A VOC odor source direction recognition system based on sensor array multi-modal fusion

Abstract

The application provides a VOC odor source direction recognition system based on sensor array multi-modal fusion, and relates to the technical field of gas detection. Specifically, the application comprises a sensor array module and a data processing module. The sensor array module collects multi-channel gas response signals and digital wind speed data under the action of a VOC odor plume, and generates a data frame comprising multi-channel digital timing response data and corresponding digital wind speed data. The data processing module calibrates the multi-channel digital timing response data for response consistency, pre-processes the signals, and constructs features to obtain a principal feature tensor. The digital wind speed data is standardized to obtain an environmental auxiliary feature vector. Through multi-scale spatial feature extraction, timing dependence modeling, environmental gating fusion, and classification output, the VOC odor source direction category is obtained.

Description

Technical Field

[0001] This invention relates to the field of gas detection technology, and in particular to a VOCs odor source direction identification system based on multimodal fusion of sensor arrays. Background Technology

[0002] Volatile organic compounds (VOCs) are widely present in petrochemical, hazardous chemical storage and transportation, industrial production, and public safety settings. VOC leaks not only cause air pollution and health hazards but also easily lead to poisoning, explosions, and other accidents. Therefore, rapid and accurate identification of VOC diffusion direction is of significant application value for leak tracing, pollution localization, robot navigation, and safety early warning.

[0003] Currently, electronic nose technology based on gas sensor arrays has become an important means of odor source direction identification due to its advantages such as small size, low cost, and ease of deployment. Existing direction identification methods mainly include methods based on wind field information inversion, response time difference of multiple gas sensors, and array data-driven models. For example, some devices utilize the response time difference of multiple gas sensors, multi-scale features, or the guiding and rotating structure of an odor compass to achieve direction indication.

[0004] However, existing methods still have significant shortcomings under complex operating conditions. There are objective differences in manufacturing processes and sensitivity between gas sensors of the same model, and baseline drift is easily generated. Without rigorous response calibration and effective signal preprocessing, hardware differences can mask the true response differences caused by the direction of odor propagation, leading to misjudgment of direction. Furthermore, the propagation process of VOCs odor plumes is highly susceptible to uncertainties such as wind speed disturbances, changes in propagation distance, and turbulent diffusion, resulting in a high degree of randomness and uncertainty in the response time, amplitude, and trends acquired by gas sensors.

[0005] Furthermore, existing methods often rely on manual features, single time differences, or single network models, making it difficult to fully extract spatial distribution differences, dynamic response changes, and long-range temporal dependencies in multi-channel response signals. They also fail to effectively incorporate environmental information such as wind speed into the model for adaptive modulation. Therefore, effectively overcoming individual differences and baseline drift in gas sensors, reducing the interference of wind speed disturbances and odor plume fluctuations on recognition results, and improving the accuracy and stability of direction recognition under complex conditions by optimizing multimodal feature fusion algorithms are pressing technical problems that need to be solved in this field. Summary of the Invention

[0006] To address the shortcomings of existing technologies in VOCs odor source direction identification, which are susceptible to individual differences in gas sensors, baseline drift, and environmental disturbances, and suffer from issues such as the lack of optimized design for gas sensor array spatial layout oriented towards gas diffusion direction features and insufficient extraction of spatiotemporal coupling features from multi-channel response signals, this invention provides a VOCs odor source direction identification system based on multimodal fusion of sensor arrays. This invention enhances the directional response differences of odors from different directions across multiple gas sensor channels by optimizing gas sensor array layout, response consistency calibration, signal preprocessing, principal feature tensor construction, introduction of environmental auxiliary features, and multimodal spatiotemporal feature fusion modeling. This reduces the impact of gas sensor differences, baseline drift, and environmental disturbances on the identification results, thereby improving the accuracy, robustness, and cross-condition adaptability of VOCs odor source direction identification.

[0007] This invention provides a VOCs odor source direction identification system based on multimodal fusion of sensor arrays, comprising: a sensor array module and a data processing module;

[0008] The sensor array module is used to acquire the multi-channel gas sensor response signal under the action of VOCs odor plume and the digital wind speed data under the current detection condition, and generate a data frame and transmit it to the data processing module; the data frame includes multi-channel digital time-series response data and corresponding digital wind speed data;

[0009] The data processing module is used to receive data frames transmitted by the sensor array module. The data processing module performs response consistency calibration, signal preprocessing and feature construction on the multi-channel digital time-series response data to obtain the main feature tensor. The data processing module performs standardization processing on the digital wind speed data to obtain the environmental auxiliary feature vector. Based on the main feature tensor and the environmental auxiliary feature vector, the module performs multimodal spatiotemporal feature fusion and directional classification processing to output the VOCs odor source directional category.

[0010] The sensor array module includes a gas sensor array, a wind speed sensor, a signal conditioning module, a data acquisition module, a main control processing module, and a communication module.

[0011] The gas sensor array is sequentially connected to the signal conditioning module and the data acquisition module. The signal conditioning module receives the analog voltage response signals output from each gas sensor channel and performs buffering, signal amplification, and low-pass filtering on the analog voltage response signals to obtain conditioned multi-channel analog voltage response signals. The data acquisition module performs analog-to-digital conversion on the conditioned multi-channel analog voltage response signals to obtain multi-channel digital response signals.

[0012] The main control processing module is connected to the data acquisition module and the wind speed sensor respectively. It is used to receive multi-channel digital response signals and digital wind speed data, and to perform data caching, time correlation and data frame encapsulation according to the preset sampling time sequence to generate a data frame including multi-channel digital time sequence response data and corresponding digital wind speed data.

[0013] The communication module is connected to the main control processing module and is used to transmit data frames including multi-channel digital time-series response data and corresponding digital wind speed data to the data processing module.

[0014] The gas sensor array includes multiple gas sensors arranged in different spatial locations to collect multi-channel gas response signals under the action of VOCs odor plumes. The wind speed sensor is set within a preset range of the detection area where the gas sensor array is located to collect wind speed information under the current detection conditions and output digital wind speed data. The digital wind speed data is input to the main control processing module, which performs time correlation and data frame encapsulation with the multi-channel digital response signals to generate a data frame including multi-channel digital time-series response data and corresponding digital wind speed data. The data frame is then transmitted to the data processing module through the communication module and, after standardization processing, is used as an environmental auxiliary feature for pattern recognition processing.

[0015] The data processing module includes a pattern recognition module and a visualization module;

[0016] The pattern recognition module is connected to the communication module and is used to receive data frames transmitted by the communication module. The pattern recognition module performs response consistency calibration, signal preprocessing and feature construction on the multi-channel digital time-series response data to obtain the main feature tensor; it performs standardization processing on the digital wind speed data to obtain the environmental auxiliary feature vector; and performs multimodal spatiotemporal feature fusion and direction classification processing based on the main feature tensor and the environmental auxiliary feature vector to output the VOCs odor source direction category.

[0017] The visualization module is connected to the communication module and the pattern recognition module respectively, and is used to receive, display, store and manage the multi-channel digital time-series response data, digital wind speed data and VOCs odor source direction category output by the pattern recognition module; the visualization module displays the multi-channel gas sensor response data and the wind speed information obtained from the digital wind speed data in real-time curve form, and displays, stores or sends the VOCs odor source direction category.

[0018] The pattern recognition module includes a response consistency calibration unit, a signal preprocessing unit, a feature construction unit, a multimodal spatiotemporal feature fusion direction recognition unit, and a classification output unit. The multimodal spatiotemporal feature fusion direction recognition unit and the classification output unit together constitute a VOCs odor source direction recognition model, which is used to output the VOCs odor source direction category based on the principal feature tensor and the environmental auxiliary feature vector.

[0019] The response consistency calibration unit is used to perform response consistency calibration on multi-channel digital time-series response data. Specifically, the response consistency calibration unit sequentially performs static calibration, response correction factor fitting, and gas sensor channel compensation processing. The static calibration uses a calibration gas with a preset concentration gradient to perform steady-state response tests on each gas sensor, obtaining the steady-state response values ​​of each gas sensor channel under different concentration conditions. The response correction factor fitting uses the average response value of the gas sensor array under the same concentration condition as a reference response to calculate the response correction factor corresponding to each gas sensor channel. Then, a polynomial fitting equation is used to establish a mapping relationship between the gas sensor response value and the response correction factor. In the actual detection stage, the gas sensor channel compensation processing determines the response correction factor corresponding to each gas sensor channel according to the mapping relationship and performs channel compensation on the digital time-series response data of each gas sensor channel to obtain calibrated multi-channel digital time-series response data.

[0020] The signal preprocessing unit is used to sequentially perform combined filtering for noise reduction, adaptive triggering and dynamic effective response segment extraction, time scale alignment and resampling, and baseline relative change rate normalization on the calibrated multi-channel digital time-series response data to obtain the normalized response sequence of each gas sensor channel. Specifically, the combined filtering uses median filtering and Savitzky-Golay smoothing. The adaptive triggering and dynamic effective response segment extraction determines the odor contact trigger time based on the initial baseline of each gas sensor channel. When the transient response value of any gas sensor channel drops below a preset ratio relative to its initial baseline value, the odor contact trigger time is determined, and the effective response segment is extracted based on the odor contact trigger time. The time scale alignment and resampling are used to unify effective response segments of different lengths to a preset time length. The baseline relative change rate normalization is used to normalize the resampled response sequence based on the baseline value of each gas sensor channel to obtain the normalized response sequence of each gas sensor channel.

[0021] The feature construction unit is used to stack the normalized response sequences of each gas sensor channel along the channel dimension to obtain a main feature tensor; wherein, the main feature tensor includes the normalized response sequences of multiple gas sensor channels within a uniform time length. The feature construction unit is also used to extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and perform standardization processing on the wind speed value to obtain standardized wind speed features, so as to construct an environmental auxiliary feature vector.

[0022] Furthermore, the multimodal spatiotemporal feature fusion direction recognition unit includes a multi-scale spatial feature extraction subunit, a temporal dependency modeling subunit, and an environment-gated fusion subunit.

[0023] The multimodal spatiotemporal feature fusion direction recognition unit takes the principal feature tensor and the environment auxiliary feature vector as input; the multi-scale spatial feature extraction subunit extracts the response differences and local dynamic response features between gas sensor channels, the temporal dependency modeling subunit extracts long-range temporal dependency features, and the environment gating fusion subunit generates environmental gating weights based on the environment auxiliary feature vectors, and adaptively modulates the temporal aggregation features and amplitude response features to obtain the fused direction features;

[0024] Specifically, the multimodal spatiotemporal feature fusion direction recognition unit takes a principal feature tensor and an environmental auxiliary feature vector as input. The principal feature tensor is denoted as:

[0025] ;

[0026] in, For the sample size, This refers to the number of gas sensor channels. The length of the unified time series after effective response segment truncation, resampling, and normalization.

[0027] The environmental auxiliary feature vector The system is constructed based on digital wind speed data. Specifically, the data processing module extracts the digital wind speed data corresponding to the current sample from the data frame, obtains the wind speed value W corresponding to the current sample based on the digital wind speed data, standardizes the wind speed value W to obtain the standardized wind speed feature Z(W), and then constructs the environmental auxiliary feature vector X. aux Represented as:

[0028] ;

[0029] Furthermore, the multi-scale spatial feature extraction subunit includes at least two parallel one-dimensional convolutional branches: a small receptive field one-dimensional convolutional branch and a large receptive field one-dimensional convolutional branch, which are used to extract short-term dynamic response features and long-term response trend features, respectively; the multi-scale spatial feature extraction process is expressed as follows:

[0030] ;

[0031] ;

[0032] in, Indicates the kernel size as The small receptive field one-dimensional convolution operation is used to extract short-time-scale dynamic response features; Indicates the kernel size as The large receptive field one-dimensional convolution operation is used to extract long-term response trend features; and These represent the convolutional features output by the small receptive field convolutional branch and the large receptive field convolutional branch, respectively. After channel-dimensional convolutional features are concatenated, batch normalized, and non-linearly activated, they are input to a channel attention enhancement unit for channel recalibration, and then processed by pooling downsampling to obtain primary spatial features. These primary spatial features are then processed by residual convolution and pooling downsampling to obtain the spatial features. ;

[0033] The temporal dependency modeling subunit outputs spatial features. The features are converted into temporal features suitable for input to a bidirectional long short-term memory (LSTM) network and then input into the LSM network for temporal modeling to obtain a bidirectional temporal feature matrix:

[0034] ;

[0035] in, This represents a dimension rearrangement operation, used to rearrange spatial features. The dimension is changed from channel-first to time-step-first to meet the input format of bidirectional long short-term memory networks; This represents a bidirectional long short-term memory network, used to extract the temporal dependency features of downsampled temporal feature sequences from both the forward and reverse time directions; This represents a bidirectional temporal feature matrix containing forward and backward temporal context information;

[0036] The environmental gating fusion subunit uses a bidirectional temporal feature matrix. Perform two-way feature aggregation; based on bidirectional temporal feature matrix Temporal aggregation features were obtained respectively. and amplitude response characteristics :

[0037] ;

[0038] ;

[0039] in, This represents temporal attention aggregation. This represents global max pooling along the time dimension.

[0040] Furthermore, the environment gating fusion subunit, based on the environment auxiliary feature vector... Generation environment gating weights:

[0041] ;

[0042] in, Generate a network for environment gating. It is the Sigmoid activation function. Environmental gating weights.

[0043] Furthermore, the environmental gating weights are used to perform element-wise modulation of the temporal aggregation features and amplitude response features respectively:

[0044] , ;

[0045] in, This indicates element-wise multiplication. The modulated temporal aggregation features, This represents the amplitude response characteristics after modulation.

[0046] Furthermore, the environment auxiliary feature vector is mapped through an environment embedding branch to obtain environment embedding features:

[0047] ;

[0048] in, This represents an environment-assisted feature extraction network, which is a multilayer perceptron network used to process environmental input branches. Mapped to context embedding features The modulated temporal aggregation features, modulated amplitude response features, and environmental embedding features are concatenated to obtain the fused directional features. :

[0049] ;

[0050] in, This indicates a feature concatenation operation, used to combine modulated temporal aggregate features. Modulated amplitude response characteristics and environmental embedding features Connect along the feature dimension to obtain the fusion direction feature. ,in ;

[0051] The classification output unit is used to output the directional category of VOCs odor sources based on the fusion directional features; specifically, the classification output unit includes a fully connected layer, a batch normalization layer, a LeakyReLU activation function layer, a Dropout layer, and a Softmax classification layer; the directional categories of VOCs odor sources include multiple azimuth categories covering the horizontal plane from 0° to 360°;

[0052] The classification process of the classification output unit is represented as follows:

[0053] ;

[0054] in, This represents the fully connected mapping process in the classification output unit, used to fuse directional features. The mapping is to the unnormalized score corresponding to each directional category; This represents the normalization exponential function, used to convert the unnormalized scores corresponding to each directional category into a probability distribution; ,in Indicates the number of input samples. The number of directional categories is used; the directional category with the highest probability is output as the VOCs odor source direction identification result.

[0055] ;

[0056] in, Indicates the first The predicted probabilities corresponding to each directional category; Represents category index in all directions Selected The category index that yields the maximum value; This defines the direction category of the VOCs odor source in the output.

[0057] The pattern recognition module includes two processes: a model training process and a direction recognition process.

[0058] In the model training process, the pattern recognition module acquires training samples under different odor source directions, wind speed conditions, and propagation distances. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. The odor source direction category serves as the supervised training label, the standardized digital wind speed data serves as an environmental auxiliary feature, and the propagation distance is used as a condition variable in the training sample construction to cover different gas diffusion conditions; it is not used as input to the environmental auxiliary feature vector. Specifically, the model training process includes the following steps:

[0059] Step T1: Acquire training samples with odor source direction labels. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. The multi-channel digital time-series response data is obtained by a gas sensor array collecting multi-channel gas response signals under different odor source directions, wind speed conditions, and propagation distances under the influence of VOCs odor plumes. The multi-channel gas response signals are processed by a signal conditioning module and converted from analog to digital by a data acquisition module to obtain multi-channel digital response signals. The digital wind speed data is synchronously output by a wind speed sensor. The multi-channel digital response signals and corresponding digital wind speed data are time-correlated and data frame encapsulated by the main control processing module to form a data frame including multi-channel digital time-series response data and corresponding digital wind speed data, which is then transmitted to the pattern recognition module.

[0060] Step T2: Perform response consistency calibration on the multi-channel digital timing response data through the response consistency calibration unit to obtain calibrated multi-channel digital timing response data;

[0061] Step T3: The calibrated multi-channel digital time-series response data is processed sequentially by the signal preprocessing unit, including combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time-series scale alignment and resampling, and baseline relative change rate normalization, to obtain the normalized response sequence of each gas sensor channel.

[0062] Step T4: Stack the normalized response sequences of each gas sensor channel along the channel dimension using the feature construction unit to construct the main feature tensor; extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and standardize the wind speed value to construct the environmental auxiliary feature vector; wherein, the propagation distance is used as a working condition variable in the model training process to construct training samples covering different gas diffusion conditions, and is not used as the input of the environmental auxiliary feature vector;

[0063] Step T5: Convert the principal feature tensor and environmental auxiliary feature vectors As input to the model, the corresponding odor source direction category is used as the supervised training label to train the VOCs odor source direction recognition model, thus obtaining the trained VOCs odor source direction recognition model.

[0064] In the direction recognition process, the pattern recognition module receives data frames collected in real time in the environment to be detected. The data frames include multi-channel digital time-series response data and corresponding digital wind speed data. The module then uses the trained VOCs odor source direction recognition model to output the VOCs odor source direction category.

[0065] The orientation recognition process includes the following steps:

[0066] Step S1: Receive the data frame transmitted by the communication module;

[0067] Step S2: Perform response consistency calibration on the multi-channel digital timing response data using the response consistency calibration unit;

[0068] Step S3: The signal preprocessing unit sequentially performs combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time scale alignment and resampling, and baseline relative change rate normalization on the calibrated multi-channel digital time-series response data to obtain the normalized response sequence of each gas sensor channel.

[0069] Step S4: Stack the normalized response sequences of each gas sensor channel along the channel dimension through the feature construction unit to construct the main feature tensor; extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and standardize the wind speed value to construct the environmental auxiliary feature vector.

[0070] Step S5: Input the main feature tensor and the environmental auxiliary feature vector into the trained VOCs odor source direction recognition model. The multimodal spatiotemporal feature fusion direction recognition unit sequentially performs multi-scale spatial feature extraction, temporal dependency modeling, and environmental gating fusion to obtain fused direction features. The classification output unit outputs the VOCs odor source direction category based on the fused direction features.

[0071] Step S6: Send the VOCs odor source direction category to the visualization module, and the visualization module displays, stores or sends the VOCs odor source direction category.

[0072] The beneficial effects of adopting the above technical solution are as follows:

[0073] This invention provides a VOCs odor source direction identification system based on multimodal fusion of sensor arrays. Compared with the prior art, this invention has the following advantages:

[0074] 1. This invention optimizes the spatial arrangement of the gas sensor array, constructing an equilateral triangular array structure and setting an optimal array spacing. This allows VOC odor plumes from different directions to form distinguishable differences in response time, response amplitude, and dynamic changes between multiple gas sensor channels. Compared to gas sensor arrays that rely on empirical arrangement, this invention enhances the physical separability of directional features, providing a clearer multi-channel response basis for subsequent odor source direction identification.

[0075] 2. This invention introduces response consistency calibration, combined filtering, adaptive effective response segment truncation, temporal scale alignment and resampling, and baseline relative change rate normalization processing before direction recognition. This reduces the impact of individual differences between gas sensors of the same model, channel output deviations, background noise, inconsistent response segment lengths, and baseline drift on the recognition results. Compared to recognition methods that directly use the original gas sensor response signals, this invention improves the consistency, comparability, and stability of multi-channel response data.

[0076] 3. This invention constructs a principal feature tensor and employs a combination of multi-scale spatial feature extraction and temporal dependency modeling to jointly extract inter-channel response differences, local dynamic responses, response time differences, and long-range temporal dependencies from multi-channel gas sensor response data. Compared to orientation recognition methods that rely on artificial features, single features, or single network structures, this invention can more fully characterize the multi-channel spatiotemporal coupling features in the VOCs odor plume diffusion process.

[0077] 4. This invention introduces digital wind speed data and uses the standardized wind speed features obtained from this data as environmental auxiliary features. Through an environmental gating fusion mechanism, it adaptively modulates the temporal aggregation features and amplitude response features, enabling the VOCs odor source direction identification model to determine direction by combining the odor plume propagation characteristics under different wind speed conditions. Compared to direction identification methods that do not consider environmental conditions, this invention improves the accuracy, robustness, and cross-condition adaptability of VOCs odor source direction identification under complex operating conditions. Attached Figure Description

[0078] Figure 1 This is a diagram illustrating the overall architecture of the VOCs odor source direction identification system according to an embodiment of the present invention.

[0079] Figure 2 This is a schematic diagram of the gas sensor array structure according to an embodiment of the present invention;

[0080] Figure 3 This is a circuit diagram of a gas sensor signal acquisition system according to an embodiment of the present invention;

[0081] Figure 4This is a hardware platform structure diagram of an embodiment of the present invention;

[0082] Figure 5 This is a comparison chart of the response curves of the gas sensor before and after calibration in an embodiment of the present invention;

[0083] Among them, (a) - response curve of gas sensor before calibration, (b) - response curve of gas sensor after calibration;

[0084] Figure 6 This is a comparison chart of the gas sensor response curves under different wind speed conditions according to an embodiment of the present invention;

[0085] Among them, (a) sensor response curve under wind speed of -0.5m / s, (b) sensor response curve under wind speed of -2m / s, and (c) sensor response curve under wind speed of -4m / s.

[0086] Figure 7 This is a structural diagram of the VOCs odor source direction recognition model according to an embodiment of the present invention;

[0087] Figure 8 This is a normalized confusion matrix diagram according to an embodiment of the present invention. Detailed Implementation

[0088] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0089] Example 1:

[0090] This invention provides a VOCs odor source direction identification system based on multimodal fusion of sensor arrays. It is used to identify the direction of VOCs odor sources based on multi-channel gas response signals under the action of VOCs odor plumes and digital wind speed data output by an anemometer. Figure 1 As shown, it includes: a sensor array module and a data processing module;

[0091] The sensor array module is used to acquire the multi-channel gas sensor response signal under the action of VOCs odor plume and the digital wind speed data under the current detection condition, and generate a data frame and transmit it to the data processing module; the data frame includes multi-channel digital time-series response data and corresponding digital wind speed data;

[0092] The data processing module is used to receive data frames transmitted by the sensor array module. The data processing module performs response consistency calibration, signal preprocessing and feature construction on the multi-channel digital time-series response data to obtain the main feature tensor. The data processing module performs standardization processing on the digital wind speed data to obtain the environmental auxiliary feature vector. Based on the main feature tensor and the environmental auxiliary feature vector, the module performs multimodal spatiotemporal feature fusion and directional classification processing to output the VOCs odor source directional category.

[0093] The sensor array module includes a gas sensor array, a wind speed sensor, a signal conditioning module, a data acquisition module, a main control processing module, and a communication module.

[0094] The gas sensor array is sequentially connected to a signal conditioning module and a data acquisition module. The signal conditioning module receives the analog voltage response signals output from each gas sensor channel and performs buffering, signal amplification, and low-pass filtering on the analog voltage response signals to obtain conditioned multi-channel analog voltage response signals. The data acquisition module performs analog-to-digital conversion on the conditioned multi-channel analog voltage response signals to obtain multi-channel digital response signals. In this embodiment, the data acquisition module uses a multi-channel A / D conversion method to synchronously sample multiple gas sensor channels.

[0095] The main control processing module is connected to both the data acquisition module and the wind speed sensor. It receives multi-channel digital response signals and digital wind speed data, and performs data buffering, time correlation, and data frame encapsulation according to a preset sampling sequence to generate a data frame containing multi-channel digital timing response data and corresponding digital wind speed data. In this embodiment, the main control processing module uses an STM32 series microcontroller, preferably an STM32F103C8T6 microcontroller; the sampling frequency of the multi-channel digital response signal is preferably 30 Hz.

[0096] The communication module is connected to the main control processing module and is used to transmit data frames, including multi-channel digital timing response data and corresponding digital wind speed data, to the data processing module. In this embodiment, the communication module adopts USART serial communication, and the preferred communication baud rate is 115200 bit / s; in other embodiments, the communication module may also adopt USB, RS485, CAN, Bluetooth, Wi-Fi, or Ethernet communication.

[0097] The gas sensor array includes multiple gas sensors arranged at different spatial locations to collect multi-channel gas response signals under the influence of VOCs odor plumes. Due to the spatial differences in the multiple gas sensors, when the VOCs odor plumes propagate to the gas sensor array from different directions, each gas sensor channel exhibits response time, response amplitude, and dynamic changes with directional differences. The wind speed sensor is positioned within a preset range of the detection area where the gas sensor array is located to collect wind speed information under the current detection conditions and output digital wind speed data. The digital wind speed data is input to the main control processing module, which performs time correlation and data frame encapsulation with the multi-channel digital response signals. The data is then transmitted to the data processing module via a communication module and, after standardization, is used as an environmental auxiliary feature in pattern recognition processing.

[0098] Preferably, such as Figure 2 As shown, the gas sensor array includes three identical metal-oxide-semiconductor (MOS) gas sensors. In this embodiment, the gas sensors are MOS gas sensors, specifically the TGS2620. The three gas sensors are designated S1, S2, and S3. They are arranged in the same detection plane, forming an equilateral triangular array structure. The distance between adjacent gas sensors is preferably 10 cm, and the array center serves as a reference center for VOCs odor source direction identification. The equilateral triangular array structure is used to obtain multi-channel response time difference, response amplitude difference, and dynamic change differences within a 360° horizontal range.

[0099] Preferably, the gas sensor employs a voltage divider signal acquisition circuit, such as... Figure 3 As shown; the voltage divider signal acquisition circuit includes a gas sensor sensitive resistor and a load resistor, which together form a voltage divider detection loop; when VOCs gas acts on the gas sensor, the sensitive resistor changes, causing a change in the output voltage across the load resistor, and the output voltage serves as the analog voltage response signal for the corresponding gas sensor channel. In this embodiment, the sensitive resistor... With load resistance The series connection forms a voltage divider network, through which the output voltage is obtained from the load resistor. Characterizes the gas response signal. Based on the partial pressure relationship, the first... The sensitive resistance of a gas sensor can be expressed as:

[0100] ;

[0101] in, To detect the operating voltage of the circuit, The gas sensor's heating circuit and detection circuit are independent of each other, which is the output voltage. The heating circuit is used to provide a stable operating temperature for the gas-sensitive material, and the detection circuit is used to collect the resistance change signal under the action of gas.

[0102] Preferably, the wind speed sensor is positioned within a preset range of the detection area where the gas sensor array is located, for real-time acquisition of wind speed information under the current detection conditions, and outputting digital wind speed data. The digital wind speed data is time-correlated and data frame-encapsulated by the main control processing module before being transmitted to the data processing module, where it is standardized and used to construct an environmental auxiliary feature vector. The digital wind speed data output by the wind speed sensor includes the wind speed value under the current detection conditions, in m / s.

[0103] Preferably, the data acquisition module uses a multi-channel A / D conversion method to synchronously sample multiple gas sensor channels.

[0104] Preferably, the main control processing module adopts an STM32 series microcontroller, and more preferably an STM32F103C8T6 microcontroller; the sampling frequency of the digital response signal of each gas sensor channel in the gas sensor array is preferably 30 Hz.

[0105] Preferably, the communication module adopts USART serial communication, and the preferred communication baud rate is 115200 bit / s.

[0106] The data processing module includes a pattern recognition module and a visualization module;

[0107] The pattern recognition module is connected to the communication module and is used to receive data frames transmitted by the communication module. The pattern recognition module performs response consistency calibration, signal preprocessing and feature construction on the multi-channel digital time-series response data to obtain the main feature tensor; it performs standardization processing on the digital wind speed data to obtain the environmental auxiliary feature vector; and performs multimodal spatiotemporal feature fusion and direction classification processing based on the main feature tensor and the environmental auxiliary feature vector to output the VOCs odor source direction category.

[0108] The visualization module is connected to the communication module and the pattern recognition module respectively, and is used to receive, display, store and manage the multi-channel digital time-series response data, digital wind speed data and VOCs odor source direction category output by the pattern recognition module; the visualization module displays the multi-channel gas sensor response data and the wind speed information obtained from the digital wind speed data in real-time curve form, and displays, stores or sends the VOCs odor source direction category.

[0109] In this embodiment, as shown Figure 4 As shown, the VOCs odor source direction identification hardware platform includes a gas sensor array, a signal conditioning module, a data acquisition module, a main control processing module, a communication module, a wind speed sensor, and a data processing module. The multi-channel analog voltage response signal output by the gas sensor array is processed by the signal conditioning module and then input to the data acquisition module, which converts it into a multi-channel digital response signal. The wind speed sensor outputs digital wind speed data. The main control processing module receives the multi-channel digital response signal and the digital wind speed data, performs time correlation, buffering, and data frame encapsulation, and then uploads them to the data processing module via the communication module.

[0110] The pattern recognition module includes a response consistency calibration unit, a signal preprocessing unit, a feature construction unit, a multimodal spatiotemporal feature fusion direction recognition unit, and a classification output unit. The multimodal spatiotemporal feature fusion direction recognition unit and the classification output unit together constitute a VOCs odor source direction recognition model. The VOCs odor source direction recognition model is used to take the main feature tensor and the environmental auxiliary feature vector as input and output the VOCs odor source direction category.

[0111] The response consistency calibration unit is used to perform response consistency calibration on multi-channel digital time-series response data. Specifically, the response consistency calibration unit sequentially performs static calibration, response correction factor fitting, and gas sensor channel compensation processing. The static calibration uses a calibration gas with a preset concentration gradient to perform steady-state response tests on each gas sensor, obtaining the steady-state response values ​​of each gas sensor channel under different concentration conditions. The response correction factor fitting uses the average response value of the gas sensor array under the same concentration condition as a reference response to calculate the response correction factor corresponding to each gas sensor channel. Then, a polynomial fitting equation is used to establish a mapping relationship between the gas sensor response value and the response correction factor. In the actual detection stage, the gas sensor channel compensation processing determines the response correction factor corresponding to each gas sensor channel according to the mapping relationship and performs channel compensation on the digital time-series response data of each gas sensor channel to obtain calibrated multi-channel digital time-series response data.

[0112] The signal preprocessing unit is used to sequentially perform combined filtering for noise reduction, adaptive triggering and dynamic effective response segment extraction, time scale alignment and resampling, and baseline relative change rate normalization on the calibrated multi-channel digital time-series response data to obtain the normalized response sequence of each gas sensor channel. Specifically, the combined filtering uses median filtering and Savitzky-Golay smoothing. The adaptive triggering and dynamic effective response segment extraction determines the odor contact trigger time based on the initial baseline of each gas sensor channel. When the transient response value of any gas sensor channel drops below a preset ratio relative to its initial baseline value, the odor contact trigger time is determined, and the effective response segment is extracted based on the odor contact trigger time. The time scale alignment and resampling are used to unify effective response segments of different lengths to a preset time length. Preferably, linear interpolation is used to resample the effective response segment to 450 sampling points. The baseline relative change rate normalization is performed on the resampled response sequence based on the baseline value of each gas sensor channel to obtain the normalized response sequence of each gas sensor channel.

[0113] The feature construction unit is used to stack the normalized response sequences of each gas sensor channel along the channel dimension to obtain a main feature tensor; wherein, the main feature tensor includes the normalized response sequences of multiple gas sensor channels within a uniform time length. In this embodiment, the number of gas sensor channels is preferably 3, and the uniform time length is preferably 450 sampling points. The feature construction unit is also used to extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, standardize the wind speed value to obtain standardized wind speed features, and construct an environmental auxiliary feature vector; during the model training phase, the propagation distance is used as a working condition variable in the construction of training samples to cover different gas diffusion conditions, and is not used as input to the environmental auxiliary feature vector.

[0114] Furthermore, the multimodal spatiotemporal feature fusion direction recognition unit includes a multi-scale spatial feature extraction subunit, a temporal dependency modeling subunit, and an environment-gated fusion subunit.

[0115] The multimodal spatiotemporal feature fusion direction recognition unit takes the principal feature tensor and the environment auxiliary feature vector as input; the multi-scale spatial feature extraction subunit extracts the response differences and local dynamic response features between gas sensor channels, the temporal dependency modeling subunit extracts long-range temporal dependency features, and the environment gating fusion subunit generates environmental gating weights based on the environment auxiliary feature vectors, and adaptively modulates the temporal aggregation features and amplitude response features to obtain the fused direction features;

[0116] Specifically, the multimodal spatiotemporal feature fusion direction recognition unit takes a principal feature tensor and an environmental auxiliary feature vector as input. The principal feature tensor is denoted as:

[0117] ;

[0118] in, For the sample size, This refers to the number of gas sensor channels. This refers to the length of the unified time series after effective response segment truncation, resampling, and normalization. Preferably, , .

[0119] The environmental auxiliary feature vector The system is constructed based on digital wind speed data. Specifically, the data processing module extracts the digital wind speed data corresponding to the current sample from the data frame, obtains the wind speed value W corresponding to the current sample based on the digital wind speed data, standardizes the wind speed value W to obtain the standardized wind speed feature Z(W), and then constructs the environmental auxiliary feature vector X. aux Represented as:

[0120] ;

[0121] Furthermore, during the model training phase, the propagation distance D is used as a condition variable in the construction of training samples to build training samples covering different gas diffusion conditions. The propagation distance D is not used as input to the environmental auxiliary feature vector.

[0122] Furthermore, the multi-scale spatial feature extraction subunit includes at least two parallel one-dimensional convolutional branches: a small receptive field one-dimensional convolutional branch and a large receptive field one-dimensional convolutional branch, which are used to extract short-term dynamic response features and long-term response trend features, respectively; the multi-scale spatial feature extraction process is expressed as follows:

[0123] ;

[0124] ;

[0125] in, Indicates the kernel size as The small receptive field one-dimensional convolution operation is used to extract short-time-scale dynamic response features; Indicates the kernel size as The large receptive field one-dimensional convolution operation is used to extract long-time scale response trend features; and These represent the convolutional features output by the small receptive field convolutional branch and the large receptive field convolutional branch, respectively. After channel-dimensional convolutional features are concatenated, batch normalized, and non-linearly activated, they are input to a channel attention enhancement unit for channel recalibration, and then processed by pooling downsampling to obtain primary spatial features. These primary spatial features are then processed by residual convolution and pooling downsampling to obtain the spatial features. In this embodiment, the kernel size of the small receptive field convolution branch is... The step size is 1, and the fill factor is 1. The kernel size of the large receptive field convolution branch is... The step size is 1, and the fill factor is 1. .

[0126] The temporal dependency modeling subunit outputs spatial features. The features are converted into temporal features suitable for input to a bidirectional long short-term memory (LSTM) network and then input into the LSM network for temporal modeling to obtain a bidirectional temporal feature matrix:

[0127] ;

[0128] in, This represents a dimension rearrangement operation, used to rearrange spatial features. The dimension is changed from channel-first to time-step-first to meet the input format of bidirectional long short-term memory networks; This represents a bidirectional long short-term memory network, used to extract the temporal dependency features of downsampled temporal feature sequences from both the forward and reverse time directions; This represents a bidirectional temporal feature matrix containing forward and backward temporal context information. Preferably, the unidirectional hidden layer dimension of the bidirectional long short-term memory network is 64, outputting a bidirectional temporal feature matrix:

[0129] ;

[0130] The environmental gating fusion subunit uses a bidirectional temporal feature matrix. Perform two-way feature aggregation; based on bidirectional temporal feature matrix Temporal aggregation features were obtained respectively. and amplitude response characteristics :

[0131] ;

[0132] ;

[0133] in, This represents temporal attention aggregation. This represents global max pooling along the time dimension.

[0134] Furthermore, the environment gating fusion subunit, based on the environment auxiliary feature vector... Generation environment gating weights:

[0135] ;

[0136] in, Generate a network for environment gating. It is the Sigmoid activation function. Environmental gating weights.

[0137] Furthermore, the environmental gating weights are used to perform element-wise modulation of the temporal aggregation features and amplitude response features respectively:

[0138] , ;

[0139] in, This indicates element-wise multiplication. The modulated temporal aggregation features, This represents the amplitude response characteristics after modulation.

[0140] Furthermore, the environment auxiliary feature vector is mapped through an environment embedding branch to obtain environment embedding features:

[0141] ;

[0142] in, This represents an environment-assisted feature extraction network, which is a multilayer perceptron network used to process environmental input branches. Mapped to context embedding features The modulated temporal aggregation features, modulated amplitude response features, and environmental embedding features are concatenated to obtain the fused directional features. :

[0143] ;

[0144] in, This indicates a feature concatenation operation, used to combine modulated temporal aggregate features. Modulated amplitude response characteristics and environmental embedding features Connect along the feature dimension to obtain the fusion direction feature. ,in ;

[0145] The classification output unit is used to output the VOCs odor source direction category based on the fusion direction features. Specifically, the classification output unit includes a fully connected layer, a batch normalization layer, a LeakyReLU activation function layer, a Dropout layer, and a Softmax classification layer. The VOCs odor source direction category includes multiple azimuth categories covering a horizontal plane from 0° to 360°. Preferably, the VOCs odor source direction category includes eight direction categories divided at 45° intervals, namely 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°.

[0146] The classification process of the classification output unit is represented as follows:

[0147] ;

[0148] in, This represents the fully connected mapping process in the classification output unit, used to fuse directional features. The mapping is to the unnormalized score corresponding to each directional category; This represents the normalization exponential function, used to convert the unnormalized scores corresponding to each directional category into a probability distribution; ,in Indicates the number of input samples. This represents the number of direction categories. Preferably, The corresponding eight directional categories are 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. The directional category with the highest probability is output as the VOCs odor source direction identification result.

[0149] ;

[0150] in, Indicates the first The predicted probabilities corresponding to each directional category; Represents category index in all directions Selected The category index that yields the maximum value; This defines the direction category of the VOCs odor source in the output.

[0151] The pattern recognition module includes two processes: a model training process and a direction recognition process.

[0152] In the model training process, the pattern recognition module acquires training samples under different odor source directions, wind speed conditions, and propagation distances. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. The odor source direction category serves as the supervised training label, the standardized digital wind speed data serves as an environmental auxiliary feature, and the propagation distance is used as a condition variable in the training sample construction to cover different gas diffusion conditions; it is not used as input to the environmental auxiliary feature vector. Specifically, the model training process includes the following steps:

[0153] Step T1: Acquire training samples with odor source direction labels. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. The multi-channel digital time-series response data is obtained by a gas sensor array collecting multi-channel gas response signals under different odor source directions, wind speed conditions, and propagation distances from VOCs odor plumes. The multi-channel gas response signals are processed by a signal conditioning module and converted from analog to digital by a data acquisition module to obtain multi-channel digital response signals. The digital wind speed data is synchronously output by a wind speed sensor. The multi-channel digital response signals and corresponding digital wind speed data are time-correlated and data frame encapsulated by the main control processing module to form a data frame including multi-channel digital time-series response data and corresponding digital wind speed data, which is then transmitted to the pattern recognition module.

[0154] Step T2: Perform response consistency calibration on the multi-channel digital timing response data through the response consistency calibration unit to obtain calibrated multi-channel digital timing response data;

[0155] Step T3: The calibrated multi-channel digital time-series response data is processed sequentially by the signal preprocessing unit, including combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time-series scale alignment and resampling, and baseline relative change rate normalization, to obtain the normalized response sequence of each gas sensor channel.

[0156] Step T4: Stack the normalized response sequences of each gas sensor channel along the channel dimension using the feature construction unit to construct the main feature tensor; extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and standardize the wind speed value to construct the environmental auxiliary feature vector; wherein, the propagation distance is used as a working condition variable in the model training process to construct training samples covering different gas diffusion conditions, and is not used as the input of the environmental auxiliary feature vector;

[0157] Step T5: Convert the principal feature tensor and environmental auxiliary feature vectors As input to the model, the corresponding odor source direction category is used as the supervised training label to train the VOCs odor source direction recognition model, resulting in the trained VOCs odor source direction recognition model.

[0158] In the direction recognition process, the pattern recognition module receives data frames collected in real time in the environment to be detected. The data frames include multi-channel digital time-series response data and corresponding digital wind speed data. The module then uses the trained VOCs odor source direction recognition model to output the VOCs odor source direction category.

[0159] The orientation recognition process includes the following steps:

[0160] Step S1: Receive the data frame transmitted by the communication module;

[0161] Step S2: Perform response consistency calibration on the multi-channel digital timing response data using the response consistency calibration unit;

[0162] Step S3: The calibrated multi-channel digital time-series response data is processed sequentially by the signal preprocessing unit through combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time-series scale alignment and resampling, and baseline relative change rate normalization to obtain the normalized response sequence of each gas sensor channel.

[0163] Step S4: Stack the normalized response sequences of each gas sensor channel along the channel dimension through the feature construction unit to construct the main feature tensor; extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and standardize the wind speed value to construct the environmental auxiliary feature vector.

[0164] Step S5: Input the main feature tensor and the environmental auxiliary feature vector into the trained VOCs odor source direction recognition model. The multimodal spatiotemporal feature fusion direction recognition unit sequentially performs multi-scale spatial feature extraction, temporal dependency modeling, and environmental gating fusion to obtain fused direction features. The classification output unit outputs the VOCs odor source direction category based on the fused direction features.

[0165] Step S6: Send the VOCs odor source direction category to the visualization module, and the visualization module displays, stores or sends the VOCs odor source direction category.

[0166] Example 2:

[0167] This embodiment describes the response signal processing and feature construction process performed by the pattern recognition module in the data processing module described in Embodiment 1. The process includes response consistency calibration, signal preprocessing, and feature construction to obtain the main feature tensor and the environment auxiliary feature vector.

[0168] Before executing the response signal processing and feature construction process, the sensor array module first acquires multi-channel gas response signals under the action of VOCs odor plume, and simultaneously acquires digital wind speed data. Specifically, the sensor array is arranged in the area to be detected or the experimental test area, and the heating circuit of each gas sensor is turned on to stabilize the output of each gas sensor channel to the baseline state. When the VOCs odor plume propagates to the area where the gas sensor array is located, each gas sensor generates a change in sensitive resistance due to contact with VOCs gas, and forms an analog voltage response signal for the corresponding gas sensor channel. The analog voltage response signal is processed by the signal conditioning module and then input into the data acquisition module, which performs analog-to-digital conversion to obtain multi-channel digital response signals; at the same time, the wind speed sensor outputs digital wind speed data. The main control processing module receives the multi-channel digital response signals and digital wind speed data, and performs data buffering, time correlation, and data frame encapsulation according to the preset sampling sequence to generate a data frame including multi-channel digital time-series response data and corresponding digital wind speed data, which is then transmitted to the data processing module through the communication module. Preferably, the sampling frequency of the digital response signal of each gas sensor channel in the gas sensor array is 30 Hz; during the model training phase, the single sampling time is preferably 3 min.

[0169] After the multi-channel digital timing response data is transmitted to the data processing module, the pattern recognition module performs the following processing steps:

[0170] Step A1: Perform response consistency calibration on the multi-channel digital timing response data. Since there may still be differences in manufacturing processes, sensitivity, and baseline between metal-oxide-semiconductor gas sensors of the same model, a static calibration experiment is needed to establish the gas sensor channel correction relationship before formally identifying the odor source direction.

[0171] In one specific implementation, the gas sensor array is placed in a closed static gas calibration platform, with ethanol gas as the calibration gas, and static response tests are performed at multiple concentration gradients to record the steady-state response values ​​of each gas sensor channel at different concentrations.

[0172] Let the concentration be Under the calibration gas conditions, the first The steady-state response value of each gas sensor is The average response value of the array is:

[0173] ;

[0174] in, The number of gas sensors is specified in this embodiment. .

[0175] Furthermore, the first The response correction factor for each gas sensor is:

[0176] ;

[0177] A polynomial is used to establish the mapping relationship between the gas sensor response value and the response correction factor:

[0178] ;

[0179] in, , , For the first The fitting coefficients corresponding to each gas sensor.

[0180] In the actual detection process, the mapping relationship is used to detect the first... The response signals of each gas sensor channel are compensated:

[0181] ;

[0182] in, This is the gas sensor channel response signal before calibration. This is the calibrated gas sensor channel response signal. Through the above response consistency calibration, individual differences between different gas sensors can be reduced, improving the consistency and comparability of multi-channel array outputs.

[0183] like Figure 5 As shown in (a), without calibration, the response curves of each gas sensor at the same concentration show significant differences; after compensation by the correction factor, as shown in (b), the multi-channel response curves tend to be consistent.

[0184] Step A2: After completing the response consistency calibration, perform signal preprocessing on the calibrated multi-channel digital timing response data. The signal preprocessing includes combined filtering for noise reduction, adaptive triggering and dynamic effective response segment truncation, timing scale alignment and resampling, and baseline relative rate of change normalization.

[0185] The combined filtering denoising is used to remove impulse noise and high-frequency disturbances from the original response signal. Specifically, median filtering is used to remove instantaneous outliers, Savitzky-Golay smoothing is used to reduce high-frequency noise, while preserving the dynamic trend of the gas sensor response curve.

[0186] The adaptive triggering and dynamic effective response segment extraction are used to extract effective response segments from the complete sampling sequence. Assume the sample before... The sampling point is the baseline interval, and the... Baseline value of each gas sensor channel for:

[0187] ;

[0188] in, For the first Baseline values ​​for each gas sensor channel This represents the number of sampling points in the baseline interval. For the first The gas sensor channel in the first The response value of each sampling point.

[0189] When any gas sensor channel satisfies When the odor contact is triggered, it is determined to be the moment when the odor is triggered. This is a relative threshold coefficient. Preferably, After determining the odor contact trigger moment, a certain length is moved forward from that trigger moment as a reference, and the effective response segment is extracted to retain the dynamic change information before and after the gas arrives.

[0190] The temporal scale alignment and resampling are used to standardize the length of the effective response segment. Since the duration and number of sampling points of the gas sensor response segment may differ under different wind speeds, propagation distances, and odor source directions, the truncated effective response segment is uniformly resampled to a preset time length to ensure consistent model input data dimensions and to preserve the dynamic trend of the original response signal during resampling. Preferably, the resampling method uses linear interpolation, and the preset time length is 450 sampling points; in other embodiments, cubic spline interpolation or other time series interpolation methods may also be used.

[0191] The baseline relative rate of change normalization is used to reduce the impact of baseline drift in the gas sensor and preserve the response amplitude gradient between channels. Specifically, the baseline relative rate of change normalization method is as follows:

[0192] ;

[0193] in, For the first The gas sensor channel in the first Normalized response at each sampling point This is the resampled response value. This is a very small constant used to avoid the denominator being zero.

[0194] After response consistency calibration and signal preprocessing, the multi-channel digital timing response data yields the normalized response sequence of each gas sensor channel.

[0195] Step A3: After normalization, stack the normalized response sequences of each gas sensor channel along the channel dimension to construct the principal feature tensor:

[0196] ;

[0197] in, This refers to the number of gas sensor channels. To ensure a uniform time length, in this embodiment, , .

[0198] Simultaneously, the corresponding digital wind speed data is extracted from the data frame, and the wind speed value of the current sample is obtained based on the digital wind speed data. Regarding wind speed values Standardize the data and construct an environment-aided feature vector:

[0199] ;

[0200] ;

[0201] in, The wind speed value is obtained from digital wind speed data. The standardized wind speed characteristics, and These are the mean and standard deviation of the wind speed values ​​in the training set, respectively.

[0202] Step A4: Convert the principal feature tensor and environmental auxiliary feature vectors As input features for the VOCs odor source direction recognition model, they are used in the subsequent model training or direction recognition stages.

[0203] Example 3: Model training process and orientation recognition process of the pattern recognition module;

[0204] This embodiment describes the model training process, model structure, and orientation recognition process in the pattern recognition module.

[0205] In the model training process, the pattern recognition module acquires training samples under different odor source directions, wind speed conditions, and propagation distances. These training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. Specifically, the odor source direction category serves as the supervised training label, the standardized digital wind speed data serves as an environmental auxiliary feature, and the propagation distance is used as a condition variable in the training sample construction to cover different gas diffusion conditions; it is not used as input to the environmental auxiliary feature vector.

[0206] In this embodiment, the gas sensor array is fixed at the center of the test area, the odor source is arranged in different directions around the array, and a stable airflow is formed by adjusting the working state of the fan.

[0207] In this embodiment, the odor source direction is set to 8 categories: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°; wind speed conditions are set to 3 types: 0.5 m / s, 2 m / s, and 4 m / s, corresponding to low, medium, and high wind speed conditions; propagation distance is set to 6 levels: 0.5 m, 1.0 m, 1.5 m, 2.0 m, 2.5 m, and 3.0 m. Each condition is collected three times, constructing a dataset containing 432 samples.

[0208] ;

[0209] Each sample group consists of three-channel time series data, with the corresponding odor source direction category used as the supervised training label.

[0210] like Figure 6As shown, the response curves of the gas sensors differ under different wind speed conditions. As shown in (a)-(c), under low wind speed conditions, the response time difference between different gas sensor channels is more obvious; under higher wind speed conditions, the amplitude difference between channels is more prominent. It can be seen that wind speed affects the response timing and amplitude response characteristics during the propagation of odor plumes. Therefore, in this embodiment, the digital wind speed data is standardized and constructed into an environmental auxiliary feature vector, which is then input into the model to participate in environmental gating fusion processing.

[0211] Specifically, the model training process includes the following steps:

[0212] Step B1: Obtain training samples with odor source direction labels. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels; wherein, the training samples are collected by the sensor array module under different odor source directions, different wind speed conditions, and different propagation distances.

[0213] Step B2: Perform response signal processing and feature construction on the training samples. Following the response signal processing and feature construction process described in Example 2, the multi-channel digital time-series response data in the training samples undergoes response consistency calibration, signal preprocessing, and feature construction sequentially to obtain the principal feature tensor. The wind speed value corresponding to the current sample is obtained based on the digital wind speed data, and the wind speed value is standardized to construct an environmental auxiliary feature vector. The propagation distance is used as a working condition variable in the model training phase to construct training samples covering different gas diffusion conditions, and is not used as input to the environmental auxiliary feature vector.

[0214] Step B3: Convert the principal feature tensor and environmental auxiliary feature vectors As input to the model, the odor source direction label is used as a supervised training label to train the VOCs odor source direction recognition model, thus obtaining the trained VOCs odor source direction recognition model.

[0215] Specifically, a VOCs odor source direction identification model is as follows: Figure 7 As shown, the VOCs odor source direction recognition model includes a multimodal spatiotemporal feature fusion direction recognition unit and a classification output unit.

[0216] The multimodal spatiotemporal feature fusion direction recognition unit includes a multi-scale spatial feature extraction subunit, a temporal dependency modeling subunit, and an environment-gated fusion subunit.

[0217] The multi-scale spatial feature extraction subunit uses at least two parallel one-dimensional convolutional branches to convolve the main feature tensor. The at least two parallel one-dimensional convolutional branches have different convolutional kernel sizes or different receptive fields. The convolutional branch with a small receptive field extracts short-term dynamic response features, while the convolutional branch with a large receptive field extracts long-term response trend features. The two convolutional features are concatenated along the channel dimension and then sequentially processed by batch normalization, nonlinear activation, channel attention enhancement, and pooling downsampling to obtain multi-scale spatial features.

[0218] The temporal dependency modeling subunit is used to extract the inter-channel response time difference features and global temporal dependency features in the dynamic response process of gas diffusion. Preferably, the temporal dependency modeling subunit includes a residual convolution unit, a downsampling unit, and a bidirectional long short-term memory network (Bi-LSTM) unit.

[0219] The environment-gated fusion subunit is used to generate environment-gated weights based on environmental auxiliary features, and to use the environment-gated weights to adaptively modulate and fuse temporal aggregation features and amplitude response features to obtain fusion direction features.

[0220] The classification output unit is used to output VOC odor source direction categories based on the fusion direction features. Specifically, the classification output unit includes a fully connected layer, a batch normalization layer, a LeakyReLU activation function layer, a Dropout layer, and a Softmax classification layer. The classification output unit outputs VOC odor source direction categories based on the fusion direction features. The VOC odor source direction categories include multiple azimuth categories covering a horizontal plane from 0° to 360°; preferably, the VOC odor source direction categories include eight direction categories divided at 45° intervals, namely 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°.

[0221] In one specific implementation, the multimodal spatiotemporal feature fusion direction recognition unit uses the principal feature tensor and environmental auxiliary feature vectors As input. The principal feature tensor is denoted as:

[0222] ;

[0223] in, For the sample size, This refers to the number of gas sensor channels. This refers to the length of the unified time series after effective response segment truncation, resampling, and normalization. Preferably, , .

[0224] The environmental auxiliary feature vector is constructed from digital wind speed data and is represented as follows:

[0225] ;

[0226] ;

[0227] in, This represents the current sample wind speed value obtained from digital wind speed data. The standardized wind speed characteristics, and These represent the mean and standard deviation of wind speed values ​​in the training set, respectively. Propagation distance. It is used only as a condition variable in the construction of training samples, and not as input to the environmental auxiliary feature vector.

[0228] Specifically, the multi-scale spatial feature extraction subunit uses two sets of one-dimensional convolutional branches with different kernel sizes to extract short-term and long-term features, as shown below:

[0229] ;

[0230] ;

[0231] in, Indicates the kernel size as Small receptive field one-dimensional convolutional branch; Indicates the kernel size as A large receptive field one-dimensional convolutional branch. Preferably, both branches have 32 output channels. , The step size is 1, and the fill factor is respectively , Therefore, the two-way convolutional features satisfy:

[0232] ;

[0233] The two convolutional features are concatenated along the channel dimension, and then batch normalization and non-linear activation are performed sequentially to obtain the concatenated multi-scale convolutional features:

[0234] ;

[0235] in, Indicates channel-dimensional splicing. Indicates batch normalization. This represents a nonlinear activation function. The concatenated features satisfy:

[0236] ;

[0237] Furthermore, the concatenated multi-scale convolutional features are input into the channel attention enhancement unit for channel recalibration to obtain channel-weighted multi-scale convolutional features:

[0238] ;

[0239] in, This represents a Squeeze-and-Excitation channel attention enhancement operation, used to generate channel weights based on the importance of different channel features, and adaptively weight the multi-scale convolutional features; the channel-weighted multi-scale convolutional features satisfy:

[0240] ;

[0241] Furthermore, the channel-weighted multi-scale convolutional features are subjected to a first one-dimensional max-pooling downsampling to obtain primary spatial features:

[0242] ;

[0243] in, This represents a one-dimensional max pooling operation, where the primary space features satisfy:

[0244] ;

[0245] Furthermore, the residual convolutional unit for Perform feature enhancement:

[0246] ;

[0247] ;

[0248] in, and Both represent one-dimensional convolution operations with a kernel size of 5, and both the input and output channels are 64.

[0249] Furthermore, a second max-pooling downsampling is performed on the residual-enhanced features to obtain spatial features:

[0250] ;

[0251] ;

[0252] Output spatial features Perform a dimensionality transformation to obtain temporal features suitable for input to a bidirectional long short-term memory network:

[0253] ;

[0254] ;

[0255] Furthermore, Temporal modeling of a bidirectional long short-term memory network:

[0256] ;

[0257] Preferably, the input dimension of the Bi-LSTM is 64, the dimension of the unidirectional hidden layer is 64, and the forward and backward hidden states are concatenated to obtain 128-dimensional temporal features:

[0258] ;

[0259] in, It is a two-way time series feature matrix.

[0260] Furthermore, a temporal attention mechanism is employed to... The key time steps in the process are weighted. Specifically, the first... Attention scores for each time step:

[0261] ;

[0262] in, Indicates that Bi-LSTM is in the first... The temporal characteristics output at each time step , , and These are trainable parameters in the temporal attention mechanism. The attention weights for each time step are:

[0263] ;

[0264] in, Represents the total number of time steps, preferably Temporal aggregated features are obtained based on attention weights:

[0265] ;

[0266] At the same time, Max pooling along the time dimension yields the amplitude response features:

[0267] ;

[0268] , ;

[0269] The environmental gating fusion subunit is based on An environment-gated weight is generated, and the time-series aggregation feature and amplitude response feature are dynamically modulated using the environment-gated weight:

[0270] ;

[0271] in, , , , These are the trainable parameters in the gated generative network. The activation function is Sigmoid. Preferably, the environment gating weights satisfy:

[0272] ;

[0273] Furthermore, element-wise multiplication is used to dynamically modulate the temporal aggregation features and amplitude response features:

[0274] , ;

[0275] in, This indicates element-wise multiplication. The modulated temporal aggregation features, This represents the amplitude response characteristics after modulation.

[0276] Furthermore, the environment-aided feature vector is used to obtain environment-embedded features through nonlinear mapping:

[0277] ;

[0278] in, , , and For the trainable parameters in the environment embedding branch, preferably, the environment embedding features satisfy:

[0279] ;

[0280] Furthermore, the modulated temporal aggregation features, modulated amplitude response features, and environmental embedding features are concatenated to obtain the fused directional features:

[0281] ;

[0282] ;

[0283] The classification process of the classification output unit is represented as follows:

[0284] ;

[0285] in, , This represents the number of direction categories. Preferably, The corresponding eight directional categories are 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. The directional category with the highest probability is output as the VOCs odor source direction identification result.

[0286] ;

[0287] in, This defines the direction category of the VOCs odor source in the output.

[0288] In the direction recognition process, the sensor array module is first placed in the environment to be detected, and the gas sensor array is positioned in the area where the VOCs odor plume to be identified may pass. When the VOCs odor plume reaches the area where the gas sensor array is located, the sensor array module acquires multi-channel gas response signals and simultaneously obtains digital wind speed data output by the wind speed sensor. After signal conditioning and data acquisition, the multi-channel gas response signals obtain multi-channel digital time-series response data, which is then time-correlated and data frame-encapsulated with the digital wind speed data by the main control processing module to form a data frame including multi-channel digital time-series response data and corresponding digital wind speed data. The data frame is transmitted to the data processing module through the communication module, and the pattern recognition module in the data processing module uses the trained VOCs odor source direction recognition model to output the VOCs odor source direction category.

[0289] Specifically, the direction recognition process includes the following steps:

[0290] Step C1: Receive data frames transmitted by the communication module, the data frames including multi-channel digital time-series response data under the action of the odor plume of VOCs to be identified and the corresponding digital wind speed data;

[0291] Step C2: Following the response signal processing and feature construction process described in Example 2, the multi-channel digital time-series response data is processed to obtain the principal feature tensor corresponding to the sample to be identified; and the wind speed value corresponding to the current sample is obtained based on the digital wind speed data, and the wind speed value is standardized to construct an environmental auxiliary feature vector;

[0292] Step C3: Input the main feature tensor and the environmental auxiliary feature vector into the trained VOCs odor source direction recognition model. After multi-scale spatial feature extraction, temporal dependency modeling, environmental gating fusion and classification output, the VOCs odor source direction category is obtained.

[0293] Step C4: Send the VOCs odor source direction category to the visualization module, which will then display, store, or send it.

[0294] It should be noted that during the model training phase, the odor source direction is a known label used to train the model to learn the mapping relationship between multi-channel response features, standardized wind speed features, and odor source direction categories; during the direction recognition phase, the odor source direction is the result to be predicted, which is predicted and output by the trained model based on the real-time collected data frames.

[0295] Example 4: Experimental Verification:

[0296] To verify the effectiveness of the VOCs odor source direction identification system provided by this invention, multiple odor source directions, multiple wind speed conditions, and multiple propagation distance conditions were set in the experimental platform. Multi-channel gas sensor response data and corresponding digital wind speed data were collected, and model training and testing were performed.

[0297] In one specific implementation, the odor source direction was set to eight categories: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°; the wind speed conditions were set to three types: 0.5 m / s, 2 m / s, and 4 m / s; and the propagation distance was set to six levels: 0.5 m, 1.0 m, 1.5 m, 2.0 m, 2.5 m, and 3.0 m. Each set of conditions was collected three times to construct a multi-condition odor direction dataset containing 432 samples.

[0298] During the experiment, the gas sensor array was first statically calibrated and its response consistency was checked. Then, the acquired multi-channel digital time-series response data underwent combined filtering for noise reduction, adaptive triggering and dynamic effective response segment truncation, time-scale alignment and resampling, and baseline relative change rate normalization to construct the principal feature tensor. Simultaneously, the wind speed value of the current sample was obtained from the corresponding digital wind speed data, and this wind speed value was standardized to construct an environmental auxiliary feature vector. Finally, the dataset was divided into a training set and a test set. The training set was used to train the VOCs odor source direction recognition model, and the test set was used to test its recognition performance.

[0299] like Figure 8 As shown, the normalized confusion matrix is ​​used to illustrate the correspondence between different true direction categories and the model-predicted direction categories. Higher values ​​on the main diagonal of the matrix indicate higher accuracy in identifying the corresponding direction category.

[0300] Furthermore, the VOCs odor source direction recognition model described in this invention is compared with methods such as support vector machines, convolutional neural networks, Bi-LSTM, and CNN-LSTM. As shown in Table 1, in this embodiment, the direction recognition accuracy of the VOCs odor source direction recognition model described in this invention reaches 92.31%, and the average angle error is 7.6°.

[0301] Table 1. Comparison of orientation recognition performance of different models: SVM Artificially constructed time series statistical features 75.38% 27.0° Bi-LSTM Bidirectional temporal dependency modeling 74.15% 23.68° CNN Local temporal convolution feature extraction 72.31% 22.43° CNN-LSTM Fusion of convolutional feature extraction and temporal modeling 86.23% 11.8° This invention model Multi-scale spatial feature extraction, temporal dependency modeling and environment gating fusion 92.31% 7.6°

[0302] Experimental results show that the VOCs odor source direction identification system of the present invention has good identification accuracy, robustness and environmental adaptability under different wind speeds, different propagation distances and different directions.

[0303] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0304] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0305] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of the methods disclosed herein and their equivalents, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A VOCs odor source direction identification system based on multimodal fusion of sensor arrays, characterized in that, include: Sensor array module and data processing module; The sensor array module is used to acquire the multi-channel gas sensor response signal under the action of VOCs odor plume and the digital wind speed data under the current detection condition, and generate a data frame and transmit it to the data processing module; the data frame includes multi-channel digital time-series response data and corresponding digital wind speed data; The data processing module is used to receive data frames transmitted by the sensor array module. The data processing module performs response consistency calibration, signal preprocessing and feature construction on the multi-channel digital time-series response data to obtain the main feature tensor. The data processing module performs standardization processing on the digital wind speed data to obtain the environmental auxiliary feature vector. Based on the main feature tensor and the environmental auxiliary feature vector, the module performs multimodal spatiotemporal feature fusion and directional classification processing to output the VOCs odor source directional category. The sensor array module includes a gas sensor array, a wind speed sensor, a signal conditioning module, a data acquisition module, a main control processing module, and a communication module. The gas sensor array is sequentially connected to the signal conditioning module and the data acquisition module. The signal conditioning module receives the analog voltage response signals output from each gas sensor channel and performs buffering, signal amplification, and low-pass filtering on the analog voltage response signals to obtain conditioned multi-channel analog voltage response signals. The data acquisition module performs analog-to-digital conversion on the conditioned multi-channel analog voltage response signals to obtain multi-channel digital response signals. The main control processing module is connected to the data acquisition module and the wind speed sensor respectively. It is used to receive multi-channel digital response signals and digital wind speed data, and to perform data caching, time correlation and data frame encapsulation according to the preset sampling time sequence to generate a data frame including multi-channel digital time sequence response data and corresponding digital wind speed data. The communication module is connected to the main control processing module and is used to transmit data frames including multi-channel digital time-series response data and corresponding digital wind speed data to the data processing module. The gas sensor array includes multiple gas sensors arranged in different spatial locations to collect multi-channel gas response signals under the action of VOCs odor plumes. The wind speed sensor is set within a preset range of the detection area where the gas sensor array is located to collect wind speed information under the current detection conditions and output digital wind speed data. The digital wind speed data is input to the main control processing module, which performs time correlation and data frame encapsulation with the multi-channel digital response signals to generate a data frame including multi-channel digital time-series response data and corresponding digital wind speed data. This data frame is then transmitted to the data processing module through the communication module and, after standardization processing, is used as an environmental auxiliary feature for pattern recognition processing. The data processing module includes a pattern recognition module and a visualization module; The pattern recognition module is connected to the communication module and is used to receive data frames transmitted by the communication module. The pattern recognition module performs response consistency calibration, signal preprocessing and feature construction on the multi-channel digital time-series response data to obtain a main feature tensor; it performs standardization processing on the digital wind speed data to obtain an environmental auxiliary feature vector; and it performs multimodal spatiotemporal feature fusion and directional classification processing based on the main feature tensor and the environmental auxiliary feature vector to output the VOCs odor source directional category. The visualization module is connected to the communication module and the pattern recognition module respectively, and is used to receive, display, store and manage the multi-channel digital time-series response data, digital wind speed data and VOCs odor source direction category output by the pattern recognition module; the visualization module displays the multi-channel gas sensor response data and the wind speed information obtained from the digital wind speed data in real-time curve form, and displays, stores or sends the VOCs odor source direction category; The pattern recognition module includes a response consistency calibration unit, a signal preprocessing unit, a feature construction unit, a multimodal spatiotemporal feature fusion direction recognition unit, and a classification output unit; wherein, the multimodal spatiotemporal feature fusion direction recognition unit and the classification output unit together constitute a VOCs odor source direction recognition model, which is used to output the VOCs odor source direction category based on the main feature tensor and the environmental auxiliary feature vector; The response consistency calibration unit is used to perform response consistency calibration on multi-channel digital time-series response data. Specifically, the response consistency calibration unit sequentially performs static calibration, response correction factor fitting, and gas sensor channel compensation processing. The static calibration uses a calibration gas with a preset concentration gradient to perform steady-state response tests on each gas sensor, obtaining the steady-state response values ​​of each gas sensor channel under different concentration conditions. The response correction factor fitting uses the average response value of the gas sensor array under the same concentration condition as a reference response to calculate the response correction factor corresponding to each gas sensor channel. Then, a polynomial fitting equation is used to establish a mapping relationship between the gas sensor response value and the response correction factor. In the actual detection stage, the gas sensor channel compensation processing determines the response correction factor corresponding to each gas sensor channel according to the mapping relationship and performs channel compensation on the digital time-series response data of each gas sensor channel to obtain calibrated multi-channel digital time-series response data. The signal preprocessing unit is used to sequentially perform combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time scale alignment and resampling, and baseline relative change rate normalization on the calibrated multi-channel digital time-series response data to obtain the normalized response sequence of each gas sensor channel. Specifically, the combined filtering uses median filtering and Savitzky-Golay smoothing. The adaptive triggering and dynamic effective response segment extraction determines the odor contact trigger time based on the initial baseline of each gas sensor channel. When the transient response value of any gas sensor channel drops below a preset ratio relative to its initial baseline value, the odor contact trigger time is determined, and the effective response segment is extracted based on the odor contact trigger time. The time scale alignment and resampling are used to unify effective response segments of different lengths to a preset time length. The baseline relative change rate normalization is used to normalize the resampled response sequence based on the baseline value of each gas sensor channel to obtain the normalized response sequence of each gas sensor channel. The feature construction unit is used to stack the normalized response sequences of each gas sensor channel along the channel dimension to obtain the main feature tensor; wherein, the main feature tensor includes the normalized response sequences of multiple gas sensor channels within a uniform time length; the feature construction unit is also used to extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and perform standardization processing on the wind speed value to obtain standardized wind speed features in order to construct an environmental auxiliary feature vector; The multimodal spatiotemporal feature fusion direction recognition unit includes a multi-scale spatial feature extraction subunit, a temporal dependency modeling subunit, and an environment-gated fusion subunit; The multimodal spatiotemporal feature fusion direction recognition unit takes the principal feature tensor and the environment auxiliary feature vector as input; the multi-scale spatial feature extraction subunit extracts the response differences and local dynamic response features between gas sensor channels, the temporal dependency modeling subunit extracts long-range temporal dependency features, and the environment gating fusion subunit generates environmental gating weights based on the environment auxiliary feature vectors, and adaptively modulates the temporal aggregation features and amplitude response features to obtain the fused direction features; Specifically, the multimodal spatiotemporal feature fusion direction recognition unit takes a principal feature tensor and an environmental auxiliary feature vector as input; wherein, the principal feature tensor is denoted as: ； in, For the sample size, This refers to the number of gas sensor channels. The length of the unified time series after effective response segment truncation, resampling, and normalization; The environmental auxiliary feature vector The system is constructed based on digital wind speed data; specifically, the data processing module extracts the digital wind speed data corresponding to the current sample from the data frame, and obtains the wind speed value corresponding to the current sample based on the digital wind speed data. W Regarding wind speed values W Standardization processing is performed to obtain standardized wind speed characteristics. Z ( W Then the environmental auxiliary feature vector X aux Represented as: ； The multi-scale spatial feature extraction subunit includes at least two parallel one-dimensional convolutional branches: a small receptive field one-dimensional convolutional branch and a large receptive field one-dimensional convolutional branch, which are used to extract short-term dynamic response features and long-term response trend features, respectively; the multi-scale spatial feature extraction process is represented as follows: ； ； in, Indicates the kernel size as The small receptive field one-dimensional convolution operation is used to extract short-time-scale dynamic response features; Indicates the kernel size as The large receptive field one-dimensional convolution operation is used to extract long-time scale response trend features; and These represent the convolutional features output by the small receptive field one-dimensional convolutional branch and the large receptive field one-dimensional convolutional branch, respectively. After channel-dimensional convolutional features are concatenated, batch normalized, and non-linearly activated, they are input to a channel attention enhancement unit for channel recalibration, and then subjected to pooling downsampling to obtain primary spatial features. These primary spatial features are then subjected to residual convolution and pooling downsampling to obtain the spatial features. ; The temporal dependency modeling subunit will incorporate spatial features The features are converted into temporal features suitable for input to a bidirectional long short-term memory (LSTM) network and then input into the LSM network for temporal modeling to obtain a bidirectional temporal feature matrix: ； in, This represents a dimension rearrangement operation, used to rearrange spatial features. The dimension is changed from channel-first to time-step-first to meet the input format of bidirectional long short-term memory networks; This represents a bidirectional long short-term memory network, used to extract the temporal dependency features of downsampled temporal feature sequences from both the forward and reverse time directions; This represents a bidirectional temporal feature matrix containing forward and backward temporal context information; The environmental gating fusion subunit uses a bidirectional temporal feature matrix. Perform two-way feature aggregation; based on bidirectional temporal feature matrix Temporal aggregation features were obtained respectively. and amplitude response characteristics : ； ； in, This represents temporal attention aggregation. This represents global max pooling along the time dimension; The environmental gating fusion subunit is based on the environmental auxiliary feature vector. Generation environment gating weights: ； in, Generate a network for environment gating. It is the Sigmoid activation function. Environmental gating weights; The environmental gating weights are used to perform element-wise modulation of the temporal aggregation features and amplitude response features respectively: , ； in, This indicates element-wise multiplication. The modulated temporal aggregation features, The amplitude response characteristics after modulation; The environment auxiliary feature vector is mapped through the environment embedding branch to obtain the environment embedding feature: ； in, This represents an environment-assisted feature extraction network, which is a multilayer perceptron network used to process environmental input branches. Mapped to context embedding features The modulated temporal aggregation features, modulated amplitude response features, and environmental embedding features are concatenated to obtain the fused directional features. : ； in, This indicates a feature concatenation operation, used to combine modulated temporal aggregate features. Modulated amplitude response characteristics and environmental embedding features Connect along the feature dimension to obtain the fusion direction feature. ,in ; The classification output unit is used to output the directional category of VOCs odor sources based on the fusion directional features; specifically, the classification output unit includes a fully connected layer, a batch normalization layer, a LeakyReLU activation function layer, a Dropout layer, and a Softmax classification layer; the directional categories of VOCs odor sources include multiple azimuth categories covering the horizontal plane from 0° to 360°; The classification process of the classification output unit is represented as follows: ； in, This represents the fully connected mapping process in the classification output unit, used to fuse directional features. The mapping is to the unnormalized score corresponding to each directional category; This represents the normalization exponential function, used to convert the unnormalized scores corresponding to each directional category into a probability distribution; ,in Indicates the number of input samples. The number of directional categories is used; the directional category with the highest probability is output as the VOCs odor source direction identification result. ； in, Indicates the first The predicted probabilities corresponding to each directional category; Represents category index in all directions Selected The category index that yields the maximum value; The output VOCs odor source direction category; The pattern recognition module includes two processes: a model training process and a direction recognition process. In the model training process, the pattern recognition module acquires training samples under different odor source directions, different wind speed conditions, and different propagation distances. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. The odor source direction category serves as the supervised training label, the standardized digital wind speed data serves as the environmental auxiliary feature vector, and the propagation distance serves as the operating condition variable in the training sample construction, covering different gas diffusion conditions, but is not used as input to the environmental auxiliary feature vector. Specifically, the model training process includes the following steps: Step T1: Acquire training samples with odor source direction labels. The training samples include multi-channel digital time-series response data, corresponding digital wind speed data, and odor source direction labels. The multi-channel digital time-series response data is obtained by a gas sensor array collecting multi-channel gas response signals under different odor source directions, wind speed conditions, and propagation distances under the influence of VOCs odor plumes. The multi-channel gas response signals are processed by a signal conditioning module and converted from analog to digital by a data acquisition module to obtain multi-channel digital response signals. The digital wind speed data is synchronously output by a wind speed sensor. The multi-channel digital response signals and corresponding digital wind speed data are time-correlated and data frame encapsulated by the main control processing module to form a data frame including multi-channel digital time-series response data and corresponding digital wind speed data, which is then transmitted to the pattern recognition module. Step T2: Perform response consistency calibration on the multi-channel digital timing response data through the response consistency calibration unit to obtain calibrated multi-channel digital timing response data; Step T3: The calibrated multi-channel digital time-series response data is processed sequentially by the signal preprocessing unit, including combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time-series scale alignment and resampling, and baseline relative change rate normalization, to obtain the normalized response sequence of each gas sensor channel. Step T4: Stack the normalized response sequences of each gas sensor channel along the channel dimension using the feature construction unit to construct the main feature tensor; extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and standardize the wind speed value to construct the environmental auxiliary feature vector; wherein, the propagation distance is used as a working condition variable in the model training process to construct training samples covering different gas diffusion conditions, and is not used as the input of the environmental auxiliary feature vector; Step T5: Convert the principal feature tensor and environmental auxiliary feature vectors As input to the model, the corresponding odor source direction category is used as the supervised training label to train the VOCs odor source direction recognition model, and the trained VOCs odor source direction recognition model is obtained. In the direction recognition process, the pattern recognition module receives data frames collected in real time in the environment to be detected. The data frames include multi-channel digital time-series response data and corresponding digital wind speed data, and outputs the VOCs odor source direction category using the trained VOCs odor source direction recognition model. The orientation recognition process includes the following steps: Step S1: Receive the data frame transmitted by the communication module; Step S2: Perform response consistency calibration on the multi-channel digital timing response data using the response consistency calibration unit; Step S3: The calibrated multi-channel digital time-series response data is processed sequentially by the signal preprocessing unit through combined filtering and noise reduction, adaptive triggering and dynamic effective response segment extraction, time-series scale alignment and resampling, and baseline relative change rate normalization to obtain the normalized response sequence of each gas sensor channel. Step S4: Stack the normalized response sequences of each gas sensor channel along the channel dimension through the feature construction unit to construct the main feature tensor; extract the corresponding digital wind speed data from the data frame, obtain the wind speed value corresponding to the current sample based on the digital wind speed data, and standardize the wind speed value to construct the environmental auxiliary feature vector. Step S5: Input the main feature tensor and the environmental auxiliary feature vector into the trained VOCs odor source direction recognition model. The multimodal spatiotemporal feature fusion direction recognition unit sequentially performs multi-scale spatial feature extraction, temporal dependency modeling, and environmental gating fusion to obtain fused direction features. The classification output unit outputs the VOCs odor source direction category based on the fused direction features. Step S6: Send the VOCs odor source direction category to the visualization module, and the visualization module displays, stores or sends the VOCs odor source direction category.

Images