Intelligent gas detection method, terminal and readable storage medium based on multi-sensor fusion

Abstract

This invention discloses a multi-sensor fusion-based intelligent gas detection method, terminal, and readable storage medium. The method includes: acquiring signal data from a multi-sensor array; processing the sensor array signal data to obtain model input data; inputting the model input data into an analysis model; the analysis model extracting signal features from the model input data; and the analysis model identifying the gas type and concentration based on the signal features. This invention uses a sensor array comprising multiple sensor units to collect spatial and temporal information of a mixed gas at the sensor dimension. Then, an analysis model is used to learn the relationship between this spatial and temporal information and the type and concentration of the mixed gas, thus taking into account the complex spatiotemporal characteristics of the mixed gas to identify the type and concentration of the mixed gas. Its detection effect on mixed gases is far superior to existing technologies.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, and in particular to a smart gas detection method, terminal, and readable storage medium based on multi-sensor fusion. Background Technology

[0002] Gas detection is mainly used in the monitoring of toxic and harmful gases, food quality and freshness inspection, and breathalyzer testing for diseases. Among these applications, mixtures of multiple gases, such as mixtures of volatile organic compounds (VOCs), are particularly important. VOCs are organic compounds with boiling points between 50°C and 250°C that are easily volatilized at room temperature. Due to their high volatility and the toxicity of some gases, they pose a potential threat to the ecological environment and human health. Therefore, their monitoring and control have received considerable attention in recent years.

[0003] VOCs mainly originate from industrial emissions (such as printing, coating, and surface treatment), solvent use (such as paints, cleaning agents, and printing inks), chemical production (such as plastics, rubber, and fertilizers), daily necessities (such as furniture, carpets, and cleaning agents), as well as vehicle exhaust and building materials. Besides their wide range of sources, VOCs also have a complex and diverse composition, including common compounds such as alcohols (methanol, ethanol, propanol); aldehydes (formaldehyde, acetaldehyde); ketones (acetone); and hydrocarbons (methane, ethane, ethylene, propylene, benzene, toluene, etc.). The hazards of excessive VOCs mainly include the following aspects:

[0004] Firstly, VOCs pose a threat to human health. Since most VOCs are irritating, high concentrations can severely irritate the eyes, nose, and throat, potentially causing redness, tearing, and coughing. Prolonged exposure can lead to dizziness, headaches, and fatigue. VOCs also harm internal organs. For example, long-term exposure to toluene and xylene can affect the nervous system, liver, and kidneys, while long-term exposure to carbon tetrachloride may damage the heart. More seriously, some VOCs have been proven to be carcinogenic, such as benzene, formaldehyde, dichlorodiethyl sulfide, and vinyl chloride. Long-term exposure to these gases may increase the risk of cancer.

[0005] Secondly, VOCs pollute the atmospheric environment. VOCs cause various harms to air quality, mainly including the following: First, VOCs are important precursors to the formation of ozone (O3) and fine particulate matter (PM2.5). Ozone is a strong oxidant; excessively high concentrations affect air quality, while fine particulate matter is the main culprit for reducing visibility. Second, VOCs can undergo complex chemical reactions in the atmosphere, generating toxic substances such as carcinogens, organic acids, and aldehydes. Third, the photochemical reaction between VOCs and nitrogen oxides can produce photochemical smog, thus leading to environmental air pollution.

[0006] Thirdly, they endanger production safety. Some VOCs are flammable and explosive, such as methane, ethylene, ethanol, toluene, and acetone. When their concentration in the air reaches a certain limit, they can easily explode and cause fires when exposed to a source of ignition or high temperatures. In addition, some VOCs are corrosive and may damage production equipment and facilities, thus posing safety risks to people's production and manufacturing.

[0007] In general, VOCs have diverse sources and are abundant in types. Their concentration and types in the environment fluctuate with time and location, making VOCs detection extremely complex. Existing gas detection methods do not take into account the complex spatiotemporal characteristics of VOCs, resulting in poor detection performance. Summary of the Invention

[0008] The purpose of this invention is to provide a smart gas detection method, terminal, and readable storage medium based on multi-sensor fusion, aiming to solve the problem of poor detection performance of mixed gases in existing technologies.

[0009] The technical solution adopted by this invention to solve the technical problem is as follows:

[0010] This invention provides a multi-sensor fusion-based intelligent gas detection method, which includes:

[0011] Acquire signal data from a multi-sensor array, process the sensor array signal data, and obtain model input data;

[0012] The model input data is input into the analysis model, and the analysis model extracts the signal features of the model input data.

[0013] The analysis model identifies the gas type and concentration based on the signal characteristics.

[0014] Furthermore, the step of acquiring multi-sensor array signal data and processing the sensor array signal data to obtain model input data specifically includes:

[0015] The signals from multiple sensor units are acquired, and the signals from the multiple sensor units are preprocessed to align the signals of each sensor unit to obtain the multi-sensor array signal data.

[0016] The model input data is obtained by denoising and standardizing the signal data from the multi-sensor array.

[0017] Furthermore, the analysis model includes a feature extraction network, which includes a spatial extraction branch, a temporal extraction branch, and a gating unit;

[0018] The analysis model extracts signal features from the model input data, specifically including:

[0019] The spatial extraction branch extracts the sensor dimension features of the model input data;

[0020] The time extraction branch extracts the time dimension features of the model input data;

[0021] The gating unit integrates the sensor dimensional features and the time dimensional features to obtain the signal features.

[0022] Furthermore, the spatial extraction branch sequentially includes a spatial coding network and one or more spatial encoder modules, each of which includes a multi-head attention layer and a spatial feedforward network layer, wherein the spatial coding network includes a spatial linear embedding layer;

[0023] The spatial extraction branch extracts sensor dimension features from the model input data, specifically including:

[0024] The spatial linear embedding layer projects the model input data into a high-dimensional space to obtain spatial data, and then inputs the spatial data into the spatial encoder module.

[0025] The multi-head attention layer extracts features from the input;

[0026] The spatial feedforward network enhances the features extracted by the multi-head attention layer of the spatial encoder module;

[0027] The output of the last spatial encoder module is used as the sensor dimensional feature.

[0028] Furthermore, the spatial encoder module also includes a residual connection and normalization layer after the multi-head attention layer and a residual connection and normalization layer after the spatial feedforward network layer;

[0029] The residual connection and normalization layer are used to perform residual connection and regularization processing on the features extracted by the multi-head attention layer or the features enhanced by the spatial feedforward network.

[0030] Furthermore, the time extraction branch sequentially includes a time coding network and one or more time encoder modules. Each time encoder module includes a masked multi-head attention layer and a time feedforward network layer. The time coding network includes a time linear embedding layer and a position embedding layer.

[0031] The time extraction branch extracts sensor dimension features from the model input data, specifically including:

[0032] The time-linear embedding layer projects the model input data into a high-dimensional space to obtain intermediate time data.

[0033] The location embedding layer performs time step encoding on the intermediate time data to obtain time data, and inputs the time data into the time encoder module;

[0034] The masked multi-head attention layer extracts features from the input;

[0035] The time feedforward network enhances the features extracted by the masked multi-head attention layer of the time encoder module;

[0036] The output of the last time encoder module is used as the sensor dimensional feature.

[0037] Furthermore, the time encoder module also includes a residual connection and normalization layer after the masked multi-head attention layer and a residual connection and normalization layer after the time feedforward network layer;

[0038] The residual connection and normalization layer is used to perform residual connection and regularization processing on features extracted by the masked multi-head attention layer or features enhanced by the temporal feedforward network.

[0039] Furthermore, the analysis model includes a classification network, which comprises a linear layer and an output layer;

[0040] The analysis model identifies the gas type and concentration based on the signal characteristics, specifically including:

[0041] The linear layer performs a linear transformation on the signal features to obtain linearly transformed data.

[0042] The output layer converts the linear transformation data into gas type and concentration, and then outputs it.

[0043] In addition, to achieve the above objectives, the present invention also provides a terminal, the terminal comprising: a memory, a processor, and a multi-sensor fusion intelligent gas detection program stored in the memory and executable on the processor, wherein when the multi-sensor fusion intelligent gas detection program is executed by the processor, the terminal controls the terminal to implement the steps of the multi-sensor fusion intelligent gas detection method as described above.

[0044] Furthermore, to achieve the above objectives, the present invention also provides a readable storage medium storing a multi-sensor fusion intelligent gas detection program, which, when executed by a processor, implements the steps of the multi-sensor fusion intelligent gas detection method as described above.

[0045] The present invention, by employing the above technical solution, has the following effects:

[0046] This invention acquires spatial information of mixed gas in a sensor array comprising multiple sensor units. Simultaneously, the signals from multiple time steps of the sensor array naturally possess temporal information. Therefore, the sensor array signal data includes both spatial and temporal information of the mixed gas. An analytical model is then used to learn the relationship between this spatial and temporal information and the type and concentration of the mixed gas. Specifically, it learns the characteristic that the concentration and type of the mixed gas in the environment fluctuate with changes in time and location, thus taking into account the complex spatiotemporal characteristics of the mixed gas to identify its type and concentration. Its detection performance for mixed gases is far superior to existing technologies. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the steps of a method for dimensional reduction representation of a three-dimensional property in a preferred embodiment of the present invention.

[0048] Figure 2 This is a flowchart illustrating the structure of the analysis model in a preferred embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the described embodiments are merely some embodiments of this disclosure, and not all embodiments. The specific embodiments described herein are only used to explain the invention and are not intended to limit the invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0051] Example 1

[0052] Please see Figure 1 Embodiment 1 of the present invention is a multi-sensor fusion intelligent gas detection method, which can be applied to a chip integrating a multi-sensor array to directly output the gas type and concentration obtained from the analysis. It can also run on a terminal or server. The terminal or server collects signal data through the multi-sensor array, analyzes the signal data to obtain the gas type and concentration, and displays it to the user accordingly. The terminal includes mobile devices, user terminals, handheld devices, computing devices, and wearable devices, etc. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data and artificial intelligence platforms.

[0053] This embodiment of a multi-sensor fusion intelligent gas detection method includes the following steps:

[0054] S1. Acquire multi-sensor array signal data, process the sensor array signal data to obtain model input data.

[0055] The multi-sensor array includes multiple sensor units distributed at different locations. Each sensor unit includes multiple chemical sensors that can detect all kinds of gases in the mixed gas to be detected. In this embodiment, each sensor unit includes sensors for multiple gas components such as formaldehyde, toluene, and methanol.

[0056] The processing of the sensor array signal data specifically includes performing signal preprocessing, noise removal, and standardization on the sensor array signal data sequentially. This results in cleaner and more reliable data, laying the foundation for subsequent signal identification.

[0057] The signal preprocessing involves aligning the time points of all sensor units. This process involves selecting a reference sensor unit, identifying the characteristic time point sequences of all sensor units, and aligning the characteristic time point sequences of the remaining sensor units with those of the quasi-sensor unit.

[0058] The standardization process is specifically based on the Z-score data normalization method, which standardizes the data to eliminate the influence of different scales and dimensions.

[0059] Specifically, finding the characteristic time point sequence of all sensor units involves analyzing the original data sequence of the corresponding sensor unit in the sensor array signal data to obtain the rise and fall time points of the original data sequence, which are then used as the characteristic time point sequence.

[0060] S2. Input the model input data into the analysis model, and the analysis model extracts the signal features of the model input data.

[0061] Specifically, in this embodiment, the powerful feature extraction and modeling capabilities of the Transformer model are used to develop a targeted gas component identification analysis model to adaptively fuse and analyze the multidimensional feature data acquired by the chemical sensor array, thereby achieving rapid and accurate detection of VOCs and harmful gases.

[0062] It uses the Python programming language in the PyCharm integrated development environment for programming and running, and uses the mainstream PyTorch deep learning development framework to realize data analysis and processing of gas sensor signals under different categories and the construction of related classification models.

[0063] Specifically, this embodiment uses a Transformer-based dual-tower GTN (Gated Transformer Networks) gas classification and recognition model as the analysis model. This invention plans to utilize deep learning technology to improve the accuracy of VOCs detection. Specifically, this invention will begin by establishing a database, collecting a large amount of VOCs data using gas sensors, and then constructing a deep learning model. This model will automatically learn and identify the characteristics and behavioral patterns of VOCs by learning from a large amount of VOCs data, thereby achieving accurate and rapid detection and differentiation of VOCs.

[0064] It is worth noting that in other alternative embodiments, classic machine learning algorithms such as SVM and random forest can also be used in conjunction with feature data acquired by the gas sensor array to identify gas components. Through feature engineering and model optimization, rapid detection of common gases can be achieved. Classic algorithm schemes are relatively simple, but they have high requirements for feature extraction and model design, and their recognition performance and adaptability are generally poor.

[0065] Alternatively, deep learning models such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) can be used to directly learn the identification features of gas components from sensor data. Compared with classic traditional machine learning methods, deep learning methods have stronger feature extraction capabilities and adaptability. However, this approach usually requires a large amount of labeled data for training, places high demands on hardware resources, and lacks interpretability of the model. This makes the overall design and improvement of the model more difficult, as well as training and implementation.

[0066] Therefore, in this embodiment, it is preferable to use the powerful feature extraction and modeling capabilities of the Transformer model to develop a targeted gas component identification analysis model as a feature extraction and classification scheme.

[0067] Please refer to Figure 2The analysis model in this embodiment includes a feature extraction network and a classification network. In this step, the feature extraction network is used.

[0068] The feature extraction network adopts a dual-tower structure, including a spatial extraction branch, a temporal extraction branch, and a gating unit. The spatial extraction branch and the temporal extraction branch are both encoder blocks of a Transformer model to encode the input data into signal features.

[0069] The spatial extraction branch sequentially includes a spatial coding network and multiple spatial encoder modules. Each spatial encoder module includes a multi-head attention layer, a residual connection and normalization layer, a spatial feedforward network layer, and a residual connection and normalization layer.

[0070] The spatial coding network is used to project the model input data into a high-dimensional space so that the subsequent feature extraction network can better understand the complex relationships in the model input data. The spatial coding network includes a spatial linear embedding layer (Embedding(Linear)), which projects the model input data into a high-dimensional space as spatial data for subsequent extraction of spatial relationships between various sensor units.

[0071] Among them, the multi-head attention layer is used to calculate the relationship between different input channels. It maps the input spatial data to different subspaces, extracts different relationships between the spatial data in each subspace, and finally merges the information of all subspaces to find the correlation between different angles on the overall sequence of spatial data.

[0072] The spatial feedforward network layer includes two linear transformations and a ReLU activation function between the linear transformations. The role of the feedforward network layer is to further process and transform the spatial features extracted by the multi-head attention layer in order to further extract and fuse feature information, thereby enhancing the nonlinearity and expressiveness of the model.

[0073] Residual connections and normalization layers are used to perform post-processing for each step in the Transformer module. These include residual connections and regularization. The purpose of residual connections is to help avoid the gradient vanishing problem during the training of deep networks, while the purpose of regularization is to accelerate the training process and improve the stability of the model.

[0074] The spatial extraction branch finally outputs the sensor dimension features of the model input data. These sensor dimension features can model the correlations and differences between different sensors and utilize the complementarity of information from multiple sensors. Therefore, sensor dimension features help to better distinguish different types of gases.

[0075] As can be seen, since the spatial extraction branch only extracts features in the sensor dimension and does not consider features in the time dimension, the position encoding of the time step can be omitted during spatial encoding, which reduces the overall latency and resource requirements of the model.

[0076] The temporal extraction branch includes a temporal coding network and multiple temporal encoder modules. Each temporal encoder module includes a masked multi-head attention layer, a residual connection and normalization layer, a temporal feedforward network layer, and a residual connection and regularization layer.

[0077] Temporal encoding networks consist of temporal linear embedding layers and positional embedding layers. Temporal linear embedding layers are also used to project model input data into a high-dimensional space. However, the model input data after temporal linear embedding layers also needs to pass through positional embedding layers. Positional embedding layers embed the positional information of sequence data in the sequence, thereby encoding time steps to form temporal data, in order to maintain the sequential relationship of sequence data so as to extract the temporal features of the model input data in the future.

[0078] Masked Multi-Head Attention calculates the relationships between different time steps and between different input channels. Masked attention is used in the model's prediction section to ensure that future data is not seen when generating predictions for the current time step, which helps maintain the causal relationships in the time series.

[0079] The temporal feedforward network layer functions similarly to the spatial feedforward network layer. It includes two linear transformations and a ReLU activation function between the linear transformations. This is used to further process and transform the temporal features extracted by the masked multi-head attention layer, so as to further extract and fuse feature information and enhance the nonlinearity and expressiveness of the model.

[0080] Residual connections and normalization layers are also used to perform post-processing for each step in the Transformer module. These include residual connections and regularization. The purpose of residual connections is to help avoid the gradient vanishing problem during deep network training, and the purpose of regularization is to accelerate the training process and improve the stability of the model.

[0081] The time extraction branch ultimately outputs the time dimension features of the model input data. This captures the characteristics of the signal as it changes over time, including rapidly changing instantaneous features and long-term trend features. These time-dimensional features enable better modeling of the dynamic behavior of gas signals.

[0082] The temporal and sensor-dimensional features of the model input data are fed into the gating unit. The gating unit controls the passage of feature information to fuse sensor and temporal features into signal features. This gating mechanism, through fusion, transmits more important information. It accepts input from the downstream linear layer and transforms the linear layer input using a sigmoid activation function, resulting in a set of values ​​between 0 and 1. These values ​​are called "gating values." The gating mechanism can learn how to adaptively weigh the importance of temporal and sensor-dimensional features, dynamically adjusting the contribution of the two dimensions to the final classification result. These "gating values" are used to perform element-wise multiplication with the upstream sensor and temporal features, selectively controlling which information can pass through and which should be blocked, and controlling the importance of each piece of information. Through this gating mechanism, the model can dynamically learn which information is most important for the current task. This fusion mechanism can fully utilize the advantages of both dimensions, thus enabling more targeted feature extraction and information transmission. It helps the model better capture long-term dependencies in time-series data, improving prediction accuracy and robustness.

[0083] It can be seen that adaptive data fusion based on the Transformer model can effectively improve the accuracy and robustness of gas component identification.

[0084] S3. The analysis model identifies the gas type and concentration based on the signal characteristics.

[0085] This step uses the classification network in the analysis model, which consists of a linear layer and an output layer.

[0086] The linear layer is designed to perform a linear transformation, mapping the input signal features to a new dimensional space. It performs this linear transformation using a learnable weight matrix and bias vector. This linear transformation helps the model learn the linear relationships and patterns in the input data.

[0087] The output layer specifically uses the Softmax activation function, which transforms the input data into a probability distribution and normalizes the model's output values ​​so that the sum of the model's output probabilities is 1, which can be interpreted as the probabilities of each class. The Softmax layer is used as the output layer for multi-class classification tasks, helping the model output probability predictions for each class, thus making more reliable classification decisions.

[0088] Example 2

[0089] Please see Figure 3Based on the above method, the present invention also provides a terminal, the terminal comprising: a memory 10, a processor 20, and a multi-sensor fusion intelligent gas detection program stored on the memory 10 and executable on the processor 20, wherein when the multi-sensor fusion intelligent gas detection program is executed by the processor 20, the terminal controls the terminal to implement the steps of the multi-sensor fusion intelligent gas detection method as described above.

[0090] In some embodiments, the memory 10 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 10 may be an external storage device of the terminal, such as a plug-in hard disk, SmartMediaCard (SMC), SecureDigital (SD) card, or FlashCard. Furthermore, the memory 10 may include both internal and external storage devices. The memory 10 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 10 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 10 stores a multi-sensor fusion intelligent gas detection program, which can be executed by the processor 20 to implement the multi-sensor fusion intelligent gas detection method of this application.

[0091] In some embodiments, the processor 20 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 10 or process data, such as executing the intelligent gas detection method of multi-sensor fusion.

[0092] Example 3

[0093] This embodiment provides a storage medium that stores a multi-sensor fusion intelligent gas detection program. When the multi-sensor fusion intelligent gas detection program is executed by a processor, it implements the steps of the multi-sensor fusion intelligent gas detection method described above.

[0094] In summary, this invention acquires spatial information of the mixed gas sensor dimension through a sensor array comprising multiple sensor units. Simultaneously, the signals from multiple time steps of the sensor array naturally possess temporal information. Therefore, the sensor array signal data includes both spatial and temporal information of the mixed gas. Subsequently, an analytical model is used to learn the relationship between this spatial and temporal information and the type and concentration of the mixed gas. This involves learning the characteristic that the concentration and type of the mixed gas in the environment fluctuate with changes in time and location, thus taking into account the complex spatiotemporal characteristics of the mixed gas to identify its type and concentration. Its detection performance for mixed gases is far superior to existing technologies.

[0095] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0096] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The storage medium can be a memory, magnetic disk, optical disk, etc.

[0097] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A multi-sensor fusion intelligent gas detection method, characterized in that, The intelligent gas detection method based on multi-sensor fusion includes: Acquire signal data from a multi-sensor array, process the sensor array signal data, and obtain model input data; The model input data is input into the analysis model, and the analysis model extracts the signal features of the model input data. The analysis model identifies the gas type and concentration based on the signal characteristics; The analysis model includes a feature extraction network, which includes a spatial extraction branch, a temporal extraction branch, and a gating unit. The analysis model extracts signal features from the model input data, specifically including: The spatial extraction branch extracts the sensor dimension features of the model input data; The time extraction branch extracts the time dimension features of the model input data; The gating unit fuses the sensor dimensional features and the time dimensional features to obtain the signal features; The spatial extraction branch sequentially includes a spatial coding network and multiple spatial encoder modules. Each spatial encoder module includes a multi-head attention layer and a spatial feedforward network layer. The spatial coding network includes a spatial linear embedding layer. The spatial extraction branch extracts sensor dimension features from the model input data, specifically including: The spatial linear embedding layer projects the model input data into a high-dimensional space to obtain spatial data, and then inputs the spatial data into the spatial encoder module. The multi-head attention layer is used to calculate the relationship between different input channels. It maps the input spatial data to different subspaces, extracts different relationships between the spatial data in each subspace, and finally merges the information of all subspaces to find the correlation relationship from different angles on the overall sequence of spatial data. The multi-head attention layer extracts features from the input; The spatial feedforward network enhances the features extracted by the multi-head attention layer of the spatial encoder module; The output of the last spatial encoder module is used as the sensor dimensional feature.

2. The intelligent gas detection method based on multi-sensor fusion according to claim 1, characterized in that, The process of acquiring multi-sensor array signal data and processing the sensor array signal data to obtain model input data specifically includes: The signals from multiple sensor units are acquired, and the signals from the multiple sensor units are preprocessed to align the signals of each sensor unit to obtain the multi-sensor array signal data. The signal data from the multi-sensor array is denoised and standardized to obtain the model input data.

3. The intelligent gas detection method based on multi-sensor fusion according to claim 1, characterized in that, The spatial encoder module also includes residual connections and normalization layers after the multi-head attention layer and residual connections and normalization layers after the spatial feedforward network layer. The residual connection and normalization layer are used to perform residual connection and regularization processing on the features extracted by the multi-head attention layer or the features enhanced by the spatial feedforward network.

4. The intelligent gas detection method based on multi-sensor fusion according to claim 1, characterized in that, The time extraction branch includes a time coding network and multiple time encoder modules in sequence. Each time encoder module includes a masked multi-head attention layer and a time feedforward network layer. The time coding network includes a time linear embedding layer and a position embedding layer. The time extraction branch extracts sensor dimension features from the model input data, specifically including: The time-linear embedding layer projects the model input data into a high-dimensional space to obtain intermediate time data. The location embedding layer performs time step encoding on the intermediate time data to obtain time data, and inputs the time data into the time encoder module; The masked multi-head attention layer extracts features from the input; The time feedforward network enhances the features extracted by the masked multi-head attention layer of the time encoder module; The output of the last time encoder module is used as the sensor dimensional feature.

5. The intelligent gas detection method based on multi-sensor fusion according to claim 4, characterized in that, The time encoder module also includes residual connections and normalization layers after the masked multi-head attention layer and residual connections and normalization layers after the time feedforward network layer. The residual connection and normalization layer is used to perform residual connection and regularization processing on features extracted by the masked multi-head attention layer or features enhanced by the temporal feedforward network.

6. The intelligent gas detection method based on multi-sensor fusion according to claim 1, characterized in that, The analysis model includes a classification network, which includes a linear layer and an output layer. The analysis model identifies the gas type and concentration based on the signal characteristics, specifically including: The linear layer performs a linear transformation on the signal features to obtain linearly transformed data. The output layer converts the linear transformation data into gas type and concentration, and then outputs it.

7. A terminal, characterized in that, The terminal includes: a memory, a processor, and a multi-sensor fusion intelligent gas detection program stored in the memory and executable on the processor. When the multi-sensor fusion intelligent gas detection program is executed by the processor, it controls the terminal to implement the steps of the multi-sensor fusion intelligent gas detection method as described in any one of claims 1-6.

8. A readable storage medium, characterized in that, The readable storage medium stores a multi-sensor fusion intelligent gas detection program, which, when executed by a processor, implements the steps of the multi-sensor fusion intelligent gas detection method as described in any one of claims 1-6.

Images