Gas concentration continuous online detection method, system, device and medium

Abstract

The gas concentration continuous online detection method, system, device and medium relate to the gas detection technical field. The method comprises: continuously sampling the to-be-detected gas to obtain multi-source time sequence environmental data including a target gas concentration; inputting the multi-source time sequence environmental data into a preset calibration range generation model, the calibration range generation model outputs and optimizes a floating calibration range for gas concentration state judgment; judging the current target gas concentration value based on the floating calibration range, if the judgment result is abnormal, triggering the corresponding alarm or control response. By adopting the scheme, the continuous online detection of the gas concentration is realized, the floating calibration range dynamically adapted to the current monitoring environment is generated based on the multi-source time sequence environmental data, and the alarm or control response is triggered in time when the concentration is abnormal, so that the adaptability and safety of the detection are improved.

Description

Technical Field

[0001] This application relates to the field of gas detection technology, specifically to a method, system, device, and medium for continuous online detection of gas concentration. Background Technology

[0002] With the increasing demand for online oxygen content detection in various industries, the need for accurate oxygen content detection is crucial for automated optimization and control of production processes, while ensuring the safety of production equipment and personnel. Oxygen concentration analysis systems play a key role in these scenarios, and the reliability and practicality of related online detection technologies directly affect the safety and efficiency of industry production.

[0003] In the current process of online detection of gas oxygen content, the direct extraction method is generally used to carry out the detection work, that is, to directly extract the gas to be tested and measure the oxygen content and related parameters, in order to meet the basic needs of the industry for monitoring gas oxygen content.

[0004] However, during the implementation of the direct extraction method, the measurement results are easily affected by impurities such as the temperature, pressure, dust, and moisture content of the gas being measured. This can lead to an excessive amount of unclean gas entering the monitoring stage, affecting the accuracy and stability of the detection. Furthermore, there is currently a lack of effective detection and alarm solutions to address this issue, making it impossible to provide timely warnings and handle related anomalies, and thus failing to fully guarantee the reliability of online detection and the safety of subsequent production processes. Summary of the Invention

[0005] This application provides a method, system, device, and medium for continuous online detection of gas concentration, which solves the problems of inaccurate judgment, lack of effective adaptive calibration, and abnormal response mechanism in traditional gas concentration detection.

[0006] In a first aspect, this application provides a method for continuous online detection of gas concentration, the method comprising: Continuous sampling of the gas to be tested is performed to obtain multi-source time-series environmental data, including the concentration of the target gas. Multi-source time-series environmental data is input into a preset calibration range generation model, and the calibration range generation model outputs and optimizes the floating calibration range used for gas concentration state judgment. The current target gas concentration value is judged based on the floating calibration range. If the judgment result is abnormal, the corresponding alarm or control response is triggered.

[0007] By adopting the above technical solution, continuous online detection of gas concentration can be achieved. A floating calibration range that dynamically adapts to the current monitoring environment is generated based on multi-source time-series environmental data, and alarms or control responses are triggered in a timely manner when the concentration is abnormal, thereby improving the adaptability and safety of detection.

[0008] In one specific implementation scheme, before performing the step of inputting multi-source time-series environmental data into a predefined calibration range generation model, a data fusion and characterization step is also included: Spatiotemporal alignment and normalization are performed on multi-source temporal environmental data to construct a unified environmental situation tensor; By using a self-attention mechanism to perform feature weighting on the environmental situation tensor, core environmental factors that have potential driving force on gas concentration fluctuations are extracted. The core environmental factors and the historical sequence of target gas concentrations are jointly encoded to generate a joint feature vector for model input.

[0009] By adopting the above technical solution, multi-source time-series environmental data are fused and characterized, core environmental factors and joint feature vectors are effectively extracted, providing high-quality input data for the calibration range generation model and laying the foundation for accurately generating floating calibration ranges.

[0010] In a specific feasible implementation, the construction and training methods of the calibration range generation model include: A training sample set was constructed by collecting multi-source environmental data and corresponding gas concentration data during historical normal operation periods and abnormal event periods. A dual-channel deep network structure is constructed, in which the first channel network is used to learn the nonlinear mapping between environmental features and concentration safety boundaries, and the second channel network is used to learn the temporal dependence of normal concentration fluctuations under different environmental conditions. With the goal of minimizing the deviation between the predicted concentration safety boundary and the measured concentration, and combined with temporal consistency constraints, a dual-channel deep network is jointly trained to obtain a calibration range generation model.

[0011] By adopting the above technical solution, a calibration range generation model with a dual-channel deep network structure is constructed and trained. This model learns the nonlinear mapping between environmental characteristics and concentration safety boundaries, as well as the temporal dependence of normal concentration fluctuations, thus ensuring the reliability of the generated floating calibration range.

[0012] In one specific implementation, the calibration range generation model performs the following steps at runtime to output a floating calibration range: Receive the joint feature vector and calculate the baseline concentration safety range under the current environment through the first channel network; Meanwhile, by analyzing the inherent trends and periodic characteristics of recent concentration sequences through a second-channel network, the expected fluctuation range for the next moment can be predicted. By performing Bayesian fusion of the baseline concentration safety range and the expected fluctuation range, a probabilistic calibration range distribution that combines environmental adaptability and temporal continuity is generated. Extract the upper and lower bounds of the interval at the preset confidence level from the probabilistic calibration range distribution, and use them as the floating calibration range output at the current moment.

[0013] By adopting the above technical solution, the calibration range generation model can output a probabilistic calibration range distribution that combines environmental adaptability and temporal continuity, and extract a precise floating calibration range, thereby improving the accuracy of concentration state judgment.

[0014] In one specific implementation, the method further includes an online reliability assessment of the floating calibration range and an adaptive calibration step: Real-time monitoring of the dispersion of the probabilistic calibration range distribution and historical backtracking deviation; If the dispersion of the probabilistic calibration range exceeds the first threshold, or the historical backtracking deviation exceeds the second threshold, then the reliability of the current floating calibration range is determined to be reduced. When the reliability decreases, the model fine-tuning process is initiated. Based on the data collected in the latest time period, the local parameters of the model generated within the calibration range are learned online incrementally to calibrate the model output results.

[0015] By adopting the above technical solution, the reliability of the floating calibration range can be monitored in real time. When the reliability decreases, the model parameters can be calibrated in a timely manner through online incremental learning, so as to continuously ensure the accuracy and reliability of the floating calibration range.

[0016] In a specific feasible implementation, the mechanism for generating the floating calibration range also incorporates external knowledge guidance, specifically including: Establish a domain knowledge graph. The nodes of the knowledge graph include environmental parameters, equipment status, and characteristics of historical accident cases. The edges represent the causal or statistical relationships between the nodes. When generating the floating calibration range, subgraph structures similar to the current environment characteristics are retrieved from the domain knowledge graph, and the associated historical security threshold range data is extracted as prior knowledge. The prior knowledge is weighted and fused with the initial floating calibration range of the model output.

[0017] By adopting the above technical solution, prior knowledge provided by domain knowledge graphs is incorporated into the process of generating the floating calibration range, so that the final calibration range conforms to the data-driven rules and is constrained by domain experience, thereby improving the rationality and security of the calibration range.

[0018] In one specific feasible implementation, the method also includes a closed-loop optimization phase, comprising: Collect all relevant data for each alarm or control response event throughout its entire lifecycle, including the floating calibration range at the time of triggering, environmental data, concentration data, and post-event verification results; By constructing a feedback dataset using event data, the generalization error and safety redundancy of the current calibration range generation model in practical applications are calculated. Regularly use the feedback dataset to train the calibration range generation model using reinforcement learning, and optimize the model strategy.

[0019] By adopting the above technical solutions, the model reinforcement learning training is carried out using full-cycle data of alarm or control response events, continuously optimizing the model strategy and improving the model's generalization ability and detection performance in long-term practical applications.

[0020] A second aspect of this application provides a continuous online gas concentration detection system, the system comprising: The data acquisition module is used to continuously sample the gas to be tested and acquire multi-source time-series environmental data, including the concentration of the target gas. The data processing and fusion module is used to fuse and represent multi-source time-series environmental data and generate joint feature vectors; The calibration range generation module includes a preset calibration range generation model, which is a jointly trained dual-channel deep network. The judgment and response module is used to judge the current target gas concentration value based on the floating calibration range, and trigger the corresponding alarm or control response when the judgment is abnormal.

[0021] A third aspect of this application provides an electronic device, comprising: a processor and a memory; wherein the memory stores a computer program adapted to be loaded by the processor and to execute the above-described method steps.

[0022] A fourth aspect of this application provides a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the method steps described above. Attached Figure Description

[0023] Figure 1 This is a schematic flowchart of a method for continuous online detection of gas concentration provided in an embodiment of this application. Detailed Implementation

[0024] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0025] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0026] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0027] Please refer to Figure 1 A flowchart illustrating a method for continuous online gas concentration detection is presented. This method can be implemented using a computer program, a microcontroller, or run on a continuous online gas concentration detection system. The computer program can be integrated into a computer device or run as a standalone application. Specifically, the method includes steps S100 to S300, as follows: S100: Continuously sample the gas to be tested to acquire multi-source time-series environmental data, including the concentration of the target gas. The continuous online detection system in this embodiment can be adapted to gas detection scenarios in multiple industries such as metallurgy, power, building materials, and chemicals. For example, it can be used to detect the oxygen concentration during the material discharge process after drying in a rake dryer. The gas to be tested refers to a specific gas whose concentration needs to be detected. In this embodiment, oxygen is the main gas to be tested, but other gas types can also be adapted according to actual needs.

[0028] Multi-source time-series environmental data refers to time-series data collected synchronously during the sampling process, including target gas concentration and various related environmental parameters. These related environmental parameters include the temperature, pressure, dust content, and moisture content of the gas being measured, as well as the ambient temperature, humidity, and pressure at the sampling point—parameters relevant to the detection results. The target gas concentration refers to the concentration value of the gas to be measured, obtained directly from the detection equipment.

[0029] In some embodiments, the continuous online detection system performs continuous sampling of the gas to be tested using a preset sampling device. The sampling device initiates the sampling process according to a preset sampling frequency. During sampling, the gas sequentially passes through a sampling tube, a sintered filter element, and a stainless steel ball valve before entering the detection stage. Simultaneously, the detection equipment collects target gas concentration data and various related environmental parameter data. The related environmental parameter data are recorded sequentially according to the acquisition time, forming multi-source time-series environmental data. The continuous online detection system stores the related environmental parameter data in a preset storage module, providing data support for subsequent processing steps.

[0030] In some embodiments, when the target gas is oxygen, the concentration detection range can be 0 to 25%.

[0031] S200: Input multi-source time-series environmental data into the preset calibration range generation model, and output and optimize the floating calibration range used for gas concentration state judgment. In this embodiment of the application, the calibration range generation model is a pre-built and trained intelligent model that can output a dynamic concentration judgment range based on the input multi-source time-series environmental data, and is used to generate a calibration range that is adapted to the current scene by combining environmental features.

[0032] The floating calibration range is a concentration safety judgment interval output by the calibration range generation model, which can be dynamically adjusted based on real-time environmental data. Unlike the traditional fixed rated calibration range, the floating calibration range can adapt to changes in different monitoring environments. Gas concentration status judgment refers to the process of comparing the actual detected gas concentration with the floating calibration range to determine whether the gas concentration is in a safe state.

[0033] In some embodiments, the continuous online monitoring system extracts the multi-source time-series environmental data collected in step S100 from the storage module and inputs this data into a preset calibration range generation model via the data transmission module. The data transmission process employs an industry-standard communication protocol to ensure data transmission stability and real-time performance. After receiving the data, the calibration range generation model performs feature extraction and analysis on the input multi-source time-series environmental data based on the pre-learned correlation between environmental data and the safe range of gas concentration. After outputting the initial floating calibration range, the model continuously optimizes this range in real time based on the new data input subsequently. This optimization process is achieved by adjusting relevant parameters within the model, ensuring that the floating calibration range always maintains a high degree of adaptability to the current monitoring environment, providing an accurate basis for subsequent gas concentration status judgment.

[0034] Based on the above embodiments, as another optional embodiment, before performing the step of inputting multi-source time-series environmental data into a preset calibration range generation model, a data fusion and characterization step is also included: S201. Perform spatiotemporal alignment and normalization on multi-source temporal environmental data to construct a unified environmental situation tensor. In this embodiment of the application, the environmental situation tensor is a tensor structure formed by arranging multi-source temporal environmental data, which has undergone spatiotemporal alignment and normalization, according to a preset dimension.

[0035] In some embodiments, the continuous online detection system performs spatiotemporal alignment processing on the multi-source time-series environmental data acquired in step S100. In the time dimension, the system extracts the acquisition timestamps of each data item, correlates and matches different types of data acquired at the same time, and uses linear interpolation to complete the alignment for data with slight discrepancies in acquisition time, ensuring that all data are arranged in a unified time series.

[0036] In the spatial dimension, data from different sampling points or different detection devices are calibrated according to a preset spatial coordinate mapping relationship to ensure that the data can accurately reflect the environmental state under the same monitoring scenario. After spatiotemporal alignment is completed, the system performs normalization processing on the data. This embodiment adopts the minimum-maximum normalization method, which maps each data to the same numerical range through preset calculation logic to eliminate the influence of differences in dimensions.

[0037] Finally, the system constructs a unified environmental situation tensor from the processed multi-source time-series environmental data according to preset dimensions such as data type and time series. The dimensions of the environmental situation tensor are determined by the number of data types and the length of the time series.

[0038] The calculation logic for min-max normalization can include: X'=(XX min ) / (X max -X min ); Where X represents the original data before normalization, and X' represents the processed data after normalization. min X is the minimum value of this type of data in the historical data collection. max This represents the maximum value of this type of data in the historical data collection.

[0039] In some embodiments, the normalized numerical range can be [0,1]; the environmental situation tensor can be 3-dimensional, corresponding to the data type dimension, time series dimension and feature attribute dimension, respectively.

[0040] S202. By using a self-attention mechanism to perform feature weighting on the environmental situation tensor, core environmental factors that have potential driving force on gas concentration fluctuations are extracted. In this embodiment of the application, the core environmental factor refers to the key environmental parameters that have a significant potential impact on gas concentration fluctuations after feature weighting processing, which can directly or indirectly drive changes in gas concentration.

[0041] In some embodiments, the continuous online detection system invokes a preset self-attention mechanism algorithm to perform feature weighting processing on the environmental situation tensor constructed in step S201. The self-attention mechanism algorithm first calculates the correlation similarity between each feature in the environmental situation tensor and all other features. The correlation similarity calculation is based on the inner product operation of feature vectors. Based on the calculated correlation similarity, the algorithm assigns a corresponding attention weight to each feature; features with higher correlation similarity receive larger weight values, and vice versa.

[0042] Subsequently, the system weights the features in the environmental situation tensor according to the assigned attention weights, enhancing the representational power of high-weight features and weakening the influence of low-weight features. After weighting, the system extracts core environmental factors through preset screening rules. The screening rules are based on feature weight values ​​and set weight thresholds. Environmental parameters with weight values ​​exceeding the threshold are identified as core environmental factors. Core environmental factors typically include parameters that significantly affect gas concentration fluctuations, such as the temperature, moisture content, and dust content of the measured gas.

[0043] In some embodiments, the attention weights are calculated using the scaled dot product attention algorithm; the weight threshold can be set to 0.6; the number of core environmental factors can be 5, namely, the temperature of the gas being measured, the water content of the gas being measured, the dust content of the gas being measured, the ambient temperature at the sampling point, and the pressure at the sampling point.

[0044] S203. Jointly encode the historical sequences of core environmental factors and target gas concentrations to generate a joint feature vector for model input.

[0045] In this embodiment, the joint feature vector is a high-dimensional vector generated after joint encoding, which integrates core environmental information and historical concentration change information, and can provide more comprehensive input data for the calibration range generation model.

[0046] In some embodiments, the continuous online detection system extracts historical sequences of target gas concentrations from the storage module. The duration of the historical sequence is set according to the concentration change rate of the actual detection scenario to ensure that the sequence can cover representative historical change patterns.

[0047] Subsequently, the system performs joint encoding processing on the core environmental factors extracted in step S202 and the historical sequence. The encoding process is implemented using a pre-defined encoder. The encoder first converts the core environmental factors into fixed-dimensional feature vectors, and simultaneously extracts temporal features from the historical sequence of the target gas concentration to obtain the corresponding temporal feature vectors. Then, the encoder integrates the two types of feature vectors according to a pre-defined concatenation rule. This concatenation rule ensures the effective fusion of core environmental information and temporal feature information, ultimately generating a high-dimensional joint feature vector.

[0048] In some embodiments, the historical sequence of the target gas concentration can be 60 seconds long, containing 60 consecutive concentration data points; the encoder can be a Transformer encoder; the feature vector after the core environmental factor transformation can be 128-dimensional, the temporal feature vector of the historical sequence of the target gas concentration can be 256-dimensional, and the joint feature vector can be 384-dimensional.

[0049] Based on the above embodiments, as another optional embodiment, the method for constructing and training the calibration range generation model includes: S204. Collect multi-source environmental data and corresponding gas concentration data during historical normal operation periods and abnormal event periods to form a training sample set; In this embodiment of the application, the abnormal event period refers to the period during which abnormal situations such as excessive gas concentration or equipment failure occur during system operation.

[0050] In some embodiments, the system filters relevant data from historical databases for periods of normal operation and periods of abnormal events. The filtering process first determines the time ranges for the two types of periods based on historical operation records, and then extracts multi-source environmental data and corresponding gas concentration data within these time ranges. Subsequently, the extracted data undergoes preprocessing operations, including data cleaning and data standardization. Data cleaning removes missing values, outliers, and other invalid data, while data standardization transforms data of different magnitudes to the same range. After preprocessing, the data from the two types of periods are divided according to a preset ratio to form a training sample set. The training sample set contains input data and corresponding label data. The input data is multi-source environmental data, and the label data is the corresponding gas concentration safety range label, used for supervised learning during model training.

[0051] S205. Construct a dual-channel deep network structure, wherein the first channel network is used to learn the nonlinear mapping between environmental features and concentration safety boundaries, and the second channel network is used to learn the temporal dependence of normal concentration fluctuations under different environmental conditions. The dual-channel deep network structure is the core architecture of the calibration range generation model in this embodiment. It consists of two independent and collaborative deep network channels, each undertaking different feature learning tasks.

[0052] The first channel network is a network channel in a dual-channel deep network structure that focuses on learning the correlation between environmental features and concentration safety boundaries, and is used to uncover the complex relationship between environmental parameters and concentration safety boundaries. Nonlinear mapping refers to a mapping relationship between environmental features and concentration safety boundaries that does not have a simple linear correlation and needs to be learned through a complex network model.

[0053] The second-channel network focuses on learning the temporal fluctuation patterns of concentration, used to capture the temporal variation characteristics of gas concentration under different environmental conditions. Environmental condition refers to the detection scenario state composed of multiple environmental parameters; different combinations of environmental parameters correspond to different environmental states. The temporal dependence of normal concentration fluctuations refers to the inherent correlation between gas concentration and time under normal conditions, such as periodic fluctuations and trend changes.

[0054] In some embodiments, the system constructs a dual-channel deep network structure. The first channel network adopts a Convolutional Neural Network (CNN) architecture, including multiple network layers such as an input layer, convolutional layers, pooling layers, and fully connected layers. The input layer receives multi-source environmental data, the convolutional layers extract spatial features from the environmental data through preset convolutional kernels, the pooling layers perform dimensionality reduction on the features output by the convolutional layers, retaining key features, and the fully connected layers map the dimensionality-reduced features to the output space of the concentration safety boundary, realizing the learning of the nonlinear mapping between environmental features and the concentration safety boundary.

[0055] The second channel network employs a Long Short-Term Memory (LSTM) architecture, comprising an input layer, an LSTM layer, and an output layer. The input layer receives time-series data of the target gas concentration. The LSTM layer uses a gating mechanism to capture long-term dependencies in the concentration data, learning the time-series dependence patterns of normal concentration fluctuations under different environmental conditions. The output layer outputs the corresponding time-series feature representations. The network architecture parameters for both channels are initially set based on the size of the training sample set and the feature complexity, and are subsequently adjusted and optimized during the training process.

[0056] S206. With the goal of minimizing the deviation between the predicted concentration safety boundary and the measured concentration, and combined with temporal consistency constraints, a dual-channel deep network is jointly trained to obtain a calibration range generation model.

[0057] In some embodiments, the system first constructs an objective function for model training. The objective function focuses on minimizing the deviation between the predicted concentration safety boundary and the measured concentration, while also introducing a temporal consistency constraint term. The core deviation term is calculated using mean squared error, and the temporal consistency constraint term is constructed by calculating the difference in the model output calibration range between adjacent time points. The objective function is a weighted sum of the core deviation term and the temporal consistency constraint term.

[0058] Subsequently, the training sample set is input into the dual-channel deep network structure, and joint training is performed according to the preset training batches and number of iterations. During training, the gradient of the objective function with respect to each network parameter is calculated using the backpropagation algorithm. Based on the gradient information, the network parameters are adjusted using stochastic gradient descent to reduce the value of the objective function. After each iteration batch is completed, the prediction error of the model is calculated using the validation set. If the prediction error reaches a preset convergence threshold, or the number of iterations reaches a preset upper limit, training is stopped. At this point, the dual-channel deep network structure is the trained calibration range generation model.

[0059] The expression for the objective function can include: L = α × L1 + β × L2; Where L is the total objective function value, L1 is the loss function for the deviation between the predicted concentration safety boundary and the measured concentration, L2 is the time series consistency constraint loss function, and α and β are weighting coefficients used to adjust the importance of the two loss terms in the total objective function.

[0060] In some embodiments, α can be 0.7 and β can be 0.3.

[0061] Based on the above embodiments, as another optional embodiment, the calibration range generation model performs the following steps at runtime to output the floating calibration range: S207. Receive the joint feature vector and calculate the baseline concentration safety range under the current environment through the first channel network; In some embodiments, the first channel network receives the joint feature vector generated in step S203. The joint feature vector is first input to the input layer of the first channel network, and after normalization by the input layer, it is transmitted to the convolutional layer. The convolutional layer uses a preset convolutional kernel to perform depth extraction of environmental features from the joint feature vector, capturing the complex correlation features between environmental parameters. The extracted features are then dimensionality-reduced by a pooling layer and transmitted to a fully connected layer. The fully connected layer maps the dimensionality-reduced features to the output space of the concentration safety boundary, and calculates the upper and lower limits of the concentration safety boundary in the current environment through an activation function. The interval formed by these upper and lower limits is the baseline concentration safety interval. During the calculation process, the first channel network calls the network parameters optimized during training to ensure the accuracy of the baseline concentration safety interval calculation.

[0062] In some embodiments, the activation function of the first channel network can be the ReLU function; for the oxygen concentration detection scenario, the calculated baseline concentration safety range can be 1% to 20%; the joint feature vector is normalized before being input into the network using the Batch Norm method.

[0063] S208. At the same time, the inherent trend and periodic characteristics of the recent concentration sequence are analyzed through the second channel network to predict the expected fluctuation range at the next moment. In some embodiments, the second-channel network operates synchronously with the first-channel network, receiving the joint feature vector generated in step S203 and extracting the recent concentration sequence from it. The recent concentration sequence is input to the input layer of the second channel and then transmitted to the LSTM layer. The LSTM layer performs temporal feature mining on the recent concentration sequence through the synergistic action of the forget gate, input gate, and output gate. The forget gate filters and retains key historical information in the recent concentration sequence, the input gate receives new concentration data and updates the cell state, and the output gate outputs the temporal features of the current moment based on the cell state.

[0064] By analyzing time-series characteristics, the second-channel network identifies the inherent trends in recent concentration sequences, such as determining the trend direction by calculating the sequence slope; it also detects the periodic characteristics of the sequence, such as calculating the period duration using the autocorrelation function. Based on the identified inherent trends and periodic characteristics, the second-channel network calculates the expected fluctuation range of the gas concentration at the next time step using the prediction function of the output layer.

[0065] S209. Perform Bayesian fusion between the baseline concentration safety range and the expected fluctuation range to generate a probabilistic calibration range distribution that combines environmental adaptability and temporal continuity. In this embodiment, the probabilistic calibration range distribution is a concentration safety range distribution with probabilistic attributes formed after fusion. It can quantify the probability that different concentration values ​​belong to the safety range, reflecting the uncertainty of the concentration safety range. Environmental adaptability refers to the ability of the fusion result to adapt to the characteristics of the current monitoring environment, and temporal continuity refers to the ability of the fusion result to maintain consistency in changes over time and avoid abrupt changes.

[0066] In some embodiments, the system invokes a preset Bayesian fusion algorithm to fuse the baseline concentration safety range obtained in step S207 and the expected fluctuation range obtained in step S208. First, the algorithm determines the confidence weights of the two ranges. The confidence weight of the baseline concentration safety range is set based on the training accuracy of the first channel network, and the confidence weight of the expected fluctuation range is set based on the training accuracy of the second channel network. The sum of the two weight values ​​is 1.

[0067] Subsequently, based on Bayes' theorem, the baseline concentration safety range is used as the prior distribution, and the expected fluctuation range is used as the likelihood function to calculate the posterior distribution. The posterior distribution is the fused probabilistic calibration range distribution, which combines the environmental adaptability of the baseline concentration safety range with the temporal continuity of the expected fluctuation range, and can more comprehensively reflect the safe range of gas concentration at the current moment.

[0068] In some embodiments, the confidence weight of the baseline concentration safety range can be 0.6, and the confidence weight of the expected fluctuation range can be 0.4; the prior distribution in the Bayesian fusion algorithm adopts the assumption of uniform distribution; the likelihood function adopts a normal distribution model, with its mean being the center value of the expected fluctuation range and its variance being the square of the half-width of the expected fluctuation range.

[0069] S210. Extract the upper and lower bounds of the interval at the preset confidence level from the probabilistic calibration range distribution, and use them as the floating calibration range output at the current moment.

[0070] In this embodiment of the application, the upper limit of the interval refers to the maximum concentration value corresponding to the probabilistic calibration range distribution under a preset confidence level; the lower limit of the interval refers to the minimum concentration value corresponding to the probabilistic calibration range distribution under a preset confidence level; the floating calibration range is the specific concentration range formed by the upper and lower limits of the interval.

[0071] In some embodiments, the system presets a fixed confidence level based on the safety requirements of the actual detection scenario. Based on the confidence level, the system performs interval extraction on the probabilistic calibration range distribution generated in step S209. The extraction process is achieved by calculating the cumulative distribution function of the probabilistic calibration range distribution, finding the concentration value corresponding to the cumulative probability reaching the lower limit of the confidence level as the lower bound of the interval, and finding the concentration value corresponding to the cumulative probability reaching the upper limit of the confidence level as the upper bound of the interval.

[0072] For example, when the confidence level is set to 95%, the concentration value corresponding to a cumulative probability of 2.5% is the lower bound of the interval, and the concentration value corresponding to a cumulative probability of 97.5% is the upper bound of the interval. The extracted upper and lower bounds constitute the floating calibration range at the current moment. The system transmits this floating calibration range to the subsequent judgment module and stores it in the historical database for subsequent model optimization and data traceability.

[0073] In some embodiments, the preset confidence level can be 95%; the cumulative distribution function is calculated using a numerical integration method; for oxygen concentration detection scenarios, the extracted floating calibration range can be 0.8% to 20.3%; the output format of the floating calibration range is a numerical range format, retaining two decimal places.

[0074] S300: Based on the floating calibration range, the current target gas concentration value is judged. If the judgment result is abnormal, the corresponding alarm or control response is triggered.

[0075] In some embodiments, the judgment module of the continuous online detection system receives the floating calibration range output in step S210 and extracts the latest collected current target gas concentration value from step S100. The judgment module compares the current target gas concentration value with the upper and lower boundaries of the floating calibration range.

[0076] If the current target gas concentration is greater than the upper limit of the interval or less than the lower limit, the judgment module outputs an abnormal judgment result. Upon receiving the abnormal judgment result, the system simultaneously triggers alarm and control responses. The alarm response is implemented through a pre-set audible and visual alarm device, which emits a preset decibel sound and flashes lights at a specific frequency to alert on-site personnel. The control response includes direct intervention to shut down the equipment; the system sends a shutdown signal to the associated production equipment to stop its operation and prevent safety accidents caused by abnormal concentrations. After the personnel have completed the handling, they can manually restore equipment operation.

[0077] Based on the above embodiments, as another optional embodiment, the method further includes an online reliability assessment and adaptive calibration step for the floating calibration range: S301. Real-time monitoring of the dispersion of the probabilistic calibration range distribution and historical backtracking deviation; In this embodiment of the application, the historical backtesting bias is the degree of difference between the floating calibration range output by the index-defined range generation model and the actual safe concentration range under similar conditions in the same historical period, which can reflect the accuracy of the model prediction.

[0078] In some embodiments, the continuous online monitoring system initiates a real-time monitoring process to continuously monitor the probabilistic calibration range distribution generated in step S209. The system quantifies the dispersion of the probabilistic calibration range distribution by calculating its variance. The variance calculation is based on the probability density function of the distribution; a larger variance value indicates a higher degree of dispersion. Simultaneously, the system extracts the actual safe concentration range under similar historical conditions from the historical database. Historical timeframe refers to the same period as the current time, and similar environment refers to a situation where the similarity between historical and current environmental parameters exceeds a preset similarity threshold. The historical backtracking deviation is obtained by averaging the absolute values ​​of the boundary differences between the current floating calibration range and the historical actual safe concentration range. The system performs the above calculations at a preset monitoring frequency to continuously track changes in dispersion and historical backtracking deviation.

[0079] In some embodiments, the similarity threshold between historical environmental parameters and current environmental parameters can be 0.8.

[0080] S302. If the dispersion of the probabilistic calibration range distribution exceeds the first threshold, or the historical backtracking deviation exceeds the second threshold, then the reliability of the current floating calibration range is determined to have decreased. The first threshold in this embodiment is a pre-set critical value used to determine whether the dispersion exceeds a reasonable range, determined based on historical monitoring data and actual safety requirements; the second threshold is a pre-set critical value used to determine whether the historical backtracking deviation exceeds a reasonable range, also set based on historical data and safety requirements; the decrease in credibility means that the probability that the current floating calibration range can accurately reflect the actual safe concentration range is reduced, and the judgment result triggers the subsequent model calibration process.

[0081] In some embodiments, the continuous online detection system compares the dispersion calculated in step S301 with a preset first threshold, and simultaneously compares the historical backtracking deviation with a preset second threshold. If the dispersion value is greater than the first threshold, it indicates that the probabilistic calibration range is too dispersed and the stability of the floating calibration range is insufficient; if the historical backtracking deviation value is greater than the second threshold, it indicates that the floating calibration range output by the model deviates significantly from the historical actual situation, and the prediction accuracy is insufficient. If either of these two conditions is met, the system determines that the reliability of the current floating calibration range has decreased, generates a corresponding judgment signal, and transmits it to the model calibration module, providing a basis for initiating the model fine-tuning process.

[0082] In some embodiments, the first threshold can be 0.04; the second threshold can be 0.5%; the first threshold and the second threshold are set using statistical methods, and are determined by calculating the 95th percentile based on monitoring data from the past 3 months.

[0083] S303. When the credibility decreases, the model fine-tuning process is initiated. Based on the data collected in the latest time period, the local parameters of the model generated within the calibration range are learned online incrementally to calibrate the model output results.

[0084] In some embodiments, after receiving the confidence decline determination signal in step S302, the system automatically initiates the model fine-tuning process. First, the system collects data from the latest time period, the duration of which is set according to the rate of environmental change to ensure the data reflects the current environmental characteristics. Then, the collected latest data is preprocessed, including spatiotemporal alignment and normalization in step S201, feature weighting and core environmental factor extraction in step S202, and joint encoding in step S203, generating a joint feature vector for fine-tuning. Afterward, the system initiates online incremental learning, inputting the joint feature vector from the fine-tuning into the calibration range to generate the model, adjusting only some local parameters of the fully connected layers and LSTM layers. The adjustment process uses a small-batch iterative update method, with a preset batch size and the number of iterations set according to the confidence decline level. The parameter gradient is calculated using the backpropagation algorithm to adjust the parameter values. After fine-tuning, the model outputs the calibrated floating calibration range, restoring its confidence.

[0085] Based on the above embodiments, as another optional embodiment, the floating calibration range generation mechanism also incorporates external knowledge guidance, specifically including: S304. Establish a domain knowledge graph. The nodes of the knowledge graph include environmental parameters, equipment status, and characteristics of historical accident cases. The edges represent the causal or statistical relationships between the nodes. In this embodiment of the application, the domain knowledge graph is a structured knowledge carrier built based on professional knowledge in the field of gas concentration detection. It is used to store entities in the field and the relationships between entities. Nodes are the basic units in the domain knowledge graph, representing various entities in the field. Environmental parameter nodes cover all environmental parameters related to detection. Equipment status nodes include the operating status of sampling equipment, analysis equipment, etc. Historical accident case feature nodes include environmental characteristics and concentration anomalies at the time of the accident. Edges are the links connecting nodes and are used to represent the relationships between nodes.

[0086] In some embodiments, the system constructs a domain knowledge graph by collecting professional data, historical operational data, and accident records in the field of gas concentration detection. First, the types of nodes in the knowledge graph are determined, including environmental parameter nodes, equipment status nodes, and historical accident case feature nodes. Environmental parameter nodes specifically include the temperature, pressure, dust content, and moisture content of the gas being measured; equipment status nodes specifically include normal sampling equipment, faulty sampling equipment, normal analysis equipment, and faulty analysis equipment; historical accident case feature nodes specifically include abnormal concentrations caused by high temperatures and detection deviations caused by equipment malfunctions. Then, based on professional knowledge and historical data, the relationships between the nodes are determined. For example, there is a causal relationship between excessively high gas temperature and a rapid increase in oxygen concentration, and a statistical relationship between sampling equipment malfunction and abnormal detection concentration. The information of nodes and edges is stored in a graph database to construct a complete domain knowledge graph, providing support for subsequent external knowledge guidance.

[0087] S305. When generating the floating calibration range, retrieve subgraph structures similar to the current environment features from the domain knowledge graph and extract the associated historical security threshold range data as prior knowledge. In this embodiment of the application, the subgraph structure is a local graph structure in the domain knowledge graph that corresponds to the current environmental features, including nodes related to the current environmental features and associated nodes.

[0088] In some embodiments, during the generation of the probabilistic calibration range distribution in step S209, the system simultaneously initiates a domain knowledge graph retrieval process. First, the current environmental features are converted into node combinations in the knowledge graph, forming the current subgraph structure. The current subgraph structure includes various environmental parameter nodes and device status nodes in the current environment. Then, a retrieval algorithm based on similarity calculation is used to retrieve historical subgraph structures similar to the current subgraph structure from the domain knowledge graph. Similarity calculation is based on features such as node type, node attributes, and edge type of the subgraph. A similarity threshold is set, and historical subgraph structures with similarity exceeding the threshold are selected. Corresponding historical safety threshold range data is extracted from these historical subgraph structures; this data constitutes prior knowledge. The system transmits this prior knowledge to the fusion module for subsequent optimization of the floating calibration range.

[0089] In some embodiments, the similarity threshold can be 0.8; the retrieval algorithm based on similarity calculation can use a graph embedding algorithm to convert the subgraph into a vector and then calculate the cosine similarity; the extraction of historical security threshold range data is implemented using a structured query language; the format of prior knowledge is consistent with the floating calibration range, which is an interval form.

[0090] S306. Weighted fusion of prior knowledge and the initial floating calibration range of the model output.

[0091] In some embodiments, the system first determines the fusion weights of prior knowledge and the initial floating calibration range. The weight of the prior knowledge is determined based on the similarity between the corresponding historical subgraph structure and the current subgraph structure, as well as the number of historical scenes; the higher the similarity and the more historical scenes, the greater the weight. The weight of the initial floating calibration range is determined based on the real-time prediction accuracy of the calibration range generation model; the higher the prediction accuracy, the greater the weight. The sum of the two weight values ​​is 1.

[0092] Subsequently, the system employs a weighted average method to fuse prior knowledge and the initial floating calibration range. For the upper bound of the interval, a weighted sum of the upper bound of the prior knowledge and the upper bound of the initial floating calibration range is calculated; for the lower bound of the interval, a weighted sum of the lower bound of the prior knowledge and the lower bound of the initial floating calibration range is calculated. The weighted sum calculation results constitute the final floating calibration range, which integrates model calculations and domain knowledge, resulting in higher rationality and security.

[0093] In some embodiments, the fusion weight of prior knowledge can be 0.3, and the fusion weight of the initial floating calibration range can be 0.7.

[0094] Based on the above embodiments, as another optional embodiment, the method further includes a closed-loop optimization phase: S307. Collect all relevant data for each alarm or control response event throughout its entire lifecycle, including the floating calibration range at the time of triggering, environmental data, concentration data, and post-event verification results. In some embodiments, after each alarm or control response event occurs, the system automatically initiates a full-cycle data collection process. The system first records the floating calibration range at the moment the event is triggered, which is extracted and stored from the real-time output data. Subsequently, it extracts multi-source time-series environmental data and target gas concentration data from 30 minutes before the event trigger to 60 minutes after the trigger, a time period that fully covers the environmental and concentration changes before and after the event. After the event is handled, staff enter the post-event verification results through the system's preset input interface, including the actual gas concentration value obtained through manual detection, the identified cause of the event, the measures taken, and the effectiveness verification of the measures. The system stores these data in association by event number, constructing a complete full-cycle event dataset.

[0095] S308. Construct a feedback dataset using event data, and calculate the generalization error and safety redundancy of the current calibration range generation model in practical applications. In some embodiments, the system integrates and processes the full-cycle relevant data of multiple events collected in step S307. First, data filtering is performed to remove duplicate, invalid, and outlier data, retaining only representative event data. Then, data standardization is performed to convert data from different events to a unified format and numerical range, ensuring data consistency. After processing, training and validation sets are divided according to a preset ratio to construct a feedback dataset. Based on the feedback dataset, the system calculates the generalization error of the model generated within the current calibration range, calculated as the average deviation between the model's output floating calibration range and the actual safe concentration range in the feedback dataset. Simultaneously, the safety redundancy is calculated as the average difference between the interval width of the floating calibration range and the interval width of the actual safe concentration range. These two metrics comprehensively evaluate the model's performance in practical applications.

[0096] S309. Regularly use the feedback dataset to perform reinforcement learning training on the calibration range generation model and optimize the model strategy.

[0097] In some embodiments, the system initiates the reinforcement learning training process according to a preset training cycle. First, the feedback dataset constructed in step S308 is input into the calibration range generation model, which outputs a predicted floating calibration range value based on the current policy. Then, the system sets the reward function for reinforcement learning, which comprehensively considers two indicators: generalization error and safety redundancy. When the generalization error of the model's output floating calibration range is less than a preset error threshold and the safety redundancy is within a preset reasonable range, the model is given a positive reward; when the generalization error is greater than the error threshold or the safety redundancy exceeds the reasonable range, the model is given a negative reward. Based on the calculation results of the reward function, the internal parameters and decision-making strategy of the model are adjusted through the reinforcement learning algorithm, gradually optimizing the model towards a direction with smaller generalization error and reasonable safety redundancy. After training is completed, the online calibration range generation model is updated to use the optimized policy to output the floating calibration range.

[0098] The expression for the reward function can include: R = γ1 × R1 + γ2 × R2; Where R is the total reward value, R1 is the reward item based on generalization error, R2 is the reward item based on safety redundancy, and γ1 and γ2 are reward weight coefficients used to adjust the importance of the two reward items.

[0099] In some embodiments, the training period can be 30 days; the preset threshold for generalization error can be 0.3%; the preset reasonable range for safety redundancy can be 0.3% to 0.8%; and the reward weight coefficients γ1 and γ2 can be 0.6 and 0.4, respectively.

[0100] Based on the above embodiments, as another optional embodiment, this application also provides a continuous online gas concentration detection system, the system comprising: The data acquisition module is used to continuously sample the gas to be tested and acquire multi-source time-series environmental data, including the target gas concentration. The data processing and fusion module is used to fuse and characterize the multi-source time-series environmental data to generate a joint feature vector. The calibration range generation module includes a preset calibration range generation model, which is a jointly trained dual-channel deep network. The judgment and response module is used to judge the current target gas concentration value based on the floating calibration range and trigger the corresponding alarm or control response when the judgment is abnormal.

[0101] In some embodiments, the system may include the Fumi-GW online multi-component detection alarm or the SGS-808 chemical process gas analysis system.

[0102] In some embodiments, the calibration range generation module is configured to: receive a joint feature vector; calculate the baseline concentration safety range under the current environment through a first-channel network; analyze the inherent trend and periodic characteristics of recent concentration sequences through a second-channel network to predict the expected fluctuation range at the next moment; perform Bayesian fusion of the baseline concentration safety range and the expected fluctuation range to generate a probabilistic calibration range distribution; and extract the upper and lower bounds of the interval at a preset confidence level from the probabilistic calibration range distribution as the floating calibration range output.

[0103] In some embodiments, the data processing and fusion module includes: a spatiotemporal alignment unit for performing timestamp alignment and interpolation processing on multi-source time-series environmental data; a normalization unit for normalizing each environmental parameter to a unified dimension range; a tensor construction unit for constructing an environmental situation tensor from the normalized multi-parameter time-series data; an attention weighting unit for performing feature weighting on the environmental situation tensor through a self-attention mechanism to extract core environmental factors; and a joint encoding unit for jointly encoding the core environmental factors and the historical sequence of the target gas concentration to generate a joint feature vector.

[0104] In some embodiments, the calibration range generation module further includes: a model storage unit for storing the network parameters and structure definition of the calibration range generation model; an online inference unit for loading the model and performing real-time inference calculations; and a Bayesian fusion unit for performing Bayesian fusion calculations of the baseline concentration safety range and the expected fluctuation range to generate a probabilistic calibration range distribution.

[0105] In some embodiments, the system further includes: a credibility assessment and calibration module, used to monitor in real time the dispersion and historical backtracking deviation of the probabilistic calibration range distribution, and to initiate an online incremental learning process for the calibration range generation model to calibrate the model output when the dispersion exceeds a first threshold or the backtracking deviation exceeds a second threshold.

[0106] In some embodiments, the system further includes: a knowledge graph module for storing a domain knowledge graph, wherein the nodes of the knowledge graph include environmental parameters, device status, and historical accident case features, and the edges represent causal or statistical correlations between nodes; and a knowledge retrieval and fusion unit for retrieving subgraph structures similar to the current environmental features from the knowledge graph when generating a floating calibration range, extracting the associated historical safety threshold range as prior knowledge, and performing weighted fusion of the prior knowledge with the initial floating calibration range output by the model.

[0107] In some embodiments, the system further includes a feedback optimization module, which collects full-cycle data for each alarm or control response event, constructs a feedback dataset, and performs periodic policy optimization on the calibration range generation model based on the dataset through reinforcement learning.

[0108] In some embodiments, the judgment and response module includes: a threshold comparison unit, used to compare the current target gas concentration value with the upper and lower boundaries of the floating calibration range; and a response strategy selection unit, used to select at least one response method among triggering an audible and visual alarm, an equipment shutdown protection command, or uploading data to a remote monitoring center, based on the anomaly level, environmental risk category, or equipment operating status.

[0109] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0110] Based on the above embodiments, as another optional embodiment, the present application embodiment may further include a computer storage medium, which may store multiple instructions adapted for loading by a processor and executing a method of the above embodiments. For the specific execution process, please refer to the detailed description of the above embodiments, which will not be repeated here.

[0111] Based on the above embodiments, as another optional embodiment, this application embodiment may further include an electronic device. The electronic device may include: at least one processor, at least one communication bus, a user interface, at least one network interface, and a memory.

[0112] The communication bus is used to enable communication between these components.

[0113] The user interface may include a display screen and a camera. Optional user interfaces may also include standard wired interfaces and wireless interfaces.

[0114] The network interface may include standard wired interfaces and wireless interfaces (such as Wi-Fi interfaces).

[0115] The processor may include one or more processing cores. It connects to various parts of the server via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in memory, and accessing data stored in memory to perform various server functions and process data. Optionally, the processor may be implemented using at least one of the following hardware forms: Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor may integrate one or more of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor.

[0116] The memory may include random access memory (RAM) or read-only memory. Optionally, the memory may include a non-transitory computer-readable storage medium. The memory can be used to store instructions, programs, code, code sets, or instruction sets. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. As a computer storage medium, the memory may include an operating system, a network communication module, a user interface module, and an application program of one method.

[0117] In electronic devices, the user interface is primarily used to provide an input interface for users and to acquire user input data; while the processor can be used to call an application program stored in memory that represents a method. When executed by one or more processors, this causes the electronic device to perform one or more methods as described in the above embodiments. It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0118] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0119] In the various embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0122] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0123] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Other embodiments of this disclosure will readily conceive of those skilled in the art upon consideration of the specification and the disclosure of practical truths.

[0124] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for continuous online detection of gas concentration, characterized in that, The method includes: Continuous sampling of the gas to be tested is performed to obtain multi-source time-series environmental data, including the concentration of the target gas. The multi-source time-series environmental data is input into a preset calibration range generation model, and the calibration range generation model outputs and optimizes a floating calibration range for gas concentration state judgment. The current target gas concentration value is judged based on the floating calibration range. If the judgment result is abnormal, the corresponding alarm or control response is triggered.

2. The method for continuous online detection of gas concentration according to claim 1, characterized in that, Before the step of inputting the multi-source time-series environmental data into the preset calibration range generation model, the method further includes a data fusion and representation step: The multi-source temporal environmental data is spatiotemporally aligned and normalized to construct a unified environmental situation tensor. By using a self-attention mechanism to perform feature weighting on the environmental situation tensor, core environmental factors that have potential driving force on gas concentration fluctuations are extracted. The core environmental factors and the historical sequence of target gas concentrations are jointly encoded to generate a joint feature vector for model input.

3. The method for continuous online detection of gas concentration according to claim 2, characterized in that, The method for constructing and training the calibration range generation model includes: A training sample set was constructed by collecting multi-source environmental data and corresponding gas concentration data during historical normal operation periods and abnormal event periods. A dual-channel deep network structure is constructed, in which the first channel network is used to learn the nonlinear mapping between environmental features and concentration safety boundaries, and the second channel network is used to learn the temporal dependence of normal concentration fluctuations under different environmental conditions. With the goal of minimizing the deviation between the predicted concentration safety boundary and the measured concentration, and incorporating temporal consistency constraints, the dual-channel deep network is jointly trained to obtain the calibration range generation model.

4. The method for continuous online detection of gas concentration according to claim 3, characterized in that, The calibration range generation model performs the following steps at runtime to output a floating calibration range: Receive the joint feature vector and calculate the baseline concentration safety range under the current environment through the first channel network; Meanwhile, by analyzing the inherent trends and periodic characteristics of recent concentration sequences through the second channel network, the expected fluctuation range at the next moment can be predicted. The baseline concentration safety range and the expected fluctuation range are fused using Bayesian methods to generate a probabilistic calibration range distribution that combines environmental adaptability and temporal continuity. Extract the upper and lower bounds of the interval at the preset confidence level from the probabilistic calibration range distribution, and use them as the floating calibration range output at the current moment.

5. The method for continuous online detection of gas concentration according to claim 4, characterized in that, The method also includes online reliability assessment and adaptive calibration steps for the floating calibration range: Real-time monitoring of the dispersion of the probabilistic calibration range distribution and historical backtracking deviation; If the dispersion of the probabilistic calibration range distribution exceeds the first threshold, or the historical backtracking deviation exceeds the second threshold, then the reliability of the current floating calibration range is determined to have decreased. When the reliability decreases, the model fine-tuning process is initiated. Based on the data collected in the latest time period, the local parameters of the model generated within the calibration range are learned online incrementally to calibrate the model output results.

6. The method for continuous online detection of gas concentration according to claim 5, characterized in that, The mechanism for generating the floating calibration range also incorporates external knowledge guidance, specifically including: Establish a domain knowledge graph, wherein the nodes of the knowledge graph include environmental parameters, equipment status, and historical accident case features, and the edges represent the causal or statistical correlation between the nodes; When generating the floating calibration range, subgraph structures similar to the current environment features are retrieved from the domain knowledge graph, and associated historical security threshold range data are extracted as prior knowledge. The prior knowledge is weighted and fused with the initial floating calibration range output by the model.

7. The method for continuous online detection of gas concentration according to claim 1, characterized in that, The method also includes a closed-loop optimization phase: Collect all relevant data for each alarm or control response event throughout its entire lifecycle, including the floating calibration range at the time of triggering, environmental data, concentration data, and post-event verification results; By constructing a feedback dataset using event data, the generalization error and safety redundancy of the current calibration range generation model in practical applications are calculated. The calibration range generation model is periodically trained using the feedback dataset to optimize the model strategy through reinforcement learning.

8. A continuous online gas concentration detection system, characterized in that, The system includes: The data acquisition module is used to continuously sample the gas to be tested and acquire multi-source time-series environmental data, including the concentration of the target gas. The data processing and fusion module is used to fuse and characterize the multi-source time-series environmental data to generate a joint feature vector; The calibration range generation module includes a preset calibration range generation model, wherein the calibration range generation model is a jointly trained dual-channel deep network; The judgment and response module is used to judge the current target gas concentration value based on the floating calibration range, and trigger the corresponding alarm or control response when the judgment is abnormal.

9. An electronic device, characterized in that, It includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted to be loaded by a processor and executed as described in any one of claims 1-7.

Images