviors and normal emission patterns and combining environmental conditions and historical abnormal data to generate new abnormal emission pattern labels;

[0064] S76. The system generates an emission anomaly detection report based on the learned normal emission patterns and abnormal labels, and issues early warnings for new abnormal emission behaviors. The report analyzes in detail the degree to which the emission data deviates from the normal pattern, identifies new abnormal patterns, and assesses potential risks.

[0065] The beneficial effects of the present invention are:

[0066] First, the present invention introduces cross-temporal and spatial event inference technology. Through graph neural networks and causal inference models, the system can effectively construct a correlation network between the emission behavior of pollution facilities and external events, and capture the complex event propagation effects in time and space dimensions. This technology enables the system to not only monitor current emission behavior, but also predict the potential impact of external events on emissions in advance. For example, how changes in meteorological conditions, natural disasters, or industrial activities affect the emission behavior of facilities, the system can give early warnings before these external events occur, improving the foresight and accuracy of the early warning. Traditional monitoring systems usually ignore the impact of external events, and the early warning lag is strong, while the cross-temporal and spatial event inference technology of the present invention significantly makes up for this shortcoming, providing managers with more time to deal with potential emission problems.

[0067] Secondly, the introduction of the self-supervised learning model greatly enhances the system's ability to learn normal emission patterns from unlabeled data. In the prior art, anomaly detection relies on a large amount of labeled data, while the present invention can automatically extract characteristic patterns of emission behavior from unlabeled data through a self-supervised learning model. Self-supervised learning can accurately detect anomalies in emission behavior even when data is scarce by designing pre-tasks, allowing the system to operate efficiently even without a large amount of labeled data, reducing the cost of data labeling and improving the system's autonomous learning ability.

[0068] Furthermore, the few-sample learning technology enables the present invention to quickly identify abnormal emission behaviors from limited abnormal samples. Traditional anomaly detection models require a large amount of abnormal data for training, while the few-sample learning technology can effectively solve the problem of scarce abnormal data. Even when there are only a few abnormal samples, the system can still quickly generate new abnormal emission pattern labels by comparing and learning with normal emission patterns. This fast learning capability enables the system to flexibly respond to complex and changing emission environments and enhances the ability to detect unknown or new abnormal emission behaviors.

[0069] In addition, the Bayesian optimization algorithm based on the uncertainty estimation of the strategy provides great flexibility for the dynamic adjustment of the system's warning threshold. Traditional early warning systems usually rely on static thresholds for judgment, and this method is prone to false alarms or missed alarms when the external environment changes. By introducing the Bayesian optimization algorithm, the present invention can dynamically adjust the warning threshold according to the real-time data of facility emissions and the feedback of the external environment, so that the system can flexibly adapt to changes in the environment and facility status. In particular, the introduction of uncertainty trade-offs in the decision-making process further improves the system's responsiveness and warning accuracy when facing complex and dynamic environments.

[0070] Finally, explainable artificial intelligence technology brings great transparency and operability to the present invention. Although the traditional black box model has improved the detection accuracy to a certain extent, it lacks clear interpretability, and it is difficult for managers to understand the reasons behind the warning results. Through explainable artificial intelligence technology, the present invention can generate a detailed abnormal source tracing report for each warning, explain the key data features that cause the abnormality and the basis for model decision-making, and help managers quickly identify problems and take effective measures. This not only improves the transparency of the system, but also greatly enhances the value of the early warning system as a decision support tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0072] Figure 1 This is a flow chart of the early warning method for abnormal emissions from pollution facilities based on big data analysis proposed by the present invention;

[0073] Figure 2 This is a schematic diagram of the abnormal emission detection and early warning system architecture of the abnormal emission early warning method for pollution facilities based on big data analysis proposed in the present invention. DETAILED DESCRIPTION

[0074] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0075] refer to Figure 1 and Figure 2 , an early warning method for abnormal emissions from pollution facilities based on big data analysis, including the following steps:

[0076] S1. Collect emission data, environmental data and external event data of pollution facilities through a sensor network and construct a data set. The sensor network transmits data through the Internet of Things technology;

[0077] S2, storing the data set in a big data storage platform, and using the Hadoop big data processing framework to store and preprocess the data set, wherein the preprocessing includes data cleaning, missing value filling, noise removal, and time series feature extraction;

[0078] S3. Use graph neural network models to model the complex relationship between external events and the emission behavior of pollution facilities, build a spatiotemporal propagation network of external events, and analyze the potential impact of external events on the emissions of pollution facilities;

[0079] S4. Use a hybrid deep generative model, combined with attention mechanism and variational inference, to model and predict time series features and predict external events that may affect emissions from pollution facilities in the future;

[0080] S5. Analyze the causal relationship between external events and facility emissions through causal inference models, determine the potential impact of external events on facility emissions, and generate a forecast report;

[0081] S6. Adopting the Bayesian optimization algorithm based on strategy uncertainty estimation, introducing uncertainty trade-offs in the decision-making process, and dynamically adjusting the emission warning threshold;

[0082] S7. Use a self-supervised learning model to learn normal emission patterns from unlabeled data, identify abnormal emission behaviors that deviate from the normal pattern, and quickly learn from limited abnormal samples through few-shot learning technology to generate new abnormal emission pattern labels;

[0083] S8. Use explainable artificial intelligence technology to explain the warning results and generate anomaly source tracing reports, showing the key data features that cause the anomalies, the basis for model decision-making and the tracing path, and provide response suggestions and action plans.

[0084] In this implementation, S3 specifically includes:

[0085] S31. Construct a heterogeneous graph structure based on the emission data of pollution facilities, environmental data and external event data. The nodes include pollution facilities, environmental variables and external events. The edges represent the associations between different nodes. The weight of each edge is calculated based on the spatiotemporal association:

[0086]

[0087] Among them, W ij represents the association weight between node i and node j, exp represents the exponential function, x i represents the feature vector of node i, x j represents the feature vector of node j, p i represents the spatial coordinates of node i, p j represents the spatial coordinate of node j, y i represents the timestamp of node i, y j represents the timestamp of node j, e i represents the external event characteristics of node i, e j represents the external event characteristics of node j, σ, α, γ, β and η represent the adjustment factors;

[0088] S32. Use graph neural network to process heterogeneous graphs, and update the features of nodes through the message passing mechanism of neighbor nodes:

[0089]

[0090] in, represents the feature representation of node i in the k+1 layer, σ represents the activation function, represents the feature representation of node i at layer k, represents the feature representation of node m at layer k, represents the feature representation of node j at the kth layer, N(i) represents the neighbor set of node i, λ represents the regularization coefficient, represents the bias term of node i in the kth layer, λ represents the regularization coefficient, and W im represents the association weight between node i and node m;

[0091] S33. Through the message passing mechanism of the spatiotemporal graph neural network, combined with historical external event data and environmental data, the association weights between nodes in the graph are dynamically adjusted. The state of the node at each moment is updated according to the historical propagation path, and the node features are updated synchronously in the spatial and temporal dimensions.

[0092] S34, based on the spatiotemporal model constructed by the spatiotemporal propagation network, further adjust the feature representation of each node and update it through multiple iteration cycles;

[0093] S35. Generate the final high-dimensional representation of each node through the iterative propagation mechanism of the multi-layer spatiotemporal graph neural network, use the dimensionality reduction algorithm to extract the main influencing factors, construct the spatiotemporal propagation path of external events, and predict the external event behavior that may affect the pollution facilities in the future;

[0094] S36. Based on the analysis of space-time propagation networks, combined with historical data and external environmental changes, the propagation paths of future external events are predicted and the potential impact on emissions from pollution facilities is assessed.

[0095] In this implementation, S4 specifically includes:

[0096] S41. Use a hybrid deep generative model to model and predict time series features. The generative model is jointly trained with a variational autoencoder and a generative adversarial network. In the hybrid deep generative model, the generator generates simulated emission data by sampling from a high-dimensional latent space. The discriminator is used to distinguish between real emission data and generated data, and optimizes the representation of the latent space through variational inference. The loss function of the generator is defined as:

[0097]

[0098]

[0099] Among them, L VAE-GAN represents the loss function of the generator, E represents the mathematical expectation, q(z|x) represents the approximate posterior, x represents the real emission data, z represents the random variable sampled from the latent space, p(x|z) represents the conditional probability of the emission data generated by the generator according to the random variable z, represents the simulated emission data generated by the generator, D(G(z)) represents the discriminant output of the discriminator on the generated data, ξ represents the adjustment factor of the time dimension, μ represents the mean in the latent space, KL(q(z|x)||p(z)) represents the KL divergence, t i and t j A timestamp representing the time step;

[0100] S42. Introduce a multi-dimensional attention mechanism to capture the contribution of each time step in the time series to other time steps and calculate the relevance weight of each time step:

[0101]

[0102] Among them, a t,t′ represents the attention weight between time step t and time step t′, T represents the total length of the time step, and W q represents the query matrix, W k represents the key matrix, h t represents the hidden state vector at time step t, h t′represents the hidden state vector at time step t′, σ1 represents the distance adjustment factor, and p t represents the spatial coordinates of time step t, p t′ represents the spatial coordinates of time step t′, β1 represents the spatial dependence adjustment factor, and ζ represents the regularization parameter;

[0103] S43, combining the relevance weights of each time step generated by the multi-dimensional attention mechanism, by modeling the dependency of the generation of future emission data on the historical time series features;

[0104] S44, based on the hybrid deep generative model and attention mechanism, the performance of the generator and discriminator is continuously optimized through iterative training;

[0105] S45. The future time series features generated by the generator and the feedback of the discriminator are optimized through multiple iterations, and the generated time series features are finally output to predict external events that may affect the emissions of pollution facilities in the future.

[0106] In this implementation manner, S5 specifically includes:

[0107] S51. Model the causal relationship between external events and facility emission behaviors through a causal inference model, define external event feature vectors and facility emission feature vectors, and use a causal graph model to capture the causal relationship between nodes, where nodes represent external events and emission behaviors;

[0108] S52. Use hybrid causal inference models to combine short-term and long-term causal effects and capture the effects of external events on different time scales;

[0109] S53. The causal path analysis between external events and facility emissions is modeled in a complex manner through a multidimensional weight mechanism:

[0110]

[0111] Among them, C global (e i ) represents an external event e i Emission behavior j The global causal effect, P(p j ∣e i ,C ij ,G ij Represents a given external event e i , Causal Path C ij and dependency path G ij The conditional probability of emission behavior under ij ,G ij ,t) represents the causal path C at time t ij and dependency path G ijcausal function, α2 represents the path sensitive adjustment factor, σ2 represents the smoothing parameter, exp represents the exponential function, T i and T j Represents external event e i and emission behavior p j In the time dimension, the timestamp, β2 represents the time adjustment factor, and γ2 represents the space adjustment factor;

[0112] S54. Based on causal counterfactual inference, the predicted value of facility emission behavior when the external event does not occur is:

[0113]

[0114] in, represents the counterfactual probability of facility emissions when the external event does not occur, E obs represents the actual observed external event characteristics, E cf represents the characteristics of external events under the counterfactual hypothesis, P obs represents the observed emission behavior, P cf represents the emission behavior under the counterfactual scenario, β3, σ3 and γ3 represent adjustment factors, and η3 and λ3 represent regularization parameters;

[0115] S55. Based on the results of the hybrid causal inference model and causal counterfactual inference, a forecast report is generated to analyze in detail the potential causal paths of external events, counterfactual impacts, and future warnings on facility emission behavior.

[0116] In this implementation manner, S6 specifically includes:

[0117] S61. Dynamically adjust the emission warning threshold using a Bayesian optimization algorithm based on strategy-based uncertainty estimation;

[0118] S62, Bayesian optimization algorithm dynamically adjusts the warning threshold through uncertainty estimation:

[0119]

[0120] Among them, V(t) represents the dynamic adjustment threshold of the current time step t, v low represents the initial low threshold, v mid represents the initial middle threshold, v high represents the initial high threshold, δ1, δ2 and δ3 represent adjustment factors, exp represents the exponential function, P(t) represents the emission observation data at the current time step t, E(t) represents the external event characteristics at the current time step t, V prev(t) represents the impact of the historical threshold on the current threshold, ρ1, ρ2, ρ3, ρ4 and ρ5 represent smoothing parameters, θ1, θ2 and θ3 represent nonlinear adjustment factors of different parts in the threshold adjustment process, represents the sensitivity of the nonlinear distance metric in the process of adjusting the threshold change, ξ1, ξ2 and ξ3 represent the adjustment factors, and p threshold represents the set threshold for pollution emission warning, P(t) <p threshold Indicates that the emission is within the normal range; p critical represents the emission critical point, P(t)>p critical Indicates that the emission behavior seriously exceeds the standard;

[0121] S63, updating the probability distribution of the current threshold in real time through Gaussian process regression, adjusting the threshold using the changes in P(t) and E(t), and generating a new threshold based on historical and current data;

[0122] S64, introduce uncertainty factors, adjust the threshold strategy by estimating the risk under the current threshold strategy;

[0123] S65. The system monitors emission behavior and external event characteristics in real time, and makes adjustments based on the predicted emission behavior and current thresholds, triggering warning signals when the thresholds are exceeded.

[0124] In this implementation manner, the S7 specifically includes:

[0125] S71. Learn normal emission patterns from unlabeled emission data through self-supervised learning models. The self-supervised learning model extracts the potential pattern of normal emissions from unlabeled data by designing “pre-tasks” to predict the next time step or reconstruct some of the missing data;

[0126] S72. In the process of self-supervised learning, the self-supervised learning model embeds the features of normal emission behavior, maps high-dimensional emission data into low-dimensional latent space, and gradually establishes the time dependency of normal emission behavior and the distribution of emission features by comparing data at different time steps;

[0127] S73. Using the generated normal emission pattern, the self-supervised learning model detects behaviors that deviate from the normal emission pattern. When new emission data flows in, the current emission behavior is compared with the learned normal emission pattern. If a significant difference is detected between the current data and the normal emission pattern, it is marked as a potential abnormal emission behavior.

[0128] S74. After initially detecting abnormal emission behavior, use a few-sample learning technique to quickly learn from a limited number of abnormal emission behavior samples. The few-sample learning technique quickly identifies the characteristics of abnormal emission behavior by comparing a small number of abnormal emission behaviors with normal emission patterns, and constructs a new abnormal emission behavior feature vector.

[0129] S75. During the few-shot learning process, the self-supervised learning model automatically generates new abnormal emission pattern labels by analyzing the differences between abnormal emission behaviors and normal emission patterns and combining environmental conditions and historical abnormal data to generate new abnormal emission pattern labels;

[0130] S76. The system generates an emission anomaly detection report based on the learned normal emission patterns and abnormal labels, and issues early warnings for new abnormal emission behaviors. The report analyzes in detail the degree to which the emission data deviates from the normal pattern, identifies new abnormal patterns, and assesses potential risks.

[0131] Embodiment 1:

[0132] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the pollution monitoring system of a large chemical plant located in a certain industrial park. The emissions of the plant mainly include sulfur dioxide, nitrogen oxides and particulate matter. Due to the large production scale and complex process of the factory, the emission volume fluctuates greatly, and the surrounding meteorological conditions often change. The traditional static threshold monitoring system can no longer effectively cope with complex emission behaviors. In order to improve the sensitivity and accuracy of the early warning, the factory decided to adopt the early warning method of abnormal emissions from pollution facilities based on big data analysis of the present invention, aiming to enhance the detection capability of abnormal emission behaviors, and realize dynamic adjustment of emission warning thresholds, so as to ensure timely and accurate prediction and processing of abnormal emission events.

[0133] In this embodiment, the factory collects emission data such as sulfur dioxide, nitrogen oxides and particulate matter in real time through a sensor network covering the entire factory, as well as meteorological data and external event information (such as emissions from other factories in the region, sudden natural disasters and traffic pollution events). The sensor network uses Internet of Things technology for data transmission, and the amount of data collected per hour is as high as hundreds of GB. All of this data is first cleaned, missing values ​​are filled and features are extracted through the Hadoop big data processing platform to form a comprehensive data set that includes emission behavior, environmental conditions and external event impacts.

[0134] After data preprocessing, the graph neural network model is used to model the complex spatiotemporal correlation between factory emission data and external events. This model not only captures the changes in factory emission behavior over time and space, but also analyzes how environmental factors such as wind speed and wind direction affect the diffusion of pollutants. This model constructs a spatiotemporal propagation network of external events, and combines variational inference to predict external events that may affect emissions in the next few hours. For example, the wind speed suddenly increases on a certain day. Combining meteorological and emission data, the system predicts that changes in wind speed will intensify the spread of PM2.5 and cause excessive emissions.

[0135] Next, the causal inference model further analyzes the causal relationship between external events and facility emissions. For example, through causal analysis, the system found that when air pressure suddenly drops, sulfur dioxide concentrations tend to rise rapidly. The establishment of this causal inference model enables the system to generate detailed forecast reports to help managers take measures in advance to deal with emission changes caused by external events.

[0136] The setting of the warning threshold is particularly important in this embodiment. The traditional method uses a fixed threshold to determine whether the emissions exceed the standard, which makes it difficult to dynamically respond to fluctuations in facility emissions. The Bayesian optimization algorithm introduced in the present invention can dynamically adjust the warning threshold according to real-time emission data and external environmental changes. The system introduces uncertainty estimation of the strategy, and adjusts the emission warning threshold in real time according to the factory's historical emission data and environmental changes. Take a certain day as an example. Due to the sudden high temperature weather, the factory's sulfur dioxide emissions rose by more than 50% in a short period of time. The traditional monitoring system may fail to issue a timely warning due to the limitation of the fixed threshold, but in the system of the present invention, through Bayesian optimization, the system adjusts the sulfur dioxide warning threshold in real time, and issues a warning signal 3 hours before the emissions exceed the standard.

[0137] In terms of detecting abnormal emission behavior, the self-supervised learning model automatically learns normal emission patterns from the factory's daily emission data and detects deviations from the normal pattern. The system successfully established a time series feature model of normal emission behavior by comparing and analyzing emission data in various meteorological conditions, factory process changes, and production scale adjustments. For example, in a certain week, the factory's normal sulfur dioxide emissions were between 100-120ppm, and the PM2.5 concentration remained at 40-50 micrograms per cubic meter. The self-supervised learning model can automatically learn this normal range without labeled data, and send out abnormal signals when new data deviates from this range.

[0138] In addition, the system also uses few-sample learning technology, which can learn and generate new abnormal emission pattern labels from very few abnormal samples even when the abnormal data is limited. In actual application, one night, the system detected a sudden increase in nitrogen oxide emissions in a short period of time, but due to the lack of labeled data for similar emergencies, traditional systems are difficult to deal with. The few-sample learning technology in the present invention can quickly generate new abnormal pattern labels from very few historical abnormal data, successfully identify the sudden increase in nitrogen oxides as abnormal emission behavior, and respond within 5 minutes after the incident, issuing a timely warning.

[0139] The present invention also generates abnormal source tracing reports through explainable artificial intelligence technology to help factory managers understand the cause of each abnormal emission. For example, in a certain abnormal warning, the system found that the excessive sulfur dioxide was caused by aging process equipment by analyzing historical data and current abnormal emission patterns. In another scenario, the system identified the causal relationship between external traffic pollution events and excessive PM2.5 concentrations. Through this transparent explanation mechanism, managers can accurately determine the root cause of the problem and take effective measures in a timely manner.

[0140] In order to verify the effectiveness of the present invention, the factory implemented a 6-month monitoring and testing. The following table shows the experimental test data:

[0141] Table 1 Comparison data between the traditional monitoring system and the abnormal emission warning system of pollution facilities of the present invention

[0142]

[0143] Through the analysis of Table 1, it can be seen that the present invention has significant advantages in detection accuracy, false alarm control and overall warning performance.

[0144] First, from the perspective of total monitoring time, both systems were operated within the same 6-month period, ensuring the comparability of the comparison results. During this period, the conventional monitoring system detected 25 abnormal emission behaviors, while the monitoring system of the present invention detected 33 abnormal emissions, an increase of 8 times, indicating that the system of the present invention is more sensitive in abnormal emission detection and can capture more potential risks. This difference shows that the system of the present invention has higher sensitivity when dealing with complex and dynamic emission behaviors.

[0145] In the detection of abnormal emissions caused by equipment problems, the traditional monitoring system detected 17 times, while the system of the present invention detected 22 times. By adopting self-supervised learning and few-sample learning technology, the present invention can quickly identify emission problems caused by equipment failures under limited abnormal data, effectively improving the response speed and detection capability of equipment problems.

[0146] For abnormal emissions caused by external events (such as weather changes, emissions from other facilities, etc.), the traditional system detected 8 times in total, while the system of the present invention detected 11 times, an increase of about 38%. This is directly related to the cross-temporal and spatial event inference technology used in the present invention. By analyzing the causal relationship between external events and facility emission behavior, the present invention can predict the potential impact of external events on emissions in advance, making the system more forward-looking and early warning.

[0147] In terms of false alarm rate, the traditional monitoring system generated 10 false alarms, while the present invention only had 3. This significant difference is mainly due to the Bayesian optimization algorithm of the present invention, which avoids the false alarm problem in the traditional static threshold setting by dynamically adjusting the warning threshold. At the same time, based on the uncertainty estimation of the strategy, the system can flexibly adapt to changes in a complex and changing environment, reducing the occurrence of false alarms. This not only improves the accuracy of the warning, but also reduces the waste of resources and the burden on managers caused by false alarms.

[0148] Overall, the warning accuracy of the monitoring system of the present invention has increased by 45%, and the false alarm rate has decreased by 28%. This result shows that the present invention effectively improves the performance of the pollution monitoring system through a variety of advanced algorithms and models (such as self-supervised learning, few-sample learning, Bayesian optimization, and cross-temporal and spatial event inference). The system can not only detect more abnormal emission behaviors, but also reduce false alarms, ensuring the timeliness and accuracy of warnings. This improvement significantly optimizes the management efficiency of pollution facilities and helps to improve the ability to respond to abnormal emission events.

[0149] In general, the application of the present invention in the factory significantly improves the detection efficiency and accuracy of abnormal emission behavior, optimizes the warning threshold adjustment mechanism, and provides managers with more transparent and operational decision support. This application not only reduces the risk of environmental violations caused by abnormal emissions, but also provides a more flexible and intelligent solution for the factory in dynamic emission monitoring.

[0150] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An early warning method for abnormal emissions from pollution facilities based on big data analysis, characterized in that: The steps include: S1. Collect emission data, environmental data and external event data of pollution facilities through a sensor network and construct a data set. The sensor network transmits data through the Internet of Things technology; S2, storing the data set in a big data storage platform, and using the Hadoop big data processing framework to store and preprocess the data set, wherein the preprocessing includes data cleaning, missing value filling, noise removal, and time series feature extraction; S3. Use graph neural network models to model the complex relationship between external events and the emission behavior of pollution facilities, build a spatiotemporal propagation network of external events, and analyze the potential impact of external events on the emissions of pollution facilities; S4. Use a hybrid deep generative model, combined with attention mechanism and variational inference, to model and predict time series features and predict external events that may affect emissions from pollution facilities in the future; S5. Analyze the causal relationship between external events and facility emissions through causal inference models, determine the potential impact of external events on facility emissions, and generate a forecast report; S6. Adopting the Bayesian optimization algorithm based on strategy uncertainty estimation, introducing uncertainty trade-offs in the decision-making process, and dynamically adjusting the emission warning threshold; S7. Use a self-supervised learning model to learn normal emission patterns from unlabeled data, identify abnormal emission behaviors that deviate from the normal pattern, and quickly learn from limited abnormal samples through few-shot learning technology to generate new abnormal emission pattern labels; S8. Use explainable artificial intelligence technology to explain the warning results and generate anomaly source tracing reports, showing the key data features that cause the anomalies, the basis for model decision-making and the tracing path, and provide response suggestions and action plans.

2. The method for early warning of abnormal emissions from pollution facilities based on big data analysis according to claim 1 is characterized in that: The S3 specifically includes: S31. Construct a heterogeneous graph structure based on the emission data of pollution facilities, environmental data and external event data. The nodes include pollution facilities, environmental variables and external events. The edges represent the associations between different nodes. The weight of each edge is calculated based on the spatiotemporal association: Among them, W ij represents the association weight between node i and node j, exp represents the exponential function, x i represents the feature vector of node i, x j represents the feature vector of node j, p i represents the spatial coordinates of node i, p j represents the spatial coordinate of node j, y i represents the timestamp of node i, y j represents the timestamp of node j, e i represents the external event characteristics of node i, e j represents the external event characteristics of node j, σ, α, γ, β and η represent the adjustment factors; S32. Use graph neural network to process heterogeneous graphs, and update the features of nodes through the message passing mechanism of neighbor nodes: in, represents the feature representation of node i in the k+1 layer, σ represents the activation function, represents the feature representation of node i at layer k, represents the feature representation of node m at layer k, represents the feature representation of node j at the kth layer, N(i) represents the neighbor set of node i, λ represents the regularization coefficient, represents the bias term of node i in the kth layer, λ represents the regularization coefficient, and W im represents the association weight between node i and node m; S33. Through the message passing mechanism of the spatiotemporal graph neural network, combined with historical external event data and environmental data, the association weights between nodes in the graph are dynamically adjusted. The state of the node at each moment is updated according to the historical propagation path, and the node features are updated synchronously in the spatial and temporal dimensions. S34, based on the spatiotemporal model constructed by the spatiotemporal propagation network, further adjust the feature representation of each node and update it through multiple iteration cycles; S35. Generate the final high-dimensional representation of each node through the iterative propagation mechanism of the multi-layer spatiotemporal graph neural network, use the dimensionality reduction algorithm to extract the main influencing factors, construct the spatiotemporal propagation path of external events, and predict the external event behavior that may affect the pollution facilities in the future; S36. Based on the analysis of space-time propagation networks, combined with historical data and external environmental changes, the propagation paths of future external events are predicted and the potential impact on emissions from pollution facilities is assessed.

3. The method for early warning of abnormal emissions from pollution facilities based on big data analysis according to claim 1 is characterized in that: The S4 specifically includes: S41. Use a hybrid deep generative model to model and predict time series features. The generative model is jointly trained with a variational autoencoder and a generative adversarial network. In the hybrid deep generative model, the generator generates simulated emission data by sampling from a high-dimensional latent space. The discriminator is used to distinguish between real emission data and generated data, and optimizes the representation of the latent space through variational inference. The loss function of the generator is defined as: Among them, L VAE-GAN represents the loss function of the generator, E represents the mathematical expectation, q(z|x) represents the approximate posterior, x represents the real emission data, z represents the random variable sampled from the latent space, p(x|z) represents the conditional probability of the emission data generated by the generator according to the random variable z, represents the simulated emission data generated by the generator, D(G(z)) represents the discriminant output of the discriminator on the generated data, ξ represents the adjustment factor of the time dimension, μ represents the mean in the latent space, KL(q(z|x)||p(z)) represents the KL divergence, t i and t j A timestamp representing the time step; S42. Introduce a multi-dimensional attention mechanism to capture the contribution of each time step in the time series to other time steps and calculate the relevance weight of each time step: Among them, a t,t′ represents time step t and time step t ′ The attention weight between them, T represents the total length of the time step, W q represents the query matrix, W k represents the key matrix, h t represents the hidden state vector at time step t, h t′ represents the hidden state vector at time step t′, σ1 represents the distance adjustment factor, and p t represents the spatial coordinates of time step t, p t′ represents the time step t ′ The spatial coordinates of , β1 represents the spatial dependence adjustment factor, and ζ represents the regularization parameter; S43, combining the relevance weights of each time step generated by the multi-dimensional attention mechanism, by modeling the dependency of the generation of future emission data on the historical time series features; S44, based on the hybrid deep generative model and attention mechanism, the performance of the generator and discriminator is continuously optimized through iterative training; S45. The future time series features generated by the generator and the feedback of the discriminator are optimized through multiple iterations, and the generated time series features are finally output to predict external events that may affect the emissions of pollution facilities in the future.

4. The method for early warning of abnormal emissions from pollution facilities based on big data analysis according to claim 1 is characterized in that: The S5 specifically includes: S51. Model the causal relationship between external events and facility emission behaviors through a causal inference model, define external event feature vectors and facility emission feature vectors, and use a causal graph model to capture the causal relationship between nodes, where nodes represent external events and emission behaviors; S52. Use hybrid causal inference models to combine short-term and long-term causal effects and capture the effects of external events on different time scales; S53. The causal path analysis between external events and facility emissions is modeled in a complex manner through a multidimensional weight mechanism: Among them, C global (e i ) represents an external event e i Emission behavior j The global causal effect, P(p j ∣e i ,C ij ,G ij Represents a given external event e i , Causal Path C ij and dependency path G ij The conditional probability of emission behavior under ij ,G ij ,t) represents the causal path C at time t ij and dependency path G ij causal function, α2 represents the path sensitive adjustment factor, σ2 represents the smoothing parameter, exp represents the exponential function, T i and T j Represents external event e i and emission behavior p j In the time dimension, the timestamp, β2 represents the time adjustment factor, and γ2 represents the space adjustment factor; S54. Based on causal counterfactual inference, the predicted value of facility emission behavior when the external event does not occur is: in, represents the counterfactual probability of facility emissions when the external event does not occur, E obs represents the actual observed external event characteristics, E cf represents the characteristics of external events under the counterfactual hypothesis, P obs represents the observed emission behavior, P cf represents the emission behavior under the counterfactual scenario, β3, σ3 and γ3 represent adjustment factors, and η3 and λ3 represent regularization parameters; S55. Based on the results of the hybrid causal inference model and causal counterfactual inference, a forecast report is generated to analyze in detail the potential causal paths of external events, counterfactual impacts, and future warnings on facility emission behavior.

5. The method for early warning of abnormal emissions from pollution facilities based on big data analysis according to claim 1 is characterized in that: The S6 specifically includes: S61. Dynamically adjust the emission warning threshold using a Bayesian optimization algorithm based on strategy-based uncertainty estimation; S62, Bayesian optimization algorithm dynamically adjusts the warning threshold through uncertainty estimation: Among them, V(t) represents the dynamic adjustment threshold of the current time step t, v low represents the initial low threshold, v mid represents the initial middle threshold, v high represents the initial high threshold, δ1, δ2 and δ3 represent adjustment factors, exp represents the exponential function, P(t) represents the emission observation data at the current time step t, and E ( t) represents the external event characteristics of the current time step t, V prev ( t ) represents the influence of historical threshold on current threshold, ρ1, ρ2, ρ3, ρ4 and ρ5 represent smoothing parameters, θ1, θ2 and θ3 represent nonlinear adjustment factors of different parts in threshold adjustment process, represents the sensitivity of the nonlinear distance metric in the process of adjusting the threshold change, ξ1, ξ2 and ξ3 represent the adjustment factors, and p threshold represents the set threshold for pollution emission warning, P(t) <p threshold Indicates that the emission is within the normal range; p critical represents the emission critical point, P(t)>p critical Indicates that the emission behavior seriously exceeds the standard; S63, updating the probability distribution of the current threshold in real time through Gaussian process regression, adjusting the threshold using the changes in P(t) and E(t), and generating a new threshold based on historical and current data; S64, introduce uncertainty factors, adjust the threshold strategy by estimating the risk under the current threshold strategy; S65. The system monitors emission behavior and external event characteristics in real time, and makes adjustments based on the predicted emission behavior and current thresholds, triggering warning signals when the thresholds are exceeded.

6. The method for early warning of abnormal emissions from pollution facilities based on big data analysis according to claim 1 is characterized in that: The S7 specifically includes: S71. Learn normal emission patterns from unlabeled emission data through a self-supervised learning model. The self-supervised learning model extracts the potential pattern of normal emissions from unlabeled data by designing "pre-tasks" to predict the next time step or reconstruct some of the missing data; S72. In the process of self-supervised learning, the self-supervised learning model embeds the features of normal emission behavior, maps high-dimensional emission data into low-dimensional latent space, and gradually establishes the time dependency of normal emission behavior and the distribution of emission features by comparing data at different time steps; S73. Using the generated normal emission pattern, the self-supervised learning model detects behaviors that deviate from the normal emission pattern. When new emission data flows in, the current emission behavior is compared with the learned normal emission pattern. If a significant difference is detected between the current data and the normal emission pattern, it is marked as a potential abnormal emission behavior. S74. After initially detecting abnormal emission behavior, use a few-sample learning technique to quickly learn from a limited number of abnormal emission behavior samples. The few-sample learning technique quickly identifies the characteristics of abnormal emission behavior by comparing a small number of abnormal emission behaviors with normal emission patterns, and constructs a new abnormal emission behavior feature vector. S75. In the process of few-shot learning, the self-supervised learning model automatically generates new abnormal emission pattern labels by analyzing the differences between abnormal emission behaviors and normal emission patterns and combining environmental conditions and historical abnormal data to generate new abnormal emission pattern labels; S76. The system generates an emission anomaly detection report based on the learned normal emission patterns and abnormal labels, and issues early warnings for new abnormal emission behaviors. The report analyzes in detail the degree to which the emission data deviates from the normal pattern, identifies new abnormal patterns, and assesses potential risks.

Images