Flammable and explosive gas anti-explosion safety early warning method and system based on artificial intelligence

Abstract

The invention discloses a flammable and explosive gas anti-explosion safety early warning method and system based on artificial intelligence. The method comprises the following steps: acquiring a data set acquired by a multi-modal sensor; performing dimension reduction processing on the data set by adopting a principal component analysis algorithm to obtain a target feature set; extracting an independent component set related to the gas concentration from the target feature set to generate an initial concentration feature value; classifying environmental parameter fluctuations by adopting a support vector machine algorithm according to the initial concentration characteristic value, and outputting a classification result; and judging whether the classification result exceeds a preset environment fluctuation threshold value or not, and if the classification result exceeds the preset environment fluctuation threshold value, adjusting the initial concentration characteristic value through a self-adaptive filter to generate a corrected concentration characteristic value. The method effectively eliminates the influence of environmental fluctuation on gas concentration measurement, improves the measurement precision and reliability, and is suitable for gas concentration monitoring scenes in various complex environments.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, and in particular discloses an artificial intelligence-based method and system for early warning of flammable and explosive gases. Background Technology

[0002] Flammable and explosive gas safety early warning technology plays a crucial role in high-precision manufacturing fields such as semiconductors and photovoltaics, as it directly relates to production safety, equipment stability, and product yield.

[0003] The risk of leakage or explosion of highly hazardous gases such as silanes could lead to catastrophic consequences, making efficient and reliable monitoring and control systems a core requirement for industry development. However, existing solutions have significant limitations in practical applications.

[0004] Traditional gas detection systems often rely on a single sensor, making them susceptible to environmental interference such as temperature and humidity, resulting in a high false alarm rate and difficulty in adapting to complex industrial scenarios. Fixed threshold warning mechanisms cannot be adjusted according to dynamic environments, often leading to missed alarms or frequent false alarms.

[0005] More importantly, existing systems struggle to balance detection accuracy with emergency response speed, and their integration of explosion-proof structure and inerting control is insufficient, making it difficult to meet the stringent requirements of high-purity gas environments.

[0006] These limitations reveal the core challenges in the field. The accuracy of gas detection is limited by the sensor's sensitivity to cross-interference, especially in ultra-high purity gas environments where identifying minute leaks becomes exceptionally difficult. Low-precision detection further leads to a lack of flexibility in setting warning thresholds; fixed thresholds cannot adapt to fluctuations in environmental parameters, thus increasing the risk of false alarms or missed alarms. Inappropriate threshold settings directly affect the timeliness and effectiveness of emergency response; for example, delayed inerting system activation or inaccurate flow regulation makes it difficult to quickly suppress the risk of combustion and explosion. Furthermore, the synergy between warning systems and explosion-proof structures is insufficient; traditional systems have significant shortcomings in structural compactness and functional integration, making efficient deployment difficult in demanding environments such as cleanrooms.

[0007] Therefore, the key issue of this research is how to overcome the technical challenges caused by cross-interference, environmental fluctuations, and response lag through high-precision multimodal sensing, dynamic threshold algorithms, and integrated explosion-proof inerting design, so as to achieve real-time monitoring and efficient control of hazardous gases such as silanes. Summary of the Invention

[0008] This invention provides an artificial intelligence-based method and system for early warning of flammable and explosive gases, aiming to solve at least one of the defects in the prior art.

[0009] One aspect of the present invention relates to an artificial intelligence-based method for early warning of fire and explosion prevention of flammable and explosive gases, comprising the following steps: Acquire a dataset collected by a multimodal sensor, including target gas concentration, temperature, humidity, and pressure data; Principal component analysis (PCA) is used to reduce the dimensionality of the dataset to obtain the target feature set. Extract a set of independent components related to gas concentration from the target feature set to generate initial concentration feature values; Based on the initial concentration feature values, the support vector machine algorithm is used to classify the fluctuations of environmental parameters and output the classification results; Determine whether the classification result exceeds the preset environmental fluctuation threshold. If the classification result exceeds the preset environmental fluctuation threshold, adjust the initial concentration feature value through an adaptive filter to generate a corrected concentration feature value.

[0010] Furthermore, the steps for acquiring the data set collected by the multimodal sensor include: Acquire raw data sets from multimodal sensors; The target gas concentration, temperature, humidity, and pressure data in the original dataset are calibrated using a sensor calibration algorithm. Noise is removed by mean filtering to obtain a preprocessed dataset.

[0011] Furthermore, the steps for using principal component analysis (PCA) to reduce the dimensionality of the dataset to obtain the target feature set include: From the preprocessed data set acquired by the multimodal sensor, the target gas concentration, temperature, humidity, and pressure data are normalized using a data standardization method to obtain a standardized data set. For a standardized dataset, principal component analysis is used to calculate the covariance matrix, obtain eigenvalues ​​and eigenvectors, determine the number of principal components, and obtain a set of dimensionality-reduced eigenvectors. Determine whether the feature dimension of the dimensionality reduction feature vector set is higher than a preset threshold. If the feature dimension of the dimensionality reduction feature vector set is higher than the preset threshold, then filter the feature vectors using a feature selection method to obtain a subset of key features and obtain a simplified feature set. Based on the simplified feature set, the feature data is reorganized using a data compression method to generate a compressed feature dataset, thus obtaining the target feature set.

[0012] Furthermore, the step of extracting a set of independent components related to gas concentration from the target feature set to generate initial concentration feature values ​​includes: The independent component analysis algorithm is used to calculate the statistical independence between the feature vectors in the target feature set, obtain the independent components related to the gas concentration, and obtain the initial independent component dataset. For the initial independent component dataset, the correlation coefficient between each independent component and the gas concentration is calculated using the feature selection method. It is then determined whether the correlation coefficient is higher than a preset threshold. If the correlation coefficient is higher than the preset threshold, the independent component is retained, and a filtered subset of independent components is obtained. Based on the selected subset of independent components, the independent components are linearly combined using a data recombination method to generate feature vectors directly related to gas concentration, thus obtaining an initial set of concentration feature values. By normalizing the initial concentration feature value set, calculating the standardized score of each feature value, obtaining the final feature value corresponding to the gas concentration, and determining the initial concentration feature value.

[0013] Furthermore, based on the initial concentration feature values, the support vector machine algorithm is used to classify environmental parameter fluctuations, and the steps for outputting the classification results include: From the initial set of concentration feature values, the feature vectors are standardized using data preprocessing methods. The mean and standard deviation of each feature vector are calculated, and the feature values ​​are mapped to a uniform scale through linear transformation to obtain a standardized feature dataset. Based on the standardized feature dataset, the sample set is divided into a training set and a test set. The ratio of the training set to the test set is determined by random sampling. Training samples and test samples are obtained for the support vector machine algorithm, resulting in the partitioned sample subset. For the training samples in the partitioned sample subset, the support vector machine algorithm is used to calculate the maximum margin hyperplane between feature vectors. The kernel function is used to perform nonlinear mapping on the feature vectors to determine the classification model parameters and obtain the trained classification model. The test samples are input into the trained classification model. The distance between the feature vector of the test sample and the hyperplane is calculated by the trained classification model. The category of environmental parameter fluctuation to which each test sample belongs is determined, and the classification result is output.

[0014] Further, the process of determining whether the classification result exceeds a preset environmental fluctuation threshold, and if so, adjusting the initial concentration feature value using an adaptive filter to generate a corrected concentration feature value, includes the following steps: The fluctuation category and classification accuracy are obtained from the classification results. By comparing the preset threshold setting with the fluctuation amplitude, it is determined whether the classification result exceeds the preset environmental fluctuation threshold. If it exceeds the threshold, the set of samples exceeding the threshold is obtained. For the set of samples exceeding the threshold, the initial concentration feature value and data noise are obtained. An adaptive filter is used to denoise the initial concentration feature value according to the filtering parameters to obtain the corrected concentration feature value. Based on the corrected concentration characteristic value, the characteristic distribution and sample quality are calculated, and the threshold setting is updated through a dynamic adjustment method to determine the new environmental fluctuation threshold. Determine whether the corrected concentration feature value meets the new environmental fluctuation threshold. If the corrected concentration feature value meets the new environmental fluctuation threshold, then classify the corrected concentration feature value using the support vector machine algorithm to determine the adjusted fluctuation category.

[0015] Another aspect of the present invention relates to an artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning system, used to implement the aforementioned artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning method. The artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning system includes: The acquisition module is used to acquire the data set collected by the multimodal sensor, which includes target gas concentration, temperature, humidity and pressure data; The processing module is used to perform dimensionality reduction processing on the dataset using the principal component analysis algorithm to obtain the target feature set; The extraction module is used to extract a set of independent components related to gas concentration from the target feature set to generate initial concentration feature values; The classification module is used to classify environmental parameter fluctuations based on initial concentration feature values ​​using a support vector machine algorithm, and output the classification results. The judgment module is used to determine whether the classification result exceeds the preset environmental fluctuation threshold. If the classification result exceeds the preset environmental fluctuation threshold, the initial concentration feature value is adjusted through an adaptive filter to generate a corrected concentration feature value.

[0016] Furthermore, the acquisition module includes: The first acquisition unit is used to acquire a raw data set from the multimodal sensor; The preprocessing unit is used to calibrate the target gas concentration, temperature, humidity, and pressure data in the original dataset using a sensor calibration algorithm, and remove noise through mean filtering to obtain a preprocessed dataset.

[0017] Furthermore, the processing module includes: The second acquisition unit is used to normalize the target gas concentration, temperature, humidity, and pressure data from the preprocessed data set acquired by the multimodal sensor using a data standardization method to obtain a standardized data set. The third acquisition unit is used to calculate the covariance matrix of the standardized dataset using the principal component analysis algorithm, obtain eigenvalues ​​and eigenvectors, determine the number of principal components, and obtain a set of dimensionality-reduced eigenvectors. The judgment unit is used to determine whether the feature dimension of the dimensionality reduction feature vector set is higher than a preset threshold. If the feature dimension of the dimensionality reduction feature vector set is higher than the preset threshold, the feature vectors are filtered by the feature selection method to obtain a subset of key features and obtain a simplified feature set. The fourth acquisition unit is used to reorganize the feature data according to the simplified feature set using a data compression method, generate a compressed feature dataset, and obtain the target feature set.

[0018] Furthermore, the extraction module includes: The fifth acquisition unit is used to calculate the statistical independence between the feature vectors in the target feature set by using the independent component analysis algorithm, to obtain the independent components related to the gas concentration, and to obtain the initial independent component dataset. The sixth acquisition unit is used to calculate the correlation coefficient between each independent component and the gas concentration using a feature selection method for the initial independent component dataset, and to determine whether the correlation coefficient is higher than a preset threshold. If the correlation coefficient is higher than the preset threshold, the independent component is retained, and a filtered subset of independent components is obtained. The seventh acquisition unit is used to linearly combine the independent components according to the filtered subset of independent components using a data recombination method to generate a feature vector directly related to the gas concentration, and obtain an initial set of concentration feature values. The determination unit is used to normalize the initial concentration feature value set, calculate the standardized score of each feature value, obtain the final feature value corresponding to the gas concentration, and determine the initial concentration feature value.

[0019] The beneficial effects achieved by this invention are as follows: This invention provides an artificial intelligence-based method and system for early warning of flammable and explosive gases. Addressing the issue that gas concentration, temperature, humidity, and pressure data collected by multimodal sensors are easily affected by environmental fluctuations, this invention employs principal component analysis (PCA) to reduce the dimensionality of the data, extracting independent components related to gas concentration to generate initial concentration feature values. Then, a support vector machine (SVM) algorithm is used to classify environmental parameter fluctuations, determining whether the classification results exceed a preset environmental fluctuation threshold. If so, an adaptive filter is used to adjust the initial concentration feature values, generating corrected concentration feature values. This invention, through the combined application of multiple algorithms, effectively eliminates the influence of environmental fluctuations on gas concentration measurement, improving measurement accuracy and reliability, and is suitable for gas concentration monitoring scenarios in various complex environments. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an embodiment of an artificial intelligence-based method for early warning of flammable and explosive gases according to the present invention. Detailed Implementation

[0021] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0022] like Figure 1As shown, the first embodiment of the present invention proposes a method for early warning of flammable and explosive gases based on artificial intelligence, including the following steps: Step S100: Obtain the data set collected by the multimodal sensor, which includes target gas concentration, temperature, humidity and pressure data.

[0023] A data collection is a collective term for a set of related data collected and organized according to specific rules or objectives. It is the basic unit for data analysis, model training, and scientific research. The core characteristics of a data collection lie in the inherent correlation between data and structured storage, rather than a simple random accumulation.

[0024] Step S200: Use principal component analysis algorithm to reduce the dimensionality of the dataset to obtain the target feature set.

[0025] Principal Component Analysis (PCA) is an unsupervised dimensionality reduction algorithm based on orthogonal transformation. It aims to project high-dimensional data into a low-dimensional space while maximizing the preservation of variance information and eliminating redundant correlations between features. Its core principle is to construct a set of uncorrelated principal components (PCs) by linearly combining the original variables, ranking them according to their variance contribution, thereby achieving data compression and feature extraction.

[0026] Dimensionality reduction is a technique that uses mathematical transformations to map high-dimensional data to a lower-dimensional space, aiming to reduce the number of features (dimensions) while preserving as much key information as possible. The core objective of dimensionality reduction is to address the "curse of dimensionality" caused by high-dimensional data, improving computational efficiency, visualization capabilities, and model generalization ability.

[0027] The target feature set is a whole consisting of distinct and clearly defined features that serve a specific analytical task or prediction goal. The core of the target feature set lies in accurately characterizing the attributes and relationships of the target object through structured screening and combination.

[0028] Step S300: Extract the set of independent components related to gas concentration from the target feature set to generate initial concentration feature values.

[0029] An independent component set (ICA) is a collection of statistically independent components extracted from a mixed signal using Independent Component Analysis (ICA). The core objective of ICA is to decompose observed high-dimensional mixed data into several independent latent source signals, each representing an independent feature or source, satisfying the principles of maximizing statistical independence and non-Gaussianity.

[0030] Initial concentration eigenvalues ​​are quantitative indicators used in signal separation or component analysis (such as independent component analysis, ICA) to describe the original intensity or weight of source signals in a mixed signal. These indicators are extracted using mathematical methods (such as matrix factorization and eigenvalue calculation) and reflect the contribution ratio and distribution characteristics of different source signals during the mixing process.

[0031] Step S400: Based on the initial concentration feature values, use the support vector machine algorithm to classify the fluctuations of environmental parameters and output the classification results.

[0032] Support Vector Machine (SVM) is a supervised learning algorithm primarily used for classification and regression analysis. Its core idea is to construct an optimal hyperplane in a high-dimensional space, such that data points of different classes are separated by maximizing the margin while minimizing the classification error.

[0033] Environmental parameter fluctuations refer to the regular or random changes in various monitored parameters (such as temperature, humidity, pollutant concentration, pH value, noise level, etc.) over time or space within an environmental system. These fluctuations reflect the dynamic interactions between internal and external factors within the environmental system and are a key analytical object for environmental quality assessment, pollution source apportionment, and ecological process research.

[0034] Step S500: Determine whether the classification result exceeds the preset environmental fluctuation threshold. If the classification result exceeds the preset environmental fluctuation threshold, adjust the initial concentration feature value through an adaptive filter to generate a corrected concentration feature value.

[0035] A preset environmental fluctuation threshold refers to a critical value for parameter fluctuations set in advance in environmental monitoring. It is used to determine whether the environmental state is within an acceptable range or to trigger an early warning mechanism. The preset environmental fluctuation threshold can be a single value or a range, representing the maximum allowable fluctuation limit of the environmental system. Exceeding this limit is considered abnormal or requires intervention.

[0036] An adaptive filter is a digital filter that can dynamically adjust its parameters according to the statistical characteristics of the input signal. The core objective of an adaptive filter is to minimize the error between the output signal and the desired signal by optimizing the filter coefficients in real time, thereby adapting to non-stationary (time-varying) signal environments or unknown system characteristics.

[0037] Corrected concentration eigenvalues ​​are baseline values ​​of substance concentration determined by eliminating measurement interference through parameter adjustment or signal processing techniques. They are used to quantify the true concentration level of a target component under specific conditions. The core of corrected concentration eigenvalues ​​lies in combining experimental data with model optimization to ensure the accuracy and reliability of the concentration eigenvalues.

[0038] Furthermore, the artificial intelligence-based early warning method for flammable and explosive gas fire prevention and explosion protection proposed in this embodiment includes step S100 as follows: Step S110: Obtain the raw data set from the multimodal sensor.

[0039] This paper focuses on the technical topic of acquiring and preprocessing raw data using multimodal sensors. The analysis revolves around a single gas detection scenario, emphasizing data acquisition, sensor calibration, and mean filtering for noise reduction. Examples highlight the technical implementation and effectiveness. For instance, a multimodal sensor is used in an industrial environment to detect the concentration of a target gas, such as silane, while simultaneously collecting environmental parameters.

[0040] The sensor array includes gas sensors, temperature sensors, humidity sensors, and pressure sensors. The raw dataset might contain silane concentration values ​​such as 5 ppm, temperature 25°C, humidity 60% RH, and pressure 101 kPa. These data may be biased due to sensor accuracy or environmental interference. For example, gas sensor readings may be higher due to temperature variations, requiring calibration to ensure accuracy. This multimodal data acquisition provides comprehensive environmental information for subsequent analysis, helping to improve detection reliability.

[0041] Step S120: The target gas concentration, temperature, humidity and pressure data in the original dataset are calibrated using a sensor calibration algorithm. Noise is removed by mean filtering to obtain a preprocessed dataset.

[0042] In one possible implementation, the sensor calibration algorithm corrects for the raw data. Volumetric sensors may drift over time, causing concentration readings to deviate from the true value. During calibration, a standard gas of known concentration, such as 4 ppm silane, can be used for comparison to adjust the sensor's output curve.

[0043] Temperature sensors may display lower readings due to thermocouple aging; for example, a reading of 25°C may be actual 26°C. This can be corrected using a calibration curve. Humidity and pressure data were also calibrated using a standard reference to ensure consistency across parameters. After calibration, data accuracy was significantly improved, laying the foundation for subsequent analysis.

[0044] Specifically, mean filtering is used to remove random noise from data. Raw data may experience abrupt changes due to electromagnetic interference or fluctuations in sensor sensitivity, such as a silane concentration jumping from 5.0 ppm to 5.1 ppm in a short period. Mean filtering smooths the data curve by calculating the average of multiple consecutive sampling points, such as 10. For example, if 10 concentration values ​​are collected (4.98, 5.0, 5.10, 4.99, 5.02, 4.97, 5.0, 4.99, 5.01, 4.98 ppm), mean filtering will yield a stable value of approximately 5 ppm. Temperature, humidity, and pressure data are also processed using this method to reduce noise interference and improve data smoothness and reliability.

[0045] Preferably, the parameters of the mean filter can be adjusted according to the scenario. In highly dynamic environments, such as workshops where gas concentrations change rapidly, the number of sampling points can be reduced to 5 to preserve data response speed; in stable environments, such as laboratories, the number of sampling points can be increased to 15 to further improve the smoothing effect. This flexibility ensures that the preprocessed dataset is both accurate and adaptable to different detection needs.

[0046] It's important to note that the combination of calibration and noise reduction significantly improves data quality. Unprocessed raw data can lead to misjudgments of gas concentrations, impacting industrial safety monitoring. Calibration ensures the data accurately reflects environmental conditions, while mean filtering removes interference, making the data more suitable for trend analysis or alarm systems. For example, preprocessed data can be used to determine whether silane concentration exceeds the safe threshold of 5 ppm, thus triggering ventilation equipment in a timely manner.

[0047] In one embodiment, the preprocessed dataset can be further used for training a machine learning model. The calibrated and denoised data exhibits higher consistency, allowing the model to more accurately predict gas concentration trends, such as whether the concentration will exceed the limit within the next hour. This application demonstrates the technical value of preprocessing and provides reliable support for industrial automation.

[0048] Understandably, the implementation of the aforementioned technologies forms a complete chain from data acquisition to preprocessing. Each stage supports the others, ensuring the transformation of data from its raw state to a usable, high-quality state, providing a stable foundation for gas detection, while simultaneously improving the system's response speed and accuracy.

[0049] Furthermore, the artificial intelligence-based early warning method for flammable and explosive gases proposed in this embodiment includes step S200 as follows: Step S210: From the preprocessed data set obtained from the multimodal sensor, the target gas concentration, temperature, humidity, and pressure data are normalized using a data standardization method to obtain a standardized data set.

[0050] The overall effectiveness of multimodal data standardization can be evaluated using the following formula: (1) In formula (1), A comprehensive standardized index representing multimodal sensor data. Indicates the total number of sensor modes. Indicates the first Data values ​​for each modality, The mean represents the first The average value of each modal data point Indicates the first The standard deviation of each modality data.

[0051] For example, in an industrial environment, the preprocessed data collected by multimodal sensors includes parameters such as silane concentration, temperature, humidity, and pressure. This data needs to be further processed to meet the needs of subsequent analysis.

[0052] Data standardization is a crucial step used to eliminate the influence of different dimensions. In principle, data standardization converts each parameter into a dimensionless value with a mean of 0 and a standard deviation of 1. For example, silane concentrations range from 4 to 6 ppm, temperatures from 20 to 30°C, humidity from 50 to 70% RH, and pressures from 100 to 102 kPa. After standardization, a concentration of 5 ppm might be converted to 0.5, a temperature of 25°C to 0, a humidity of 60% RH to 0.2, and a pressure of 101 kPa to 0.3. This process makes different parameters comparable, laying the foundation for subsequent analysis.

[0053] Step S220: For the standardized dataset, use the principal component analysis algorithm to calculate the covariance matrix, obtain eigenvalues ​​and eigenvectors, determine the number of principal components, and obtain a set of dimensionality-reduced eigenvectors.

[0054] The covariance matrix of the standardized dataset is calculated as follows: (2) In formula (2), Represents the covariance matrix. Indicates the number of samples. This represents the i-th sample vector. This represents the sample mean vector.

[0055] In one possible implementation, a standardized dataset is used for principal component analysis (PCA) to achieve dimensionality reduction. PCA extracts the main directions of variation in the data by calculating the covariance matrix. The covariance matrix reflects the correlation between parameters; for example, temperature and humidity may show a positive correlation.

[0056] After calculation, eigenvalues ​​and eigenvectors are obtained. The magnitude of the eigenvalues ​​determines the importance of the principal components. For example, with four eigenvalues ​​of 2.5, 1.2, 0.2, and 0.1, the top two principal components with a cumulative contribution rate of 90% are selected, retaining the correlation between concentration and temperature while ignoring minor changes in pressure and humidity. The dimensionality of the eigenvector set is reduced from 4 to 2, simplifying the data structure.

[0057] Step S230: Determine whether the feature dimension of the reduced feature vector set is higher than a preset threshold. If the feature dimension of the reduced feature vector set is higher than the preset threshold, then filter the feature vectors using the feature selection method to obtain a subset of key features and obtain a simplified feature set.

[0058] The dimensional threshold determination result is obtained using the following formula: (3) In formula (3), This indicates the result of the dimensional threshold judgment. This represents the current feature dimension of the reduced feature vector set. This represents the preset dimensional threshold. Greater than When the judgment result is 1, feature selection needs to be performed. Less than or equal to If the result is 0, feature selection is not required.

[0059] It should be noted that if the feature dimensions after dimensionality reduction are still higher than a preset threshold, such as setting the maximum dimension to 1, further feature selection is required. Feature selection methods can be based on information gain or correlation analysis to screen key features. For example, if analysis reveals that concentration and temperature have a significant impact on subsequent analysis, while humidity and pressure have lower information gain, then the feature vectors of concentration and temperature are retained to form a simplified feature set. This screening ensures that the data focuses on key information.

[0060] Step S240: Based on the simplified feature set, the feature data is reorganized using a data compression method to generate a compressed feature dataset, thus obtaining the target feature set.

[0061] Specifically, the simplified feature set is recombined into the target feature set through data compression methods. Data compression can employ linear transformations or encoding techniques to map feature vectors to a lower-dimensional space. For example, a two-dimensional feature vector can be compressed into a one-dimensional feature value, retaining the main information.

[0062] In one embodiment, the feature vectors of concentration and temperature are weighted and combined to generate a single feature value, such as concentration accounting for 70% and temperature accounting for 30%, thus generating a compressed feature dataset. This compression reduces storage requirements and facilitates real-time transmission.

[0063] Preferably, each step in the above processing chain supports the others. Standardization ensures data consistency, principal component analysis reduces dimensionality, feature selection focuses on key information, and compression optimizes the data structure. For example, raw data may contain redundant information; minor fluctuations in humidity and pressure have little impact on concentration analysis. Dimensionality reduction and compression eliminate redundancy, generating a compact set of target features. This set is suitable for subsequent analyses, such as concentration trend prediction or anomaly detection.

[0064] Understandably, the flexibility of the technology is reflected in parameter adjustments. For example, the cumulative contribution rate threshold of principal component analysis can be adjusted according to the scenario, set to 95% in a laboratory with high precision requirements and 85% in a workshop with rapid response.

[0065] Feature selection can also be adjusted according to the analytical objective; for example, when focusing on concentration prediction, concentration features can be prioritized. This flexibility ensures that the processing workflow can adapt to different needs.

[0066] In one embodiment, the target feature set can be directly used for industrial monitoring. For example, the compressed feature values ​​reflect the combined changes in concentration and temperature and can be used to determine whether an alarm should be triggered.

[0067] Compared to the raw data, the target feature set is more concise, computationally efficient, and suitable for real-time applications. The entire processing flow, from standardization to compression, forms a complete chain, ensuring efficient data utilization.

[0068] Furthermore, the artificial intelligence-based early warning method for flammable and explosive gases proposed in this embodiment includes step S300 as follows: Step S310: Using the independent component analysis algorithm, calculate the statistical independence between each feature vector in the target feature set, obtain the independent components related to the gas concentration, and obtain the initial independent component dataset.

[0069] For example, in industrial gas monitoring scenarios, the target feature set typically contains multi-dimensional features, such as a comprehensive representation of parameters like gas concentration and temperature. Independent component analysis (ICA) algorithms are used to separate statistically independent signals from these features, focusing on extracting independent components directly related to gas concentration. The principle is to decompose the mixed signal into uncorrelated components, assuming that each component signal source is independent. For instance, data collected by sensors may contain the combined effects of concentration and temperature; ICA separates the concentration-dominant signal by maximizing non-Gaussianity.

[0070] In one possible implementation, the target feature set contains a 4-dimensional feature vector, which, after analysis, yields four independent components. One component is highly correlated with concentration, such as a signal of silane concentration change, with a value range of 0.4-0.8, while the other components may reflect temperature or noise.

[0071] Step S320: For the initial independent component dataset, calculate the correlation coefficient between each independent component and the gas concentration using the feature selection method, and determine whether the correlation coefficient is higher than a preset threshold. If the correlation coefficient is higher than the preset threshold, retain the independent component to obtain the filtered subset of independent components.

[0072] The Pearson correlation coefficient between independent components and gas concentration is calculated using the following formula: (4) In formula (4), Indicates the first The correlation coefficients between each independent component and the gas concentration Indicates the first The independent component in the first The value of each sample point Indicates the first The mean of each independent component, Indicates the first Gas concentration values ​​at each sample point This represents the mean gas concentration. This represents the total number of samples.

[0073] The set of independent component indices retained after filtering is as follows: (5) In formula (5), This represents the set of independent component indices retained after filtering. Indicates the first The correlation coefficients between each independent component and the gas concentration This represents the preset correlation coefficient threshold. This represents the total number of initial independent components. Formula (5) defines the criterion for feature selection: only independent components with an absolute correlation coefficient greater than or equal to the threshold will be retained.

[0074] The final filtered subset of independent components takes the following form: (6) In formula (6), This represents the filtered subset matrix of independent components. Indicates the first The retained independent component vectors, This indicates the number of independent components retained after filtering. This represents the number of elements in the set of independent component indices that satisfy the threshold condition.

[0075] It should be noted that feature selection uses correlation coefficients to filter independent components. The correlation coefficient measures the degree of association between each component and concentration, with a preset threshold of 0.7. For example, if calculations show that the coefficient for the concentration-related component is 0.85 and the coefficient for the temperature-related component is 0.3, then the former will be retained.

[0076] Specifically, the filtered subset of independent components contains only highly relevant components, reducing the dimension from 4 to 1 or 2 to minimize redundancy. For example, the retained components might represent concentration trends, such as fluctuations of 4-6 ppm. This screening focuses on key information, improving the targeting of subsequent processing.

[0077] Step S330: Based on the selected subset of independent components, the independent components are linearly combined using a data recombination method to generate a feature vector directly related to the gas concentration, thus obtaining an initial set of concentration feature values.

[0078] In one embodiment, data recombination integrates independent components into a feature vector directly related to concentration through linear combination. The linear combination can be based on weight allocation, such as 80% for concentration-related components and 20% for minor components, generating a single feature vector.

[0079] For example, the combined feature vector may map the concentration signal to a range of 0-1, reflecting the intensity of concentration changes. This reorganization simplifies the data structure and facilitates subsequent analysis.

[0080] Step S340: Normalize the initial concentration feature value set, calculate the standardized score of each feature value, obtain the final feature value corresponding to the gas concentration, and determine the initial concentration feature value.

[0081] Preferably, the initial set of concentration characteristic values ​​needs to be normalized to ensure that the characteristic values ​​are comparable. For example, the characteristic value 0.6 may be normalized to 0.5 to eliminate dimensional differences.

[0082] Understandably, the normalized final eigenvalues ​​directly reflect concentration characteristics. For example, standardized scores might map concentration eigenvalues ​​to a range of -1 to 1, where 0.8 represents a high concentration. This processing makes the eigenvalues ​​more intuitive and suitable for real-time monitoring.

[0083] In one embodiment, the final feature value can be used for concentration anomaly detection, such as triggering an alarm when the fraction exceeds 0.9. The advantage of this method is its compact data, high computational efficiency, and suitability for the rapid response needs of industrial sites.

[0084] Through a complete process from independent component analysis to normalization, the generated eigenvalue set accurately focuses on concentration information, providing a reliable foundation for subsequent analysis.

[0085] Furthermore, the artificial intelligence-based early warning method for flammable and explosive gases proposed in this embodiment includes step S400 as follows: Step S410: From the initial set of concentration feature values, the feature vectors are standardized using data preprocessing methods. The mean and standard deviation of each feature vector are calculated, and the feature values ​​are mapped to a uniform scale through linear transformation to obtain a standardized feature dataset.

[0086] For example, in industrial gas monitoring scenarios, the initial concentration feature set typically contains multiple feature vectors, such as numerical representations of parameters like silane concentration and humidity.

[0087] Standardization in data preprocessing aims to eliminate differences in the dimensions of different features.

[0088] In principle, standardization linearly transforms data to a scale with a mean of 0 and a standard deviation of 1 by calculating the mean and standard deviation of the feature vector. For example, if a feature vector has a concentration range of 4-6 ppm, a mean of 5 ppm, and a standard deviation of 0.5 ppm, after standardization, the values ​​are mapped to a range of -2 to 2; for instance, 0.5 ppm might be converted to 1. This process ensures that subsequent algorithms treat the weights of each feature fairly.

[0089] Step S420: Based on the standardized feature dataset, divide the sample set into a training set and a test set. Determine the ratio of the training set to the test set using a random sampling method, and obtain the training samples and test samples for the support vector machine algorithm to obtain the partitioned sample subset.

[0090] In one possible implementation, the standardized feature dataset needs to be divided into training and test sets to support model training and validation. Random sampling is often used for the division ratio, such as 80% training set and 20% test set. Specifically, assuming the dataset contains 1000 samples, 800 are randomly selected as the training set, and the remaining 200 are used as the test set.

[0091] It should be noted that random sampling must ensure a uniform sample distribution, for example, ensuring that the proportion of high-concentration and low-concentration samples is consistent, to avoid the model being biased towards one type of sample. This partitioning method provides reliable training and testing samples for the support vector machine algorithm.

[0092] Step S430: For the training samples in the partitioned sample subset, the support vector machine algorithm is used to calculate the maximum margin hyperplane between feature vectors. The kernel function is used to perform nonlinear mapping on the feature vectors to determine the classification model parameters and obtain the trained classification model.

[0093] The formula for calculating the maximum margin in a support vector machine is: (7) In formula (7), The normal vector of the classification hyperplane. This represents the dimension of the feature space. The normal vector is represented by the first normal vector. Each component. Formula (7) calculates the geometric margin between two categories. The support vector machine obtains the optimal classification boundary by maximizing this margin.

[0094] Preferably, the Support Vector Machine (SVM) algorithm classifies feature vectors by finding the maximum margin hyperplane. In principle, the hyperplane separates samples of different classes, maximizing the margin between classes. For example, in gas concentration monitoring, it is necessary to distinguish between normal fluctuations and abnormal high-concentration states.

[0095] Kernel functions, such as radial basis functions, can map eigenvectors to a high-dimensional space, solving nonlinear separability problems.

[0096] In one embodiment, the training samples contain standardized concentration feature vectors, and the algorithm calculates the hyperplane parameters to generate a classification model.

[0097] During model training, it is assumed that the feature values ​​of normal concentration samples are between -1 and 1, and the feature values ​​of high concentration abnormal samples are between 1 and 2. The hyperplane may be located near the feature value of 1 to distinguish between the two types of samples.

[0098] Step S440: Input the test samples into the trained classification model, calculate the distance between the feature vector of the test samples and the hyperplane through the trained classification model, determine the category of environmental parameter fluctuation to which each test sample belongs, and output the classification result.

[0099] The distance from the test sample to the hyperplane is: (8) In formula (8), Indicates the first The distance from each test sample to the hyperplane. Describes the normal vector of the hyperplane. Indicates the first The feature vector of each test sample The bias term represents the hyperplane. This represents the magnitude of the normal vector.

[0100] The output feature vector of the classification model is: (9) In formula (9), This represents the output feature vector of the classification model. This represents the input test sample. This represents the trained weight matrix. This represents the bias vector.

[0101] The classification results are as follows: (10) In formula (10), Indicates the first The classification results of each test sample This represents the total number of categories of environmental parameter fluctuations. Represents a given feature vector Under the condition of belonging to the first The probability of a class This indicates that the category with the highest probability is selected as the classification result.

[0102] Understandably, after a test sample is input into the classification model, the model determines the category by calculating the distance between the sample's feature vector and the hyperplane. For example, if a test sample has a feature value of 1.5, the model calculates that its distance from the hyperplane is positive, thus classifying it as a high-concentration anomaly.

[0103] Specifically, the classification results can be output as category labels, such as "abnormal" or "normal," which facilitates the real-time monitoring system to identify fluctuations in environmental parameters.

[0104] In one embodiment, the classification model can further refine its output, such as distinguishing between mild and severe anomalies, based on distance thresholds, for example, a distance greater than 2 indicates a severe anomaly. This refinement improves the accuracy of monitoring. For example, standardization makes feature values ​​comparable, a reasonable partitioning of the training and test sets ensures the model's generalization ability, and the kernel function mapping and hyperplane computation of the support vector machine achieve efficient classification.

[0105] These steps together constitute a complete process from data preprocessing to classification, suitable for the real-time needs of industrial gas monitoring. The implementation methods of each step support each other, ensuring the reliability and practicality of the classification results.

[0106] Furthermore, the artificial intelligence-based early warning method for flammable and explosive gases proposed in this embodiment includes step S500 as follows: Step S510: Obtain the fluctuation category and classification accuracy from the classification results. By comparing the preset threshold setting with the fluctuation amplitude, determine whether the classification result exceeds the preset environmental fluctuation threshold. If it exceeds the threshold, obtain the set of samples exceeding the threshold.

[0107] The classification accuracy of the classification results is calculated using the following formula: (11) In formula (11), Indicating classification accuracy, Represents the total number of samples. Indicates the first The true class label of each sample Indicates the first Predicted class label for each sample, This indicates that the indicator function is 1 if the prediction is correct and 0 otherwise.

[0108] For example, in industrial gas monitoring scenarios, classification results typically include fluctuation categories, such as "normal" or "abnormal," as well as classification accuracy, to evaluate model performance.

[0109] Classification accuracy is calculated by comparing the predicted class with the actual class. For example, if a test set contains 200 samples, and 180 samples are correctly predicted, the classification accuracy is 90%. This accuracy reflects the model's ability to distinguish between fluctuations in environmental parameters.

[0110] It should be noted that the preset environmental fluctuation threshold is usually set based on historical data or business needs. For example, a fluctuation of more than 0.5 ppm in silane concentration is considered abnormal.

[0111] The classification results are compared with a preset environmental fluctuation threshold to select a set of samples that exceed the threshold. For example, if the concentration of a sample fluctuates by 0.6 ppm, exceeding the preset environmental fluctuation threshold, it will be included in the set of samples that exceed the threshold.

[0112] Step S520: For the set of samples exceeding the threshold, obtain the initial concentration feature value and data noise, and use an adaptive filter to denoise the initial concentration feature value according to the filtering parameters to obtain the corrected concentration feature value.

[0113] The characteristic value of the corrected concentration is: (12) In formula (12), Indicates the first The characteristic value of the correction concentration, Indicates the first The corrected concentration value, The total number of concentration features. Indicates the first The weighting coefficients of each feature. Formula (12) calculates the feature value of each concentration by normalizing the corrected concentration and combining it with the weighting coefficients.

[0114] Specifically, for a sample set exceeding the threshold, the initial concentration feature values ​​may be affected by noise, such as sensor jitter or environmental interference. An adaptive filter is used to denoise the feature values.

[0115] In one embodiment, the adaptive filter smooths the concentration data based on filtering parameters, such as the time window length. For example, if the time window is set to 5 seconds, the filter weights the concentration values ​​over the past 5 seconds to reduce the impact of noise.

[0116] Suppose a sample has an initial concentration characteristic value of 5.1 ppm, including 0.1 ppm noise. After denoising, the corrected concentration characteristic value is 5 ppm. This denoising process improves data reliability.

[0117] Step S530: Calculate the characteristic distribution and sample quality based on the corrected concentration characteristic value, update the threshold setting through a dynamic adjustment method, and determine the new environmental fluctuation threshold.

[0118] The overall sample quality is assessed by calculating the weighted sum of squares of sample biases. (13) In formula (13), Indicates the sample quality assessment value. Represents the total number of samples. Indicates the first Measurements of a sample Represents the sample mean. Indicates the sample standard deviation. Indicates the first Quality weights for each sample.

[0119] The fluctuation threshold is dynamically updated by weighting the environmental parameter deviation and the original threshold: (14) In formula (14), This indicates the updated environmental fluctuation threshold. This indicates the original threshold. Indicates the dynamic adjustment coefficient. Indicates the number of environmental parameters. Indicates the first One environmental parameter value, Indicates environmental reference value, This represents the threshold decay factor.

[0120] Preferably, the characteristic distribution and sample quality are calculated based on the corrected concentration characteristic values. The characteristic distribution is determined by analyzing the range and central tendency of the corrected characteristic values; for example, the corrected concentration values ​​are distributed between 4.8 and 5.2 ppm, with a mean of 5 ppm. Sample quality is determined by assessing data consistency, such as detecting the presence of outliers.

[0121] In one possible implementation, if a sample's corrected concentration is 5.3 ppm, which deviates significantly from the mean, it may be labeled as a low-quality sample. The dynamic adjustment method updates the threshold setting based on the characteristic distribution. For example, if analysis reveals increased recent concentration fluctuations, the threshold setting might be adjusted from 0.5 ppm to 0.6 ppm to adapt to the new environmental characteristics.

[0122] Step S540: Determine whether the corrected concentration feature value meets the new environmental fluctuation threshold. If the corrected concentration feature value meets the new environmental fluctuation threshold, then classify the corrected concentration feature value using the support vector machine algorithm to determine the adjusted fluctuation category.

[0123] The following formula is used to classify and determine the characteristic values ​​of the correction concentration: (15) In formula (15), This represents the classification decision function of the support vector machine. Represents the Lagrange multipliers. Indicates the labels of the training samples. Represents the kernel function. Indicates the bias term. This indicates the number of support vectors.

[0124] The probability distribution of each volatility category is calculated using the softmax function to determine the final adjusted volatility category. (16) In formula (16), Indicates the first The probability of a wave type. Indicates the first The score of the class, Indicates the first The probability of a wave type. This indicates the total number of fluctuation categories.

[0125] If the corrected concentration feature value meets the new environmental fluctuation threshold, the support vector machine algorithm classifies the feature value to determine the adjusted fluctuation category. For example, if the corrected concentration value is 5.01 ppm and the fluctuation amplitude is 0.1 ppm, which does not exceed the new environmental fluctuation threshold of 0.6 ppm, it is classified as "normal". In one embodiment, the classification model determines the category based on the distance between the corrected feature value and the hyperplane; if the distance is less than the new environmental fluctuation threshold, it is marked as normal. This classification method ensures accurate differentiation of fluctuation categories.

[0126] The classification results can be directly used for real-time monitoring, triggering corresponding alarms or adjustment measures. For example, a scheme combining dynamically adjusted thresholds with adaptive filtering can adapt to environmental changes and reduce false alarms.

[0127] The corrected eigenvalues ​​improve the robustness of classification, while the accuracy of the classification results provides a reliable basis for subsequent decision-making. These steps, through a complete process from denoising to threshold updating, collectively support the real-time performance and accuracy of industrial gas monitoring.

[0128] Another aspect of the present invention relates to an artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning system, used to implement the aforementioned artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning method. The artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning system includes an acquisition module, a processing module, an extraction module, a classification module, and a judgment module. The acquisition module acquires a data set collected by a multimodal sensor, including target gas concentration, temperature, humidity, and pressure data. The processing module uses principal component analysis to perform dimensionality reduction on the data set to obtain a target feature set. The extraction module extracts a set of independent components related to gas concentration from the target feature set to generate initial concentration feature values. The classification module uses a support vector machine algorithm to classify environmental parameter fluctuations based on the initial concentration feature values ​​and outputs the classification results. The judgment module determines whether the classification result exceeds a preset environmental fluctuation threshold. If the classification result exceeds the preset environmental fluctuation threshold, an adaptive filter is used to adjust the initial concentration feature values ​​to generate corrected concentration feature values.

[0129] Furthermore, the artificial intelligence-based flammable and explosive gas fire and explosion prevention and early warning system provided in this embodiment includes an acquisition module comprising a first acquisition unit and a preprocessing unit. The first acquisition unit is used to acquire a raw data set from a multimodal sensor. The preprocessing unit is used to calibrate the target gas concentration, temperature, humidity, and pressure data in the raw data set using a sensor calibration algorithm, and remove noise through mean filtering to obtain a preprocessed data set.

[0130] Preferably, the artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning system provided in this embodiment includes a processing module comprising a second acquisition unit, a third acquisition unit, a judgment unit, and a fourth acquisition unit. The second acquisition unit is used to normalize the target gas concentration, temperature, humidity, and pressure data from a preprocessed data set acquired by a multimodal sensor using a data standardization method, resulting in a standardized data set. The third acquisition unit is used to calculate the covariance matrix of the standardized data set using a principal component analysis algorithm, obtain eigenvalues ​​and eigenvectors, determine the number of principal components, and obtain a dimensionality-reduced feature vector set. The judgment unit is used to determine whether the feature dimension of the dimensionality-reduced feature vector set is higher than a preset threshold. If the feature dimension is higher than the preset threshold, the feature vectors are filtered using a feature selection method to obtain a subset of key features, resulting in a simplified feature set. The fourth acquisition unit is used to reorganize the feature data according to the simplified feature set using a data compression method, generating a compressed feature dataset, resulting in a target feature set.

[0131] Furthermore, the artificial intelligence-based flammable and explosive gas fire and explosion prevention safety early warning system provided in this embodiment includes an extraction module comprising a fifth acquisition unit, a sixth acquisition unit, a seventh acquisition unit, and a determination unit. The fifth acquisition unit is used to calculate the statistical independence between each feature vector in the target feature set using an independent component analysis algorithm, to obtain independent components related to gas concentration, and to obtain an initial independent component dataset. The sixth acquisition unit is used to calculate the correlation coefficient between each independent component and the gas concentration using a feature selection method for the initial independent component dataset, and to determine whether the correlation coefficient is higher than a preset threshold. If the correlation coefficient is higher than the preset threshold, the independent component is retained, and a filtered subset of independent components is obtained. The seventh acquisition unit is used to linearly combine the independent components using a data recombination method based on the filtered subset of independent components to generate feature vectors directly related to gas concentration, and to obtain an initial concentration feature value set. The determination unit is used to normalize the initial concentration feature value set, calculate the standardized score of each feature value, obtain the final feature value corresponding to the gas concentration, and determine the initial concentration feature value.

[0132] This embodiment provides an AI-based method and system for early warning of flammable and explosive gases. Compared with existing technologies, it addresses the issue that gas concentration, temperature, humidity, and pressure data collected by multimodal sensors are easily affected by environmental fluctuations. It employs principal component analysis (PCA) to reduce the dimensionality of the data, extracting independent components related to gas concentration to generate initial concentration feature values. Then, a support vector machine (SVM) algorithm is used to classify environmental parameter fluctuations and determine whether the classification results exceed a preset environmental fluctuation threshold. If so, an adaptive filter is used to adjust the initial concentration feature values, generating corrected concentration feature values. This embodiment, through the combined application of multiple algorithms, effectively eliminates the influence of environmental fluctuations on gas concentration measurement, improves measurement accuracy and reliability, and is suitable for gas concentration monitoring scenarios in various complex environments.

[0133] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for early warning of flammable and explosive gases based on artificial intelligence, characterized in that, Includes the following steps: Acquire a data set collected by a multimodal sensor, the data set including target gas concentration, temperature, humidity and pressure data; The target feature set is obtained by dimensionality reduction of the dataset using principal component analysis algorithm; Extract an independent component set related to gas concentration from the target feature set to generate initial concentration feature values; Based on the initial concentration feature values, the support vector machine algorithm is used to classify the fluctuations of environmental parameters and output the classification results; Determine whether the classification result exceeds a preset environmental fluctuation threshold. If the classification result exceeds the preset environmental fluctuation threshold, adjust the initial concentration feature value through an adaptive filter to generate a corrected concentration feature value.

2. The artificial intelligence-based early warning method for flammable and explosive gas combustion and explosion prevention as described in claim 1, characterized in that, The steps for acquiring the data set collected by the multimodal sensor include: Acquire raw data sets from multimodal sensors; The target gas concentration, temperature, humidity, and pressure data in the original dataset are calibrated using a sensor calibration algorithm, and noise is removed by mean filtering to obtain a preprocessed dataset.

3. The artificial intelligence-based early warning method for flammable and explosive gas combustion and explosion prevention as described in claim 1, characterized in that, The step of using principal component analysis to reduce the dimensionality of the dataset to obtain the target feature set includes: From the preprocessed data set acquired by the multimodal sensor, the target gas concentration, temperature, humidity, and pressure data are normalized using a data standardization method to obtain a standardized data set. For the standardized dataset, the covariance matrix is ​​calculated using principal component analysis to obtain eigenvalues ​​and eigenvectors, determine the number of principal components, and obtain the dimensionality-reduced eigenvector set. Determine whether the feature dimension of the reduced feature vector set is higher than a preset threshold. If the feature dimension of the reduced feature vector set is higher than the preset threshold, then filter the feature vectors using a feature selection method to obtain a subset of key features and obtain the simplified feature set. Based on the simplified feature set, the feature data is recombined using a data compression method to generate a compressed feature dataset, thus obtaining the target feature set.

4. The artificial intelligence-based early warning method for flammable and explosive gas combustion and explosion prevention as described in claim 1, characterized in that, The steps for extracting a set of independent components related to gas concentration from the target feature set to generate initial concentration feature values ​​include: The independent component analysis algorithm is used to calculate the statistical independence between each feature vector in the target feature set, obtain the independent components related to the gas concentration, and obtain the initial independent component dataset; For the initial independent component dataset, the correlation coefficient between each independent component and the gas concentration is calculated using a feature selection method. It is then determined whether the correlation coefficient is higher than a preset threshold. If the correlation coefficient is higher than the preset threshold, the independent component is retained, resulting in a filtered subset of independent components. Based on the selected subset of independent components, the independent components are linearly combined using a data recombination method to generate a feature vector directly related to the gas concentration, thus obtaining an initial set of concentration feature values. By normalizing the initial concentration feature value set, calculating the standardized score of each feature value, obtaining the final feature value corresponding to the gas concentration, and determining the initial concentration feature value.

5. The artificial intelligence-based early warning method for flammable and explosive gas combustion and explosion prevention as described in claim 1, characterized in that, The steps for classifying environmental parameter fluctuations using a support vector machine algorithm based on the initial concentration feature values ​​and outputting the classification results include: From the initial set of concentration feature values, the feature vectors are standardized using data preprocessing methods. The mean and standard deviation of each feature vector are calculated, and the feature values ​​are mapped to a uniform scale through linear transformation to obtain a standardized feature dataset. Based on the standardized feature dataset, the sample set is divided into a training set and a test set. The ratio of the training set to the test set is determined by random sampling. Training samples and test samples for the support vector machine algorithm are obtained, resulting in the partitioned sample subset. For the training samples in the partitioned sample subset, the support vector machine algorithm is used to calculate the maximum margin hyperplane between feature vectors. The kernel function is used to perform nonlinear mapping on the feature vectors to determine the classification model parameters and obtain the trained classification model. The test samples are input into the trained classification model. The distance between the feature vector of the test sample and the hyperplane is calculated by the trained classification model. The category of environmental parameter fluctuation to which each test sample belongs is determined, and the classification result is output.

6. The artificial intelligence-based early warning method for flammable and explosive gas combustion and explosion prevention as described in claim 1, characterized in that, The step of determining whether the classification result exceeds a preset environmental fluctuation threshold, and if the classification result exceeds the preset environmental fluctuation threshold, adjusting the initial concentration feature value through an adaptive filter to generate a corrected concentration feature value includes: The fluctuation category and classification accuracy are obtained from the classification results. By comparing the preset threshold setting with the fluctuation amplitude, it is determined whether the classification results exceed the preset environmental fluctuation threshold. If they exceed the threshold, a set of samples exceeding the threshold is obtained. For the set of samples exceeding the threshold, initial concentration feature values ​​and data noise are obtained. An adaptive filter is used to denoise the initial concentration feature values ​​according to the filtering parameters to obtain corrected concentration feature values. Based on the corrected concentration characteristic value, the characteristic distribution and sample quality are calculated, and the threshold setting is updated through a dynamic adjustment method to determine a new environmental fluctuation threshold. Determine whether the corrected concentration feature value meets the new environmental fluctuation threshold. If the corrected concentration feature value meets the new environmental fluctuation threshold, classify the corrected concentration feature value using the support vector machine algorithm to determine the adjusted fluctuation category.

7. An artificial intelligence-based early warning system for flammable and explosive gases, used to implement the artificial intelligence-based early warning method for flammable and explosive gases as described in any one of claims 1 to 6, wherein the artificial intelligence-based early warning system for flammable and explosive gases comprises: The acquisition module is used to acquire a data set collected by the multimodal sensor, the data set including target gas concentration, temperature, humidity and pressure data; The processing module is used to perform dimensionality reduction processing on the dataset using principal component analysis algorithm to obtain the target feature set; The extraction module is used to extract a set of independent components related to gas concentration from the target feature set to generate initial concentration feature values; The classification module is used to classify environmental parameter fluctuations based on the initial concentration feature values ​​using a support vector machine algorithm, and output the classification results. The judgment module is used to determine whether the classification result exceeds a preset environmental fluctuation threshold. If the classification result exceeds the preset environmental fluctuation threshold, the initial concentration feature value is adjusted through an adaptive filter to generate a corrected concentration feature value.

8. The artificial intelligence-based flammable and explosive gas fire and explosion prevention and safety early warning system as described in claim 7, characterized in that, The acquisition module includes: The first acquisition unit is used to acquire a raw data set from the multimodal sensor; The preprocessing unit is used to calibrate the target gas concentration, temperature, humidity, and pressure data in the original dataset using a sensor calibration algorithm, and remove noise through mean filtering to obtain a preprocessed dataset.

9. The artificial intelligence-based flammable and explosive gas fire and explosion prevention and safety early warning system as described in claim 7, characterized in that, The processing module includes: The second acquisition unit is used to normalize the target gas concentration, temperature, humidity, and pressure data from the preprocessed data set acquired by the multimodal sensor using a data standardization method to obtain a standardized data set. The third acquisition unit is used to calculate the covariance matrix using the principal component analysis algorithm for the standardized data set, obtain eigenvalues ​​and eigenvectors, determine the number of principal components, and obtain the dimensionality-reduced eigenvector set. The judgment unit is used to determine whether the feature dimension of the dimensionality reduction feature vector set is higher than a preset threshold. If the feature dimension of the dimensionality reduction feature vector set is higher than the preset threshold, the feature vector is filtered by a feature selection method to obtain a subset of key features and thus obtain the simplified feature set. The fourth acquisition unit is used to reorganize the feature data according to the simplified feature set using a data compression method to generate a compressed feature dataset and obtain the target feature set.

10. The artificial intelligence-based flammable and explosive gas fire and explosion prevention and safety early warning system as described in claim 7, characterized in that, The extraction module includes: The fifth acquisition unit is used to calculate the statistical independence between each feature vector in the target feature set by using the independent component analysis algorithm, to obtain the independent components related to the gas concentration, and to obtain the initial independent component dataset. The sixth acquisition unit is used to calculate the correlation coefficient between each independent component and the gas concentration using a feature selection method for the initial independent component dataset, determine whether the correlation coefficient is higher than a preset threshold, and retain the independent component if the correlation coefficient is higher than the preset threshold to obtain a filtered subset of independent components. The seventh acquisition unit is used to linearly combine the independent components according to the filtered subset of independent components using a data recombination method to generate a feature vector directly related to the gas concentration, thereby obtaining an initial set of concentration feature values. The determining unit is used to normalize the initial concentration feature value set, calculate the standardized score of each feature value, obtain the final feature value corresponding to the gas concentration, and determine the initial concentration feature value.

Images