A handheld ai electronic nose device, system and method

Abstract

The present disclosure provides a kind of palm AI electronic nose device, system and method.The device includes: electronic nose device (901), including environmental parameter monitoring unit (101), multi-channel metal oxide gas sensor array unit (102), gas path-gas chamber integration component (103), integrated pump valve module (104), display and input module (105), acquisition and execution control board (106), high-resolution analog-digital conversion module (107), edge AI processing module (108) and battery and management module (109).The present disclosure can realize the promotion of detection efficiency, the enhancement of long-term stability, achieve edge intelligence and controllable self-adaptation, and realize portable and integrated optimization.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of gas detection devices, in particular to a palm AI electronic nose device, system and method. BACKGROUND

[0002] Metal oxide sensors (MOS) are widely used in gas detection field due to their low cost and high sensitivity. However, existing MOS-based electronic nose systems still face a series of technical challenges in practical applications, especially in portable and on-site rapid detection scenarios.

[0003] Firstly, low detection efficiency: The working cycle of MOS sensors usually includes preheating, adsorption detection, desorption cleaning and recovery stages. In existing technologies, the preheating period often takes tens of seconds to several minutes to reach a stable working temperature; the detection period needs to wait for the sensor response to tend to be stable before reliable discrimination can be made, which takes a significant amount of time; the desorption cleaning stage is inefficient, relying on natural desorption or constant temperature cleaning, which makes it difficult to remove residual gas and slows down the baseline recovery. This makes the single complete detection cycle usually more than 3 minutes, which is difficult to meet the application requirements of high time efficiency such as emergency response and medical on-site rapid screening. Although there are attempts to shorten some links by early discrimination (early stop) or optimization of temperature control through algorithms, the effect of a single strategy is limited, i.e. relying solely on algorithm early stop is prone to misjudgment, and optimizing temperature control alone cannot solve the problem of long detection period.

[0004] Secondly, significant influence of environmental and device drift: MOS sensor performance is easily affected by environmental factors (such as temperature, humidity, air pressure, and gas flow rate fluctuations) and device state (such as aging and contamination), resulting in baseline drift, sensitivity changes, response time changes, and other problems. This drift can significantly reduce the accuracy and long-term stability of the detection results, resulting in an increase in false positive or false negative rates. Traditional fixed threshold discrimination or offline calibration methods are difficult to adapt to this dynamic change in real time, especially in complex and variable on-site environments, making it difficult to ensure the reliability of the detection results. Existing technologies lack a real-time, online running layered compensation and self-calibration mechanism to systematically address multiple sources of drift.

[0005] Thirdly, lack of intelligence and adaptability: Existing electronic nose systems rely on cloud for complex model reasoning and updating. This makes it difficult for the device to work effectively or update the model in network-limited or network-free environments (such as disaster sites, outdoors, and security isolation areas). The cloud model update cycle is long and difficult to quickly adapt to sensor individual differences, new environmental background gases, or sensor drift. In addition, uploading raw gas response data to the cloud for processing also poses a privacy and security risk of sensitive information (such as medical breathing characteristics) being leaked. Therefore, there is an urgent need for a mechanism that can achieve efficient and reliable reasoning and controllable self-adaptive learning on the device's local edge.

[0006] Finally, the system integration and portability are poor: the existing system is mostly laboratory or desktop structure, large volume, high power consumption, difficult to meet the demand of portable, handheld application. The gas path design is often complex, there are large dead volume, uneven flow of each sensor gas channel problems, leading to different time of gas synchronization, affecting the accuracy of early feature extraction and multi-channel fusion effect, and increasing the risk of residual gas cross contamination. The scattered layout of multiple modules is easy to be affected by electromagnetic interference (such as the influence of pump valve drive on weak signal acquisition), and lacks of safety isolation design for high temperature working area, which is not conducive to the safety and user experience of handheld operation. The power consumption optimization is also insufficient, which limits the endurance of battery powered equipment.

[0007] In summary, the MOS electronic nose system in the prior art has many deficiencies in detection speed, anti-drift ability, environmental adaptability, intelligence level, portability and safety, which restricts its wide application in on-site rapid and reliable detection scenarios. SUMMARY

[0008] (I) Technical problems to be solved

[0009] In view of the above problems, the present disclosure provides a palm AI electronic nose device, system and method to at least partially solve the problems of low efficiency of traditional gas detection, significant influence of environment and device drift, insufficient intelligence and adaptability, poor system integration and portability.

[0010] (II) Technical solutions

[0011] The first aspect of the present disclosure provides a palm AI electronic nose device, comprising: an electronic nose device, including an environmental parameter monitoring unit, a multi-channel metal oxide gas sensor array unit, a gas path-gas chamber integrated assembly, an integrated pump valve module, a display and input module, a collection and execution control board, a high-resolution analog-to-digital conversion module, an edge AI processing module and a battery and management module; the environmental parameter monitoring unit is located at the central installation position of the multi-channel metal oxide gas sensor array unit, the multi-channel metal oxide gas sensor array unit is arranged in the gas path-gas chamber integrated assembly, the integrated pump valve module is in communication with the gas path-gas chamber integrated assembly, the collection and execution control board is connected with the environmental parameter monitoring unit and the multi-channel metal oxide gas sensor array unit, the high-resolution analog-to-digital conversion module is connected with the collection and execution control board above, converts the analog data collected by the collection and execution control board into digital data, and transmits the converted digital data to the edge AI processing module connected below, the edge AI processing module is connected and sends the control instruction back to the collection and execution control board; the battery and management module provides energy and safety management for the electronic nose device; the display and input module displays the mode, temperature control curve, progress, result and prompt of the electronic nose device.

[0012] The second aspect of the present disclosure provides a palm AI electronic nose system, which comprises the palm AI electronic nose device described above. The system is divided into a perception layer, a decision layer and a control layer according to functions, and the three layers interact through a unified message / event bus.

[0013] According to the embodiments of the present disclosure, the perception layer is composed of the environmental parameter monitoring unit, the multi-channel metal oxide gas sensor array unit, the gas path-gas chamber integrated assembly, the integrated pump valve module and the high-resolution analog-to-digital conversion module, which are used to complete gas sampling and signal acquisition; gas introduction and distribution: external gas enters through the integrated pump valve module, in which the gas pump and the gas valve are combined in a compact structure, the gas pump is located on the left side of the integrated pump valve module, and the gas valve is located on the right side of the integrated pump valve module; the gas pump provides the driving force for gas flow, and the gas valve functions as a switch and a flow path switching; under the action of the gas pump, the external gas enters the gas inlet of the gas path-gas chamber integrated assembly through the gas valve, passes through the central channel and then enters the sampling microcavity through the inlet of each sampling microcavity, and then reacts with the sensitive layer of the multi-channel metal oxide gas sensor array unit located at the sensor mounting interface in the sampling microcavity, and is collected from the outlet of the sampling microcavity into the central channel and then discharged through the gas outlet; signal and environmental quantity collection: the high-resolution analog-to-digital conversion module performs high-precision synchronous sampling on the multi-channel sensor signals of the multi-channel metal oxide gas sensor array unit; the environmental parameter monitoring unit outputs the environmental parameters including temperature, humidity, flow rate and air pressure in real time, which are used for subsequent online compensation and self-calibration; the perception layer outputs the multi-channel sensor signals and environmental parameters formed to the decision layer.

[0014] According to the embodiments of the present disclosure, the decision layer is composed of the dynamic temperature control and early stop discrimination unit, the anti-drift and self-calibration unit and the edge AI and local incremental learning unit, which run under the support of the temperature control / execution of the acquisition and execution control board and the data of the high-resolution analog-to-digital conversion module; the dynamic temperature control and early stop discrimination unit sequentially executes the processes of pulse rapid preheating, low-temperature adsorption detection, two-stage early stop sequential probability ratio test and multi-outlet time convolution network and high-temperature desorption cleaning according to the temperature control curve issued by the acquisition and execution control board; when the threshold is met, a discrimination conclusion and a confidence level are generated, and a trigger signal for cutting in cleaning is sent to the control layer; the anti-drift and self-calibration unit takes the environmental quantity of the environmental parameter monitoring unit and the historical baseline of the multi-channel metal oxide gas sensor array unit / high-resolution analog-to-digital conversion module as inputs, performs online environmental compensation, anchor point baseline calibration, channel health score and gating / shielding, confidence level calibration, shape rejection and outlier distribution identification, and outputs parameter updates and resampling instructions; the edge AI and local incremental learning unit runs in the edge AI processing module, performs model reasoning, sample screening, constrained incremental update and version rollback on the preprocessed multi-channel data, and generates model / log versions.

[0015] According to an embodiment of the present disclosure, the control layer is composed of a collection and execution control board, a high-resolution analog-to-digital conversion module, an edge AI processing module, a battery and management module, and a display and input module, responsible for execution, collection, energy consumption and interaction; execution and collection: the collection and execution control board controls the integrated pump valve module and the multi-channel metal oxide gas sensor array unit heating and temperature measurement, and cooperates with the high-resolution analog-to-digital conversion module to realize multi-channel synchronous sampling and preprocessing; local calculation: the edge AI processing module completes inference, confidence calibration and incremental learning, and returns the control instruction to the collection and execution control board; power supply and interaction: the battery and management module provides energy and safety management; the display and input module displays mode, temperature control curve, progress, result and prompt; data and control path: the target link is the connection of the high-resolution analog-to-digital conversion module, the collection and execution control board, the collection and execution control board, the edge AI processing module, and the edge AI processing module, the display and input module; the control loop is the connection of the edge AI processing module and the collection and execution control board, used for triggering pump valve / temperature control action.

[0016] The third aspect of the present disclosure provides a palm AI electronic nose method, comprising the palm AI electronic nose system described above, the method comprising: step S1: performing initialization operation of the system and establishing baseline file; step S2: performing sampling starting point detection and data preprocessing; step S3: performing pulse type fast preheating process in the first stage; step S4: performing low temperature adsorption detection in the second stage, and executing multi-stage early stop discrimination; step S5: realizing online environment compensation and data normalization processing; step S6: performing baseline tracking, anchor point calibration and channel health state evaluation; step S7: performing threshold self-adaptive adjustment, uncertainty control and unknown class identification; step S8: performing constrained incremental learning and rollback operation at the edge AI end; step S9: realizing closed-loop cooperative control of dynamic temperature control and early stop strategy; step S10: performing structure optimization and power consumption management of the system; and step S11: adjusting parameter range and working mode according to running state.

[0017] According to an embodiment of the present disclosure, the initialization operation of the system is performed, and the baseline file is established, comprising: step S1.1: clean gas baseline collection: after the equipment is first deployed or powered on, the clean gas is automatically switched, a multi-channel signal in a preset period is collected, and each channel initial baseline , noise floor , dynamic range scale and rise time constant are obtained, and the reference environment vector is recorded synchronously. ; Step S1.2: Environment compensation initial value fitting: taking the reference environment vector e0 as reference, fitting the environment sensitive coefficient of each channel; Step S1.3: Version management of calibration file: generating a calibration file with a time stamp and storing it in a database.

[0018] According to an embodiment of the present disclosure, the low-temperature adsorption detection in the second stage is performed, and multi-stage early stopping discrimination is performed, including: Step S4.1: Low-temperature adsorption: collecting a response signal at a low-temperature platform, and extracting features including early slope, amplitude and cross-channel ratio from the response signal to improve information density; Step S4.2: Performing statistical early stopping discrimination based on sequential probability ratio test: calculating a log-likelihood ratio , defined as:

[0019]

[0020] wherein, represents an alternative hypothesis that the target gas exists, represents an original hypothesis that the target gas does not exist, represents the kth observation value; when ≥logA, early acceptance , when ≤logB, acceptance ; when is in a hysteresis interval, stability is combined with sliding window consistency and double-threshold hysteresis; Step S4.3: Multi-exit deep early stopping: based on a time convolution network, a plurality of exit points are set within a preset response time interval, and when the discrimination confidence of any exit point reaches a threshold value, early output is performed; the position of the exit point is adjusted according to the scene; Step S4.4: Performing bottom-up complete collection: if any of the above early stopping is not triggered or the confidence does not meet the preset condition, the response signal is continuously collected until the pre-defined collection period.

[0021] According to an embodiment of the present disclosure, baseline tracking, anchor calibration and channel health status evaluation are performed, including: Step S6.1: Baseline tracking: at the end of post-cleaning or during the interval between two detections, the latest baseline , noise and time constant are measured and obtained; a drift index is calculated, wherein, , , is a reference reference value; if any index exceeds an adaptive threshold band, the adaptive threshold band is determined based on quantiles of historical distribution, a target quantile interval is selected, and is marked as "deviation" state; Step S6.2: Anchor calibration: triggered by a fixed period or "deviation" condition generated by S6.1, short sampling is performed by switching a clean gas to refresh the reference reference value with the reference environment; and synchronously fine-tune in S5 the coefficients; step S6.3: health score and gating: define the health score of channel i at time t as:

[0022]

[0023] wherein, is a normalization function mapping the deviation to the interval [0,1]; is the cross-correlation consistency score of this channel with the rest of the channels; w k is a configurable weight; the channel is weighted before feature fusion / model input ; channels with health score continuously below a preset threshold enter the candidate shielding state and report an event.

[0024] (Three) beneficial effects

[0025] 1. Detection efficiency is improved: through pulse rapid preheating, multi-stage early stop discrimination (SPRT and multi-outlet TCN) in the low-temperature adsorption period, and high-temperature desorption cleaning after early stop triggering, the system can shorten the time of preheating, adsorption and cleaning three stages, while ensuring the reliability of discrimination, and realizing the shortening of the overall detection period.

[0026] 2. Long-term stability is enhanced: through online environment compensation, anchor point baseline calibration, and hierarchical anti-drift and self-calibration mechanism of channel health gating / shielding, the influence of temperature and humidity, flow rate and device aging can be reduced, thereby maintaining the long-term consistency and reliability of the detection results.

[0027] 3. Edge intelligence and controllable self-adaptation: local inference and constrained incremental learning are performed on the device, and combined with versioning and rollback mechanism, even in the absence of network or weak network environment, new scenarios and individual differences can be adapted, while avoiding the uploading of privacy data, improving the independence and security of the system.

[0028] 4. Portable and integrated optimization: a one-to-many flow small dead space gas chamber optimized by CFD, low adsorption inner wall and pump valve integrated packaging, combined with analog / digital / power three-domain isolation circuit and power management, improve the response synchronization and signal-to-noise ratio of multiple channels, reduce cross-contamination, and meet the needs of handheld applications in volume, power consumption and thermal safety. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more completely understand the present disclosure and its advantages, the following description will now be made in connection with the accompanying drawings, in which:

[0030] Figure 1 The structure of the electronic nose device provided by the embodiment of the present disclosure is schematically shown in a perspective view;

[0031] Figure 2 A schematic diagram of the air path-air chamber integrated assembly structure provided by the embodiments of the present disclosure is shown;

[0032] Figure 3 A flowchart of the rapid detection provided by the embodiments of the present disclosure is shown;

[0033] Figure 4 A flowchart of the anti-drift and self-calibration provided by the embodiments of the present disclosure is shown;

[0034] Figure 5 A flowchart of the edge AI and local incremental learning provided by the embodiments of the present disclosure is shown;

[0035] Figure 6 A principle diagram of the early stop discrimination provided by the embodiments of the present disclosure is shown;

[0036] Figure 7 An experimental data diagram of the first embodiment provided by the embodiments of the present disclosure is shown;

[0037] Figure 7 A shows a classification confusion matrix for the pest Bactrocera dorsalis at a normal decision point in the first embodiment of the present disclosure, showing that the overall classification accuracy is 95.5%;

[0038] Figure 7 B shows a classification confusion matrix for the same pest at an earliest decision point, showing that the overall classification accuracy is 90.9%;

[0039] Figure 7 C shows a bar chart of the average time required for each infestation group of Bactrocera dorsalis to reach a classification decision provided by the embodiments of the present disclosure;

[0040] Figure 7 D shows a classification confusion matrix for the pest Bactrocera correcta at a normal decision point, showing that the overall classification accuracy is 93.3%;

[0041] Figure 7 E shows a classification confusion matrix for the same pest at an earliest decision point, showing that the overall classification accuracy is 88.9%;

[0042] Figure 7 F shows a bar chart of the average time required for each infestation group of Bactrocera correcta to reach a classification decision provided by the embodiments of the present disclosure;

[0043] Figure 8An early stop decision schematic diagram provided by the embodiment of the present disclosure is schematically shown.

[0044] Figure 9 A customs conveyor belt and mechanical arm cooperative detection schematic diagram provided by the embodiment of the present disclosure is schematically shown.

[0045] Legend of reference signs:

[0046] 101 - environmental parameter monitoring unit; 102 - multi-channel metal oxide gas sensor array unit; 103 - gas path-gas chamber integrated assembly; 104 - integrated pump valve module; 105 - display and input module; 106 - acquisition and execution control board; 107 - high-resolution analog-to-digital conversion module; 108 - edge AI processing module; 109 - battery and management module; 201 - air inlet; 202 - sampling microcavity; 203 - sampling microcavity inlet; 204 - sampling microcavity outlet; 205 - sensor mounting interface; 206 - air outlet; 901 - electronic nose device; 902 - mechanical arm; 903 - conveyor belt; 904 - sample; 905 - sample box. DETAILED DESCRIPTION

[0047] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary, and are not intended to limit the scope of the present disclosure. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. However, it will be apparent to those skilled in the art that one or more embodiments can be practiced without these specific details. In addition, in the following description, descriptions of well-known structures and techniques have been omitted to avoid unnecessarily obscuring the concept of the present disclosure.

[0048] The terms used herein are merely used to describe specific embodiments, and are not intended to limit the present disclosure. The terms "include", "comprise" and the like used herein indicate the presence of the features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0049] All terms used herein (including technical and scientific terms) have meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0050] 1. System overall structure

[0051] As Figure 1As shown, the handheld AI electronic nose device 901 disclosed herein includes an environmental parameter monitoring unit 101, a multi-channel metal oxide gas sensor array unit 102, an integrated gas path-gas chamber assembly 103, an integrated pump and valve module 104, a display and input module 105, an acquisition and execution control board 106, a high-resolution analog-to-digital conversion module 107, an edge AI processing module 108, and a battery and management module 109.

[0052] The environmental parameter monitoring unit 101 is located in the central mounting position of the multi-channel metal oxide gas sensor array unit 102. The multi-channel metal oxide gas sensor array unit 102 is set inside the gas path-gas chamber integrated component 103. The integrated pump valve module 104 is connected to the gas path-gas chamber integrated component 103. The acquisition and execution control board 106 is connected to the environmental parameter monitoring unit 101 and the multi-channel metal oxide gas sensor array unit 102. The high-resolution analog-to-digital conversion module 107 is connected to the acquisition and execution control board 106 above and converts the analog data it acquires into digital data. Then, the converted digital data is transmitted to the edge AI processing module 108 connected below. The edge AI processing module 108 is connected and sends control commands back to the acquisition and execution control board 106. The battery and management module 109 provides energy and safety management for the electronic nose device 901. The display and input module 105 displays the mode, temperature control curve, progress, results and prompts of the electronic nose device 901.

[0053] The environmental parameter monitoring unit 101 is located in the central mounting position of the multi-channel metal oxide gas sensor array unit 102. It collects environmental parameters such as temperature, humidity, flow rate and air pressure as inputs for online compensation and self-calibration.

[0054] The multi-channel metal oxide gas sensor array unit 102 can preferably have 8 channels, each channel having independent heating and independent temperature measurement, for performing differentiated dynamic temperature control.

[0055] like Figure 2 As shown, the integrated gas path-gas chamber assembly 103 includes an air inlet 201, a central channel, at least one sampling microcavity 202, and an air outlet 206. The sampling microcavity 202 is provided with a sampling microcavity inlet 203 and a sampling microcavity outlet 204. A sensor mounting interface 205 is provided on the lower inner side of the sampling microcavity 202. A sensitive layer of a multi-channel metal oxide gas sensor array unit 102 is provided at the sensor mounting interface 205 to ensure that the gas acts on the sensitive layer of the multi-channel metal oxide gas sensor array unit 102 within the sampling microcavity 202.

[0056] The integrated pump and valve module 104 is connected to the integrated gas path-gas chamber component 103, supporting constant flow sampling, pulse sampling and cleaning modes.

[0057] The acquisition and execution control board 106 is responsible for the preliminary acquisition of sensor signals, array heating curve execution, and pump valve drive control, and is synchronized with the high-resolution analog-to-digital conversion module 107.

[0058] The high-resolution analog-to-digital conversion module 107 is used for high-precision (multi-channel synchronization) sampling of the output of each channel of the multi-channel metal oxide gas sensor array unit 102, ensuring the quality of early weak signals.

[0059] The edge AI processing module 108 is used for local model inference, confidence calibration, and constrained incremental learning, and versioning and rollback.

[0060] The battery and management module 109 is used to provide power, charge and discharge, and safety management, and supports low-power sleep.

[0061] The display and input module 105 is used for mode selection and result display.

[0062] 2. System function architecture

[0063] Based on the above device, the electronic nose system is divided into a perception layer, a decision layer, and a control layer, and the three layers interact through a unified message / event bus.

[0064] The perception layer is composed of the environmental parameter monitoring unit 101, the multi-channel metal oxide gas sensor array unit 102, the gas path-gas chamber integrated assembly 103, the integrated pump valve module 104, and the high-resolution analog-to-digital conversion module 107, and is used to complete gas sampling and signal acquisition.

[0065] Specifically, first, the perception layer can be used for gas introduction and distribution: external gas first enters via the integrated pump valve module 104, in which the gas pump and the gas valve are integrated in a compact structure, with the gas pump on the left side of the integrated pump valve module 104 and the gas valve on the right side of the integrated pump valve module 104. The gas pump provides the driving force for gas flow, and the gas valve functions as an on-off switch and flow path switch. The external gas enters the gas inlet 201 of the gas path-gas chamber integrated assembly 103 under the action of the gas pump, then passes through the central channel and enters the sampling microcavities 202 through the respective sampling microcavity inlets 203, and then fully interacts with the sensitive layer of the multi-channel metal oxide gas sensor array unit 102 at the sensor mounting interface 205, and then flows into the central channel through the sampling microcavity outlets 204 and is discharged through the gas outlet 206. Second, it can also be used for signal and environmental quantity acquisition: the high-resolution analog-to-digital conversion module 107 performs high-precision synchronous sampling on the multi-channel electrical signals (i.e., sensor signals) of the multi-channel metal oxide gas sensor array unit 102; the environmental parameter monitoring unit 101 outputs real-time environmental parameters such as temperature, humidity, flow rate, and air pressure, which are used for subsequent online compensation and self-calibration. Finally, the perception layer can output to the decision layer: form an input of “multi-channel sensor signals + environmental parameters”, which corresponds to S2 in the following full-link).

[0066] The decision layer is composed of three functional units, namely, the dynamic temperature control and early stop discrimination unit, the anti-drift and self-calibration unit, and the edge AI and local incremental learning unit, which operate under the support of the temperature control / execution of the acquisition and execution control board 106 and the data of the high-resolution analog-to-digital conversion module 107.

[0067] Among them, as shown in Figure 3 , the dynamic temperature control and early stop discrimination unit executes the process of pulse rapid preheating → low-temperature adsorption detection → two-stage early stop (sequential probability ratio test (SPRT) and multi-outlet time convolution network (TCN)) → high-temperature desorption cleaning according to the temperature control curve issued by the acquisition and execution control board 106. When the threshold is met, a discrimination conclusion and a confidence level are generated, and a trigger signal for cutting in cleaning is sent to the control layer (corresponding to S3-S4).

[0068] As shown in Figure 4 , the anti-drift and self-calibration unit takes the environmental quantities of the environmental parameter monitoring unit 101 and the historical baseline of the multi-channel metal oxide gas sensor array unit 102 / high-resolution analog-to-digital conversion module 107 as inputs, and performs online environmental compensation, anchor point baseline calibration, channel health score and gating / shielding, confidence level calibration, conformal rejection, and out-of-distribution (OOD) identification, and outputs parameter updates and resampling instructions (corresponding to S5).

[0069] As shown in Figure 5As shown, edge AI and local incremental learning unit run in edge AI processing module 108, model inference, sample screening, constrained incremental update and version rollback are performed on pre-processed multi-channel data, and model / log version is generated (corresponding to S6).

[0070] The control layer is composed of acquisition and execution control board 106, high-resolution analog-to-digital conversion module 107, edge AI processing module 108, battery and management module 109, and display and input module 105, responsible for execution, acquisition, energy consumption and interaction.

[0071] Execution and acquisition: acquisition and execution control board 106 controls integrated pump valve module 104 and multi-channel metal oxide gas sensor array unit 102 heating / temperature measurement, and cooperates with high-resolution analog-to-digital conversion module 107 to realize multi-channel synchronous sampling and preprocessing.

[0072] Local computing: edge AI processing module 108 completes inference, confidence calibration and incremental learning, and sends control instructions (such as cutting in cleaning / re-sampling) back to acquisition and execution control board 106.

[0073] Power supply and interaction: battery and management module 109 provides energy and safety management; display and input module 105 displays mode, temperature control curve, progress, result and prompt.

[0074] Data and control path: typical target link is high-resolution analog-to-digital conversion module 107→acquisition and execution control board 106→edge AI processing module 108→display and input module 105; control loop is edge AI processing module 108→acquisition and execution control board 106 (trigger pump valve / temperature control action).

[0075] 3. Full-link workflow

[0076] Step S1: initialization and baseline file establishment (deployment / power-on).

[0077] Step S1.1: clean gas baseline acquisition: after the device is first deployed or powered on, switch to clean gas automatically, acquire multi-channel signals for a short period (i.e. preset period), get initial baseline of each channel , noise floor , dynamic range scale and rise time constant , and record reference environment vector .

[0078] Step S1.2: environment compensation initial value fitting: with e0 as reference, fit the environment sensitivity coefficient of each channel (linear or second-order polynomial; optional Kalman / Bayesian filter parameter initial value).

[0079] Step S1.3: Calibration profile version management: generate "calibration profile v0" (containing channel parameters, environmental regression coefficients, threshold initial values, rejection strategy parameters), store and timestamp.

[0080] Step S2: Sampling start point detection and data preprocessing (every detection start).

[0081] Start point detection: determine the inhalation start point by cross-correlation of pump flow, valve switching and channel signals.

[0082] Preprocessing: perform denoising (preferably low-order Savitzky-Golay or Kalman filtering) on the original sequence, zero-mean / dimension normalization, and channel screening to form a preprocessed data stream; the acquisition link can include a high-resolution analog-to-digital conversion module 107 (such as ADS1256) to ensure the quality of weak early signals.

[0083] Step S3: Perform pulsed rapid preheating (preheating acceleration) in the first stage, i.e., stage I.

[0084] Through short-time high-power heating, the sensor quickly enters the low-temperature detection platform (located in the medium-low temperature section of the safe working range, preferably reaching 200-300°C within 1-5s); closed-loop temperature measurement is used to prevent overshoot; the preheating timing and amplitude can be self-adaptive according to the channel heat capacity and array consistency.

[0085] Step S4: Perform low-temperature adsorption detection + multi-level early stop discrimination in the second stage, i.e., stage II (mid-section acceleration).

[0086] Step S4.1: Low-temperature adsorption: collect response signals at the low-temperature platform, and extract early slope, amplitude, and inter-channel ratio from the response signals to improve information density.

[0087] Step S4.2: Perform statistical early stop discrimination based on sequential probability ratio test, i.e., statistical early stop (Level-1, SPRT): sequential probability ratio test calculates the log-likelihood ratio , defined as:

[0088] where, represents the alternative hypothesis that the target gas exists, represents the original hypothesis that the target gas does not exist, represents the kth observation value.

[0089] When ≥ logA, accept early (enter cleaning), when ≤ logB, accept ; when When in the hysteresis interval, the sliding window consistency and double-threshold hysteresis are combined for stabilization.

[0090] Step S4.3: Multi-exit depth early stop (Level-2, TCN): Based on the time convolution network (TCN), multiple exit points (preferably 2-5 s) are set within a preset response interval (e.g., 1-6 s), and when the discrimination confidence of any exit reaches the threshold, it is output in advance; the exit position can be adapted according to the scene, rather than fixed at a specific number of seconds.

[0091] Step S4.4: Bottom-up, i.e., execute bottom-up complete acquisition: If any of the above early stops does not trigger or the confidence is insufficient, i.e., the preset conditions are not met, continue to collect the response signal until the complete cycle (predefined acquisition period).

[0092] Step S5: Online environmental compensation and data normalization processing (throughout the whole process).

[0093] As shown in Figure 4 , dynamic compensation is performed for each channel, and the compensated signal is calculated by the following formula:

[0094]

[0095] wherein, is the original signal of the i-th sensing channel at time t; β i,T , β i,H , β i,P , β i,Q represents the first-order compensation coefficient of the i-th channel for different environmental parameters; represents the high-order compensation coefficient of the i-th channel for the cross action of environmental parameters; ΔT, ΔH, ΔP, ΔQ... represent the difference or product of difference between the current environmental parameter and the reference environmental parameter; , which represents the difference between the current environmental parameter value and the reference environmental parameter value; the coefficient β is adaptively updated at a slow rate according to the sliding window or exponential weighting to avoid following the instantaneous noise; for , normalization and denoising are performed, and the multi-channel sequence is output for early stop / TCN and subsequent modules to share.

[0096] As shown in Figure 4 , step S6: baseline tracking, anchor point calibration, and channel health status evaluation (period / condition trigger).

[0097] Step S6.1: Baseline tracking: At the end of post-cleaning or during the interval between two detections, the latest baseline , noise and time constant are measured and obtained; the drift index is calculated, wherein, 、 、 Reference value; if any index exceeds the adaptive threshold band (determined based on the quantile of historical distribution, preferably 90-99% (target quantile) interval), it is marked as "deviation" state.

[0098] Step S6.2: Anchor calibration: switch to short sampling of clean gas to refresh with reference environment and synchronize fine-tuning of coefficients in S5 triggered by fixed period or "deviation" condition generated in S6.1.

[0099] Step S6.3: Health score and gating: define the health score of channel i at time t as:

[0100]

[0101] where, is a normalization function that maps the deviation amount to the [0,1] interval; is the cross-correlation consistency score of this channel with the rest of the channels; is a configurable weight; the channels are weighted by before feature fusion / model input; channels with continuously lower health score than the preset threshold enter the candidate shielding state and report events.

[0102] As shown in Figure 4 , step S7: threshold adaptation, uncertainty control, and unknown class (OOD) identification are performed.

[0103] Step S7.1: Confidence calibration: temperature calibration or minimum expected calibration error (ECE / MMCE) calibration is performed on the model output (SPRT / TCN) to make the confidence interpretable.

[0104] Step S7.2: Conformal dynamic threshold: the non-conformal score is calculated within the sliding window W, which can be defined in multiple ways, including but not limited to: in classification tasks, it can be defined as the maximum class probability, which is an uncertainty measure of the prediction result; in regression or continuous value prediction tasks, it can be defined as the prediction residual, such as the difference between the predicted value and the reference value.

[0105] The system selects the appropriate non-conformal score definition according to the detection task and calculates its distribution quantile in the window W. When the non-conformal score of the current sample is greater than or equal to , it is determined to be unreliable or to be rejected, thereby realizing a dynamic threshold discrimination with controllable coverage.

[0106] ​Step S7.3: OOD identification: maintain class centers in the intermediate feature space With covariance , compute Mahalanobis distance or energy score, where z is the feature representation of the sample to be identified in the model's intermediate layer; k is the class index. When exceeds the historical quantile threshold, it is identified as OOD. OOD samples directly go through the "full cycle -> resampling" path and do not enter incremental learning.

[0107] As shown in Figure 4 , step S7.4: resampling control and bottom-up.

[0108] Triggering conditions: consecutive events of rejection / OOD / health score anomaly, or key indicators crossing adaptive thresholds (historical quantile setting).

[0109] Resampling strategy: after entering resampling, the sampling time or pump flow can be moderately extended within the safety / compliance range; rejudge after completing the full cycle.

[0110] Collaborative control: in the resampling stage, issue the "prohibit early stopping / increase threshold" instruction to the early stopping unit to prioritize reliable conclusions; if necessary, recommend extending the post-cleaning or increasing the high-temperature cleaning intensity.

[0111] As shown in Figure 5 , step S8: perform constrained incremental learning and rollback (offline / semi-online) at the edge AI end.

[0112] Step S8.1: sample screening: only samples with "high confidence + passing S1-S8 quality control" are included in the candidate set; use a replay buffer to maintain class balance and old knowledge.

[0113] Step S8.2: light update: prefer to fine-tune only the final classification head / normalize statistics; if full network fine-tuning is required, introduce elastic weight consolidation (EWC) constraints, and the loss function is

[0114]

[0115] where L CE is the cross-entropy loss; λ is a hyperparameter; θ j represents the value of the jth parameter in the current model; represents the original value of the jth parameter after the old task training is completed.

[0116] Step S8.3: local verification and rollback: after increment, monitor accuracy, ECE, and average early stopping time on the local validation set; when ΔAcc is lower than the threshold (such as ) or the rejection rate rises more than the threshold (such as >10%), rollback to and empty the recent update sample pool.

[0117] Step S8.4: Version and audit: model and calibration configuration versioning (v1, v2…), record timestamp, trigger reason and index change and encrypted storage.

[0118] Data path (example): ADS1256 → feature extraction → AI model inference → output prediction and confidence; when confidence ≥ threshold (such as 0.90) trigger early stop, otherwise continue sampling.

[0119] As shown in Figure 3 Step S9: closed-loop collaborative control of dynamic temperature control and early stop strategy.

[0120] Time dimension collaboration: dynamic temperature control is used to compress the preheating and cleaning period, and early stop is used to compress the adsorption detection period; both complement each other to achieve full-cycle acceleration.

[0121] Trigger collaboration: when early stop outputs in advance, dynamic temperature control immediately switches to the high-temperature desorption stage; if early stop is not up to standard, the complete temperature control curve is maintained.

[0122] Safety collaboration: when the health score is low or the rejection / OOD occurs frequently, automatically raise the early stop threshold or temporarily disable early stop; when the environment and health recover to stability, restore the default threshold.

[0123] Step S10: structure optimization and power management of the system.

[0124] Shunting and uniformity: the inlet end is divided into eight shunts, and CFD optimizes the shunt angle and pipe resistance, so that the flow difference is preferably not more than 5%, to improve the response synchronicity of multiple channels and enhance the availability of early features.

[0125] Low adsorption and small dead space: the valve is close to the air chamber inlet, shortening the pipeline; the inner wall of the gas path is coated with a low adsorption coating; the dead space volume is preferably not more than 0.5 mL.

[0126] Modularization and miniaturization: pump, valve, air chamber, array and battery integrated packaging; array modularization facilitates replacement or upgrade.

[0127] PCB three-domain isolation: analog acquisition, digital AI and power drive are physically partitioned to improve anti-interference and signal-to-noise ratio.

[0128] Power consumption and safety: battery management monitors power consumption and supports low-power mode; high-temperature area implements heat insulation, and the whole machine meets the handheld volume and weight; landscape UI supports one-handed operation.

[0129] Step S11: adjust the parameter range and working mode according to the running state.

[0130] Temperature control platform: Low-temperature detection platform and high-temperature desorption platform use "range + preferred + safety upper limit" setting (for example, low temperature is located at 200-300℃, high temperature is located at 350-500℃ and does not damage the device).

[0131] Early exit: Set multiple candidate points in the 1-6s interval (preferably 2-5s), and adapt to the scene.

[0132] Confidence threshold (optional mode): The emergency rapid mode threshold can be set to 0.75; the medical screening mode can be set to 0.90.

[0133] Conformal window and confidence level: The window W is of the order of several tens to several hundreds of frames, and ε is of the order of several percentage points; typical settings are given by the examples.

[0134] Health score weight: ∈[0,1] and verified to determine.

[0135] Anchor point period: Use a dual strategy of "fixed interval + event trigger".

[0136] Cleaning strategy dynamic: Dynamically match temperature (300-500℃) and time (10-20s) according to pollution index Q.

[0137] Resampling: Extend the sampling time or adjust the pump flow within the safety / compliance range to enhance the features.

[0138] Example 1: Entry port fruit fly infestation rapid screening system

[0139] 1. Application scenario

[0140] This example is applied to the customs entry port, aiming at the on-site rapid screening of fruit fly pests (Bactrocera dorsalis, Bactrocera correcta) in tropical fruits (mango, lotus, guava).

[0141] As shown in Figure 9 , in the batch quarantine scene, the electronic nose device 901 can work with the existing conveyor belt 903 and mechanical arm 902: the fruit sample 904 enters the sampling position of the electronic nose device 901 with the conveyor belt 903, the system completes the rapid discrimination and transmits the result to the mechanical arm 902, and the mechanical arm 902 sorts the sample 904 into different sample boxes 905, realizing the closed-loop process of "automatic conveying- rapid detection- immediate sorting".

[0142] 2. System configuration

[0143] Sensor array: 8-channel MOS gas sensor (preferably TGS2602, TGS2620);

[0144] Dynamic temperature control unit: detection interval is preferably 250±10℃, and pulse heating at 450℃ is used in the cleaning stage, with a duty cycle of about 1:3;

[0145] Edge AI model: based on a multi-outlet TCN structure, with outlet points set at three key feature points:

[0146] Feature point one: maximum change rate point Max (ds / dt), i.e. the inflection point of rapid signal rise, as shown in Figure 6 ;

[0147] Feature point two: response peak point, i.e. the point at which the signal amplitude is maximum;

[0148] Feature point three: minimum change rate point Min (ds / dt), i.e. the slow change point in the falling stage;

[0149] Drift compensation module: dynamic coefficient update based on Kalman filtering, for real-time correction of temperature and humidity fluctuations.

[0150] Automatic collaboration: the electronic nose device 901 outputs detection results through the edge AI processing module 108, communicates with the mechanical arm 902 and the conveyor belt 903, and drives the mechanical arm 902 to sort the sample 904 into different sample boxes 905.

[0151] 3. Detection process

[0152] Step S1: initialization;

[0153] Fast preheating: after the device is powered on, the sensor is quickly heated to the detection temperature zone (about 250±10℃) by pulse heating (preferably 1-3 seconds), shortening the preheating time;

[0154] Baseline calibration: after the sensor reaches the detection temperature zone, air is introduced as a clean gas, and the initial baseline signal is collected and a baseline file is established (including noise level and dynamic response time constant).

[0155] Step S2: sample detection and early stop judgment;

[0156] The fruit sample 904 enters the sampling position of the electronic nose device 901 with the conveyor belt 903; the air pump of the electronic nose device 901 operates in pulse mode (preferably 2000 mL / min), and the signal is input into the AI model in real time; the system outputs a judgment result at feature point one (about 1-2 s), feature point two (about 3-5 s), or feature point three / complete cycle; the judgment result is transmitted to the mechanical arm 902, and the mechanical arm 902 performs sorting according to the result.

[0157] Step S3: cleaning and compensation;

[0158] Immediately after the discrimination is completed, switch to high-temperature pulse cleaning (preferably 450°C), while air is introduced to flush the sensor cavity, remove residual gas and restore the baseline. The drift compensation module updates the parameters synchronously.

[0159] Step S4: Sorting coordination, as shown in Figure 9 ,

[0160] Normal result: sample 904 continues to flow with the conveyor belt 903, or is placed in the qualified sample box 905 by the mechanical arm 902; suspected result: the mechanical arm 902 rejects the sample 904 to the isolated sample box 905; uncertain result: the system triggers the bottom, and the sample 904 enters the sample box 905 for rechecking.

[0161] (1) Experimental verification data

[0162] As shown in Figure 7 , the accuracy rate of the Bactrocera dorsalis regular decision point is 95.5%; the earliest decision point accuracy rate is 90.9%; and the average early discrimination time is shortened by 3 / 4 of the cycle.

[0163] As shown in Figure 7 , the accuracy rate of the Bactrocera correcta regular decision point is 93.3%; the earliest decision point accuracy rate is 88.9%; and the average early discrimination time is shortened by about 1 / 2-3 / 4 of the cycle.

[0164] The histogram results show that the average discrimination time of the control group is 4.5s, while the infected group is extended to 7-9.6s in the early, middle and late stages, and the system can trigger early discrimination in 1-2 seconds.

[0165] (2) Rapid detection and coordination schematic diagram

[0166] As shown in Figure 8 , the cycle shortening ratio corresponding to different discrimination points is: about 1 / 4 for feature point one, about 1 / 2 for feature point two, and about 3 / 4 for feature point three.

[0167] As shown in Figure 9 , the electronic nose device 901 cooperates with the conveyor belt 903, the mechanical arm 902 and the sample box 905 to realize the beat detection and automatic sorting of batch fruit samples 904.

[0168] 4. Technical effects

[0169] Full cycle acceleration: the average detection time is shortened from the traditional 182 seconds to less than 38 seconds;

[0170] Automatic processing: through the cooperation of the conveyor belt 903 and the mechanical arm 902, automatic transportation, detection and sorting of samples 904 are realized;

[0171] Reliability assurance: The accuracy rate of the earliest decision point remains above 88%, combined with a complete cycle fallback mechanism and the 905 sample box to be verified, to ensure safety;

[0172] Application flexibility: Sample 904 is not limited to apples and can be a variety of fruits; the number of sample boxes 905 can also be configured according to on-site needs and is not limited by the number of schematic diagrams.

[0173] Example 2: Medical Breath Analysis System

[0174] 1. Application Scenarios

[0175] This embodiment is applied in a medical setting to detect key biomarkers in human exhaled breath, including acetone (aided diagnosis of diabetes), hydrogen sulfide (oral / liver function monitoring), and methane (gut microbiota status). Through individualized incremental learning, multimodal environment compensation, and three-level safety protection, this system achieves high reliability, speed, and clinical adaptability.

[0176] 2. System Configuration

[0177] (1) Sensor array

[0178] The preferred filters are TGS822 (acetone), TGS825 (hydrogen sulfide), and TGS2600 (methane), with an optional version, TGS2611-E00, featuring an anti-alcohol interference filter to reduce ethanol interference with breathalyzer testing.

[0179] (2) Dynamic temperature control unit

[0180] Detection range: 200–230℃, to avoid high-temperature decomposition of biogas;

[0181] Rapid preheating: The temperature rises to the target temperature zone within 2 seconds, ensuring that the patient can be tested after a short exhalation;

[0182] Cleaning temperature: 360–400℃. The cleaning time should be adjusted adaptively according to the frequency of use (e.g., 360℃ / 12 seconds for ≤5 times / day, 400℃ / 18 seconds for >10 times / day) to extend the sensor life.

[0183] (3) AI discrimination and learning module

[0184] Model structure: Multi-exit TCN, with exit points set at 2–4 seconds;

[0185] Incremental learning mechanism: healthy population: fine-tune classification head parameters only (<1% parameter amount); chronic disease patients: fine-tune classification head and BN layer (<5% parameter amount); abnormal samples: trigger gold sample labeling, batch update during idle period. Patient profile binding mechanism: generate patient ID at first detection, associate age, BMI, and basic disease information, and link with incremental learning module to form individualized adaptive model.

[0186] (4) Multi-modal environment compensation module

[0187] Body temperature compensation: correct sensor signal according to patient's sublingual temperature, ;

[0188] Breathing pattern classifier: smooth breathing: use regular sampling; shallow breathing or wheezing: start segmented integration and extend sampling to 8 seconds; cough interruption: automatically discard the cycle and prompt resampling.

[0189] (5) Three-level safety protection system

[0190] Primary protection: dynamically adjust confidence threshold according to use mode (emergency mode threshold 0.75, screening mode threshold 0.90);

[0191] Secondary arbitration: when confidence is between 0.80-0.89, automatically trigger 3-second pulse retest, and use 2 / 3 consistent principle among three detections to output results;

[0192] Ultimate protection: establish pathological abnormality code library, when the result matches the abnormal feature spectrum, the system is forced to enter the manual review process.

[0193] 3. Detection process

[0194] Step S1: initialization;

[0195] System quickly preheats to 200-230°C, purges baseline with air, and calls patient profile parameters (such as body temperature, basic disease information) for individualized baseline correction.

[0196] Step S2: sampling detection;

[0197] The subject exhales through a disposable mouthpiece, with pump flow controlled at 100-150 mL / min, and the sensor collects signals and inputs them into the AI model in real time.

[0198] Step S3: early stop and rejection;

[0199] If the model determines the confidence of the result at the 2-4 second exit point is ≥ threshold, the result is output in advance; if the confidence is in the middle interval, retest is triggered; if the confidence is insufficient, the complete detection cycle (10-12 seconds) is entered.

[0200] Step S4: cleaning recovery;

[0201] After detection is completed, according to the frequency of use, a 360-400℃ pulse cleaning mode is selected, air is introduced to remove residual gas and restore the baseline.

[0202] 4. Technical effects

[0203] Through the above technical solutions, the embodiment achieves:

[0204] Individual adaptation: based on patient records and hierarchical incremental learning strategies, the detection model can complete individual adjustment within 24 hours, significantly improving the adaptability of exhaled breath analysis for chronic disease patients;

[0205] Multi-modal compensation: combined with body temperature and breathing pattern correction, avoid interference from shallow breathing, wheezing or cough interruption, improve detection success rate;

[0206] Three-level safety protection: dynamic threshold, retest arbitration and abnormal feature library triple mechanism, reduce the risk of misjudgment, ensure the clinical safety of medical scenes;

[0207] Hardware life optimization: through alcohol interference filter cover and hierarchical thermal management strategy, sensor life is improved from 1800 times to more than 3500 times, reducing equipment maintenance cost.

[0208] Embodiment three: environmental emergency detection system

[0209] 1. Application scenario

[0210] The embodiment is used for public safety and emergency monitoring, including subway stations, factories, warehouses, tunnels and chemical parks, etc. closed or semi-closed environments, for rapid screening and hierarchical warning of dangerous gases such as carbon monoxide (CO), benzene series (such as benzene, toluene, xylene) and ammonia (NH3) in a closed or semi-closed environment, adapting to sudden leakage, dangerous operation inspection and fire rescue scenes.

[0211] 2. System configuration

[0212] (1) Sensor array

[0213] CO / Channel: preferably TGS2610, TGS2611, with additional molecular sieve filter membrane in the CO channel to improve selectivity for ethanol / formaldehyde;

[0214] Benzene series channel: preferably PID (photoionization) sensor, achieving ppb level detection capability;

[0215] Or other toxic gas channel: can be equipped with MOS or electrochemical channel;

[0216] ​Channel differential maintenance: MOS channel supports high-temperature cleaning; PID channel performs light source calibration and zero-point calibration, and does not participate in high-temperature cleaning.

[0217] (2) Dynamic temperature control and cleaning life management

[0218] MOS detection interval: 280-320℃; Fast preheating time is preferably 1-2 seconds;

[0219] Adaptive pulse heat cleaning protocol:

[0220] Routine cleaning: 480℃, about 15 seconds, continuous mode;

[0221] Severe pollution: 500℃, about 8 seconds, pulse mode (preferably 5Hz);

[0222] Daily maintenance: 450℃, about 20 seconds, gradient heating;

[0223] Health status monitoring: Real-time calculation of quality factor Q, maintenance alarm is triggered when Q continuously decreases by more than 30%.

[0224] (3) Extreme environment adaptation module

[0225] Hardware protection:

[0226] The gas chamber is equipped with a PTC heating sheet (maintaining 60±5℃) to prevent condensation;

[0227] The pipeline uses a nano-hydrophobic coating (contact angle preferably >150°) to reduce moisture adsorption;

[0228] Environmental sensing: Integrated temperature and humidity sensor, sampling rate preferably ≥10Hz;

[0229] Algorithm compensation:

[0230] High humidity conditions (RH>80%): Characteristic vector ×[1–α·(RH–80)], where α is a slowly varying compensation coefficient;

[0231] Low temperature conditions (T<10℃): Start preheating compensation mode, first heat to 50±5℃ constant temperature, then enter normal sampling.

[0232] (4) Anti-crossing and interference module

[0233] Hardware suppression: CO channel molecular sieve filter membrane; benzene series channel uses PID instead of MOS;

[0234] Algorithm deduction: Construct an interference feature library (ethanol, formaldehyde, acetone, etc.), and subtract according to the cross-sensitivity coefficient k i Real-time correction

[0235]

[0236] where C i k is the estimated interference concentration i is the experimental calibration value.

[0237] (5) AI model and hierarchical early warning coordination

[0238] Discrimination model: SPRT + multi-outlet TCN, early stop outlet preferred in 1-3 seconds;

[0239] Hierarchical early warning mechanism:

[0240] Level III (warning): TLV < concentration < STEL → device flicker prompt;

[0241] Level II (medium risk): STEL ≤ concentration < IDLH → local audible and visual alarm + cloud push;

[0242] Level I (high risk): concentration ≥ IDLH → strong audible and visual alarm + linkage exhaust or power off;

[0243] Networking coordination: based on positioning information to automatically generate pollution diffusion model, and issue adjacent devices to enter "high sensitivity mode".

[0244] 3. Detection process

[0245] Step S1: initialization and health check;

[0246] MOS channel fast preheating to 280-320℃; PID channel performs light source stabilization and zero point calibration; air is introduced to refresh the baseline, temperature and humidity are collected and initial compensation is performed.

[0247] Step S2: high flow sampling;

[0248] Start high flow pump (preferably 300 mL / min), gas enters the gas chamber, PTC keeps 60±5℃, hydrophobic coating reduces condensation residue.

[0249] Step S3: anti-interference deduction and early stop discrimination;

[0250] The system performs real-time correction according to the interference library, and sends the net response to SPRT / TCN:

[0251] If the confidence threshold is reached in 1-3 seconds, the hierarchical early warning is issued in advance;

[0252] Otherwise, enter the complete cycle (10-15 seconds) to complete the discrimination.

[0253] Step S4: hierarchical early warning and coordination n;

[0254] According to the TLV / STEL / IDLH rule output warning level, and can be linked to exhaust / power; while the data upload cloud, drive adjacent equipment into high sensitivity monitoring.

[0255] Step S5: cleaning and life management;

[0256] MOS channel performs adaptive pulse thermal cleaning; PID channel performs zero calibration; update Q value and write file, if the threshold trigger maintenance alarm if falling.

[0257] 4. Technical effects

[0258] Anti-interference: hardware filter + PID + interference deduction algorithm cooperation, greatly reduce the false report of ethanol, formaldehyde background.

[0259] Extreme environmental adaptability: PTC heating, hydrophobic coating and high humidity / low temperature compensation algorithm, to ensure that the system can still work stably in high humidity and low temperature environment.

[0260] Fast warning ability: early stop export can trigger warning in 1-3 seconds, meet the real-time requirements of public safety.

[0261] Hierarchical response and linkage: three-level warning mechanism based on TLV / STEL / IDLH, and support multi-device cooperation, realize diffusion tracking and emergency disposal.

[0262] Life extension and maintenance convenience: through pulse thermal cleaning and Q value monitoring, MOS channel life is improved to about 5000 cycles, and maintenance prompt is provided.

[0263] Those skilled in the art can understand that the features described in various embodiments of the present disclosure can be combined and / or combined in various ways, even if such combinations and / or combinations are not explicitly described in the present disclosure. In particular, the features described in various embodiments of the present disclosure can be combined and / or combined in various ways without departing from the spirit and teachings of the present disclosure. All these combinations and / or combinations fall within the scope of the present disclosure.

[0264] Although the present disclosure has been shown and described with reference to certain exemplary embodiments thereof, it should be understood by those skilled in the art that various changes in form and detail can be made therein without departing from the spirit and scope of the present disclosure as defined by the appended claims and their equivalents. Therefore, the scope of the present disclosure should not be limited to the above-described embodiments, but should be determined only by the appended claims, and should be limited by the equivalents of the appended claims.

Claims

1. A handheld AI electronic nose system, characterized in that, The invention includes a handheld AI electronic nose device, comprising: an electronic nose device (901), including an environmental parameter monitoring unit (101), a multi-channel metal oxide gas sensor array unit (102), an integrated gas path-gas chamber assembly (103), an integrated pump valve module (104), a display and input module (105), an acquisition and execution control board (106), a high-resolution analog-to-digital conversion module (107), an edge AI processing module (108), and a battery and management module (109). The environmental parameter monitoring unit (101) is located in the central mounting position of the multi-channel metal oxide gas sensor array unit (102). The multi-channel metal oxide gas sensor array unit (102) is installed inside the gas path-gas chamber integrated assembly (103). The integrated pump valve module (104) is connected to the gas path-gas chamber integrated assembly (103). The acquisition and execution control board (106) is connected to the environmental parameter monitoring unit (101) and the multi-channel metal oxide gas sensor array unit (102). The high-resolution analog-to-digital conversion module (107) is located above the acquisition... Connected to the execution control board (106), the analog data acquired by the acquisition and execution control board (106) is converted into digital data, and the converted digital data is transmitted to the edge AI processing module (108) connected below. The edge AI processing module (108) connects to and sends control commands back to the acquisition and execution control board (106). The battery and management module (109) provides energy and safety management for the electronic nose device (901). The display and input module (105) displays the mode, temperature control curve, progress, results and prompts of the electronic nose device (901). The system is functionally divided into a perception layer, a decision-making layer, and a control layer, which interact through a unified message / event bus. The decision-making layer consists of a dynamic temperature control and early stop discrimination unit, an anti-drift and self-calibration unit, and an edge AI and local incremental learning unit. The dynamic temperature control and early stop discrimination unit, based on the temperature control curve issued by the acquisition and execution control board (106), sequentially executes the pulse rapid preheating, low-temperature adsorption detection, two-stage early stop, and high-temperature desorption cleaning processes; when the threshold is met, it generates a discrimination conclusion and confidence level, and sends a trigger signal for starting cleaning to the control layer; including: The response signal was acquired on a cryogenic platform, and features including early slope, amplitude, and cross-channel ratio were extracted from the response signal. Perform statistical early stopping detection based on sequential probability ratio test: calculate log-likelihood ratio in real time. Defined as: in, This represents the alternative assumption that the target gas exists. This indicates the null hypothesis that the target gas does not exist. This represents the k-th observation. when Accept in advance when ≥logA ,when Accept when ≤logB ;when When in the hysteresis interval, stabilization is achieved by combining sliding window consistency and double threshold hysteresis. Multiple exit deep early stop: Based on a temporal convolutional network, multiple exit points are set within a preset response time interval. When the discrimination confidence of any exit reaches a threshold, the output is output in advance; the position of the exit is adjusted according to the scenario. Perform fallback complete acquisition: If any of the above early stop is not triggered or the confidence level does not meet the preset conditions, continue to acquire the response signal until the predefined acquisition period.

2. The handheld AI electronic nose system according to claim 1, characterized in that, The environmental parameter monitoring unit (101) is used to collect environmental parameters as inputs for online compensation and self-calibration. The environmental parameters include temperature, humidity, flow rate and air pressure. Each channel of the multi-channel metal oxide gas sensor array unit (102) has independent heating and independent temperature measurement, which is used to perform differentiated dynamic temperature control; The integrated gas path-gas chamber assembly (103) includes an air inlet (201), a central channel, an air outlet (206), and at least one sampling microcavity (202). The sampling microcavity (202) is provided with a sampling microcavity inlet (203) and a sampling microcavity outlet (204). A sensor mounting interface (205) is provided on the lower inner side of the sampling microcavity (202). A sensitive layer of a multi-channel metal oxide gas sensor array unit (102) is provided at the sensor mounting interface (205) to ensure that the gas acts on the sensitive layer in the sampling microcavity (202). The integrated pump and valve module (104) is connected to the integrated gas path-gas chamber assembly (103) and supports constant flow sampling, pulse sampling and cleaning modes; The acquisition and execution control board (106) is responsible for the initial acquisition of sensor signals, the execution of array heating curves and the control of pump valve drive, and is synchronized with the high-resolution analog-to-digital conversion module (107); The high-resolution analog-to-digital converter module (107) is used to perform high-precision multi-channel synchronous sampling of the output of each channel of the multi-channel metal oxide gas sensor array unit (102) to ensure the quality of weak signals in the early stage. The edge AI processing module (108) is used to perform model inference, confidence calibration and constrained incremental learning locally, and to perform versioning and rollback. The battery and management module (109) is used to provide power, charge / discharge and safety management, and supports low-power sleep mode; The display and input module (105) is used for mode selection and result display.

3. The handheld AI electronic nose system according to claim 2, characterized in that, The sensing layer consists of the environmental parameter monitoring unit (101), the multi-channel metal oxide gas sensor array unit (102), the gas path-gas chamber integrated component (103), the integrated pump valve module (104), and the high-resolution analog-to-digital conversion module (107), and is used to complete gas sampling and signal acquisition. Gas introduction and distribution: External gas enters through the integrated pump and valve module (104), wherein the gas pump and the gas valve are integrated into a compact structure. The gas pump is located on the left side of the integrated pump and valve module (104), and the gas valve is located on the right side of the integrated pump and valve module (104). The gas pump provides the driving force for gas flow, and the gas valve plays the role of switching and flow path switching. Under the action of the gas pump, the external gas enters the inlet (201) of the gas path-gas chamber integrated component (103) through the gas valve, and then enters the sampling microcavity (202) through the central channel and through the inlet (203) of each sampling microcavity. In the sampling microcavity (202), it fully interacts with the sensitive layer of the multi-channel metal oxide gas sensor array unit (102) located at the sensor mounting interface (205), and flows into the central channel from the sampling microcavity outlet (204) and is discharged through the outlet (206). Signal and environmental quantity acquisition: The high-resolution analog-to-digital conversion module (107) performs high-precision synchronous sampling of the multi-channel sensor signals of the multi-channel metal oxide gas sensor array unit (102); The environmental parameter monitoring unit (101) outputs environmental parameters including temperature, humidity, flow rate and air pressure in real time for subsequent online compensation and self-calibration; The perception layer outputs the generated multi-channel sensor signals and environmental parameters to the decision layer.

4. The handheld AI electronic nose system according to claim 1, characterized in that, The decision-making layer operates under the temperature control / execution of the acquisition and execution control board (106) and with the data support of the high-resolution analog-to-digital conversion module (107); The anti-drift and self-calibration unit takes the environmental parameters of the environmental parameter monitoring unit (101) and the historical baseline of the multi-channel metal oxide gas sensor array unit (102) / high-resolution analog-to-digital conversion module (107) as inputs, performs online environmental compensation, anchor baseline calibration, channel health scoring and gating / masking, confidence calibration, conformal rejection and outlier distribution identification, and outputs parameter update and resampling instructions. The edge AI and local incremental learning unit run on the edge AI processing module (108) to perform model inference, sample screening, constrained incremental updates and version rollback on the preprocessed multi-channel data, and generate model / log versions.

5. The handheld AI electronic nose system according to claim 1, characterized in that, The control layer consists of an acquisition and execution control board (106), a high-resolution analog-to-digital conversion module (107), an edge AI processing module (108), a battery and management module (109), and a display and input module (105), and is responsible for execution, acquisition, energy consumption, and interaction. Execution and Acquisition: The acquisition and execution control board (106) controls the heating and temperature measurement of the integrated pump valve module (104) and the multi-channel metal oxide gas sensor array unit (102), and works in conjunction with the high-resolution analog-to-digital conversion module (107) to achieve multi-channel synchronous sampling and preprocessing; Local computation: The edge AI processing module (108) completes inference, confidence calibration and incremental learning, and sends the control commands back to the acquisition and execution control board (106). Energy consumption and interaction: The battery and management module (109) provides energy and safety management; the display and input module (105) displays the mode, temperature control curve, progress, results and prompts; Data and control path: The target link is a high-resolution analog-to-digital conversion module (107) connected to the acquisition and execution control board (106), the acquisition and execution control board (106) connected to the edge AI processing module (108), and the edge AI processing module (108) connected to the display and input module (105); the control loop is an edge AI processing module (108) connected to the acquisition and execution control board (106), used to trigger pump valve / temperature control actions.

6. A handheld AI electronic nose method, characterized in that, The method comprising the handheld AI electronic nose system according to any one of claims 1 to 5, wherein the method includes: Step S1: Perform the initialization operation of the system and establish a baseline profile; Step S2: Perform sampling start point detection and data preprocessing; Step S3: Perform a pulsed rapid preheating process in the first stage; Step S4: In the second stage, perform low-temperature adsorption detection and execute two-stage early stop judgment; Step S5: Implement online environmental compensation and data normalization processing; Step S6: Perform baseline tracking, anchor point calibration, and channel health status assessment; Step S7: Perform threshold adaptive adjustment, uncertainty control, and unknown class identification; Step S8: Perform constrained incremental learning and rollback operations in the edge AI processing module (108); Step S9: Achieve closed-loop coordinated control of dynamic temperature control and early stop strategy; Step S10: Perform system structure optimization and power consumption management; Step S11: Adjust the parameter range and working mode according to the operating status.

7. The method according to claim 6, characterized in that, The initialization operation of the system and the establishment of a baseline profile include: Step S1.1: Clean Gas Baseline Acquisition: After the device is initially deployed or powered on, the clean gas is automatically switched, and multi-channel signals are acquired for a preset time period to obtain the initial baseline for each channel. Noise flooring Dynamic range scale With rise time constant And simultaneously record the reference environment vector ; Step S1.2: Initial value fitting for environmental compensation: Using the reference environmental vector e0 as a reference, fit the environmental sensitivity coefficients of each channel; Step S1.3: Version management of calibration files: Generate calibration files with timestamps and store them in the database.

8. The method according to claim 6, characterized in that, The execution of baseline tracking, anchor point calibration, and channel health status assessment includes: Step S6.1: Baseline Tracking: Measure and acquire the latest baseline during the interval between the end of post-washing or between two detections. ,noise and time constant ; Calculate drift index ,in, , , The reference benchmark value is used; if any indicator exceeds the adaptive threshold band, the adaptive threshold band is determined based on the quantiles of the historical distribution, the target quantile interval is selected, and it is marked as "deviation" state; Step S6.2: Anchor point calibration: Triggered by a fixed period or the "deviation" condition generated in S6.1, switch to clean gas for short sampling to refresh the reference value. Compared with the reference environment; Step S6.3: Health Score and Gating: Define the health score of channel i at time t. for: in, This is a normalization function that maps the deviation to the interval [0,1]. The cross-correlation consistency score between this channel and the other channels; w1, w2, w3, and w4 are configurable weights; before feature fusion / model input. Channels are weighted; channels whose health scores are continuously below a preset threshold are placed into candidate blocking status and the event is reported.

Images