A gas identification method, system and device based on humanoid robot olfaction and a storage medium

Abstract

The application relates to a gas identification method, system and device based on human-shaped robot olfaction and a storage medium, and relates to the technical field of human-shaped robots. The gas identification method comprises the following steps: monitoring the gas in the robot operating environment according to the gas suction frequency, collecting and cleaning the gas component concentration; analyzing the gas component concentration according to a transient response window combined with a fluid dynamics model to extract a gas transient feature sequence; identifying and classifying the gas transient feature sequence according to a preset multi-modal semantic model to reason an entity reasoning scene; comparing the amplitude difference and the phase difference of the gas transient feature sequence of each cavity according to the entity reasoning scene to update and position a target gas source; analyzing the gas concentration based on fluid dynamics and a transient feature extraction algorithm, combining a multi-modal semantic model to identify a scene, and accurately positioning the gas source by comparing the amplitude difference and the phase difference of each cavity, which significantly improves the real-time performance and traceability accuracy of robot environment perception.

Description

Technical Field

[0001] This application relates to the field of humanoid robot technology, and in particular to a gas identification method, system, device and storage medium based on the olfactory sense of a humanoid robot. Background Technology

[0002] Humanoid robots, as the core physical carriers of embodied intelligence, possess a key advantage in deeply integrating into human-designed social environments through their anthropomorphic body structure—bipedal locomotion and dexterous hand manipulation. Within this closed loop of perception, cognition, decision-making, and action, multimodal fusion perception forms the foundation for environmental understanding. Olfactory perception, as a key dimension simulating human senses, can capture volatile chemical information (such as toxic gases, fire smoke, food odors, and human metabolic odors), providing humanoid robots with decision-making support that transcends vision and hearing for safety assessments and refined services in complex scenarios.

[0003] Current cutting-edge humanoid robot technologies mainly revolve around three dimensions: perception and decision-making (multimodal fusion, autonomous navigation), motion control (high dynamic balance, dexterous manipulation), and energy and hardware (high energy density drive, lightweight joints). In industrial scenarios, they can undertake tasks such as flexible material handling in automobile manufacturing, precision electronic assembly, and photovoltaic panel installation; in service scenarios, they can be deployed in commercial performances, logistics sorting, hotel guidance, rehabilitation training, and family companionship; in special operations, they can replace humans in high-risk environments such as nuclear power plant inspections, disaster rubble search and rescue, and space station maintenance.

[0004] A multi-scene fusion perception system and method for humanoid robot olfactory sensors, relating to the field of intelligent sensor technology, is disclosed. The system includes: an olfactory perception optimization module for obtaining an olfactory perception optimization network; a first matrix module for monitoring gas pipelines to obtain a first matrix; a second matrix module for interference compensation based on the real-time environment; a third matrix module for correction based on sensor status; a leak detection module for detecting pipeline leaks; and a suppression verification module for false alarm suppression verification of the pipeline leak detection results to obtain a pipeline early warning signal. This application addresses the technical problem in existing technologies where high false alarm rates due to susceptibility to environmental interference lead to limited accuracy in gas leak detection. By adjusting the multi-node olfactory perception network in real-time using historical data, performing multi-level corrections on the monitoring data, and introducing false alarm suppression verification, the accuracy of gas leak detection in various scenarios is improved.

[0005] The existing technical solutions mentioned above have the following drawbacks: 1. Most existing artificial olfactory systems (electronic noses) are independent external modules and have not yet achieved deep integration with the humanoid robot body in a miniaturized and biomimetic manner. They are difficult to actively sniff using the natural airflow structure of the head like the human nasal cavity, resulting in low sampling efficiency and directional positioning ability; 2. Even if gas information is captured, its interpretation is still limited to the level of a single threshold alarm and lacks the semantic understanding ability of a large visual model. It is difficult to establish an accurate correlation between gas concentration maps and object states (such as the stage of food spoilage), human physiological information (such as wound infection), or environmental events (such as early smoldering). Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a gas identification method, system, device, and storage medium based on the olfactory sense of a humanoid robot. By combining the amplitude and phase differences of the transient feature sequences of the dual-cavity system with fluid dynamics compensation and particle filtering algorithms, and fusing the semantic generation of the entity scene to generate the posterior probability distribution, high-precision positioning and dynamic updating of the target gas source are achieved.

[0007] The objective of this application is achieved through the following technical solution: In a first aspect, this application provides a gas recognition method based on the olfactory sense of a humanoid robot, comprising: Monitor the gas in the robot's operating environment based on the gas pumping frequency, and collect and clean the gas composition concentration; Based on the transient response window and fluid dynamics model, the gas component concentration is analyzed, and the transient characteristic sequence of the gas is extracted. Based on a pre-defined multimodal semantic model, the transient feature sequences of gases are identified and classified to infer entity reasoning scenarios. Based on the entity reasoning scenario, the amplitude difference and phase difference of the transient gas characteristic sequence of each cavity are compared to update and locate the target gas source.

[0008] By adopting the above technical solution, gas concentration is analyzed based on fluid dynamics and transient feature extraction algorithms, and scene is identified by combining multimodal semantic models. By comparing the amplitude difference and phase difference of each cavity, the gas source is accurately located, which significantly improves the real-time performance and source tracing accuracy of the robot's environmental perception.

[0009] This application further includes: monitoring the gas in the robot's operating environment based on the gas extraction frequency, and collecting and cleaning the gas component concentration, including: The robot's olfactory detection is graded and set according to the dynamic level of the environment to determine the gas suction frequency. The gas in the robot's operating environment is detected in a dual-channel manner based on the gas pumping frequency, and the original component concentration of each channel is collected. Based on the sniffing timestamp, the original component concentration is filtered to remove noise and duplicate values ​​to obtain the standard component concentration of each cavity. Statistical analysis of the standard component concentrations based on the suction cycle was performed to calculate the average peak-to-peak value, and the gas concentration signal-to-noise ratio was obtained by combining the noise variance during the gas-free period. If the gas concentration signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, the current standard component concentration is determined to be an effective value, and the standard component concentrations of each cavity are aggregated and correlated to obtain the gas component concentration.

[0010] By adopting the above technical solution, the suction frequency is adaptively set based on the dynamic level of the environment. Through dual-channel synchronous acquisition and Kalman filtering for noise reduction, combined with the signal-to-noise ratio threshold to verify the effective concentration, highly reliable and low-noise data acquisition for robot olfactory perception is achieved, significantly improving the accuracy and anti-interference capability of gas detection.

[0011] This application further specifies: analyzing gas component concentration based on a transient response window combined with a fluid dynamics model, and extracting a transient feature sequence of the gas, including: The robot's operating status is monitored according to the sniffing cycle, the robot's motion parameters are collected, and the relative airflow speed is calculated in combination with the ambient wind speed and direction. Based on the robot's head structure, fluid simulation fitting is performed on the relative airflow velocity and robot motion parameters to generate a fluid dynamics model; Based on the fluid dynamics model, the gas component concentration is stripped to obtain the effective component concentration, and the spurious response concentration increment is calculated. If the ratio of the false response concentration increment to the peak concentration of the gas component is greater than the preset percentage threshold, the current effective component concentration is marked as a strong motion interference bit. No, then the effective component concentration is corrected point by point according to the transient response window, and the component correction compensation concentration is calculated. By performing feature extraction and segmented polymerization approximation on the component correction compensation concentration, a gas time-series feature vector is obtained; The gas temporal feature vector and robot motion parameters are fused and spliced ​​to generate a gas transient feature sequence.

[0012] By adopting the above technical solution, motion disturbances are corrected based on the fluid dynamics model, component concentrations are compensated point by point using the transient response window, and transient gas feature sequences are generated by fusing robot motion parameters. This significantly improves the anti-interference capability and feature fidelity of gas perception in complex motion environments.

[0013] This application further specifies: fluid simulation fitting of relative airflow velocity and robot motion parameters based on the robot head structure to generate a fluid dynamics model, including: The robot's head structure is analyzed, the outer shell and nasal cavity are extracted, and combined with the 3D map, a head component model and a nasal cavity component model are generated. A watershed is constructed on the head component model to generate a hemispherical virtual wind tunnel, and the nasal cavity component model is simplified and smoothed to obtain a simplified nasal cavity model. The head component model is prism meshed to divide the near-wall response area, and the simplified nasal cavity model is filled with spheres to obtain the encrypted spherical area. The near-wall response region and the dense spherical region were detected and analyzed, and the mesh tilt and mesh orthogonality were calculated. If both are greater than the corresponding quality threshold, then the current near-wall response region and the current encrypted sphere region are deemed compliant. The hemispherical virtual wind tunnel is parameterized based on the relative airflow velocity to obtain the airflow dynamic wind tunnel; The near-wall response region is simulated and quantized based on the robot's motion parameters to obtain the near-wall motion space. The dynamic wind tunnel is driven by preset simulation variable parameters, and turbulence simulation is performed on the densified spherical region in combination with the near-wall motion space to collect gas convection mass transfer parameters and generate an initial dynamic model. The gas convection mass transfer parameters are cleaned and solved serially to calculate the flow field data at the detection point. Based on the static reference state, the flow field data at the detection point is subjected to reference difference calculation to extract the airflow disturbance value and pressure distortion value; If both are within the corresponding tolerance threshold range, then the initial dynamic model is lightweighted and pruned to obtain the fluid dynamic model; If not, the corresponding residual value is calculated and fused to generate a linear error regression correction term to correct the initial dynamic model, thus obtaining the fluid dynamic model.

[0014] By adopting the above technical solution, based on computational fluid dynamics (CFD) and mesh quality assessment algorithms, and through geometric modeling, prism / sphere mesh generation and turbulence simulation, combined with benchmark difference and linear regression correction, a fluid dynamics model adapted to the robot head structure and motion parameters was constructed, which significantly improved the accuracy and real-time performance of gas diffusion prediction in complex airflow environments.

[0015] This application is further configured to: identify and classify transient gas feature sequences based on a preset multimodal semantic model, and reason about entity reasoning scenarios, including: The transient feature sequences of gases are parsed and transformed to extract odor embedding vectors and map the semantics of chemical components. The robot's motion parameters are vectorized and multi-layered, and then concatenated with the odor embedding vector to generate an environmental odor semantic vector. Based on a pre-defined multimodal semantic model, the environmental odor semantic vector is cross-identified to determine the odor component entity objects; Based on the scene odor component weights, odor component entity objects are classified to generate scene prior vectors; Based on the knowledge graph, entity objects of odor components are retrieved and associated to obtain entity event nodes; The scenario prior vectors are bound to and decoded with entity event nodes to generate an entity reasoning scenario.

[0016] By adopting the above technical solution, based on the multi-layer projection fusion of odor embedding vector and motion parameters, and utilizing multimodal cross recognition and knowledge graph retrieval, semantic mapping and reasoning of gas components to physical scenes are realized, which significantly improves the accuracy of robot environmental understanding and scene adaptability.

[0017] This application is further configured as: a multimodal semantic model, including: The robot's environmental parameters are analyzed based on the odor transient window, and visual image frames, microphone audio segments, and tactile / force readings are extracted. Lightweight encoding is performed on visual image frames to extract global object scene feature vectors; Perform temporal convolution on the microphone audio clips to extract the audio embedding feature vectors; Based on the transient feature sequence of gas, the global object scene feature vector and the audio embedding feature vector are gated and weighted and fused to generate a multimodal temporal feature vector; Entity semantic encoding is performed on the multimodal temporal feature vectors to obtain the multimodal entity weight matrix; The multimodal entity weight matrix is ​​bound to the odor type code based on the cross-attention weight matrix to form a semantic ternary weight matrix; Perform Monte Carlo simulation on the semantic ternary weight matrix, output the scene reasoning statement, and calculate the posterior distribution variance; If the posterior distribution variance is less than the preset event entity threshold, the current semantic ternary weight matrix is ​​determined to be a multimodal semantic model and stored in the real-time environment memory.

[0018] By adopting the above technical solutions, and through lightweight visual coding, audio temporal convolution and gated weighted fusion of gas temporal features, combined with cross attention and Monte Carlo simulation to construct semantic triples, the real-time performance, semantic relevance and scene reasoning robustness of the robot's multimodal environment perception are improved.

[0019] This application is further configured to: compare the amplitude difference and phase difference of the gas transient characteristic sequences of each cavity according to the entity reasoning scenario, update and locate the target gas source, including: Transient window truncation and hydrodynamic compensation are performed on the gas transient characteristic sequences of each cavity to obtain the corresponding independent transient compensation sequences; Time alignment is performed on the independent transient compensation sequences, the peak amplitude of the corresponding cavity is calculated, and normalized to the amplitude difference; Peak time detection is performed on independent transient compensation sequences, the time difference of characteristic inflection points is calculated, and converted into phase difference; Based on the cavity spacing, the azimuth calculation of the characteristic inflection point time difference and the effective diffusion velocity is performed to obtain the time difference incident angle. The amplitude difference is mapped according to the preset dual-cavity directivity sensitivity curve, and the amplitude ratio azimuth angle is calculated. The relative gas source azimuth is obtained by weighting the incident angle of the time difference and the azimuth of the amplitude ratio. The entity reasoning scenario is analyzed to obtain the entity scenario semantics and entity boundary coordinates; Based on the physical properties of gas diffusion, the coordinates of the entity boundary are diffused to generate a Gaussian prior probability distribution map. The relative gas source azimuth angle and the Gaussian prior probability distribution map are projected and fused to obtain the posterior gas source probability density distribution map. Particle resampling is performed on the posterior gas source probability density distribution map to calculate the particle cloud convergence variance; If the particle cloud convergence variance is less than the preset particle deviation threshold, then the entity corresponding to the current particle cloud is determined to be the target gas source based on the semantics of the entity scene.

[0020] By adopting the above technical solution, based on the amplitude difference and phase difference of the transient feature sequence of the dual-cavity system, combined with fluid dynamics compensation and particle filtering algorithms, and fused with the semantic generation posterior probability distribution of the entity scene, high-precision positioning and dynamic updating of the target gas source were achieved, significantly improving the environmental detection and source tracing capabilities of the robot's olfactory system.

[0021] Secondly, this application also provides a gas recognition system based on the olfactory sense of a humanoid robot, employing the following technical solution: A gas recognition system based on the olfactory sense of a humanoid robot, comprising the following gas recognition methods: The gas analysis module is used to monitor the gas in the robot's operating environment based on the gas pumping frequency, and to collect and clean the gas composition concentration. The feature extraction module is used to analyze the gas component concentration based on the transient response window and the fluid dynamics model, and extract the transient feature sequence of the gas. The multi-modal reasoning module is used to identify and classify transient gas feature sequences based on a preset multi-modal semantic model, and to reason about entity reasoning scenarios. The target localization module is used to compare the amplitude difference and phase difference of the transient gas characteristic sequence of each cavity according to the entity reasoning scenario, and update and locate the target gas source.

[0022] By adopting the above technical solutions, based on adaptive suction and Kalman filtering to clean gas concentration, using fluid dynamics compensation to extract transient features, using weighted fusion of multimodal semantic models and Monte Carlo reasoning to identify the scene, and combining dual-cavity amplitude difference, phase difference and particle filtering to locate the gas source, the robot's environmental perception accuracy, anti-interference ability and source tracing real-time performance are significantly improved.

[0023] Thirdly, this application also provides an electronic device, comprising: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement any of the methods in the above scheme.

[0024] Fourthly, this application also provides a storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the gas recognition method based on the humanoid robot's sense of smell as described above.

[0025] In summary, the beneficial technical effects of this application are as follows: Based on adaptive suction and Kalman filtering to clean gas concentration, transient features are extracted using fluid dynamics compensation. Scene recognition is achieved through weighted fusion of multimodal semantic models and Monte Carlo reasoning. Furthermore, by combining the amplitude difference, phase difference, and particle filtering of the dual-cavity system to locate the gas source, the robot's environmental perception accuracy, anti-interference ability, and real-time source tracing are improved. Based on the amplitude and phase differences of the transient feature sequences of the dual-cavity system, combined with fluid dynamics compensation and particle filtering algorithms, and fused with the semantic generation posterior probability distribution of the entity scene, high-precision positioning and dynamic updating of the target gas source were achieved, improving the environmental detection and source tracing capabilities of the robot's olfactory system. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the gas identification method in this application; Figure 2 This is a flowchart illustrating step B in this application; Figure 3 This is a schematic diagram of the operation flow of the multimodal semantic model in this application; Figure 4 This is a schematic diagram of the gas identification system in this application. Detailed Implementation

[0027] The present application will be further described in detail below with reference to the accompanying drawings.

[0028] Firstly, referring to Figure 1 This application discloses a gas recognition method based on the olfactory sense of a humanoid robot, comprising: S1: Monitor the gas in the robot's operating environment based on the gas pumping frequency, and collect and clean the gas composition concentration; S2: Analyze the gas component concentration based on the transient response window and the fluid dynamics model, and extract the gas transient characteristic sequence; S3: Identify and classify transient gas feature sequences based on a preset multimodal semantic model, and infer entity reasoning scenarios; S4: Based on the entity reasoning scenario, compare the amplitude difference and phase difference of the gas transient feature sequence of each cavity, update and locate the target gas source.

[0029] In this embodiment, during a hazardous gas leak inspection task at a chemical plant, a humanoid robot actively extracts ambient gas at a variable suction frequency. It collects the concentrations of hydrogen, carbon monoxide, and volatile organic compounds through a built-in electrochemical and photoionization sensor array. After wavelet denoising and moving average filtering, a clean concentration sequence is obtained. Subsequently, based on a structured transient response window (sampling window 50ms, sliding step 10ms) coupled with a fluid dynamics model—including establishing turbulence and convection diffusion equations for leak diffusion—the concentration sequence is analyzed for time-space characteristics, and the transient gas feature sequence composed of rise time, peak overshoot rate, and fluctuation variance is extracted.

[0030] Next, the robot invokes a pre-trained multimodal semantic model (integrating attention mechanism and graph neural network) to perform cross-modal joint encoding and classification of feature sequences, inferring that the current entity scenario is "minor hydrogen leak at high-pressure flange interface". Simultaneously, based on the peak-valley amplitude differences (maximum difference ≥15%) and arrival phase delays (maximum delay difference 80ms) of the same gas transient feature sequences captured by sensor arrays in four independent sampling cavities (left and right nasal cavities and neck), the robot uses the bow tie orientation algorithm to calculate the gas diffusion gradient vector, updates the probability density distribution of the leak source after Bayesian filtering, and finally accurately locates the gas source coordinates (X:12.3m, Y:5.8m, Z:1.2m) on a 3D digital map. The robot dynamically adjusts its head orientation and guides the sniffing probe carried by the robotic arm to approach the leak point, achieving centimeter-level gas source tracing. All data is transmitted back to the central control room via 5G and a 3D dynamic gas cloud map is generated.

[0031] Preferably, step S1 includes: The robot's olfactory detection is graded and set according to the dynamic level of the environment to determine the gas suction frequency. The gas in the robot's operating environment is detected in a dual-channel manner based on the gas pumping frequency, and the original component concentration of each channel is collected. Based on the sniffing timestamp, the original component concentration is filtered to remove noise and duplicate values ​​to obtain the standard component concentration of each cavity. Statistical analysis of the standard component concentrations based on the suction cycle was performed to calculate the average peak-to-peak value, and the gas concentration signal-to-noise ratio was obtained by combining the noise variance during the gas-free period. If the gas concentration signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, the current standard component concentration is determined to be an effective value, and the standard component concentrations of each cavity are aggregated and correlated to obtain the gas component concentration.

[0032] In this embodiment: During a hazardous gas leak inspection task in a petrochemical plant area, the humanoid robot determines the gas suction frequency in real time based on the dynamic environmental level (e.g., normal inspection is set to "medium risk", with a gas suction frequency of 2Hz; when the ambient wind speed changes suddenly or nearby equipment alarms, it automatically switches to the "high risk" level, increasing the suction frequency to 20Hz). Then, it drives the dual-cavity (left and right nasal cavities are sampled independently) to perform active inhalation detection at this frequency. The system collects the original component concentration (e.g., methane, hydrogen sulfide, carbon monoxide) of each cavity through a built-in electrochemical sensor array. The system performs high-pass filtering on the original data based on the microsecond-level sniffing timestamp to suppress baseline drift, and combines a sliding window to remove duplicate sampling points to obtain the standard component concentration of each cavity.

[0033] Then, statistical analysis is performed on the standard concentration sequence using a fixed suction cycle (e.g., every 50ms under high risk). The peak-to-peak concentration within a single cycle is calculated (e.g., 0.8%VOL for the left cavity and 0.75%VOL for the right cavity), and the noise variance during the gas-free period (suction interval) is extracted (e.g., 0.004%VOL²). This yields the gas concentration signal-to-noise ratio (SNR = peak-to-peak² / noise variance, calculated as SNR = 160 for the left cavity and 140 for the right cavity). When the SNR of both cavities is greater than a preset threshold (e.g., 100), the current standard component concentration is determined to be valid. The data from the left and right cavities are then aligned by timestamps and weighted to aggregate a single-channel gas component concentration (final methane concentration 0.77%VOL). This is then combined with the previously extracted transient feature sequence to locate the leak source. Throughout the process, if the SNR of any cavity falls below the threshold, a self-calibration command is triggered (adjusting the suction frequency or cleaning the sensor) to ensure high reliability of the gas data.

[0034] Preferably, step S2 includes: A: Monitor the robot's operating status according to the sniffing cycle, collect the robot's motion parameters, and calculate the relative airflow speed by combining the ambient wind speed and direction; B: Based on the robot's head structure, fluid simulation fitting is performed on the relative airflow velocity and robot motion parameters to generate a fluid dynamics model; C: Based on the fluid dynamics model, the gas component concentration is stripped to obtain the effective component concentration, and the spurious response concentration increment is calculated; D: If the ratio of the false response concentration increment to the peak concentration of the gas component is greater than the preset percentage threshold, then mark the current effective component concentration as a strong motion interference bit. No, then the effective component concentration is corrected point by point according to the transient response window, and the component correction compensation concentration is calculated. E: Feature extraction and segmented polymerization approximation are performed on the component correction compensation concentration to obtain the gas time series feature vector; F: The gas time-series feature vector is fused and spliced ​​with the robot motion parameters to generate a gas transient feature sequence.

[0035] The implementation principle of this embodiment is as follows: During an inspection of a chemical plant area, a humanoid robot monitors its own IMU and joint encoders at a 0.5-second sniffing cycle to obtain motion parameters such as body speed and acceleration. Simultaneously, it integrates environmental anemometer data to calculate the composite airflow velocity relative to the robot's head. Based on the three-dimensional structure of the robot's head (including the geometric model of the nasal air inlet), CFD fluid simulation is performed to fit the aforementioned composite velocity and motion parameters, constructing a personalized fluid dynamics model. This model removes the pseudo-airflow entrainment component caused by the robot's walking and arm swinging from the original gas composition concentration, obtaining an effective... The composition concentration is calculated, and the spurious response concentration increment is determined. When the ratio of the spurious response increment to the measured peak value exceeds 30%, the gas data at that moment is marked as a strong motion interference position, and odor assimilation is paused. Otherwise, the concentration is corrected point by point using the Hanning window transient response window to obtain the composition correction compensation concentration. Finally, the compensation concentration sequence is segmented and aggregated (window width 0.2s) to extract morphological features, which are then concatenated with motion parameters to generate a gas transient feature sequence (such as the correlation vector between the steepness of the concentration rise and the fuselage acceleration), providing high-purity olfactory features after motion decoupling for subsequent scene inference and gas source localization.

[0036] Preferably, refer to Figure 2 Step B includes: The robot's head structure is analyzed, the outer shell and nasal cavity are extracted, and combined with the 3D map, a head component model and a nasal cavity component model are generated. A watershed is constructed on the head component model to generate a hemispherical virtual wind tunnel, and the nasal cavity component model is simplified and smoothed to obtain a simplified nasal cavity model. The head component model is prism meshed to divide the near-wall response area, and the simplified nasal cavity model is filled with spheres to obtain the encrypted spherical area. The near-wall response region and the dense spherical region were detected and analyzed, and the mesh tilt and mesh orthogonality were calculated. If both are greater than the corresponding quality threshold, then the current near-wall response region and the current encrypted sphere region are deemed compliant. The hemispherical virtual wind tunnel is parameterized based on the relative airflow velocity to obtain the airflow dynamic wind tunnel; The near-wall response region is simulated and quantized based on the robot's motion parameters to obtain the near-wall motion space. The dynamic wind tunnel is driven by preset simulation variable parameters, and turbulence simulation is performed on the densified spherical region in combination with the near-wall motion space to collect gas convection mass transfer parameters and generate an initial dynamic model. The gas convection mass transfer parameters are cleaned and solved serially to calculate the flow field data at the detection point. Based on the static reference state, the flow field data at the detection point is subjected to reference difference calculation to extract the airflow disturbance value and pressure distortion value; If both are within the corresponding tolerance threshold range, then the initial dynamic model is lightweighted and pruned to obtain the fluid dynamic model; If not, the corresponding residual value is calculated and fused to generate a linear error regression correction term to correct the initial dynamic model, thus obtaining the fluid dynamic model.

[0037] In this embodiment, in the application of hazardous gas inspection in a chemical plant area, the robot's head structure is first finely analyzed to extract the 3D CAD model of the head shell and nasal cavity, constructing head component models and nasal cavity component models. Then, a hemispherical virtual wind tunnel with a radius of 500mm is constructed around the head component as the computational flow domain. The nasal cavity model is simplified and smoothed to obtain a simplified nasal cavity model. The head component model is meshed using prism layers, with the mesh near the wall surface being densified to form a near-wall response zone. Simultaneously, a locally densified spherical region is generated using the spherical filling method for the simplified nasal cavity model. Mesh quality is ensured by calculating the mesh tilt (≥0.65) and orthogonality (≥0.85). Next, the virtual wind tunnel boundary conditions are parameterized based on the measured relative airflow velocity (0.5-2.5m / s) during robot movement to obtain an airflow dynamic wind tunnel. The robot's motion parameters (body speed, head deflection angle) are then mapped to the near-wall response zone to achieve motion space. The simulation was performed using a dynamic wind tunnel driven by preset simulation variables such as turbulence intensity (5%) and gas diffusion coefficient (1.6e-5m² / s). Large eddy simulation turbulence was conducted on the dense spherical region in conjunction with the near-wall motion space. Gas convection mass transfer parameters (convective heat transfer coefficient, component mass fraction gradient) at the nasal cavity inlet were collected to generate an initial fluid dynamics model. The flow field data (velocity vector, turbulent kinetic energy) obtained by iterative solution at the detection point were cleaned, and the flow field data at the same location under the static reference state (robot stationary, no wind) was subtracted to extract the airflow disturbance value (Δv≤0.1m / s) and pressure distortion value (ΔP≤2Pa). Both were within the tolerance threshold. Therefore, the initial model was lightweighted and pruned (retaining 90% energy) based on singular value decomposition. Finally, a fluid dynamics model that is usable in real time and has a 40% reduction in computational cost was obtained. This model is used to remove pseudo-gas interference introduced by the robot's movement during walking, significantly improving the accuracy of olfactory localization.

[0038] Preferably, step 3 includes: The transient feature sequences of gases are parsed and transformed to extract odor embedding vectors and map the semantics of chemical components. The robot's motion parameters are vectorized and multi-layered, and then concatenated with the odor embedding vector to generate an environmental odor semantic vector. Based on a pre-defined multimodal semantic model, the environmental odor semantic vector is cross-identified to determine the odor component entity objects; Based on the scene odor component weights, odor component entity objects are classified to generate scene prior vectors; Based on the knowledge graph, entity objects of odor components are retrieved and associated to obtain entity event nodes; The scenario prior vectors are bound to and decoded with entity event nodes to generate an entity reasoning scenario.

[0039] In this embodiment: During a leak inspection at a petrochemical plant, a humanoid robot maps the transient gas feature sequence (such as the methane rise steepness of 0.8% VOL / s and the hydrogen sulfide peak delay of 0.25s) into a 512-dimensional odor embedding vector through a nonlinear projection layer. This vector is then concatenated with a 128-dimensional vector obtained by multi-layer fully connected projection of motion parameters such as the head's three-axis angular velocity and body acceleration collected by the IMU, forming an environmental odor semantic vector that integrates motion and odor association. A pre-trained multimodal semantic model (based on Transformer cross-attention) performs cross-modal joint encoding on this vector, cross-identifying the odor component entities as "trace benzene series compounds" and "trace amounts". The system identifies "hydrogen sulfide" as the hazard level. It then weights and classifies entities based on the odor composition weights (benzene series 0.7, hydrogen sulfide 0.2, methane 0.1) to generate a prior vector indicating the hazard level [leakage probability = 0.85, toxicity = high, flammability = medium]. Simultaneously, the system uses a knowledge graph to retrieve entities associated with "benzene series," obtaining event nodes such as "tank valve corrosion" and "catalyst deactivation." Finally, the prior vector is bound to the event nodes for decoding, generating an entity reasoning scenario: "The failure of the A-12 flange seal in the tank area has caused a minor leak of benzene series compounds. Immediate repair and evacuation of the downwind area are recommended." This drives the robot to complete accurate alarms and path planning.

[0040] Preferably, refer to Figure 3 Multimodal semantic models, including: The robot's environmental parameters are analyzed based on the odor transient window, and visual image frames, microphone audio segments, and tactile / force readings are extracted. Lightweight encoding is performed on visual image frames to extract global object scene feature vectors; Perform temporal convolution on the microphone audio clips to extract the audio embedding feature vectors; Based on the transient feature sequence of gas, the global object scene feature vector and the audio embedding feature vector are gated and weighted and fused to generate a multimodal temporal feature vector; Entity semantic encoding is performed on the multimodal temporal feature vectors to obtain the multimodal entity weight matrix; The multimodal entity weight matrix is ​​bound to the odor type code based on the cross-attention weight matrix to form a semantic ternary weight matrix; Perform Monte Carlo simulation on the semantic ternary weight matrix, output the scene reasoning statement, and calculate the posterior distribution variance; If the posterior distribution variance is less than the preset event entity threshold, the current semantic ternary weight matrix is ​​determined to be a multimodal semantic model and stored in the real-time environment memory.

[0041] In this embodiment: the extracted feature tensor [number of channels × number of sub-intervals × 3] is subjected to layer normalization and position encoding to preserve the temporal order of pulse extraction.

[0042] The feature sequence is transformed into a fixed-dimensional odor embedding vector Eolf using a pre-trained 1D Transformer odor encoder or a frequency band attention convolutional network. During the training phase, the encoder has learned to map specific transient patterns (such as rapid adsorption slope + high overshoot-to-fallback ratio) to specific chemical composition semantics (such as "alcohols" or "sulfides").

[0043] Auxiliary parameters such as temperature, pH value, and motion state are projected through a multilayer perceptron and then concatenated with Eolf to form an odor semantic vector with environmental constraints.

[0044] Centered on the start time t0 of the odor transient window, visual image frames, microphone audio segments, and tactile / force readings within 100ms before and after this time are extracted.

[0045] The image is input into a lightweight vision Transformer or convolutional network with frozen parameters to extract the feature vectors Evision_regions of the global scene feature map and candidate boxes of salient object regions.

[0046] Ambient sound clips are compressed into audio embeddings (Eaudio) by a Mel spectrogram feature extractor and temporal convolution, which are used to capture accompanying events such as hissing sounds from leaks and crackling sounds from burning.

[0047] If a modal sensor is occluded or malfunctions, the corresponding embedded bit is filled with a learnable "modal vacancy" vector, and a missing mask is set. Using the odor vector Eolf as the query and visual region features as the key and value, the algorithm is computed through a multi-head cross-attention layer. The model will spontaneously increase the weight of regions strongly associated with the odor's chemical components—for example, when smelling a burnt odor, visual attention will focus on the circuit board or socket area.

[0048] The cross-attention output is gated and weighted together with the visual and audio embeddings to generate the multimodal fusion feature Ffused. The gating coefficients are dynamically generated by the confidence estimates of each modality encoder to prevent a single low-quality modality from contaminating the overall representation.

[0049] Visual global features are used by a classifier to determine the current macro scene (such as "hospital ward", "car production line paint shop", "kitchen"), and generate scene prior vectors.

[0050] Using preliminary odor component labels as search keys, related entities and events are retrieved from a pre-constructed "odor-object-scenario" knowledge graph. For example, "isoprene" is associated with "abnormal human metabolism / wound," and "hydrogen sulfide" is associated with "organic decay" or "industrial waste gas."

[0051] The entity nodes obtained from knowledge retrieval are embedded as additional nodes and input into the graph attention network to pass messages with multimodal fusion features, thus unifying data-driven perception and knowledge-driven reasoning.

[0052] A classification and localization head is applied to the candidate bounding boxes of the visual region to identify the outline of the target entity (such as "infusion bag interface", "pipe flange", "food packaging").

[0053] Based on the cross-attention weight matrix, the determined odor category (such as "ethanol volatiles") is directly bound to the corresponding entity as an attribute, forming a semantic triple ⟨Entity: pipe flange, attribute: emits methanol gas, intensity: 0.87>.

[0054] The binding relationships, sound events (if any), and scene priors are concatenated and input into a lightweight large language model decoder to generate short scene reasoning statements, such as: "A strong volatile organic compound leak was detected at the flange of the ventilation duct in the northeast corner of the paint shop. It is recommended to stop the machine and check the gasket." For control scenarios that do not require natural language output, structured event codes and coordinates are directly output.

[0055] Apply Monte Carlo Dropout or deep ensemble to the output binding relationships and event categories, and calculate the variance of the posterior distribution. If the variance exceeds a threshold, it is marked as "low-confidence inference".

[0056] For low-confidence results, the robot automatically performs verification actions—slightly adjusting its head posture to obtain a better sniffing angle, moving closer to the target object for a closer re-inspection, or turning on a spectral supplement light of a specific band for visual verification.

[0057] If proactive verification increases confidence, the confirmed scene, entity, and odor associations are stored in the real-time environment memory for continuous reasoning in the following seconds or minutes, reducing repeated false alarms and achieving persistent scene understanding.

[0058] Preferably, step S4 includes: Transient window truncation and hydrodynamic compensation are performed on the gas transient characteristic sequences of each cavity to obtain the corresponding independent transient compensation sequences; Time alignment is performed on the independent transient compensation sequences, the peak amplitude of the corresponding cavity is calculated, and normalized to the amplitude difference; Peak time detection is performed on independent transient compensation sequences, the time difference of characteristic inflection points is calculated, and converted into phase difference; Based on the cavity spacing, the azimuth calculation of the characteristic inflection point time difference and the effective diffusion velocity is performed to obtain the time difference incident angle. The amplitude difference is mapped according to the preset dual-cavity directivity sensitivity curve, and the amplitude ratio azimuth angle is calculated. The relative gas source azimuth is obtained by weighting the incident angle of the time difference and the azimuth of the amplitude ratio. The entity reasoning scenario is analyzed to obtain the entity scenario semantics and entity boundary coordinates; Based on the physical properties of gas diffusion, the coordinates of the entity boundary are diffused to generate a Gaussian prior probability distribution map. The relative gas source azimuth angle and the Gaussian prior probability distribution map are projected and fused to obtain the posterior gas source probability density distribution map. Particle resampling is performed on the posterior gas source probability density distribution map to calculate the particle cloud convergence variance; If the particle cloud convergence variance is less than the preset particle deviation threshold, then the entity corresponding to the current particle cloud is determined to be the target gas source based on the semantics of the entity scene.

[0059] In this embodiment: In a leak tracing task at a storage tank area of ​​a chemical refinery, the transient gas characteristic sequences independently collected by the robot's left and right nasal cavities were compensated by fluid dynamics to obtain independent transient compensated sequences. The peak amplitudes of the two sequences (0.8% VOL for the left cavity and 0.6% VOL for the right cavity) were calculated by time alignment and the amplitude difference was normalized to 0.25. Simultaneously, the characteristic inflection points (peak time 1.2s for the left cavity and 1.5s for the right cavity) were detected, yielding a time difference of 0.3s. Combining the 0.5m cavity spacing and the effective diffusion velocity of 0.8m / s, the incident angle of the time difference was calculated to be approximately 28°. Then, a preset dual-cavity directional sensitivity curve was used to determine the time difference. The line maps the amplitude difference to an amplitude ratio azimuth angle of 32°, and after weighted averaging, the relative gas source azimuth angle of 30° is obtained. The system analyzes the entity reasoning scenario ("benzene leakage near flange B-12 in the storage tank area"), extracts the entity boundary coordinates, and generates a prior probability distribution map based on the Gaussian plume diffusion model. The azimuth angle of 30° is projected and fused to obtain the posterior gas source probability density distribution map. Sequential Monte Carlo particle filtering is used for resampling. When the particle cloud convergence variance converges to 0.05 (less than the threshold of 0.1), the entity (flange sealing surface) corresponding to the particle cloud is determined to be the target gas source, and the robot is driven to approach and issue a leak warning.

[0060] Secondly, referring to Figure 4 This application also provides a gas recognition system based on the olfactory sense of a humanoid robot, applied to gas recognition methods, including: The gas analysis module is used to monitor the gas in the robot's operating environment based on the gas pumping frequency, and to collect and clean the gas composition concentration. The feature extraction module is used to analyze the gas component concentration based on the transient response window and the fluid dynamics model, and extract the transient feature sequence of the gas. The multi-modal reasoning module is used to identify and classify transient gas feature sequences based on a preset multi-modal semantic model, and to reason about entity reasoning scenarios. The target localization module is used to compare the amplitude difference and phase difference of the transient gas characteristic sequence of each cavity according to the entity reasoning scenario, and update and locate the target gas source.

[0061] In this embodiment, in the spherical tank area of ​​a natural gas processing plant, a humanoid inspection robot actively draws in ambient gas at a frequency of 2Hz and synchronously collects methane concentration using a dual-cavity electrochemical sensor. After median filtering and timestamp alignment, a clean concentration sequence is obtained. Subsequently, based on a 200ms transient response window coupled with a preceding fluid dynamics model (with a pre-calibrated flow field around the robot's head), a transient feature sequence consisting of rise steepness, peak overshoot rate, and fluctuation variance is extracted from the concentration sequence. A pre-trained multimodal semantic model (integrating odor embedding and motion parameters) classifies the sequence as a "flange sealing surface micro-leak" entity scenario and infers a leakage probability of 0.92 and a diffusion direction of southeast.

[0062] Finally, the target localization module compares the amplitude difference (0.35) and peak time phase difference (0.12s) of the transient feature sequences of the left and right nasal cavities. It uses the binaural orientation algorithm to calculate the relative air source azimuth angle as 28°. Combined with the equipment layout coordinates provided by the physical scene, the posterior probability distribution is updated through particle filtering. The precise location of the leak source is marked on the 3D grid map, guiding the robot to approach and issuing an audible and visual alarm. The entire response time is less than 3 seconds.

[0063] An electronic device, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement any of the methods described in the above scheme.

[0064] A storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the gas recognition method based on the humanoid robot's sense of smell as described above.

[0065] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A gas recognition method based on the olfactory sense of a humanoid robot, characterized in that, include: Monitor the gas in the robot's operating environment based on the gas pumping frequency, and collect and clean the gas composition concentration; The gas component concentration is analyzed based on the transient response window and the fluid dynamics model, and the gas transient characteristic sequence is extracted. Based on a preset multimodal semantic model, the transient feature sequence of the gas is identified and classified to infer entity reasoning scenarios; Based on the entity reasoning scenario, the amplitude difference and phase difference of the gas transient characteristic sequence of each cavity are compared to update and locate the target gas source.

2. The gas recognition method based on the olfactory sense of a humanoid robot according to claim 1, characterized in that, The process of monitoring the gas in the robot's operating environment based on the gas pumping frequency, and collecting and cleaning the gas component concentration includes: The robot's olfactory detection is graded and set according to the dynamic level of the environment to determine the gas suction frequency. The gas in the robot's operating environment is detected in a dual-channel manner based on the gas pumping frequency, and the original component concentration of each channel is collected. Based on the sniffing timestamp, the original component concentration is filtered, denoised, and duplicate values ​​are removed to obtain the standard component concentration of each cavity. The concentration of the standard component is statistically analyzed based on the suction cycle, the average peak-to-peak value is calculated, and the gas concentration signal-to-noise ratio is obtained by combining the noise variance during the gas-free period. If the signal-to-noise ratio of the gas concentration is greater than the preset signal-to-noise ratio threshold, the current standard component concentration is determined to be an effective value, and the standard component concentrations of each cavity are aggregated and correlated to obtain the gas component concentration.

3. The gas recognition method based on the olfactory sense of a humanoid robot according to claim 1, characterized in that, The step of analyzing the gas component concentration based on the transient response window and the fluid dynamics model, and extracting the gas transient feature sequence, includes: The robot's operating status is monitored according to the sniffing cycle, the robot's motion parameters are collected, and the relative airflow speed is calculated in combination with the ambient wind speed and direction. Based on the robot's head structure, a fluid dynamics model is generated by performing fluid simulation fitting between the relative airflow velocity and the robot's motion parameters. Based on the fluid dynamics model, the gas component concentration is stripped to obtain the effective component concentration, and the spurious response concentration increment is calculated. If the ratio of the false response concentration increment to the peak concentration of the gas component is greater than a preset percentage threshold, then the current effective component concentration is marked as a strong motion interference bit. No, then the concentration of the effective ingredient is corrected point by point according to the transient response window, and the ingredient correction compensation concentration is calculated; The component correction and compensation concentration is subjected to feature extraction and segmented aggregation approximation to obtain a gas time-series feature vector; The gas temporal feature vector is fused and spliced ​​with the robot motion parameters to generate a gas transient feature sequence.

4. The gas recognition method based on the olfactory sense of a humanoid robot according to claim 3, characterized in that, The step of performing fluid simulation fitting between the relative airflow velocity and the robot motion parameters based on the robot head structure to generate a fluid dynamics model includes: The robot's head structure is analyzed, the outer shell and nasal cavity are extracted, and combined with the 3D map, a head component model and a nasal cavity component model are generated. A watershed is constructed on the head component model to generate a hemispherical virtual wind tunnel, and the nasal cavity component model is simplified and smoothed to obtain a simplified nasal cavity model. The head component model is prism meshed to divide the near-wall response area, and the simplified nasal cavity model is filled with spheres to obtain a dense spherical area; The near-wall response region and the encrypted spherical region are detected and analyzed to calculate the grid tilt and grid orthogonality. If both are greater than the corresponding quality threshold, then the current near-wall response region and the current encrypted sphere region are deemed compliant. The hemispherical virtual wind tunnel is parameterized based on the relative airflow velocity to obtain the airflow dynamic wind tunnel; The near-wall response region is simulated and quantized based on the robot's motion parameters to obtain the near-wall motion space. The airflow dynamic wind tunnel is driven according to the preset simulation variable parameters, and turbulence simulation is performed on the densified spherical region in combination with the near-wall motion space to collect gas convection mass transfer parameters and generate an initial dynamic model. The gas convection mass transfer parameters are cleaned and solved serially to calculate the flow field data at the detection point; Based on the static reference state, the flow field data at the detection point is subjected to a reference difference operation to extract the airflow disturbance value and pressure distortion value; If both are within the corresponding tolerance threshold range, then the initial dynamic model is lightweighted and pruned to obtain the fluid dynamic model; If not, the corresponding residual value is calculated and fused to generate a linear error regression correction term to correct the initial dynamic model, thus obtaining the fluid dynamic model.

5. The gas recognition method based on the olfactory sense of a humanoid robot according to claim 1, characterized in that, The step of identifying and classifying the gas transient feature sequence based on a preset multimodal semantic model, and reasoning about the entity reasoning scenario, includes: The transient feature sequences of gases are parsed and transformed to extract odor embedding vectors and map the semantics of chemical components. The robot motion parameters are vectorized and multi-layered, and then concatenated with the odor embedding vector to generate an environmental odor semantic vector. The environmental odor semantic vector is cross-identified based on a preset multimodal semantic model to determine the odor component entity object; The odor component entities are classified according to the scene odor component weights to generate scene prior vectors; The entity objects of the odor components are retrieved and associated using the knowledge graph to obtain entity event nodes; The scenario prior vector is bound to and decoded with the entity event node to generate an entity reasoning scenario.

6. The gas recognition method based on the olfactory sense of a humanoid robot according to claim 1 or 5, characterized in that, The multimodal semantic model includes: The robot's environmental parameters are analyzed based on the odor transient window, and visual image frames, microphone audio segments, and tactile / force readings are extracted. Lightweight encoding is performed on the visual image frames to extract global object scene feature vectors; Perform temporal convolution on the microphone audio segment to extract the audio embedding feature vector; Based on the gas transient feature sequence, the global object scene feature vector and the audio embedding feature vector are subjected to gated weighted fusion to generate a multimodal temporal feature vector; Entity semantic encoding is performed on the multimodal temporal feature vectors to obtain a multimodal entity weight matrix; The multimodal entity weight matrix is ​​bound to the odor type code based on the cross-attention weight matrix to form a semantic ternary weight matrix; Monte Carlo simulation is performed on the semantic ternary weight matrix to output the scene reasoning statement and calculate the posterior distribution variance; If the posterior distribution variance is less than the preset event entity threshold, the current semantic ternary weight matrix is ​​determined to be a multimodal semantic model and stored in the real-time environment memory.

7. The gas recognition method based on the olfactory sense of a humanoid robot according to claim 1, characterized in that, The step of comparing the amplitude and phase differences of the transient gas characteristic sequences of each cavity based on the entity reasoning scenario, and updating and locating the target gas source, includes: Transient window truncation and hydrodynamic compensation are performed on the gas transient characteristic sequences of each cavity to obtain the corresponding independent transient compensation sequences; The independent transient compensation sequence is time-aligned, the peak amplitude of the corresponding cavity is calculated, and normalized to the amplitude difference; Peak time detection is performed on the independent transient compensation sequence, the characteristic inflection point time difference is calculated, and it is converted into a phase difference; Based on the cavity spacing, the azimuth calculation of the time difference of the characteristic inflection point and the effective diffusion velocity is performed to obtain the time difference incident angle. The amplitude difference is mapped according to the preset dual-cavity directivity sensitivity curve, and the amplitude ratio azimuth angle is calculated. The relative gas source azimuth is obtained by weighting the time difference incident angle and the amplitude ratio azimuth angle. The entity reasoning scenario is analyzed to obtain the entity scenario semantics and entity boundary coordinates; Based on the physical properties of gas diffusion, the boundary coordinates of the entity are diffused to generate a Gaussian prior probability distribution map. The relative gas source azimuth angle and the Gaussian prior probability distribution map are projected and fused to obtain the posterior gas source probability density distribution map. Particle resampling is performed on the posterior gas source probability density distribution map to calculate the particle cloud convergence variance; If the particle cloud convergence variance is less than a preset particle deviation threshold, then the entity corresponding to the current particle cloud is determined to be the target gas source based on the semantics of the entity scene.

8. A gas identification system based on the olfactory sense of a humanoid robot, used to implement the gas identification method based on the olfactory sense of a humanoid robot as described in any one of claims 1 to 7, characterized in that, include: The gas analysis module is used to monitor the gas in the robot's operating environment based on the gas pumping frequency, and to collect and clean the gas composition concentration. The feature extraction module is used to analyze the gas component concentration based on the transient response window and the fluid dynamics model, and extract the gas transient feature sequence. The multi-modal reasoning module is used to identify and classify the transient feature sequence of the gas according to the preset multi-modal semantic model, and to reason about the entity reasoning scenario. The target localization module is used to compare the amplitude difference and phase difference of the gas transient characteristic sequence of each cavity according to the entity reasoning scenario, and update and locate the target gas source.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A storage medium storing at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the gas recognition method based on the olfactory sense of a humanoid robot as described in any one of claims 1 to 7.

Images