Linear light beam type smoke and methane composite detection method and device
By integrating dual-wavelength spectral features with deep learning, a linear beam detection method has been developed to solve the problem of detecting the coexistence of smoke and methane in long-distance or high-altitude spaces. This method achieves accurate differentiation between smoke and methane, reduces false alarm rates and equipment costs, and is suitable for multi-hazard early warning in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG FIRE RES INST OF MEM
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-12
AI Technical Summary
In long-distance or high-ceilinged spaces, traditional point-type smoke/heat detectors have limited coverage and are prone to blind spots. Linear beam smoke detection technology is susceptible to interference and has limited functionality. Methane detection and smoke detection are independent systems, resulting in high equipment costs and asynchronous signals. Existing dual-wavelength detection architectures have low detection accuracy and high false alarm rates in mixed scenarios.
A linear beam detection method combining dual-wavelength spectral features and deep learning is adopted. Dual-wavelength light intensity data are acquired by sensors, and after homogenization, smoke features are fused using a weight calculation module and a dual-gated GRU network. Methane features are extracted by combining a 1D convolutional neural network and a channel attention mechanism, and the merging and judgment are performed through a fully connected layer and a Sigmoid activation function.
It achieves accurate identification of mixed smoke and methane scenarios, improves anti-interference capabilities, reduces equipment costs and false alarm rates, and meets the needs of multiple hazard warnings in complex environments.
Smart Images
Figure CN122193115A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fire safety and gas leak monitoring technology, specifically to a linear beam detection method and device that integrates dual-wavelength spectral features and deep learning. It is suitable for early and accurate detection and risk warning in long-distance or enclosed spaces such as coal mines, petrochemical parks, natural gas processing stations, large storage facilities, and underground integrated pipe corridors, where smoke and methane leaks coexist. Background Technology
[0002] Fires and methane gas leaks are core safety hazards in industrial production and daily life, and accidents caused by both are characterized by their suddenness and destructive power. Smoke in the early stages of a fire is a key early warning signal, and timely detection can buy time for personnel evacuation and fire fighting; methane is a flammable and explosive gas, and leaks can easily cause explosions when the concentration reaches a certain threshold. Therefore, early coordinated detection of both is crucial to improving safety control levels.
[0003] Existing technologies have significant limitations: On the one hand, in tall spaces and long-distance scenarios, traditional point-type smoke / heat detectors have limited coverage and are prone to blind spots. While linear beam smoke detection technology can solve the coverage problem, it is limited in function, only able to detect smoke, and is susceptible to interference from displacement, vibration, ambient light, dust, etc., leading to false alarms. On the other hand, methane detection often employs catalytic combustion and tunable diode laser absorption spectroscopy (TDLAS) technology. Although TDLAS can achieve high-precision detection by utilizing the selective absorption of methane at a wavelength of 1653.7nm, in scenarios where smoke and methane are mixed, smoke scattering can interfere with the spectral signal. Furthermore, existing methane detection and smoke detection are mostly independent systems, resulting in high equipment costs, asynchronous signals, and difficulty in coordinated analysis.
[0004] With the development of deep learning technology, gated recurrent networks have shown advantages in time-series signal processing, enhancing the anti-interference capability of smoke detection through the fusion of infrared and ultraviolet dual-wavelength data. However, existing technologies are only applied to single smoke detection and have not been combined with methane detection. Furthermore, while existing dual-wavelength detection architectures can cover both types of targets, they lack suitable models for handling mixed signals, resulting in low detection accuracy and high false alarm rates in complex scenarios. Therefore, there is an urgent need for a smoke and methane composite detection method that integrates dual-wavelength features to overcome the limitations of existing technologies. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a linear beam detection method and device that integrates dual-wavelength spectral features and deep learning, so as to solve the problem of detecting the coexistence of smoke and methane in long-distance / high-altitude spaces.
[0006] According to one aspect of the present invention, a linear beam-type smoke and methane composite detection method is provided, comprising the following steps: data acquisition and preprocessing, smoke detection, methane detection, and merging judgment, wherein, Data acquisition and preprocessing: 850nm light intensity data and 1653.7nm light intensity data were obtained by sensors, and the detected dual-wavelength light intensity data were homogenized. Smoke detection: After normalizing the dual-wavelength light intensity data, the attention weights of the two channels are obtained through the weight calculation module, input into the dual-gated GRU network to calculate the hidden state of each channel, and output the smoke feature vector after fusion. Methane detection: The 1653.7nm light intensity data is filtered, anomaly corrected, and cross-correlation located to extract the effective absorption range. Methane features are extracted through a 1D convolutional neural network and channel attention mechanism. The data is input into a GRU network to capture dynamic changes and output a methane feature vector. Merging and judging: The feature vectors of smoke and methane are concatenated, and then converted into confidence scores for three scenarios: smoke, methane, and mixture, through a fully connected layer and a sigmoid activation function, to determine whether an alarm should be triggered.
[0007] According to another aspect of the present invention, a fire and methane gas leak detection device is also provided, comprising: a data acquisition and preprocessing module, a smoke detection module, a methane detection module, and a merging and judgment module. The data acquisition and preprocessing module includes a light intensity sensor that detects light intensity data at 850nm and 1653.7nm respectively, and performs uniformization processing on the detected dual-wavelength light intensity data. The smoke detection module normalizes the dual-wavelength light intensity data, obtains the attention weights of the two channels through the weight calculation module, inputs them into the dual-gated GRU network to calculate the hidden state of each channel, and outputs the smoke feature vector after fusion. Methane detection module: Filters, corrects anomalies, and locates cross-correlation in 1653.7nm light intensity data, extracts the effective absorption range, extracts methane features through a 1D convolutional neural network and channel attention mechanism, inputs the data into a GRU network to capture dynamic changes, and outputs a methane feature vector; Merging and judging module: The feature vectors of smoke and methane are concatenated and transformed into confidence scores for three scenarios: smoke, methane, and mixture, through a fully connected layer and a sigmoid activation function, to determine whether an alarm should be triggered.
[0008] The beneficial effects of this invention are: by constructing a "dual-wavelength detection architecture + dual-gated loop network", it can accurately distinguish smoke, methane and mixed scenes, improve anti-interference ability, reduce equipment cost and false alarm rate, and meet the needs of multiple danger early warning in complex environments. Attached Figure Description
[0009] Figure 1 This is a block diagram of a linear beam-type smoke methane composite detection method according to a specific embodiment of the present invention. Detailed Implementation
[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other. To better understand this invention, it will be further described below with reference to the accompanying drawings and specific embodiments.
[0011] This invention provides a fire and methane gas leak detection device comprising: a data acquisition and preprocessing module, a smoke detection module, a methane detection module, and a merging and judgment module. The data acquisition and preprocessing module includes a light intensity sensor that detects 850nm and 1653.7nm light intensity data respectively, and performs uniformization processing on the detected dual-wavelength light intensity data. The smoke detection module, after normalizing the dual-wavelength light intensity data, obtains the attention weights of the two channels through a weight calculation module, inputs them into a dual-gated GRU network to calculate the hidden state of each channel, and outputs a smoke feature vector after fusion. The methane detection module filters, corrects anomalies, and performs cross-correlation localization on the 1653.7nm light intensity data, extracts the effective absorption range, extracts methane features through a 1D convolutional neural network and a channel attention mechanism, inputs them into a GRU network to capture dynamic changes, and outputs a methane feature vector. The merging and judgment module concatenates the smoke and methane feature vectors, converts them into confidence levels for three scenarios—smoke, methane, and a mixture—through a fully connected layer and a Sigmoid activation function, and determines whether an alarm has occurred.
[0012] In another preferred embodiment of the present invention, a linear beam-type smoke methane composite detection method is provided, which belongs to a composite detection method based on dual-wavelength characteristics. This method can utilize the aforementioned detection device for detection, such as... Figure 1 As shown, it includes the following steps: Step S10, data acquisition and preprocessing; includes: Step 101: Dual-wavelength data acquisition.
[0013] In this embodiment, beams are emitted using an 850nm laser and a 1653.7nm sawtooth-wave modulated laser, and data is acquired in real time using a photodiode sensor. The raw continuous light intensity data of the 850nm channel is denoted as... The raw data of the sawtooth wave modulated light intensity of the 1653.7nm channel is denoted as follows: .
[0014] The 850nm wavelength was chosen because experiments on smoke reduction showed that 850nm laser light has a significant effect on reducing smoke light, and it is less susceptible to interference from dust and other sources. The 1653.7nm wavelength was chosen because methane gas has a strong absorption effect on 1653.7nm laser light, which can be used to determine whether methane is present in the air.
[0015] Step 102: Data normalization processing.
[0016] The dual-wavelength data obtained in step 101 is mapped to the [0, 1] interval, as shown in the following formula:
[0017] in: The raw light intensity data at 850nm or 1653.7nm obtained in step 101. This indicates the channel type, representing 850nm and 1653nm channels respectively. The light intensity template data is collected by a pre-calibrated sensor in a smoke-free, methane-free blank scene. This is the normalized data.
[0018] Step S20, smoke detection, specifically includes the following steps: Step 201: Calculation of key parameters of GRU (Gated Recurrent Unit) network.
[0019] For each time window, the input light intensity data is normalized in step 102. ,in The light intensity data at each moment after normalization in step 102 This indicates the channel type, specifically 850nm and 1653nm, where t represents the current time and is combined with the hidden state from the previous time t-1. The update gate, reset gate, candidate hidden state, and current hidden state of GRU are calculated using the following formula.
[0020] Update gate calculation: ， Reset door calculation: , Candidate hidden state calculation: , Current hidden state calculation: , In the formula: , These represent the outputs of the reset gate and update gate at time t, respectively; h t Indicates at time The candidate hidden state; Indicates that GRU is at time [time]. The output; , , , , , This represents the corresponding weight matrix. Indicates update gate The weights preceding the variables, Indicates update gate The weights preceding the variables, This indicates the deviation of the updated gate; Indicates door reset The weights preceding the variables, Indicates door reset The weights preceding the variables, Indicates the deviation of the reset door; Indicating candidate hidden state calculation The weights preceding the variables, Indicates hidden state calculation in progress and The weights preceding the variables, This indicates the deviation in the calculation of the candidate hidden state; This represents the sigmoid activation function; Represents the hyperbolic tangent tanh(); Represents element-wise product; The channel types are 850nm and 1653nm, respectively.
[0021] Step 202: Output the smoke feature vector.
[0022] The feature vector for smoke detection is obtained by summing the hidden states of the two-channel GRU network at time t. : , in, This represents the hidden state of the 850nm channel at time t. This represents the hidden state of the 1653nm channel at time t.
[0023] Step S30, methane detection, specifically includes the following steps: Step 301: Original signal filtering and anomaly correction.
[0024] In this embodiment, the original light intensity signal is modulated by the 1653.7nm sawtooth wave of the 1653nm channel at sampling time t. To process the signal, a Butterworth low-pass filter is used to remove high-frequency electronic noise, resulting in a smooth signal. .
[0025] , in, The cutoff frequency, For Butterworth filtering operation, The raw signal collected by the sensor. The output is a smooth signal, and t is the sampling time.
[0026] Step 302: Then use The criteria include outlier correction and calculation of the mean of the smoothed signal. with standard deviation Relocate and correct to meet the requirements The anomalies were identified, and the corrected signals were obtained. The formula is as follows: , ,
[0027] Values are taken for non-abnormal and abnormal conditions, respectively. Abnormal conditions refer to points whose values differ significantly from those of adjacent points due to reasons such as high-frequency electronic noise.
[0028] Where N is the number of sampling points of the smoothed signal, and outliers are replaced by linear interpolation of adjacent points to avoid sudden interference affecting subsequent processing.
[0029] Step 303: Introduce a blank sawtooth wave template that has been pre-acquired and calibrated in a methane-free scenario. By using cross-correlation analysis to locate the synchronization starting point between the real-time correction signal and the template, the effective segment containing only the rising edge of the sawtooth wave is extracted, as shown in the following formula: , in, The time offset is taken as the synchronization start point, corresponding to the maximum cross-correlation value. L is the length of the blank template. It is the segment that contains only the rising edge of the sawtooth wave. ,
[0030] Step 304: Extract the effective segment of the rising edge of the sawtooth wave after synchronization and output a single-cycle effective signal. The obtained interval data is shown below: , Step 305: The valid signal obtained in step 301... Inputting the data into a 1DCNN transforms the time-series signal into a multi-channel feature map, encoding the differences in the curve shape of methane absorption. The formula is as follows: , F1D is the encoded methane absorption characteristic in the curve, where... For 1D convolution operations, the Conv1d function in PyTorch is used.
[0031] Step 306: Subsequently, batch normalization is used to stabilize the training process and avoid oscillations caused by differences in absolute data values. Then, ReLU activation is used to suppress invalid negative features caused by global smoke offset, as shown in the following formula: , , In the formula: BN(·) is the batch normalization operation, using the BatchNorm1d function in PyTorch, and ReLU(·) is the ReLU activation function.
[0032] FBN is the methane absorption characteristic after batch normalization, and FReLU is the methane absorption characteristic after processing with the ReLU activation function.
[0033] Step 307: Extract the global statistical features of each channel using a channel attention mechanism, and strengthen the channel weights related to methane absorption, as shown in the following formula:
[0034]
[0035] In the formula: ECA(·) is the channel attention operation, and att is the channel attention weight. The methane feature vector after attention weighting, " represents matrix multiplication.
[0036] Step 308: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input the GRU network and repeat the calculation process in step 201 to obtain the feature vector of methane detection at time t. , It is by After multiple processing steps, the output result of step S30, the methane detection step, represents the final methane absorption characteristics.
[0037] It is the eigenvector of methane detection at time t.
[0038] Step S40, merging and judgment processing, includes the following steps: Step 401: Feature vector concatenation.
[0039] In this embodiment, the output of step 202 is... The output of step 308 By directly concatenating the features along the feature dimension, the original details of both types of features are preserved to the maximum extent. The formula is as follows: , Fconcat is the output of step 202. The output of step 308 The concatenated feature vector is given by the following formula: For feature concatenation, the concat function in PyTorch is used to concatenate along the feature dimensions.
[0040] Step 402: Mapping of fully connected layers.
[0041] In this embodiment, the spliced fused feature vector The input fully connected layer is mapped to the classification space of three scenarios: "smoke, methane, and mixture," as shown in the following formula: , In the formula: FC(·) represents a fully connected operation.
[0042] Step 403: Sigmoid confidence transformation.
[0043] In this embodiment, the output of the fully connected layer is transformed into a probability vector ranging from 0 to 1 using the Sigmoid activation function, yielding the existence confidence scores for three scenarios: smoke, methane, and a mixture of smoke and methane. The formulas are as follows: , FSigmoid is a probability vector processed by the Sigmoid activation function, which can correspond to the presence of smoke, methane, or both in the scene.
[0044] In this invention, by constructing a "dual-wavelength detection architecture + dual-gated loop network", the accurate identification of smoke, methane and mixed scenes is achieved, the anti-interference capability is improved, the equipment cost and false alarm rate are reduced, and the needs of multiple danger early warning in complex environments are met.
[0045] In summary, the present invention has been described in detail with reference to the accompanying drawings and specific embodiments. However, those skilled in the art will understand that this description is exemplary and the present invention is not limited to the specific embodiments described. Various modifications and changes can be made to it, as long as they do not depart from the spirit and purpose of the present invention. All such modifications and changes should fall within the protection scope of the present invention, which is defined by the appended claims.
Claims
1. A linear beam-based method for detecting combined smoke and methane, characterized in that, Includes the following steps: Data acquisition and preprocessing, smoke detection, methane detection, and data merging and analysis are among the key components. Data acquisition and preprocessing: 850nm light intensity data and 1653.7nm light intensity data were obtained by sensors, and the detected dual-wavelength light intensity data were homogenized. Smoke detection: After normalizing the dual-wavelength light intensity data, the attention weights of the two channels are obtained through the weight calculation module, input into the dual-gated GRU network to calculate the hidden state of each channel, and output the smoke feature vector after fusion. Methane detection: The 1653.7nm light intensity data is filtered, anomaly corrected, and cross-correlation located to extract the effective absorption range. Methane features are extracted through a 1D convolutional neural network and channel attention mechanism. The data is input into a GRU network to capture dynamic changes and output a methane feature vector. Merging and judging: The feature vectors of smoke and methane are concatenated, and then converted into confidence scores for three scenarios: smoke, methane, and mixture, through a fully connected layer and a sigmoid activation function, to determine whether an alarm should be triggered.
2. The linear beam-type smoke methane composite detection method as described in claim 1, characterized in that, The specific steps for data acquisition and preprocessing are as follows: Step 101: A laser beam is emitted using an 850nm laser and a 1653.7nm sawtooth-wave modulated laser, and the beam is acquired in real time using a photodiode sensor; the raw data of the continuous light intensity of the 850nm channel is recorded as X. 850 ; The raw data of the sawtooth-modulated light intensity of the 1653.7nm channel is denoted as X. 1653 , Step 102: Data normalization processing, mapping the dual-wavelength data obtained in step 101 to the [0, 1] interval, as shown in the following formula: ， in: The raw light intensity data at 850nm or 1653.7nm obtained in step 101. This indicates the channel type, representing 850nm and 1653nm channels respectively. The light intensity template data is collected by a pre-calibrated sensor in a smoke-free, methane-free blank scene. This is the normalized data.
3. The linear beam-type smoke and methane composite detection method as described in claim 1, characterized in that, Smoke detection specifically includes the following steps: Step 201: Calculate key network parameters using gated cyclic units; For each time window, the input light intensity data is normalized in step 102. ,in The light intensity data at each moment after normalization in step 102 This indicates the channel type, specifically 850nm and 1653nm, where t represents the current time and is combined with the hidden state from the previous time t-1. The update gate, reset gate, candidate hidden state, and current hidden state of GRU are calculated using the following formulas; Update gate calculation: ， Reset door calculation: ， Candidate hidden state calculation: ， Current hidden state calculation: ， In the formula: , These represent the outputs of the reset gate and the update gate at time t, respectively. Indicates at time The candidate hidden state; Indicates that GRU is at time [time]. The output; , , , , , This represents the corresponding weight matrix. Indicates update gate The weights preceding the variables, Indicates update gate The weights preceding the variables, This indicates the deviation of the updated gate; Indicates door reset The weights preceding the variables, Indicates door reset The weights preceding the variables, This indicates the deviation of the reset door; Indicating candidate hidden state calculation The weights preceding the variables, Indicates hidden state calculation in progress and The weights preceding the variables, This indicates the deviation in the calculation of the candidate hidden state; This represents the sigmoid activation function; Represents the hyperbolic tangent tanh(); Represents element-wise product; This indicates the channel type, specifically 850nm and 1653nm. Step 202: Output the smoke feature vector; The feature vector for smoke detection is obtained by summing the hidden states of the two-channel GRU network at time t. : ， in, This represents the hidden state of the 850nm channel at time t. This represents the hidden state of the 1653nm channel at time t.
4. The linear beam-type smoke methane composite detection method as described in any one of claims 1 to 3, Its features are, The methane detection procedure specifically includes the following steps: Step 301: Original signal filtering and anomaly correction; The original light intensity signal modulated at sampling time t using a 1653.7nm sawtooth wave in a 1653nm channel. To process the signal, a Butterworth low-pass filter is used to remove high-frequency electronic noise, resulting in a smooth signal. , ， in, The cutoff frequency, For Butterworth filtering operation, The raw signal collected by the sensor. The output is a smoothed signal, where t is the sampling time. Step 302: Then use The criteria include outlier correction and calculation of the mean of the smoothed signal. with standard deviation Relocate and correct to meet the requirements The anomalies were identified, and the corrected signals were obtained. The formula is as follows: ， ， ， Values are taken for non-abnormal and abnormal conditions, respectively. Abnormal conditions refer to points whose values differ significantly from those of adjacent points due to reasons such as high-frequency electronic noise. N is the number of sampling points of the smooth signal. Abnormal points are replaced by linear interpolation of adjacent points.
5. The linear beam-type smoke methane composite detection method as described in claim 4, characterized in that, Also includes: Step 303: Introduce a blank sawtooth wave template that has been pre-acquired and calibrated in a methane-free scenario. By using cross-correlation analysis to locate the synchronization starting point between the real-time correction signal and the template, the effective segment containing only the rising edge of the sawtooth wave is extracted, as shown in the following formula: , in, The time offset is taken as the synchronization start point, corresponding to the maximum cross-correlation value. L is the length of the blank template. It is the segment that contains only the rising edge of the sawtooth wave. , Step 304: Extract the effective segment of the rising edge of the sawtooth wave after synchronization and output a single-cycle effective signal. The obtained interval data is shown below: 。 6. The linear beam-type smoke methane composite detection method as described in claim 5, characterized in that, Also includes: Step 305: The valid signal obtained in step 301... Inputting the data into a 1DCNN transforms the time-series signal into a multi-channel feature map, encoding the differences in the curve shape of methane absorption. The formula is as follows: ， F1D is the encoded methane absorption characteristic in the curve, where... For 1D convolution operations, the Conv1d function in PyTorch is used. Step 306: Subsequently, batch normalization is used to stabilize the training process and avoid oscillations caused by differences in absolute data values. Then, ReLU activation is used to suppress invalid negative features caused by global smoke offset, as shown in the following formula: , , In the formula: BN(·) is the batch normalization operation, using the BatchNorm1d function in PyTorch, and ReLU(·) is the ReLU activation function. FBN is the methane absorption characteristic after batch normalization, and FReLU is the methane absorption characteristic after processing with the ReLU activation function.
7. The linear beam-type smoke and methane composite detection method as described in claim 6, characterized in that, Also includes: Step 307: Extract the global statistical features of each channel using a channel attention mechanism, and strengthen the channel weights related to methane absorption, as shown in the following formula: ， ， In the formula: ECA(·) is the channel attention operation, and att is the channel attention weight. For the attention-weighted methane feature vector, " represents matrix multiplication.
8. The linear beam-type smoke methane composite detection method as described in claim 6, characterized in that, Also includes: Step 308: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Input the GRU network and repeat the calculation process in step 201 to obtain the feature vector of methane detection at time t. , It is by After multiple processing steps, the output result of step S30, the methane detection step, represents the final methane absorption characteristics. It is the eigenvector of methane detection at time t.
9. The linear beam-type smoke methane composite detection method as described in claim 8, wherein the merging and judgment processing step includes the following steps: Step 401: Feature vector concatenation; The output of step 202 The output of step 308 By directly concatenating the features along the feature dimension, the original details of both types of features are preserved to the maximum extent. The formula is as follows: ， Fconcat is the output of step 202. The output of step 308 The concatenated feature vector This is a feature concatenation operation, specifically the `concat` function in PyTorch, which concatenates features along their respective dimensions. Step 402: Mapping of fully connected layers; The concatenated fused feature vector The input fully connected layer is mapped to the classification space of three scenarios: "smoke, methane, and mixture," as shown in the following formula: ， In the formula: FC(·) represents a fully connected operation; The merge judgment and processing steps include the following steps: Step 403: Sigmoid confidence transformation; The output of the fully connected layer is transformed into a probability vector between 0 and 1 using the Sigmoid activation function, yielding the existence confidence scores for three scenarios: smoke, methane, and a mixture of smoke and methane, respectively. The formulas are as follows: ， FSigmoid is the probability vector processed by the Sigmoid activation function, which can correspond to the presence of smoke, methane, or both in the scene.
10. A fire and methane gas leak detection device, characterized in that, include: The system includes a data acquisition and preprocessing module, a smoke detection module, a methane detection module, and a data merging and judgment module. The data acquisition and preprocessing module includes a light intensity sensor that detects light intensity data at 850nm and 1653.7nm respectively, and performs uniformization processing on the detected dual-wavelength light intensity data. The smoke detection module normalizes the dual-wavelength light intensity data, obtains the attention weights of the two channels through the weight calculation module, inputs them into the dual-gated GRU network to calculate the hidden state of each channel, and outputs the smoke feature vector after fusion. Methane detection module: Filters, corrects anomalies, and locates cross-correlation in 1653.7nm light intensity data, extracts the effective absorption range, extracts methane features through a 1D convolutional neural network and channel attention mechanism, inputs the data into a GRU network to capture dynamic changes, and outputs a methane feature vector; Merging and judging module: The feature vectors of smoke and methane are concatenated and transformed into confidence scores for three scenarios: smoke, methane, and mixture, through a fully connected layer and a sigmoid activation function, to determine whether an alarm should be triggered.