Flue gas monitoring and early warning system based on artificial intelligence

Abstract

The invention discloses a flue gas monitoring and early warning system based on artificial intelligence, and relates to the technical field of industrial safety monitoring and environmental protection. Risk state prediction in a future time period is carried out in a fusion feature space, and a comprehensive risk value in a single scalar form is constructed by using a weighted norm and statistical aggregation; the system can pay attention to the abnormal amplitude, the duration time and the cross-channel linkage degree at the same time, so that earlier and more stable response to global risks with the development trend is achieved while accidental fluctuation false alarms are restrained; a comprehensive risk value output by the system is compared with a preset grading threshold value, a multi-grade early warning and automatic control strategy can be directly driven, a layered defense line is formed for lithium battery thermal runaway and harmful gas leakage possibly occurring in high-risk scenes such as an energy storage cabinet, operation and maintenance inspection or load reduction is guided in a light abnormal stage, and the safety of the system is improved. And protection actions such as shutdown and ventilation are linked in a severe abnormal stage.

Description

Technical Field

[0001] This invention relates to the field of industrial safety monitoring and environmental protection technology, and in particular to an artificial intelligence-based flue gas monitoring and early warning system. Background Technology

[0002] In the field of industrial safety and environmental monitoring, the application of flue gas and hazardous gas monitoring and early warning systems is gradually expanding from open environments to enclosed or semi-enclosed spaces. Typical scenarios include battery cabinets in energy storage power stations, process pipeline rooms in petrochemical plants, and underground integrated pipe corridors. These spatial structures are relatively enclosed, and once risk events such as lithium battery thermal runaway or flammable or toxic gas leaks occur, they can develop rapidly and cause great damage, placing higher demands on the response timeliness and reliability of early warning systems.

[0003] To address these needs, some existing early warning systems deploy various types of sensors at key locations, such as smoke sensors, hydrogen sensors, carbon monoxide sensors, volatile organic compound sensors, and temperature sensors, forming a multi-source sensing and monitoring network. At the data processing level, existing technologies typically establish time series analysis or prediction models for the historical monitoring data of each type of sensor. For example, they use trend filtering methods based on exponentially weighted moving averages and autoregressive algorithms to estimate the future trends of each monitoring indicator and trigger an early warning based on whether the predicted values ​​of one or more indicators exceed a preset threshold.

[0004] However, in enclosed or semi-enclosed spaces, the responses of multi-source heterogeneous sensors to the same risk event vary significantly. Different sensors exhibit considerable inconsistencies in response sensitivity, response time constant, background noise, and fluctuation characteristics. In the early stages of risk emergence, the output signals of various sensors often show only slight changes with small amplitudes, inconsistent timing, or even diverging from each other within a short period. In this situation, the existing approach of independently modeling and predicting for each type of sensor is insufficient to promptly and accurately distinguish between early signs of local interference and global risks. This often leads to two problems: on the one hand, oversensitivity to occasional fluctuations of individual sensors resulting in false alarms; on the other hand, waiting for more obvious and sustained abnormal changes from multiple sensors before triggering an early warning, resulting in uncertainty or even delay in the timing of the warning.

[0005] To alleviate the above problems, some existing solutions set up more complex multi-indicator linkage triggering logic, or assign fixed weights to different sensors based on expert experience, in order to achieve comprehensive judgment of monitoring data. However, such collaborative judgment methods based on preset rules are difficult to adaptively characterize the temporal correlation and dynamic coupling relationship between multi-sensor signals under different risk modes. In application scenarios with complex risk characteristics and variable operating conditions, it is still difficult to balance the sensitivity and reliability of early warning. Therefore, how to fully explore the inherent correlation between various monitoring quantities under multi-source sensor conditions and improve the early warning capability of smoke and harmful gas risks in closed or semi-closed spaces remains one of the technical problems to be solved in this field. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] This invention provides an artificial intelligence-based flue gas monitoring and early warning system that solves the problem that existing multi-sensor flue gas early warning systems, which mostly rely on independent modeling and threshold logic, are difficult to handle early asynchronous and weak signals.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] This invention provides an artificial intelligence-based flue gas monitoring and early warning system, comprising:

[0010] A multi-source sensor network is used to be deployed in the monitoring area. The multi-source sensor network includes at least two types of flue gas parameter sensors for collecting corresponding flue gas parameter monitoring data.

[0011] The data acquisition module is used to receive and store timestamped monitoring data from the multi-source sensor network.

[0012] A data processing and analysis engine, connected to the data acquisition module, is used to perform fusion modeling and risk prediction on the monitoring data. The data processing and analysis engine includes:

[0013] The time-series data matrix construction unit is used to construct a multivariate time-series data matrix from the monitoring data from the multi-source sensor network in chronological order;

[0014] An attention-weighted fusion unit, connected to the time-series data matrix construction unit, is used to perform fusion processing on the multivariate time-series data matrix. The attention-weighted fusion unit is built based on a neural network, dynamically calculates the weights reflecting the correlation between monitoring data from different sensors through an attention mechanism, and uses the weights to perform weighted fusion on the multivariate time-series data matrix, outputting a fused feature vector sequence.

[0015] A risk state prediction unit, connected to the attention-weighted fusion unit, is used to receive the fused feature vector sequence and perform time series prediction to obtain a risk state sequence covering a preset future time range.

[0016] A comprehensive risk calculation unit, connected to the risk state prediction unit, is used to process the risk state sequence and calculate a comprehensive risk value in scalar form.

[0017] The early warning judgment module is connected to the comprehensive risk calculation unit and has a preset risk threshold. It is used to compare the comprehensive risk value with the risk threshold and generate an early warning signal based on the comparison result.

[0018] As a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, the data processing and analysis engine further includes a sliding window management unit.

[0019] The sliding window management unit is used to maintain a dynamic data window with a fixed time length and update step size.

[0020] The time-series data matrix construction unit is used to extract monitoring data from the dynamic data window to construct the multivariate time-series data matrix.

[0021] As a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, the attention-weighted fusion unit comprises a linear projection layer, a multi-head self-attention layer, and a feedforward neural network layer connected in sequence.

[0022] The multi-head self-attention layer is used to calculate attention weights based on query vectors, key vectors, and value vectors, and to perform weighted fusion on the multivariate time-series data matrix.

[0023] In a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, the risk state sequence is a predicted concentration vector sequence for multiple consecutive future time points.

[0024] Each predicted concentration vector includes the predicted concentration values ​​corresponding to the at least two types of flue gas parameter sensors.

[0025] As a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, the comprehensive risk calculation unit is specifically used for:

[0026] The weighted norm is calculated for each predicted concentration vector in the predicted concentration vector sequence to obtain a scalar risk instantaneous value sequence;

[0027] Then, a statistical aggregation operation is performed on the scalar risk instantaneous value sequence to obtain the comprehensive risk value.

[0028] In a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, the early warning judgment module is configured to output a graded early warning signal based on the comparison result between the comprehensive risk value and the risk threshold.

[0029] The graded early warning signal includes at least two different risk levels, corresponding to different alarm display methods and / or control response strategies.

[0030] As a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, it further includes an execution control module.

[0031] The execution control module is connected to the early warning judgment module. After receiving the early warning signal, it generates control instructions according to preset control rules and sends them to the controlled process equipment to perform load reduction, shutdown or ventilation safety protection actions.

[0032] In a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system described in this invention, the monitoring area is an energy storage cabinet.

[0033] The multi-source sensor network includes a hydrogen sensor and a carbon monoxide sensor for detecting characteristic gases of thermal runaway in lithium batteries.

[0034] As a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system of the present invention, the multi-source sensor network further includes at least one of a smoke sensor, a temperature sensor, and / or a volatile organic compound sensor.

[0035] As a preferred embodiment of the artificial intelligence-based flue gas monitoring and early warning system of the present invention, the time series prediction model in the risk state prediction unit is at least one of a multivariate autoregressive integral moving average model, a temporal convolutional network model, or a recurrent neural network model.

[0036] The beneficial effects of this invention are as follows: The artificial intelligence-based flue gas monitoring and early warning system provided by this invention deploys a multi-source sensor network in a closed or semi-closed space, organizes multi-sensor time-series data in a sliding window manner, and introduces a neural network based on a multi-head self-attention mechanism to adaptively model the correlation between different sensor channels and different time steps. This fuses the originally independent sensor signals into a feature vector sequence reflecting the overall risk state, fundamentally overcoming the problem of weak, asynchronous, or even superficially contradictory early signals caused by inconsistent response sensitivity, response lag, and noise levels among different channels. Compared to traditional multi-indicator joint triggering methods that rely on fixed thresholds and empirical logic, this invention predicts the risk state for future periods within a fused feature space and constructs a comprehensive risk value in the form of a single scalar using weighted norms and statistical aggregation. This allows the system to simultaneously monitor the magnitude, duration, and cross-channel linkage of anomalies, thereby suppressing occasional false alarms while achieving an earlier and more stable response to global risks with developmental trends. The system's output comprehensive risk value, compared with preset classification thresholds, can directly drive multi-level early warning and automated control strategies. This forms a layered defense against potential lithium battery thermal runaway and hazardous gas leaks in high-risk scenarios such as energy storage cabinets. In the case of minor anomalies, it guides maintenance checks or load reduction; in the case of severe anomalies, it triggers shutdown and ventilation measures, significantly improving the sensitivity, robustness, and interpretability of the monitoring and early warning system under complex operating conditions. This helps reduce the probability of accidents and mitigate their consequences. Furthermore, the system's highly modular architecture facilitates integration and deployment within existing monitoring platforms and allows for subsequent online upgrades of model parameters. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0038] Figure 1 This is a schematic diagram of the framework of the AI-based flue gas monitoring and early warning system in the embodiment. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0040] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0041] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0042] This application proposes an artificial intelligence-based flue gas monitoring and early warning system, combined with... Figure 1 As shown, the system includes:

[0043] A multi-source sensor network is used to be deployed in the monitoring area. The multi-source sensor network includes at least two types of flue gas parameter sensors for collecting corresponding flue gas parameter monitoring data.

[0044] The data acquisition module is used to receive and store timestamped monitoring data from a multi-source sensor network.

[0045] The data processing and analysis engine, connected to the data acquisition module, is used for fusion modeling and risk prediction of monitoring data. The data processing and analysis engine includes:

[0046] The time-series data matrix construction unit is used to construct a multivariate time-series data matrix from monitoring data from a multi-source sensor network in chronological order.

[0047] The attention-weighted fusion unit, connected to the time-series data matrix construction unit, is used to fuse multivariate time-series data matrices. The attention-weighted fusion unit is built based on a neural network. It dynamically calculates the weights that reflect the correlation between monitoring data from different sensors through an attention mechanism, and uses the weights to perform weighted fusion of the multivariate time-series data matrix, outputting a fused feature vector sequence.

[0048] The risk state prediction unit is connected to the attention-weighted fusion unit, which is used to receive the fused feature vector sequence and perform time series prediction to obtain a risk state sequence covering a preset future time range.

[0049] The comprehensive risk calculation unit is connected to the risk state prediction unit and is used to process the risk state sequence to calculate the comprehensive risk value in scalar form.

[0050] The early warning judgment module is connected to the comprehensive risk calculation unit. It has a preset risk threshold and is used to compare the comprehensive risk value with the risk threshold and generate an early warning signal based on the comparison result.

[0051] In this embodiment, the flue gas parameter sensors in the multi-source sensor network can be evenly deployed in typical high-risk locations inside the energy storage cabinet or pipe gallery. The installation locations and heights are determined by on-site engineers based on equipment layout, gas specific gravity, and ventilation conditions. The sensors are connected to the data acquisition module via analog or digital bus. The embedded processor in the data acquisition module adds a timestamp based on a unified clock to each sampled data point and caches it locally or uploads it to the host computer database according to the sampling order. Specifically, the system's sampling period can be set to a fixed value between 1 and 10 seconds, preferably around 3 seconds, based on the risk development speed of the monitored object, to achieve a balance between monitoring sensitivity and communication load. The sampling period is set by the commissioning personnel through the parameter configuration interface and remains unchanged during operation. The initial risk threshold value in the early warning judgment module can be provided by safety specifications, equipment manuals, or expert experience, for example, using the safety exposure limit of the critical gas or the alarm concentration recommended by the equipment manufacturer as a reference. The threshold is then fine-tuned through simulation or on-site trial operation to control the false alarm rate and missed alarm rate. For example, when the system is in actual operation, the calculation and comparison of the comprehensive risk value can be repeated at fixed time intervals. By default, the judgment result is updated every time a new frame of time-series data arrives. If multiple consecutive judgment results are higher than the threshold of the same warning level, the warning signal of that level will be maintained for at least a preset minimum duration to avoid frequent warning reversals due to momentary jitter. Optionally, when the data acquisition module detects partial sensor communication interruption or abnormal data quality marking within a short period of time, the system can maintain the most recent reliable comprehensive risk value and simultaneously provide a data anomaly prompt on the interface. If necessary, the warning level can be raised by one level to ensure a safety margin.

[0052] In one embodiment, the data processing and analysis engine further includes a sliding window management unit.

[0053] The sliding window management unit is used to maintain a dynamic data window with a fixed time length and update step size.

[0054] The time series data matrix construction unit is used to extract monitoring data from the dynamic data window to construct a multivariate time series data matrix;

[0055] Specifically, during engineering implementation, the sliding window management unit can select the window coverage time length based on the sampling period and the target early warning lead time. For example, for rapidly evolving lithium battery thermal runaway scenarios, the window length can be set to the monitoring data within the most recent 5 to 20 minutes. Each time step within the window corresponds to a sampling result with a fixed sampling period. The ratio of the window length to the sampling period is the time step parameter of the matrix, and its size can be set in the configuration file before deployment. For instance, when the sampling period is two seconds and the window length is 10 minutes, the dynamic data window corresponds to 300 time steps in memory, and the time dimension of the time series data matrix is ​​300. After a new data point is sampled, the window management unit discards the earliest data point and appends the new data to the end, thus forming a queue arranged in ascending order of time. Furthermore, during the initial window filling phase, when the system has just started and historical data is insufficient to fill the window, the matrix can be constructed by copying the first valid data point or shortening the window according to the existing data length. At the same time, a flag indicating that the window is not fully filled is marked in the internal state so that the risk state prediction unit can use conservative weights or output lower limit constraints in the early operation phase. Optionally, the update step size of the dynamic data window can be set to be the same as the sampling period, or it can be set to an integer multiple of the sampling period, such as sliding forward once only after accumulating ten new data points, so that the computational load of the system matches the acceptable latency requirements on site.

[0056] In one embodiment, the attention-weighted fusion unit includes a linear projection layer, a multi-head self-attention layer, and a feedforward neural network layer connected in sequence.

[0057] Among them, the multi-head self-attention layer is used to calculate attention weights based on query vectors, key vectors and value vectors and to perform weighted fusion of multivariate time series data matrices;

[0058] The attention-weighted fusion steps include:

[0059] Step a: Denote the multivariate time series data stacked in ascending time order within the sliding window as a matrix. ; for the first Each attention head performs linear projection:

[0060] ,

[0061] in, Indicates the first The query matrix has dimensions of [number]. , Indicates the first The key matrix of size, with dimension [missing information]. , Indicates the first The value matrix of each head, with dimension [missing information]. , Represents a time series data matrix with dimensions of , Indicates the first The query projection matrix of the head, with dimensions of , Indicates the first The key projection matrix of the head, with dimensions of... , Indicates the first The projection matrix of the values ​​of each head, with dimensions of . , Indicates the number of time steps. Indicates the number of sensor channels. Indicates the embedding dimension of the key / query. The embedding dimension of the value. Indicates the attention head index, which is a positive integer;

[0062] Step b, calculate the unnormalized score and perform row-wise normalization to obtain the weights, for the first... The size is achieved by scaling dot product attention and overlaying a mask:

[0063] ,

[0064] in, Indicates the first The unnormalized attention score matrix for the head, with dimensions of [missing information]. , This represents the transpose operator. Represents a mask matrix with dimension . Used for causal constraints and missing test masking; minimum value is taken at unavailable locations. Indicates the first The attention weight matrix for each head, with dimensions [missing information]. , grouped by row Indicates the first The temperature coefficient of a unit is a positive scalar.

[0065] Its element-level normalization is written as:

[0066] ,

[0067] in, express The Line number Column elements, express The Line number Column elements, Indicates the time row index. Indicates a time column index. This variable represents the summation over column indices, and its value range is... Consistent;

[0068] Step c, based on right Perform time-series weighting and then linearly map the results after concatenation among multiple heads:

[0069] ,

[0070] in, Indicates the first The weighted fusion result of the size, with dimensions as follows: , This represents the result matrix concatenated according to the feature dimension, with dimension 1. , This represents the concatenation operator based on the feature dimension. This represents the number of attention heads, and is a positive integer. This represents a sequence of fused feature vectors with dimension . , This represents the output mapping matrix, with dimension 1. , This represents the output bias vector, with dimension . , Indicates the output feature dimension;

[0071] It employs numerically stable normalization, which allows for computational... First, subtract the maximum value within each row to suppress overflow. Generated from causal relationships and valid data labels, the upper triangle is shielded at a position that takes a minimum constant. When the value is large, the local attention bandwidth can be configured to limit the computation domain while keeping the module interface unchanged. During training, the local attention bandwidth can be adjusted accordingly. Set lower and upper limits to avoid an overly sharp or overly flat weight distribution;

[0072] Specifically, the above implementation method, without changing the system architecture, maps the time-series data of multi-source sensors to a learnable query, key, and value space. It obtains a probabilistic weight matrix through scaling dot products and temperature adjustment, thereby achieving adaptive weighted fusion across time and channels. Linear projection provides differentiated representation capabilities for different sensors, scaling terms avoid numerical amplification caused by high-dimensional inner products, and row normalization allocates the dependencies of each moment to historical moments, facilitating direct consumption by subsequent prediction models. Masking simultaneously handles causal constraints and missing data masking on the engineering side, reducing the risk of anomaly propagation. A multi-head mechanism captures different correlation patterns in parallel, and through splicing and linear mapping, obtains fixed-dimensional fusion features, facilitating interface integration with the risk state prediction unit. Combined with numerically stable normalization and temperature range constraints, it maintains computational stability and controllable weight distribution even under large time windows and multi-sensor scales, providing higher robustness and resolution for subsequent aggregation and hierarchical early warning of comprehensive risks.

[0073] In this embodiment, the number of sensor channels in the multivariate time-series data matrix is ​​consistent with the number of flue gas parameter sensors actually installed on-site, typically between three and sixteen channels. The number of time steps is determined by the window length and sampling period, usually ranging from tens to hundreds of steps. The number of attention heads in the multi-head self-attention layer can be set to 2-8, with the specific value determined by a comprehensive trade-off between model complexity and computational resources during offline verification. Similarly, the embedding dimensions of queries and keys, as well as the embedding dimensions of values, can be selected between 16 and 128. A unified embedding dimension and output feature dimension can be used in the same project to simplify configuration and deployment. During the model training phase, the weight matrix and bias vector of the linear projection layer are optimized by minimizing the loss function that minimizes the error between the predicted risk state and historical observations. Training data can come from long-term monitoring records in actual operation, laboratory simulated fault tests, or simulation-generated data. Normal and abnormal operating conditions are clearly marked in the dataset to facilitate the model learning the correlation between sensors under different modes. In this embodiment, the mask matrix is ​​automatically generated by the data validity markers in the sliding window during engineering implementation. For upper triangular positions that exceed the causal order and for data positions determined to be missing or obviously abnormal, a very small constant is uniformly assigned, causing the corresponding weights to approach zero after normalization, thereby avoiding interference from missing samples on the attention weights. Optionally, a maximum time step limit can be set during the model inference stage. When the number of time steps within the window exceeds this limit, only the most recent data is retained for computation, ensuring that real-time requirements are met even on edge computing devices.

[0074] In one embodiment, the risk state sequence is a sequence of predicted concentration vectors for multiple consecutive future time points.

[0075] Each predicted concentration vector includes predicted concentration values ​​corresponding to at least two types of flue gas parameter sensors;

[0076] For example, during offline training, the risk status prediction unit uses a fused feature vector sequence as input and the actual sensor concentration or a filtered approximate actual concentration from multiple future steps as the supervision signal. It employs a regression loss function to minimize the difference between the predicted vector and the target vector, thereby learning a time-series prediction model suitable for the target monitoring object. Specifically, the number of prediction steps can be set to several consecutive moments ranging from tens of seconds to several minutes, based on the expected advance warning time. For instance, it can predict the concentration change trajectory within the next 1-3 minutes every sampling period, ensuring the output risk status sequence has temporal continuity. Furthermore, for scenarios involving multiple sensor channels such as hydrogen, carbon monoxide, and smoke, each component in the predicted concentration vector corresponds to the predicted value of different types of sensors. Before inputting the model, the historical data of each channel can be normalized to zero mean and unit variance or normalized by range. After the prediction output, it is then inversely normalized to restore the physical quantity unit for direct comparison with safety thresholds and display on the monitoring interface. Optionally, when a certain sensor channel has many missing data in the training data, stronger regularization can be applied to that channel or its weight in the loss function can be appropriately reduced to reduce the dependence of the prediction model on the information of that channel, thereby improving the usability of the overall model under imperfect data conditions in the field.

[0077] In one embodiment, the comprehensive risk calculation unit is specifically used for:

[0078] The weighted norm is calculated for each predicted concentration vector in the predicted concentration vector sequence to obtain the scalar risk instantaneous value sequence.

[0079] Then, statistical aggregation operations are performed on the scalar risk instantaneous value sequence to obtain the comprehensive risk value;

[0080] In the comprehensive risk calculation unit, the weighted norm is implemented and statistical aggregation is performed as follows:

[0081] Step d, regarding the future The predicted concentration vector at each time step First, subtract from the reference baseline and normalize according to the range to obtain the element-level dimensionless deviation:

[0082] ,

[0083] in, Indicating the future At the [time]th moment Dimensionless deviation of each sensor channel This indicates the corresponding predicted concentration. This indicates the reference baseline (such as a safety limit or background value) for that channel. This indicates the scale factor (such as range or resolution) of the channel. Indicates the number of prediction steps. Indicates the number of sensor channels. For the index of the predicted time, For channel indexing;

[0084] Step e: Calculate the weighted average based on the dimensionless bias. Norm, to obtain the instantaneous value of risk:

[0085] ,

[0086] in, Indicates the first Based on the prediction time The instantaneous value of scalar risk obtained from the norm Indicates the first The non-negative weight of the channel satisfies , To denote the order of the norm, take ;

[0087] Step f: When statistical correlation exists among the channels, the quadratic norm is used.

[0088] ,

[0089] in, Indicates the first The instantaneous risk value at each prediction time is obtained based on the quadratic form. This represents the predicted concentration vector at that moment (dimension). ), Represents the reference baseline vector (dimension) ), its first Components , The covariance matrix representing the prediction error (dimensions) ), is a symmetric positive definite matrix. Indicates the transpose operator;

[0090] Step g, to enhance the response to exceeding limits, defines a piecewise penalty function and performs weighted aggregation:

[0091]

[0092]

[0093] in, Indicates the first The instantaneous risk value at each prediction time is obtained based on the threshold sensitivity norm. This represents the piecewise penalty function. Indicates the first The risk threshold of the channel This represents the scaling factor above the threshold, used to control the buffer bandwidth;

[0094] Step h, example of statistical aggregation operation (from instantaneous to comprehensive):

[0095] any instant sequence (Recommended) , or The values ​​are aggregated into a comprehensive risk value;

[0096] (h1) Index-weighted aggregation (with greater emphasis on the near future):

[0097] ,

[0098] in, Indicates the index-weighted composite risk. Represents the smoothing coefficient, taken as... , Indicates the first The instantaneous value of risk at any given moment;

[0099] (h2) Quantile tail mean (CVaR style, focusing on extreme risks):

[0100] ,

[0101] in, This indicates the overall risk of the mean at the quantile tails. This represents the set of indices that are not less than a quantile. Represents the cardinality of a set. Indicates level as quantile function, take ;

[0102] Weight It can learn based on safety importance or historical false alarm losses. It can be estimated from the validation set residuals and updated periodically. and It can be calibrated based on regulatory thresholds and acceptable overrun durations, when selected Norm It is beneficial for smoothing. Sparser, aggregation methods can be customized according to the scenario and Switching between time zones;

[0103] Specifically, the above implementation method proposes three alternative weighting norms for the comprehensive risk calculation unit: weighting based on channel importance. Norms are suitable for scenarios where channels are independent and deviations are mainly normalized. Quadratic norms based on covariance can capture cross-channel correlations and are more sensitive to multi-gas linkages. Threshold-sensitive robust norms increase the response to the out-of-limit interval through piecewise penalties, while maintaining zero contribution in the non-out-of-limit interval, reducing false alarms. All the above instantaneous measures normalize or center the prediction vector and the reference baseline to reduce the deviation caused by the difference in dimensions, and reflect the safety priority through weights. At the aggregation level from instantaneous to comprehensive, the exponential weighting method emphasizes the near future, which is more in line with the timeliness of early warning control, while the quantile tail mean highlights extreme cases and is suitable for risk aversion strategies for severe accidents. Both types of aggregation maintain scalar outputs, which are consistent with the threshold comparison interface of graded early warning, making it easy to directly replace or deploy in parallel in the existing architecture, forming a more flexible risk measurement and strategy linkage path.

[0104] In this embodiment, each component in the reference baseline vector can be set according to the safety limits of different sensors, the allowable fluctuation range of the process, or the long-term historical average. The scaling factor can be selected from the range, standard deviation, or reference amplitude given in regulations for the corresponding channel. The channel weight vector is manually set or adjusted through offline optimization, taking into account factors such as the importance of the sensor in safety protection, false alarm costs, and maintenance difficulty. Specifically, the norm order and aggregation method can be combined and tested before deployment through historical data playback and scenario simulation to observe the lead time and sensitivity of the comprehensive risk curve relative to existing alarm records under different parameter configurations, thereby determining the recommended parameter combination. For example, the smoothing coefficient in exponentially weighted aggregation can be selected in the range of 5%-30%. A larger smoothing coefficient is selected when it is necessary to highlight the risk in the most recent short period of time, and a smaller smoothing coefficient is selected when it is necessary to take into account the risk that has accumulated slowly over a longer period of time. The quantile level in the quantile tail mean can be set between 0.8 and 0.95 to focus on the high-value portion of the instantaneous risk value distribution. Similarly, when multiple instantaneous risk measures are running in parallel, the results of different measures can be weighted and summed or the maximum value can be taken for secondary aggregation within the comprehensive risk calculation unit. This ensures that both gradual risks and sudden extreme value risks can be identified. Optionally, to prevent the comprehensive risk value from getting out of control in the event of extreme anomalies or obvious data errors, upper and lower limit truncation can be added to the output to constrain the comprehensive risk value between zero and the upper limit. When there are obvious errors in the input data, the system still remains within an interpretable numerical range, and a prompt message indicating that the risk value has been truncated can be displayed on the monitoring interface.

[0105] In one embodiment, the early warning judgment module is configured to output a graded early warning signal based on the comparison result between the comprehensive risk value and the risk threshold.

[0106] The graded early warning signal includes at least two different risk levels, corresponding to different alarm display methods and / or control response strategies;

[0107] In one embodiment, it also includes an execution control module.

[0108] The execution control module is connected to the early warning judgment module. After receiving an early warning signal, it generates control commands according to preset control rules and sends them to the controlled process equipment to perform safety protection actions such as load reduction, shutdown, or ventilation.

[0109] Optionally, during engineering implementation, the execution control module can map different levels of early warning signals output by the early warning judgment module into different combinations of control commands through digital output ports, relay contacts, or communication interfaces with existing control systems. For example, under the lowest level of early warning, only local audible and visual alarms and monitoring interface highlighting are triggered; under the medium level of early warning, charging and discharging power is limited or some non-critical loads are shut down; under the highest level of early warning, emergency shutdown and forced ventilation commands are issued, and corresponding event logs and timestamps are recorded for post-event analysis. Furthermore, to avoid malfunctions due to communication interruptions or control link failures, the execution control module can be configured with heartbeat detection and fault-tolerant policies. When no new early warning signal is received for a preset time, the current control state is maintained by default, or the system switches to a predefined safe operating condition and reports the abnormal state of the control link. For example, in energy storage cabinet applications, if the comprehensive risk value remains within the medium level of early warning for a certain period without manual intervention, the execution control module can gradually reduce the charging and discharging power according to a preset load reduction curve to slow down the risk development process and allow maintenance personnel time for on-site inspection.

[0110] In one embodiment, the monitoring area is an energy storage cabinet.

[0111] The multi-source sensor network includes a hydrogen sensor and a carbon monoxide sensor for detecting characteristic gases of thermal runaway in lithium batteries.

[0112] In one embodiment, the multi-source sensor network further includes at least one of a smoke sensor, a temperature sensor, and / or a volatile organic compound sensor;

[0113] In one embodiment, the time series prediction model in the risk state prediction unit is at least one of a multivariate autoregressive integral moving average model, a temporal convolutional network model, or a recurrent neural network model.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0115] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A flue gas monitoring and early warning system based on artificial intelligence, characterized in that, include: A multi-source sensor network is used to be deployed in the monitoring area. The multi-source sensor network includes at least two types of flue gas parameter sensors for collecting corresponding flue gas parameter monitoring data. The data acquisition module is used to receive and store timestamped monitoring data from the multi-source sensor network. A data processing and analysis engine, connected to the data acquisition module, is used to perform fusion modeling and risk prediction on the monitoring data. The data processing and analysis engine includes: The time-series data matrix construction unit is used to construct a multivariate time-series data matrix from the monitoring data from the multi-source sensor network in chronological order; An attention-weighted fusion unit, connected to the time-series data matrix construction unit, is used to perform fusion processing on the multivariate time-series data matrix. The attention-weighted fusion unit is built based on a neural network, dynamically calculates the weights reflecting the correlation between monitoring data from different sensors through an attention mechanism, and uses the weights to perform weighted fusion on the multivariate time-series data matrix, outputting a fused feature vector sequence. A risk state prediction unit, connected to the attention-weighted fusion unit, is used to receive the fused feature vector sequence and perform time series prediction to obtain a risk state sequence covering a preset future time range. A comprehensive risk calculation unit, connected to the risk state prediction unit, is used to process the risk state sequence and calculate a comprehensive risk value in scalar form. The early warning judgment module is connected to the comprehensive risk calculation unit and has a preset risk threshold. It is used to compare the comprehensive risk value with the risk threshold and generate an early warning signal based on the comparison result.

2. The artificial intelligence-based flue gas monitoring and early warning system as described in claim 1, characterized in that, The data processing and analysis engine also includes a sliding window management unit. The sliding window management unit is used to maintain a dynamic data window with a fixed time length and update step size. The time-series data matrix construction unit is used to extract monitoring data from the dynamic data window to construct the multivariate time-series data matrix.

3. A flue gas monitoring and early warning system based on artificial intelligence as described in claim 1 or 2, characterized in that, The attention-weighted fusion unit comprises a linear projection layer, a multi-head self-attention layer, and a feedforward neural network layer connected in sequence. The multi-head self-attention layer is used to calculate attention weights based on query vectors, key vectors, and value vectors, and to perform weighted fusion on the multivariate time-series data matrix.

4. A flue gas monitoring and early warning system based on artificial intelligence as described in any one of claims 1 to 3, characterized in that, The risk state sequence is a sequence of predicted concentration vectors for multiple consecutive future time points. Each predicted concentration vector includes the predicted concentration values ​​corresponding to the at least two types of flue gas parameter sensors.

5. The artificial intelligence-based flue gas monitoring and early warning system as described in claim 4, characterized in that, The comprehensive risk calculation unit is specifically used for: The weighted norm is calculated for each predicted concentration vector in the predicted concentration vector sequence to obtain a scalar risk instantaneous value sequence; Then, a statistical aggregation operation is performed on the scalar risk instantaneous value sequence to obtain the comprehensive risk value.

6. A flue gas monitoring and early warning system based on artificial intelligence as described in any one of claims 1 to 5, characterized in that, The early warning judgment module is configured to output a graded early warning signal based on the comparison result between the comprehensive risk value and the risk threshold. The graded early warning signal includes at least two different risk levels, corresponding to different alarm display methods and / or control response strategies.

7. A flue gas monitoring and early warning system based on artificial intelligence as described in any one of claims 1 to 6, characterized in that, It also includes an execution control module. The execution control module is connected to the early warning judgment module. After receiving the early warning signal, it generates control instructions according to preset control rules and sends them to the controlled process equipment to perform load reduction, shutdown or ventilation safety protection actions.

8. A flue gas monitoring and early warning system based on artificial intelligence as described in any one of claims 1 to 7, characterized in that, The monitoring area is an energy storage cabinet. The multi-source sensor network includes a hydrogen sensor and a carbon monoxide sensor for detecting characteristic gases of thermal runaway in lithium batteries.

9. The artificial intelligence-based flue gas monitoring and early warning system as described in claim 8, characterized in that, The multi-source sensor network also includes at least one of a smoke sensor, a temperature sensor, and / or a volatile organic compound sensor.

10. A flue gas monitoring and early warning system based on artificial intelligence as described in any one of claims 1 to 9, characterized in that, The time series prediction model in the risk status prediction unit is at least one of the following: a multivariate autoregressive integral moving average model, a time series convolutional network model, or a recurrent neural network model.

Images