Harmful gas concentration intelligent identification and detection method

Abstract

The invention discloses an intelligent harmful gas concentration identification and detection method, and particularly relates to the field of gas concentration identification and detection.The method comprises the steps that a multi-channel gas response unit is arranged in a target area, and the gas response unit at least comprises an electrochemical channel, a spectrum channel and an ionization channel; collecting gas response values at the same moment to form a three-dimensional response tensor; meanwhile, corresponding environment tensors are collected; performing local sliding window transformation on the gas response value of each channel, calculating a first-order derivative, a second-order derivative, a response intensity variable amplitude and a recovery rate in a window, and constructing a local feature vector group; by constructing a concentration recognition closed-loop process composed of multichannel response feature extraction, concentration decoding modeling, adaptive calibration and disturbance confidence regulation, the recognition accuracy, adaptive capability and anti-interference stability of the system in a complex mixed gas environment are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of harmful gas detection, and more specifically, to a method for intelligently identifying and detecting harmful gas concentrations. Background Art

[0002] As the country's environmental protection standards and emission regulations continue to tighten, the environmental governance industry has placed higher demands on timely, high-precision, and intelligent gas concentration detection and identification methods. Traditional periodic detection and manual analysis methods are no longer able to meet the current needs for rapid identification and continuous monitoring of pollution sources, driving the evolution of related technologies towards real-time, automated, and intelligent methods. The mainstream hazardous gas detection methods currently on the market include electrochemical sensors, infrared spectroscopy, gas chromatography, and photoionization detection (PID). While these technologies can initially monitor gas concentrations under specific operating conditions, most existing technologies are based on physical / chemical sensing technologies, primarily relying on preset concentration thresholds and static rules for detection. These technologies lack the ability to collaboratively identify multi-component concentrations in complex gas mixtures. This is particularly true when faced with multiple sources of pollution, the influx of interfering gases, and fluctuations in ambient temperature and humidity. Detection accuracy, response speed, and system stability are all significantly challenged. Specifically, existing gas detection technology exposes many key problems: First, a single sensing path or processing logic based on threshold judgment is prone to misjudgment or missed judgment, and cannot accurately distinguish harmful gases with similar concentrations but different compositions; second, the system generally lacks adaptive capabilities and cannot perform modeling optimization and dynamic correction based on long-term monitoring data; third, traditional detection devices are usually unable to integrate multiple gas response characteristics and form an effective discrimination model, which limits their applicability in complex environments. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method for intelligent identification and detection of harmful gas concentrations. By constructing a concentration identification closed-loop process consisting of multi-channel response feature extraction, concentration decoding modeling, adaptive calibration and disturbance confidence control, the system's recognition accuracy, adaptability and anti-interference stability in complex mixed gas environments are improved to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligently identifying and detecting harmful gas concentrations, comprising: S1: Deploy a multi-channel gas response unit in the target area, wherein the gas response unit includes at least an electrochemical channel, a spectral channel, and an ionization channel; collect gas response values ​​at the same time , forming a three-dimensional response tensor ; At the same time, collect the corresponding environment tensor ; S2: Gas response value for each channel Perform local sliding window transformation, calculate the first-order derivative, second-order derivative, response intensity variation and recovery rate within the window, and construct a local feature vector group ; Generate multi-channel feature tensors by splicing all channels in time series ,in is the feature dimension, is the number of window center moments; S3: Multi-channel feature tensor Input into the concentration decoding network structure, build the gas concentration decoding function and output the concentration estimate ; S4: Get past periods Concentration sequences identified within the same target area , which forms a recursive consistency vector with the current valuation ; Calculate its variance, mean deviation and slope trend indicators, and input consistency analysis function , expressed as: ; definition is the preset consistency threshold, if , then call the local residual self-learning module to recalibrate The activation layer weights in ; S5: Estimating the concentration with real-time environment tensors Fusion, execution of perturbation weight function : , and output the confidence concentration value through the soft gating mechanism , ;in is the gating function; Indicates time Collected environmental parameter tensor; perturbation weight function Used to fuse concentration estimation and environmental parameters to calculate weight values; gating function Used to adjust the credibility of concentration estimates under disturbances.

[0005] In a preferred embodiment, in S1, is the channel number, is the gas type index, is the time index; is the number of sensor channels, is the total number of gas types, is the length of the sampling time series; The 3 in the figure represents temperature, humidity and air pressure.

[0006] In a preferred embodiment, in S2, the local sliding window transformation adopts a symmetrical sliding structure, and its window length is Set to an odd number, , the central time point is , the window step size is 1, is the offset of the time step; for the sequence within each window The following processes are performed in sequence: S2-1: Calculating the first derivative by central difference method With the second derivative , the specific form is:

[0007]

[0008] S2-2: Calculate the window response amplitude and recovery rate ; S2-3: Constructing a local feature vector group from the four types of derivative features , and perform zero-mean normalization and sliding median filtering on all eigenvalues ​​to reduce the impact of noise; S2-4: All the gas in each channel Concatenate to generate local feature tensor segments.

[0009] In a preferred embodiment, the gas concentration decoding function in S3 Expressed as:

[0010] The gas concentration decoding function Indicates that for The neural network structure obtained by gas-like training, Represents its parameter set and outputs concentration estimation ; Indicates from The sub-tensor selected from is used as the input for the current gas concentration estimation; in Indicates the number of channels from the 1st channel to the channels, take all sensor channels; Indicates the time of taking Total The center of the time window forms a local time series sub-segment; Indicates taking all feature dimensions, including first-order derivative, second-order derivative, response intensity variation and recovery rate.

[0011] In a preferred embodiment, in S3, the gas concentration decoding function Represents a multi-channel one-dimensional convolutional neural network with a residual structure, whose backbone network includes: Input encoding layer: performs channel fusion convolution on the input multi-dimensional feature sequence, and uses a 1D convolution layer with a kernel width of 3 to extract local temporal patterns; Residual Enhancement Module: It consists of two stacked 1D convolutional layers, each followed by batch normalization and ReLU activation function, and the output is added to the input residual through identity mapping; Channel attention module: Based on the SE mechanism, it extracts the global response of each channel through global average pooling, generates a channel weight vector, and performs a reweighting operation on the intermediate feature map; Output linear mapping layer: Use a fully connected layer to map the convolution output to the concentration estimate of the target gas The concentration prediction results.

[0012] In a preferred embodiment, in S4, is the time index; is the offset of the time step; The backtracking depth indicates the maximum number of time steps to go back to the history when performing self-checking. Indicates the starting time point to be traced back to; concentration sequence Indicates at time , looking back Time step for the All concentration estimates made by the gas class constitute a historical concentration trajectory; consistency analysis function Used to combine the statistical features of historical concentration series to quantify the degree of consistency between current concentration estimates and historical trajectories; The consistency analysis function The calculation process includes the following: Time mean shift ; Local slope trend ; Stability score ; in is the preset coefficient; Indicates the Estimated concentration of gas Over time The rate of change.

[0013] In a preferred embodiment, in S5, the disturbance weight function The build includes:

[0014] in 、 are the temperature and humidity offsets, is the perturbation weight factor, It is the reference value of temperature and humidity; The gating function is the Sigmoid function.

[0015] In a preferred embodiment, the method further includes S6: based on a preset upper limit of the concentration threshold and the lower concentration threshold Comparison, confidence concentration value Execution status segmentation: like , output normal state; like , output warning status; like , output exceeds the standard state; After each state recognition, the response characteristics, concentration output and label state of the corresponding period are written into the historical learning buffer pool to support subsequent recursive training and model weight adjustment.

[0016] In a preferred embodiment, the generated historical learning buffer pool is divided into a high-frequency sample group and a fluctuation sample group according to time periods, which are used for model short-term memory training and abnormal state playback analysis respectively.

[0017] In a preferred embodiment, for newly added gas types, independent response channels and concentration decoding networks are constructed, and supervised training is performed based on labeled concentration samples. After the trained decoding network is parameter normalized, it is connected in parallel to the existing decoding model cluster to form a concentration decoding module set that supports multiple gas types, which is used to realize the expansion of concentration recognition for newly added gas types.

[0018] Technical effects and advantages of the present invention: This invention solves the problems of misjudgment and missed judgment caused by existing systems based on single-channel and fixed threshold judgment by constructing a concentration decoding mechanism that integrates multi-channel gas responses. It converts multidimensional response sequences into a learnable structure through sliding window feature extraction and time series splicing. Combined with a deep network, it can distinguish gases with similar compositions, improving the recognition ability in complex mixed gas environments. This paper designs a recursive consistency check mechanism based on historical concentration trajectories. By looking back at the time series variation characteristics of concentration estimates, statistical indicators such as mean shift, slope trend, and response fluctuation are introduced to evaluate the stability of the current recognition results. When recognition drifts, local residual self-learning is triggered to adjust the decoding model, thereby enhancing the long-term stability and adaptability of the system. The disturbance weight function proposed in this invention dynamically calculates the confidence concentration adjustment factor by regressing the response error of ambient temperature and humidity fluctuations, thereby suppressing the impact of environmental interference on the recognition results. In particular, it maintains the credibility and continuity of the output results in high-humidity or high-temperature scenarios, meeting the concentration monitoring requirements under harsh working conditions. The concentration decoding network structure constructed by the present invention integrates a residual enhancement module and a channel attention mechanism, which can strengthen key response channels, suppress redundant channels, and enhance the perception of temporal changes during feature processing. This improves the decoding network's generalization ability for complex response patterns and the accuracy of concentration prediction. The present invention independently constructs response channels and concentration decoding networks for newly added gas types, and connects them in parallel to the original model cluster through parameter normalization to form a dynamically expandable concentration decoding module set. From the architectural level, it achieves high portability and compatibility of the system for gas type expansion, and is suitable for monitoring scenarios where pollution source types change frequently. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The figure is a flow chart of the method steps of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] Refer to the instruction manual Figure 1 According to an embodiment of the present invention, a method for intelligently identifying and detecting harmful gas concentrations includes: S1: Deploy a multi-channel gas response unit in the target area, wherein the gas response unit includes at least an electrochemical channel, a spectral channel, and an ionization channel; collect gas response values ​​at the same time , forming a three-dimensional response tensor ; At the same time, collect the corresponding environment tensor ; S2: Gas response value for each channel Perform local sliding window transformation, calculate the first-order derivative, second-order derivative, response intensity variation and recovery rate within the window, and construct a local feature vector group ; Generate multi-channel feature tensors by splicing all channels in time series ,in is the feature dimension, is the number of window center moments; S3: Multi-channel feature tensor Input into the concentration decoding network structure, build the gas concentration decoding function, and output the concentration estimate ; S4: Get past periods Concentration sequences identified within the same target area , which forms a recursive consistency vector with the current valuation ; Calculate its variance, mean deviation and slope trend indicators, and input consistency analysis function , expressed as: ; definition is the preset consistency threshold, if , then call the local residual self-learning module to recalibrate The activation layer weights in , to achieve model adaptive adjustment; S5: Estimating the concentration with real-time environment tensors Fusion, execution of perturbation weight function : , and output the confidence concentration value through the soft gating mechanism , ;in is the gating function; Indicates time Collected environmental parameter tensor; perturbation weight function Used to fuse concentration estimation and environmental parameters to calculate weight values; gating function Used to adjust the credibility of concentration estimates under disturbances.

[0022] In S1, is the channel number, is the gas type index, is the time index; is the number of sensor channels, is the total number of gas types, is the length of the sampling time series; The 3 in the figure represents temperature, humidity and air pressure.

[0023] In S2, the local sliding window transformation adopts a symmetrical sliding structure, and its window length Set to an odd number, , the central time point is , the window step size is 1, is the offset of the time step; for the sequence within each window The following processes are performed in sequence: S2-1: Calculating the first derivative by central difference method With the second derivative , the specific form is:

[0024]

[0025] S2-2: Calculate the window response amplitude and recovery rate ; S2-3: Constructing a local feature vector group from the four types of derivative features , and perform zero-mean normalization and sliding median filtering on all eigenvalues ​​to reduce the impact of noise; S2-4: All the gas in each channel Concatenate to generate local feature tensor segments.

[0026] The gas concentration decoding function described in S3 Expressed as:

[0027] The gas concentration decoding function Indicates that for The neural network structure obtained by gas-like training, Represents its parameter set and outputs concentration estimation ; Indicates from The sub-tensor selected from is used as the input for the current gas concentration estimation; in Indicates the number of channels from the 1st channel to the channels, take all sensor channels; Indicates the time of taking Total The center of the time window forms a local time series sub-segment; Indicates taking all characteristic dimensions, including first-order derivative, second-order derivative, response intensity variation and recovery rate; right For example, suppose we are currently identifying: The current moment of a gas (such as NH3) is The half-width of the sliding window is (Indicates taking 2 frames before and after) Then the tensor subset input to the network is: ; Here it means: for all channels (1 to ), only looking at the second type of gas (NH3), extracting the window features from time 48 to 52 (a total of 5 time points), each time point has D-dimensional feature input.

[0028] In S3, the gas concentration decoding function Represents a multi-channel one-dimensional convolutional neural network with a residual structure, whose backbone network includes: Input encoding layer: performs channel fusion convolution on the input multi-dimensional feature sequence, and uses a 1D convolution layer with a kernel width of 3 to extract local temporal patterns; Residual Enhancement Module: It consists of two stacked 1D convolutional layers, each followed by batch normalization and ReLU activation function, and the output is added to the input residual through identity mapping; Channel attention module: Based on the SE mechanism, it extracts the global response of each channel through global average pooling, generates a channel weight vector, and performs a reweighting operation on the intermediate feature map to achieve selective enhancement of important channel features; Output linear mapping layer: Use a fully connected layer to map the convolution output to the concentration estimate of the target gas The concentration prediction results.

[0029] In S4, is the time index; is the offset of the time step; The backtracking depth indicates the maximum number of time steps to go back to the history when performing self-checking. Indicates the starting time point to be traced back to; concentration sequence Indicates at time , looking back Time step for the All concentration estimates made by the gas class constitute a historical concentration trajectory; consistency analysis function Used to combine and score the statistical characteristics of historical concentration series (such as mean shift, slope trend, and variance fluctuation) to quantify the degree of consistency between current concentration estimates and historical trajectories; The consistency analysis function The calculation process includes the following: Time mean shift ; Local slope trend ; Stability score ; in is the preset coefficient, as the preset coefficient and The value of is set according to the sensitivity requirements of mean shift and trend change in the specific application scenario, including determining its size through parameter tuning or cross-validation method on training data, and meeting Normalization conditions to ensure consistent scoring scales; Indicates the Estimated concentration of gas Over time The rate of change.

[0030] In S5, the perturbation weight function The build includes:

[0031] in 、 are the temperature and humidity offsets, is the perturbation weight factor, is the reference value of temperature and humidity; in addition, This involves collecting the sensor concentration error under different temperature and humidity conditions, calculating the regression coefficients of the concentration error to the temperature and humidity offsets, and using them as the disturbance weight factors for the two items respectively; The gating function is the Sigmoid function.

[0032] Also includes S6: Based on a preset upper concentration threshold and the lower concentration threshold Comparison, confidence concentration value Execution status segmentation: like , output normal state; like , output warning status; like , output exceeds the standard state; After each state recognition, the response characteristics, concentration output and label state of the corresponding period are written into the historical learning buffer pool to support subsequent recursive training and model weight adjustment.

[0033] The generated historical learning buffer pool is divided into high-frequency sample groups and fluctuation sample groups according to time periods, which are used for model short-term memory training and abnormal state playback analysis respectively.

[0034] For newly added gas types, independent response channels and concentration decoding networks are constructed, and supervised training is performed based on labeled concentration samples. After parameter normalization, the trained decoding network is connected in parallel to the existing decoding model cluster to form a concentration decoding module set that supports multiple gas types, which is used to implement concentration recognition expansion for newly added gas types.

[0035] It should be noted that the development of this solution stems from an in-depth analysis of the poor adaptability, high misjudgment rate, and lack of self-learning capabilities of traditional hazardous gas detection technologies in complex environments. Based on the systematic integration of existing multi-channel sensor technology and deep model decoding mechanisms, a closed-loop intelligent recognition process covering acquisition, analysis, identification, calibration, and expansion was gradually constructed. The solution, centered around the core principle of "building a concentration decoding mechanism and disturbance adaptive capabilities," has designed a technical path based on five main steps. First, in step S1, by deploying gas response units with multiple channels, including electrochemical, spectral, and ionization, the response values ​​of different channels to various gases are collected at a unified time scale. A three-dimensional response tensor and an environmental tensor are then constructed in combination with environmental parameters (temperature, humidity, and air pressure), ensuring multi-channel, multi-dimensional, and multi-modal coverage of the data from the source. Then, in step S2, a sliding window mechanism is used to extract local features from the time series response of each channel. Specifically, these features include multi-dimensional dynamic indicators such as the first-order derivative, second-order derivative, response amplitude, and recovery rate. This design enables the system to identify continuously changing gas concentrations in the face of concentration trend changes, sudden interference, or instantaneous leaks. Subsequently, in S3, the above temporal features are structured and integrated into input tensors and passed into a specially constructed concentration decoding network. This network consists of an input encoding layer, a residual enhancement module, a channel attention mechanism, and a linear mapping layer. It has the ability to capture local response patterns, strengthen the role of key channels, and suppress redundant interference, thereby achieving nonlinear mapping between multi-channel features and concentration output; After obtaining the concentration estimate, the solution does not directly output the result. Instead, it enters the recursive consistency check in step S4. Here, by extracting statistical features (such as mean shift, slope trend, and response variance) from the historical concentration trajectory, it determines whether the current estimate has abnormal drift or trend mutation. If it deviates from the preset threshold, the local residual self-learning process is triggered, actively adjusting some parameters of the network's intermediate layer to enhance the model's adaptability to long-term changes, thereby making the entire recognition mechanism self-calibrating. Furthermore, an environmental disturbance compensation mechanism is introduced in S5. This mechanism incorporates environmental factors such as temperature and humidity into the disturbance weight function, weighting the confidence level according to their actual impact on the estimation stability. Finally, the soft gating structure outputs the corrected confidence concentration value. This mechanism improves the system's stability in extreme environments such as high humidity and drastic temperature fluctuations, and is particularly effective in scenarios where environmentally induced errors are frequent. Finally, in S6, a state label is generated through the concentration threshold judgment rule, and the corresponding identification trajectory, original features, and system output are written to the historical buffer pool to achieve continuous memory and data retention of system behavior. The buffer pool is divided into high-frequency and fluctuating groups based on data frequency and change amplitude, serving the model's short-term memory training and risk state recurrence analysis, respectively. The entire solution architecture also takes model scalability into consideration. To address the potential emergence of new hazardous gas types in the future, a pluggable decoding module expansion mechanism has been designed. This mechanism builds independent response channels for these new gases, performs supervised training based on existing samples, performs parameter normalization, and integrates them into the existing decoding network cluster, enabling upgraded concentration recognition capabilities for multiple gas types on the same platform. This design not only ensures the stability and versatility of the model structure, but also enhances the flexibility and lifecycle of the system during actual deployment. In summary, the solution forms a complete closed-loop solution covering collection, identification, evaluation, correction and update, from response construction, feature extraction, concentration decoding, self-calibration consistency judgment, disturbance weight control, result label output to model expansion mechanism. In actual deployment, it can be widely applied to various pollution source-intensive areas such as sewage treatment plants, chemical enterprises, livestock and poultry farms, and has clear engineering implementation value for meeting high-intensity emission monitoring and dynamic pollution identification tasks.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for intelligent identification and detection of harmful gas concentration, characterized in that: include: S1: Deploy a multi-channel gas response unit in the target area, wherein the gas response unit includes at least an electrochemical channel, a spectral channel, and an ionization channel; Collect gas response values ​​at the same time , forming a three-dimensional response tensor ; At the same time, collect the corresponding environment tensor ; S2: Gas response value for each channel Perform local sliding window transformation, calculate the first-order derivative, second-order derivative, response intensity variation and recovery rate within the window, and construct a local feature vector group ; Generate multi-channel feature tensors by splicing all channels in time series ,in is the feature dimension, is the number of window center moments; S3: Multi-channel feature tensor Input into the concentration decoding network structure, build the gas concentration decoding function, and output the concentration estimate ; S4: Get past periods Concentration sequences identified within the same target area , which forms a recursive consistency vector with the current valuation ; Calculate its variance, mean deviation and slope trend indicators, and input consistency analysis function , expressed as: ; definition is the preset consistency threshold, if , then call the local residual self-learning module to recalibrate The activation layer weights in ; S5: Estimating the concentration with real-time environment tensors Fusion, execution of perturbation weight function : , and output the confidence concentration value through the soft gating mechanism , ;in is the gating function; Indicates time Collected environmental parameter tensor; perturbation weight function Used to fuse concentration estimation and environmental parameters to calculate weight values; gating function Used to adjust the credibility of concentration estimates under disturbances.

2. The method for intelligent identification and detection of harmful gas concentration according to claim 1, characterized in that: In S1, is the channel number, is the gas type index, is the time index; is the number of sensor channels, is the total number of gas types, is the length of the sampling time series; The 3 in the figure represents temperature, humidity and air pressure.

3. The method for intelligent identification and detection of harmful gas concentration according to claim 2, characterized in that: In S2, the local sliding window transformation adopts a symmetrical sliding structure, and its window length Set to an odd number, , the central time point is , the window step size is 1, is the offset of the time step; for the sequence within each window The following processes are performed in sequence: S2-1: Calculating the first derivative by central difference method With the second derivative , the specific form is: ； ； S2-2: Calculate the window response amplitude and recovery rate ; S2-3: Constructing a local feature vector group from the four types of derivative features , and perform zero-mean normalization and sliding median filtering on all eigenvalues ​​to reduce the impact of noise; S2-4: All the gas in each channel Concatenate to generate local feature tensor segments.

4. The method for intelligent identification and detection of harmful gas concentration according to claim 3, characterized in that: The gas concentration decoding function described in S3 Expressed as: ； The gas concentration decoding function Indicates that for The neural network structure obtained by gas-like training, Represents its parameter set and outputs concentration estimation ; Indicates from The sub-tensor selected from is used as the input for the current gas concentration estimation; in Indicates the number of channels from the 1st channel to the channels, take all sensor channels; Indicates the time of taking Total The center of the time window forms a local time series sub-segment; Indicates taking all feature dimensions, including first-order derivative, second-order derivative, response intensity variation and recovery rate.

5. The method for intelligent identification and detection of harmful gas concentration according to claim 4, characterized in that: In S3, the gas concentration decoding function Represents a multi-channel one-dimensional convolutional neural network with a residual structure, whose backbone network includes: Input encoding layer: performs channel fusion convolution on the input multi-dimensional feature sequence, and uses a 1D convolution layer with a kernel width of 3 to extract local temporal patterns; Residual Enhancement Module: It consists of two stacked 1D convolutional layers, each followed by batch normalization and ReLU activation function, and the output is added to the input residual through identity mapping; Channel attention module: Based on the SE mechanism, it extracts the global response of each channel through global average pooling, generates a channel weight vector, and performs a reweighting operation on the intermediate feature map; Output linear mapping layer: Use a fully connected layer to map the convolution output to the concentration estimate of the target gas The concentration prediction results.

6. The method for intelligent identification and detection of harmful gas concentration according to claim 5, characterized in that: In S4, is the time index; is the offset of the time step; The backtracking depth indicates the maximum number of time steps to go back to the history when performing self-checking. Indicates the starting time point to be traced back to; concentration sequence Indicates at time , looking back Time step for the All concentration estimates made by the gas class constitute a historical concentration trajectory; consistency analysis function Used to combine the statistical features of historical concentration series to quantify the degree of consistency between current concentration estimates and historical trajectories; The consistency analysis function The calculation process includes the following: Time mean shift ; Local slope trend ; Stability score ; in is the preset coefficient; Indicates the Estimated concentration of gas Over time The rate of change.

7. The method for intelligent identification and detection of harmful gas concentration according to claim 6, characterized in that: In S5, the perturbation weight function The build includes: ； in 、 are the temperature and humidity offsets, is the perturbation weight factor, It is the reference value of temperature and humidity; The gating function is the Sigmoid function.

8. The method for intelligent identification and detection of harmful gas concentration according to claim 7, characterized in that: Also includes S6: Based on a preset upper concentration threshold and the lower concentration threshold Comparison, confidence concentration value Execution status segmentation: like , output normal state; like , output warning status; like , output exceeds the standard state; After each state recognition, the response characteristics, concentration output and label state of the corresponding period are written into the historical learning buffer pool to support subsequent recursive training and model weight adjustment.

9. The method for intelligent identification and detection of harmful gas concentration according to claim 8, characterized in that: The generated historical learning buffer pool is divided into high-frequency sample groups and fluctuation sample groups according to time periods, which are used for model short-term memory training and abnormal state playback analysis respectively.

10. A method for intelligent identification and detection of harmful gas concentration according to any one of claims 1 to 9, characterized in that: For newly added gas types, independent response channels and concentration decoding networks are constructed, and supervised training is performed based on labeled concentration samples. After parameter normalization, the trained decoding network is connected in parallel to the existing decoding model cluster to form a concentration decoding module set that supports multiple gas types, which is used to implement concentration recognition expansion for newly added gas types.

Images