Artificial intelligence olfaction system based on heterogeneous sensor array

Abstract

The invention relates to the technical field of artificial intelligence olfaction, and provides an artificial intelligence olfaction system and method based on a heterogeneous sensor array. In order to solve the problems of incomparability of cross-mechanism response time sequence and quantification and environmental drift interference, the method comprises the following steps: simultaneously acquiring sensing units of at least two mechanisms in the same sampling cavity and under a unified clock, completing denoising, baseline calibration, scale normalization and time alignment in a unified time window, and generating a concentration reference value according to a calibration relationship of quantitative channels such as PID / NDIR; characteristics such as steady state, dynamics and temperature segmentation ratios are extracted based on a standardized time sequence, health degree gating and weighted fusion of noise, baseline drift and saturation signs are combined, a joint model is input, and meanwhile, odor categories, target component concentrations and confidence coefficients are output, so that the method is suitable for complex mixed gas and low-concentration targets. And the cross-mechanism quantitative consistency and robustness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sensor data analysis, and in particular to an artificial intelligence olfactory system based on a heterogeneous sensor array. BACKGROUND

[0002] The artificial intelligence olfactory system is used for environmental monitoring, industrial safety, agriculture and other scenarios, and realizes identification and quantification of odor / gas through a sensor array and an algorithm. The industry needs a data control capability that can uniformly collect, fuse and analyze, and output standardized results, and that can balance classification and quantification for direct calling by a higher-level system.

[0003] Existing technologies are mostly single-mechanism or loosely coupled multi-mechanism solutions: for example, a metal oxide array focuses on pattern recognition, electrochemistry or photoionization is used for quantification, and a non-spectroscopic infrared device independently outputs concentration. Each device is mostly separately sampled and processed, and the time reference, dimension and field are inconsistent, making it difficult to form comparable results under the same gas sample; secondary integration and manual checking are often required for higher-level applications, and there is a lack of a “central control” that can unify classification and quantification and stably publish structured results. At the same time, environmental temperature and humidity effects and long-term drift are mostly calibrated after the fact, and there is a lack of cross-scenario reusability and interface standardization.

[0004] The above-mentioned solutions have the defect that they cannot be used as a field “central control” to uniformly provide reliable data to the outside: different channels lack synchronization in the same cavity and the same clock, and classification and quantification are disjointed in time and dimension; anomalies and drifts lack pre-control, and the output is difficult for external systems to directly consume and trace.

[0005] Therefore, there is an urgent need in the art for an artificial intelligence olfactory central control that can face a multi-mechanism sensor array in a real scenario, form comparable inputs under the same sampling cavity and a unified clock, complete integrated processing of classification and quantification, and output categories, concentrations and confidence levels to the outside in the form of standardized data packets. SUMMARY

[0006] One object of the present application is to provide an artificial intelligence olfactory system and method based on a heterogeneous sensor array, which can form comparable inputs under the same sampling cavity and a unified clock, complete integrated processing of classification and quantification, and output categories, concentrations and confidence levels to the outside in the form of standardized data packets.

[0007] An artificial intelligence olfactory system based on a heterogeneous sensor array according to an embodiment of the present application comprises:

[0008] A data acquisition and array management module is configured to introduce a gas sample and acquire an array original response sequence according to a unified clock, and record a temperature sequence, reference channel intensity, measurement channel intensity and optical path parameters;

[0009] A unified time window and scale establishment module is configured to align channels and unify dimensions to generate standardized timing data;

[0010] A concentration reference value calculation module is configured to calculate a concentration reference value and a representative quantity of a stable section thereof according to a reference channel intensity, a measurement channel intensity, an optical path and a calibration coefficient;

[0011] A fusion feature construction module is configured to extract a steady-state amplitude, a relative change, a response time, a recovery time and a temperature segmented ratio and splice them into a fusion feature vector;

[0012] A channel health assessment and weight generation module is configured to assess noise, drift, saturation and consistency alignment, generate a channel weight and output a purified fusion feature;

[0013] A joint inference and confidence generation module is configured to output an initial category and concentration result and generate a confidence in a single inference cycle;

[0014] A temperature and humidity covariant and drift compensation module is configured to determine an environmental correction coefficient according to a calibration table, implement consistency correction and fine tuning in combination with zero point and proportional correction parameters and output a compensated result;

[0015] A result packaging and communication module is configured to package a final category, a final concentration and a final confidence into a result data packet according to a time marker and output the result data packet through a communication interface;

[0016] A calibration and parameter management module is configured to uniformly manage an absorption coefficient, an optical path, a gain matching, a temperature boundary point, a threshold, a conversion coefficient and a confidence weight.

[0017] An artificial intelligence olfaction method based on a heterogeneous sensor array according to an embodiment of the application is applied to the artificial intelligence olfaction system based on the heterogeneous sensor array in claim 1, and characterized in that the method comprises:

[0018] S1, introducing a gas sample into a sampling cavity and collecting responses of the heterogeneous sensor array in the same sampling cavity with a unified clock, the array comprising at least two different mechanism gas sensing units, at least two of which are selected from metal oxide, electrochemistry, photoionization, colorimetry and non-spectroscopic infrared, wherein the metal oxide is implemented with temperature modulation and records a temperature sequence, the electrochemistry works in a preset noise threshold and a bandwidth range through a transimpedance amplifier, the non-spectroscopic infrared enables a reference channel for double-channel ratio and optical path calculation, and an array original response sequence is output;

[0019] S2, taking the array original response sequence as input, completing denoising, baseline calibration, scale normalization and time alignment in a unified time window, and obtaining a concentration reference value based on a non-spectroscopic infrared reference and measurement channel ratio in combination with an optical path parameter, and outputting standardized timing data;

[0020] S3. Using standardized time-series data as input, extract the steady-state amplitude, relative change, dynamic indicators of response and recovery, and temperature modulation segmentation ratio, and incorporate the concentration reference value to form a fusion feature vector, and output the fusion feature vector.

[0021] S4. Using the fused feature vector as input, perform abnormal channel removal and weighting based on channel health indicators including noise, baseline drift and saturation signs. Use a weighted fusion strategy that includes at least one of weighted voting or attention weighting to obtain the purified fused features, and output the purified fused features.

[0022] S5. Using the purified fusion features as input, input the fusion model to jointly output the odor category and the concentration and confidence level of at least one target component, and output the initial results of category and concentration.

[0023] S6. Using the initial results of category and concentration as input, perform consistency correction by combining drift compensation and temperature and humidity covariates, and use the concentration reference value as a quantitative benchmark for verification and fine-tuning, and output the compensated results.

[0024] S7. Using the compensated result as input, output the category, concentration, and confidence level through the communication interface to obtain the result data packet.

[0025] Optionally, the specific method of acquiring the response of the heterogeneous sensor array with a unified clock in the same sampling cavity is as follows: the gas sample is introduced into the sampling cavity, so that at least two of the sensing units in the array, including metal oxide, electrochemical, photoionization, colorimetric, and non-dispersive infrared, are simultaneously exposed under the same gas sample conditions. The acquisition controller distributes the same clock and sampling trigger to each channel, and assigns a unified time mark to each sampling, thereby forming the array's original response sequence for the same time and the same gas sample.

[0026] The array contains at least two different gas sensing units with different mechanisms. Specifically, it combines sensing units based on complementary mechanisms such as resistance change, Faraday current, photoionization current, visible or near-infrared absorption, and infrared absorption within the same sampling cavity to cover responses to different chemical reactions and optical absorption paths and provide distinguishable signal sources.

[0027] The specific method for implementing temperature modulation and recording temperature sequences for the metal oxide is as follows: a segmented or continuous temperature trajectory is applied to the heater of the metal oxide sensing unit. The trajectory consists of a starting temperature, an ending temperature, a change mode, and a duration. The temperature sequence is recorded synchronously under this trajectory, and the corresponding response is obtained, so that the adsorption and reaction processes at different temperature ranges are explicitly encoded as auxiliary inputs for subsequent steps.

[0028] Optionally, the concentration reference value is obtained by reading the intensity values ​​of the reference channel and the measurement channel within a unified time window in the non-dispersive infrared optical path, calculating the intensity ratio of the two channels, and mapping them according to the optical path parameters and device calibration coefficients to generate a concentration reference value sequence marked with time.

[0029] Optionally, the temperature modulation segmentation ratio specifically involves: dividing standardized time-series data into multiple temperature segments according to preset temperature boundaries based on the temperature sequence; extracting the mean or final value of the response of the metal oxide channel within each temperature segment; and forming a segmentation ratio set by the ratio of the responses of adjacent temperature segments or preset combination temperature segments. The segmentation is divided into multiple temperature segments with the starting temperature, ending temperature, and change mode in the temperature sequence as boundaries. The segmentation ratio is recorded synchronously with a unified time stamp. This is used to distinguish the response modes corresponding to different adsorption and reaction barriers in complex mixed gas scenarios, so that the temperature-related information of the same channel is explicitly encoded by the segmentation ratio and can be compared with other mechanism channels.

[0030] The fusion feature vector is a single input structure formed by classifying and splicing the features according to the mechanism under a unified scale. It includes the temperature modulation segment ratio and steady-state amplitude of the metal oxide channel, the steady-state amplitude and recovery time of the electrochemical channel, the response peak and recovery slope of the photoionization channel, the color change of the colorimetric channel, and the concentration reference value of the non-dispersive infrared channel. The features of each channel are composed in a fixed positional order and are labeled with the channel. The concentration reference value is embedded in the same vector as a quantitative anchor to support the subsequent classification and concentration estimation in the fusion model. This provides comparable and verifiable cross-mechanism input in business scenarios with complex mixed gases and low-concentration targets.

[0031] Optionally, the abnormal channel elimination and weight reduction based on channel health indicators specifically involves: calculating the noise level, baseline drift, and saturation signs for each channel within a unified time window to obtain a health score; comparing the health score with a preset elimination threshold and a weight reduction threshold; marking channels below the elimination threshold as abnormal and deleting them from the fused feature vector; retaining channels between the two thresholds but assigning them reduced channel weights; assigning standard weights to channels above the weight reduction threshold; and correspondingly saving the channel weights to the feature segments of the corresponding channels in the fused feature vector as channel weight vectors.

[0032] A weighted fusion strategy is adopted to weight and combine feature segments in the channel dimension: when weighted voting is used, the channel weight vector is normalized and then the same type of feature fields are weighted and aggregated to form a channel aggregation field. When attention weighting is used, the channel correlation between the channel weight vector and the fused feature vector is used as the input of the attention score. The channel attention weight is calculated and the feature segments are weighted in the channel dimension, giving priority to strengthening channel segments with high consistency with the concentration reference value.

[0033] Optionally, the fusion model is a joint model that simultaneously completes category determination and concentration estimation in one inference, comprising three parts: a shared input mapping layer, a discriminative branch, and a quantitative branch;

[0034] The shared input mapping layer receives the cleaned fusion features and maintains the channel identifiers and field order in consistency.

[0035] The discrimination branch uses the odor category set as the output space to classify the purified fusion features;

[0036] The quantitative branch uses the concentration reference value as the quantitative anchor point to establish a mapping relationship between the purified fusion characteristics and the concentration of the target component, and outputs the concentration of at least one target component;

[0037] The discrimination branch and the quantitative branch are linked through consistency constraints within the same inference process. The consistency constraints are based on the concentration reference value and the channel weight vector to check the compatibility of the outputs of the two branches.

[0038] The confidence level is generated by combining the verification results of the consistency constraints with the matching degree of the channel weights, and is output simultaneously with the category and concentration to form the initial results of the category and concentration, which are used to provide usable judgment and quantitative basis in business scenarios of complex gas mixtures and low-concentration targets.

[0039] Optionally, the consistency correction combining drift compensation and temperature and humidity covariates is as follows: taking the initial results of category and concentration as input, simultaneously reading the temperature and humidity covariates and concentration reference values, and aligning the time stamps of the three within a unified time window;

[0040] The sensitivity correction coefficients for each channel were obtained by looking up tables based on the temperature and humidity covariates, and the zero-point and proportional correction parameters of drift compensation were applied to correct the contributions of each channel in the initial results of category and concentration.

[0041] Using the concentration reference value as the quantitative benchmark, the corrected concentration is scaled and the quantitative contribution of each channel is weighted and merged according to the channel weight vector, so that the quantitative result is consistent with the concentration reference value within the preset consistency threshold.

[0042] For category determination, if the revised quantitative value is out of bounds or inconsistent with the category label, the category is adjusted or marked as requiring review according to the consistency rule, and the confidence level of the determination is updated simultaneously.

[0043] The beneficial effects of this invention are:

[0044] (1) This scheme collects at least two mechanisms, including metal oxide, electrochemical, photoionization, colorimetric, and non-dispersive infrared, within the same sampling chamber, unified clock, and unified time window, forming standardized time-series data comparable across mechanisms. It generates time-stamped concentration reference values ​​by combining the non-dispersive infrared reference / measurement dual-channel ratio with optical path parameters, and uses these values ​​as quantitative anchors in feature extraction, fusion, inference, and post-correction stages, ensuring that classification and quantification remain linked and consistent within the same process. This design avoids the secondary integration and decoupling problems caused by inconsistent dimensions and time scales across different channels in existing schemes, facilitating direct access by external hosts and monitoring platforms.

[0045] (2) At the feature level, in addition to steady-state amplitude, relative change, and response / recovery kinetics, segmented ratios modulated by metal oxide temperature are used to explicitly encode the adsorption and reaction processes, allowing temperature-related information in the same channel to be compared with other mechanisms. At the fusion level, channel health indicators such as noise, baseline drift, and saturation signs are introduced to perform abnormal channel removal and weighting. Weighted combinations are then performed at the channel dimension using weighted voting or attention weighting, prioritizing the strengthening of channel segments with high consistency with concentration reference values. Compared to fusion strategies that rely on experience or average weighting, this scheme effectively suppresses the impact of distorted channels on results in mixed gas and low-concentration scenarios, improves the consistency and usability of cross-mechanism inputs, and supports the stable operation of the central control system from the source.

[0046] (3) At the model and output level, a joint structure of shared input mapping layer, discriminative branch and quantitative branch is adopted. In one inference, the odor category, target component concentration and confidence level are output simultaneously, and consistency constraints are implemented with concentration reference value and channel weight vector. Then, consistency correction is performed by combining temperature and humidity covariates and drift compensation, and quantitative benchmark is used for verification and fine-tuning with concentration reference value. Finally, the results are packaged into a result data package carrying category, concentration and confidence level under a unified time scale. This process realizes the standardized release of artificial intelligence olfactory central control to the outside world, reduces the splicing of different mechanism data and manual verification by the upper system, and achieves scenario-based integrated processing and traceable output of classification and quantification compared with the existing technology. Attached Figure Description

[0047] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0048] Fig. 1 This is a flowchart of an artificial intelligence olfactory method based on a heterogeneous sensor array proposed in this invention;

[0049] Fig. 2 This is a flowchart illustrating the feature extraction and fusion process of an artificial intelligence olfactory method based on heterogeneous sensor arrays proposed in this invention.

[0050] Fig. 3 This is a flowchart of the joint model inference process for an artificial intelligence olfactory method based on heterogeneous sensor arrays proposed in this invention. Detailed Implementation

[0051] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0052] An artificial intelligence olfactory system based on a heterogeneous sensor array, characterized in that it includes:

[0053] Data acquisition and array management module: Introduces gas samples, unifies clock acquisition of array raw response, includes sampling chamber, flow actuator and clock distribution circuit and firmware, records temperature sequence, output of each channel and quantitative parameters; NDIR records reference / measurement intensity and optical path, PID records ionization current and lamp intensity or baseline;

[0054] The unified time window and scale establishment module is used for channel alignment and unit unification to generate standardized time series data;

[0055] The concentration reference value calculation module is used to calculate the concentration reference value and its representative quantity in the stable segment based on the reference channel intensity, measurement channel intensity, optical path length, and calibration coefficient.

[0056] The fusion feature construction module is used to extract steady-state amplitude, relative change, response time, recovery time, and temperature segmentation ratio and concatenate them into a fusion feature vector;

[0057] The channel health assessment and weight generation module is used to assess noise, drift, saturation, and consistency alignment, generate channel weights, and output the cleaned fusion features.

[0058] The joint inference and confidence generation module is used to output the initial results of category and concentration and generate confidence scores within a single inference cycle;

[0059] The temperature and humidity covariance and drift compensation module is used to determine the environmental correction coefficient according to the calibration table, and to perform consistency correction and fine-tuning in combination with zero point and proportional correction parameters, and output the compensated result.

[0060] The results encapsulation and communication module is used to encapsulate the final category, final concentration and final confidence level into a result data packet according to time stamp and output it through the communication interface;

[0061] The calibration and parameter management module is used to uniformly manage absorption coefficient, optical path length, gain matching, temperature threshold, threshold, conversion factor, and confidence weight.

[0062] refer to Figs. 1 to 3An artificial intelligence olfactory method based on heterogeneous sensor arrays, applied to an artificial intelligence olfactory system based on heterogeneous sensor arrays, characterized by comprising:

[0063] S1. Introduce a gas sample into a sampling chamber and collect the response of a heterogeneous sensor array within the same sampling chamber using a unified clock. The array contains at least two gas sensing units with different mechanisms, selected from at least two of the following: metal oxide, electrochemical, photoionization, colorimetric, and non-dispersive infrared. The metal oxide unit performs temperature modulation and records the temperature sequence, the electrochemical unit operates within a preset noise threshold and bandwidth range via a transimpedance amplifier, and the non-dispersive infrared unit enables a reference channel for dual-channel ratio and optical path calculation, outputting the array's original response sequence.

[0064] S2. Using the original response sequence of the array as input, denoising, baseline calibration, scale normalization and time alignment are completed within a unified time window. Based on the ratio of non-dispersive infrared reference and measurement channels combined with optical path parameters, the concentration reference value is obtained, and standardized time series data is output.

[0065] S3. Using standardized time-series data as input, extract the steady-state amplitude, relative change, dynamic indicators of response and recovery, and temperature modulation segmentation ratio, and incorporate the concentration reference value to form a fusion feature vector, and output the fusion feature vector.

[0066] S4. Using the fused feature vector as input, perform abnormal channel removal and weighting based on channel health indicators including noise, baseline drift and saturation signs. Use a weighted fusion strategy that includes at least one of weighted voting or attention weighting to obtain the purified fused features, and output the purified fused features.

[0067] S5. Using the purified fusion features as input, input the fusion model to jointly output the odor category and the concentration and confidence level of at least one target component, and output the initial results of category and concentration.

[0068] S6. Using the initial results of category and concentration as input, perform consistency correction by combining drift compensation and temperature and humidity covariates, and use the concentration reference value as a quantitative benchmark for verification and fine-tuning, and output the compensated results.

[0069] S7. Using the compensated result as input, output the category, concentration, and confidence level through the communication interface to obtain the result data packet.

[0070] In this embodiment, the specific implementation of step S1 is as follows:

[0071] A gas sample is introduced into the sampling chamber via a gas sampling unit. A heterogeneous sensor array is arranged within the same sampling chamber and data is collected using a unified clock. The sampling frequency of this unified clock is denoted as f. s Discrete-time notation is denoted as {t} k}, where tk Let C be the time marker for the kth sampling. The channel set of the heterogeneous sensor array is denoted as C. It contains at least two different gas sensing units with different mechanisms, selected from metal oxide, electrochemical, photoionization, colorimetric, and non-dispersive infrared. Each channel is simultaneously exposed under the same gas sample conditions and is triggered to sample according to a unified clock, forming a response record oriented towards the same time marker.

[0072] Temperature modulation is applied to the metal oxide channel, and the temperature sequence is denoted as T(t). k V represents the set of time-stamped values ​​of the heater's temperature trajectory during the sampling period. The temperature trajectory consists of a preset starting temperature, ending temperature, and change pattern. These values ​​are recorded synchronously during the sampling process and correspond one-to-one with the response of the channel. The sampling output of the metal oxide channel is denoted as V. MOS (t k ), for time marker t k The response voltage read at the location;

[0073] The electrochemical channel is read out using a transimpedance amplifier and operates within a preset noise threshold and bandwidth range. The output of the transimpedance amplifier is denoted as V. EC (t k ), for time marker t k The voltage signal obtained by electrochemical current conversion is denoted as the noise threshold. The upper limit of allowable noise is denoted as B. EC The above parameters are set during device calibration and remain stable during sampling, representing the passband range of the readout channel.

[0074] The non-dispersive infrared channel enables the reference channel for dual-channel ratio and optical path calculation. The intensity sequence of the reference channel is denoted as I. ref (t k The intensity sequence of the measurement channel is denoted as I. meas (t k Both are at time marker t k The optical path parameter is read from the sample and recorded as L, which is the calibrated value of the optical absorption path length. The intensity sequence of the reference channel and the measurement channel and the optical path parameter are recorded synchronously during sampling for subsequent steps to calculate the concentration reference value and correlate with a unified time scale.

[0075] The sampling output of the photoionization channel is denoted as I. PID (t k ), for time marker t k The ionization current read at the point is denoted as S by the sampling output of the colorimetric channel. COL (t k ), for time marker t kThe optical readout values ​​read from the above two types of channels, together with metal oxide, electrochemical, and non-dispersive infrared channels, constitute the multi-mechanism response source of the array;

[0076] The original response sequence of the array is denoted as R(t). k ), for time marker t k R(t) is a vector set composed of the sampled outputs of each channel in a fixed order. k ) and temperature sequence T(t k Reference channel intensity sequence I ref (t k ), Measurement channel intensity sequence I meas (t k The optical path parameter L is stored synchronously on the same time scale and is used as the output of step S1 for unified time window processing in the next step and subsequent concentration reference value calculation.

[0077] In this embodiment, the specific implementation of step S2 is as follows:

[0078] The original array response sequence R(t) output in step S1 k ), temperature sequence T(t) k Reference channel intensity sequence I ref (t k ), Measurement channel intensity sequence I meas (t k As input, and with the optical path parameter L, a single time scale is established within a unified time window, and channel alignment and dimensional unification are completed. The unified time window is denoted as . Where t s To analyze the starting time, t e To analyze the end time, let the time step be denoted as Δt, and let the discrete time marker set {t} be such that... k} satisfies t k =t s +kΔt, the range of index k is [k s ,k e ], where k s With k e Corresponding to t respectively s With t e ,exist Within, the response trajectories of each channel—metal oxide, electrochemical, photoionization, colorimetric, and non-spectral infrared—are arranged according to {t}. k Time alignment is performed, and standardized time-series data is generated under a unified scale, denoted as X(t). k ), which is at time marker t k A collection of channel responses organized in a fixed field order;

[0079] After establishing a unified time window and time scale, a series of concentration reference values ​​for non-spectral infrared spectroscopy is calculated to provide quantitative anchors for subsequent steps. The concentration reference values ​​are denoted as C. ref (t k Based on the intensity ratio of the reference channel to the measurement channel, combined with the optical path parameters and device calibration coefficients, the expression is as follows:

[0080]

[0081] Where κ is the effective absorption coefficient determined during device calibration, characterizing the ratio of the absorption intensity of the gas being measured to its concentration in the filter band used; L is the optical absorption path length recorded in step S1; η is the gain and spectral matching coefficient between the reference channel and the measurement channel, determined by comparing the reference intensities of the two channels under cavity or standard gas conditions; and I... ref (t k ) for time stamp t k The strength of the reference channel, I meas (t k ) for time stamp t k The intensity of the measured channel is given by ln(·), which is the natural logarithm used to map the intensity ratio to a quantity linearly related to the path length and absorption coefficient. The above C... ref (t k ) and X(t k At the same time scale {t k Synchronize storage under} and maintain synchronization with R(t) k The field and order correspondence of X(t) are established so that they can be uniformly referenced as "concentration reference value" and "standardized time series data" in subsequent steps. Through this implementation method, the output of step S2 includes X(t) k ) and C ref (t k The result set is used as a single scale input and quantitative anchor point for entering step S3.

[0082] In this embodiment, the specific implementation of step S3 is as follows:

[0083] The standardized time series data X(t) output in step S2 k ) and concentration reference value sequence C ref (t k ) is the input, and the time stamp {t} defined in step S1 is used. k}, Channel set C, Metal oxide temperature sequence T(t) k The sampling order of each channel (metal oxide, electrochemical, photoionization, colorimetric, and non-spectral infrared) is represented within a unified time window. Within, for each channel c∈C from X(t) k Read its time stamp t in )k The value at that location is denoted as X. c (t k Its baseline value is denoted as B. c The stable segment index set is denoted as the baseline output of the channel determined by the baseline calibration in step S2. To satisfy the time index set of the stability criterion, the stability criterion is based on the channel change threshold γ. c The definition is used to determine the interval within which a channel enters a steady state within a uniform time window. The temperature segmentation of the metal oxide channel is based on the temperature sequence T(t). k ) and the preset temperature boundary vector Θ=[Θ1,Θ2,…,Θ P Confirmed, Θ p For the p-th temperature boundary point, in order to gather discriminant quantities across mechanisms and under the same scale, a fusion feature vector F is constructed. Its composition consists of steady-state amplitude, relative change, dynamic indicators of response and recovery, temperature modulation segment ratio, and steady-state representative quantities of concentration reference value, which are specifically concatenated in a fixed field order.

[0084] The calculation of the fused feature vector F is expressed as follows:

[0085]

[0086] in, This represents the steady-state amplitude of channel c. Let ε represent the relative change in channel c, and let ε be a small positive number to avoid a denominator of zero. Let be the response time of channel c. To meet The minimum index, Let be the recovery time for channel c. To meet The minimum index, At the end of the exposure phase The time stamp within, This represents the segmentation ratio of the metal oxide channel across adjacent temperature ranges, where δ is a small positive number to avoid a denominator of zero. The average response within the temperature segment p. Let F be the representative value of the concentration reference value within the stable range, where F is the fused feature vector, and X is the value of the reference value within the stable range. c (t k For standardized time-series data, channel c is defined at time marker t. k The value at point B c γ is the baseline value for channel c. c The stability threshold for channel c. V is the set of stable segment indices. MOS (t k ) represents the metal oxide channel at time marker tk The response at that point, where Θ is the temperature boundary vector. Let C be the set of time indices for temperature segments p, where δ and ε are both small positive numbers. ref (t k ) represents the concentration reference value defined in step S2, and t s t e With {t k}Same as step S2;

[0087] In implementation, Θ is selected based on the device's settings for the temperature trajectory of the metal oxide to ensure that each temperature segment covers the main adsorption and reaction processes. Press γ c exist After a unified internal determination, all channels are shared to ensure consistent comparisons across mechanisms. F is organized in a fixed field order and carries channel identifiers so that it corresponds one-to-one with the purification strategy, channel weight vector, and input structure of the fusion model in subsequent steps.

[0088] In this embodiment, the specific implementation of step S4 is as follows:

[0089] Using the fused feature vector F output in step S3 as input, and following the time stamp {t} defined in steps S1 and S2... k}, Channel set C, Stable segment index set Representative amount of concentration reference value within the stable range And the number of temperature segments P of the metal oxide, calculate the channel health index for each channel c∈C within a unified time window. The health index includes noise level, baseline drift and saturation signs, and the noise level is denoted as σ. c For this channel in The root mean square fluctuation on the baseline is denoted as D. c , is the absolute value of the difference between the mean values ​​of the channel during the start and end segments of a unified time window; saturation is denoted as . The consistency alignment ρ is defined as the percentage of time within a uniform time window when the channel is at its upper limit or under-recovery. This ρ is used to measure the compatibility of the channel with the concentration reference value. c , representing the normalized correlation of the channel within a unified time window, based on X. c (t k ) and C ref (t k The time stamps are consistent, and the above health indicators are used to generate the channel weight vector w = {w c} c∈C The abnormal channel segments in F are then subjected to abnormal channel removal and weight reduction, and the purified fused feature F is output. clean Keep the field order consistent with F;

[0090] The joint calculation of channel weights and purified fusion features is expressed as follows:

[0091] For each channel c∈C:

[0092]

[0093] Where, σ max D is the normalized upper limit constant for the noise level. max S is the normalized upper bound constant for baseline drift. max α is the normalized upper limit constant for saturation signs; α1, α2, α3, α4 are weighting coefficients for health indicators and consistency alignment, used to reflect the relative importance of each factor in complex gas mixture scenarios; λ is the weighted normalization temperature parameter used to adjust the sensitivity to scoring; 1[·] is the indicator function, taking a value of one when all conditions within the parentheses are true, and a value of zero otherwise; θ σ θ D θ S These are the removal thresholds for noise level, baseline drift, and saturation indicators, respectively, used for gating the removal and deweighting of abnormal channels. A c Q is the steady-state amplitude of channel c. c This represents the relative change in channel c. Let be the response time of channel c. Let be the recovery time for channel c. w represents the segmentation ratio of the metal oxide channel at temperature p to temperature (p+1). MOS These are the channel weights corresponding to the metal oxide channels. The above F represents the concentration reference value within the stable range. clean The field order and channel identifier are consistent with F. The removed channel segments are gated by an indicator function and then processed by F. clean No contribution is made; the preserved channel fragments are in w c Weighting is performed to form a purified fusion feature and channel weight vector for entering step S5 at the same time scale.

[0094] In this embodiment, the specific implementation of step S5 is as follows:

[0095] The purified fusion feature F output in step S4 clean As input, and simultaneously read the channel weight vector w = {w} obtained in step S4. c} c∈C Representative amount of the concentration reference value defined in step S2 within the stable range The fusion model simultaneously performs category determination and concentration estimation within a single inference iteration, employing a joint structure of a shared input mapping layer, a discriminative branch, and a quantification branch: the shared input mapping layer receives F... clean Maintaining the same field order and channel identifier as in step S3, the branch is identified by odor category set. For the output space, the quantitative branch takes the number of target components M as the output dimension and constrains the output to be non-negative. To provide a reliable confidence level that can be used in complex gas mixture scenarios, the fusion model introduces a consistency measure at the inference end, which integrates the concentration reference value, channel weight distribution and category decision margin to generate a confidence level and output it simultaneously with the category and concentration, forming the initial results of category and concentration.

[0096] The expressions for joint reasoning and confidence calculation are as follows:

[0097] z=φ(F clean ;θ φ );

[0098] π = softmax(W) cls z+b cls );

[0099]

[0100] C eff ={c∈C∣w c >0};

[0101]

[0102] In the formula, z is the intermediate representation of the shared input mapping layer, φ(·; θ φ ) is the mapping function of the shared input mapping layer, θ φ Its parameters are π, where π is the class probability vector. y Let W be the probability of category y, and softmax(·) be the operation that maps a real-valued vector to a probability distribution. cls With b cls These are the coefficient matrix and bias vector for identifying the branch, respectively. Let W be a concentration estimation vector of length M, where max(0,·) ensures that each component is non-negative. reg With b reg These are the coefficient matrix and bias vector of the quantitative branch, respectively. For the predicted odor category, α is the equivalent reference quantity calculated based on the output of the quantitative branch, and it is used to estimate the concentration of the target component. Mapped to an equivalent quantity with the same dimensions as the concentration reference value. C is the representative amount of the concentration reference value defined in step S2 within the stable range. effFor a set of effective channels, satisfying the corresponding weight w c Greater than zero, |C eff | represents the number of effective channels, κ represents the confidence level, β1, β2, β3 are non-negative weight coefficients satisfying β1 + β2 + β3 = 1, τ is the scaling factor for the consistency measure, and π (1) With π (2) These are the largest and second largest components of π, respectively, and ln(·) is the natural logarithm operation;

[0103] In the above formula, the first line gives the joint generation of shared input mapping, class probability and concentration estimation; the second line gives the definition of class determination, equivalent reference quantity calculation and effective channel set; and the third line gives the calculation of confidence κ. The three terms respectively measure the consistency with the concentration reference value, the concentration of the channel weight distribution and the class determination margin.

[0104] When implemented, F clean Input the shared input mapping layer according to the fixed field order of step S3, ensuring that the channel identifier and field position are consistent with step S3. α is obtained through device calibration. Concentration reference value For comparability, w directly adopts the channel weight vector from step S4, and the entropy term... Used to reflect the discreteness of the channel weight distribution and normalize it to the range [0,1], β1, β2, β3 and τ are set as device parameters during the calibration stage and remain unchanged during the inference stage. Finally, the initial results of the category and concentration output by the fusion model are determined by... Composed of κ and the other two, and on a unified time scale {t} k The corresponding time stamp is associated with the following, and is used to proceed to step S6 for consistency correction and long-term stability compensation.

[0105] In this embodiment, the specific implementation of step S6 is as follows:

[0106] The initial results of category and concentration output from step S5 are used as input, where the predicted category is denoted as... The concentration estimation vector is denoted as (The dimension is the number of target components, arranged in a fixed order), the confidence level is denoted as κ, and the channel weight vector w = {w} obtained in step S4 is called simultaneously. c} c∈C and the effective channel set C eff The representative amount of the concentration reference value obtained in step S2 within the stable range. And the conversion coefficient vector α used in step S5 to convert to an equivalent amount, this step is within a unified time window and time stamp {t k Consistency correction and long-term stability compensation are performed under these conditions.

[0107] First, obtain the temperature and humidity covariate sequences and index the set according to the stable segment. Alignment, the ambient temperature sequence is denoted as T env (t k The environmental humidity sequence is denoted as H. env (t k The representative quantities of these quantities within the stable segment are denoted as follows: and According to the device calibration table, and Mapped to a set of environmental correction coefficients for each channel and To correct for sensitivity bias caused by temperature and humidity, the above covariates and correction coefficients are all related to the time stamp {t}. k To maintain consistency with the concentration reference value sequence C ref (t k ) and its representative quantity Perform one-to-one comparison and verification;

[0108] Next, the drift compensation parameter set is called to perform long-term correction on the zero point and ratio of each channel. The zero point correction parameter set is denoted as... The set of proportional correction parameters is denoted as The above parameters are determined during device calibration and operational period calibration, and are corrected accordingly for each channel in this step, in conjunction with S. T S H Δ 0 Δg After performing equivalent corrections for the channel source, we obtain the candidate concentration vector after environment and drift correction, denoted as... and The fields are kept consistent in order and units, making it easy to use directly with the conversion coefficient vector α;

[0109] In the consistency correction stage, the conversion method in step S5 will be applied. Mapped to an equivalent reference quantity, denoted as and Perform a difference assessment, and denote the consistency threshold as ∈ cons ,when Not exceeding ∈ cons When, directly confirm the corrected concentration and category, if it exceeds ∈ cons At that time, based on the channel weight vector w, Perform weighted fine-tuning, prioritizing the retention of C. eff The contribution of channels with higher internal weights is considered, while the contribution of channels with lower weights or mismatched weights is suppressed. Perform a consistency check; if any out-of-bounds or incompatible categories and quantities are found, then check the category set. Internally, select according to the principle of consistency first. and For compatible alternative classes, update the predicted classes, and update the confidence level synchronously in this step. The updated confidence level is denoted as κ. ＊ It is comprehensively evaluated based on the consistency of equivalent reference quantities, the concentration of channel weight distribution, and category margins, and compared with the time stamp {t}. k Maintain the corresponding relationship;

[0110] The final output after compensation is denoted as (y ＊ ,c ＊ ,κ ＊ ), where y ＊ For the final category, c ＊ For the final concentration vector, κ ＊ For the final confidence level, the compensated results are within a unified time window. With time stamp {t k The data is associated with the data in this period and maintains complete consistency with the field order and terminology of the aforementioned steps, for the result output and communication of step S7.

[0111] In this embodiment, the specific implementation of step S7 is as follows:

[0112] The result after compensation in step S6 (y) ＊ ,c ＊ ,κ ＊ ) is the input, and the time scale is uniform. k The following data is encapsulated into a result data packet according to a fixed field order. The result data packet is denoted as... Where t k y is the time stamp for this output. ＊ For the final category, c ＊ For the final concentration vector (arranged according to the predetermined order and dimensions of the target components), κ ＊ For the final confidence level, the Maintaining the same field naming and order as in the previous steps, the communication interface will... Output the completed category, concentration, and confidence level to the public, and archive it with the data for this period for subsequent query and retrieval.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An artificial intelligence olfactory system based on a heterogeneous sensor array, characterized in that, include: The data acquisition and array management module is used to introduce gas samples and acquire them according to a unified clock to form the array's original response sequence, and to record the temperature sequence, reference channel intensity, measurement channel intensity and optical path parameters. The unified time window and scale establishment module is used for channel alignment and unit unification to generate standardized time series data; The concentration reference value calculation module is used to calculate the concentration reference value and its representative quantity in the stable segment based on the reference channel intensity, measurement channel intensity, optical path length, and calibration coefficient. The fusion feature construction module is used to extract steady-state amplitude, relative change, response time, recovery time, and temperature segmentation ratio and concatenate them into a fusion feature vector; The channel health assessment and weight generation module is used to assess noise, drift, saturation, and consistency alignment, generate channel weights, and output the cleaned fusion features. The joint inference and confidence generation module is used to output the initial results of category and concentration and generate confidence scores within a single inference cycle; The temperature and humidity covariance and drift compensation module is used to determine the environmental correction coefficient according to the calibration table, and to perform consistency correction and fine-tuning in combination with zero point and proportional correction parameters, and output the compensated result. The results encapsulation and communication module is used to encapsulate the final category, final concentration and final confidence level into a result data packet according to time stamp and output it through the communication interface; The calibration and parameter management module is used to uniformly manage absorption coefficient, optical path length, gain matching, temperature threshold, threshold, conversion factor, and confidence weight.

2. An artificial intelligence olfactory method based on a heterogeneous sensor array, applied to the artificial intelligence olfactory system based on a heterogeneous sensor array as described in claim 1, characterized in that, include: S1. Introduce a gas sample into a sampling chamber and collect the response of a heterogeneous sensor array within the same sampling chamber using a unified clock. The array contains at least two gas sensing units with different mechanisms, selected from at least two of the following: metal oxide, electrochemical, photoionization, colorimetric, and non-dispersive infrared. The metal oxide unit performs temperature modulation and records the temperature sequence, the electrochemical unit operates within a preset noise threshold and bandwidth range via a transimpedance amplifier, and the non-dispersive infrared unit enables a reference channel for dual-channel ratio and optical path calculation, outputting the array's original response sequence. S2. Using the original response sequence of the array as input, denoising, baseline calibration, scale normalization and time alignment are completed within a unified time window. Based on the ratio of non-dispersive infrared reference and measurement channels combined with optical path parameters, the concentration reference value is obtained, and standardized time series data is output. S3. Using standardized time-series data as input, extract the steady-state amplitude, relative change, dynamic indicators of response and recovery, and temperature modulation segmentation ratio, and incorporate the concentration reference value to form a fusion feature vector, and output the fusion feature vector. S4. Using the fused feature vector as input, perform abnormal channel removal and weighting based on channel health indicators including noise, baseline drift and saturation signs. Use a weighted fusion strategy that includes at least one of weighted voting or attention weighting to obtain the purified fused features, and output the purified fused features. S5. Using the purified fusion features as input, input the fusion model to jointly output the odor category and the concentration and confidence level of at least one target component, and output the initial results of category and concentration. S6. Using the initial results of category and concentration as input, perform consistency correction by combining drift compensation and temperature and humidity covariates, and use the concentration reference value as a quantitative benchmark for verification and fine-tuning, and output the compensated results. S7. Using the compensated result as input, output the category, concentration, and confidence level through the communication interface to obtain the result data packet.

3. The artificial intelligence olfactory method based on heterogeneous sensor array according to claim 2, characterized in that: The specific steps of acquiring the response of the heterogeneous sensor array with a unified clock in the same sampling chamber are as follows: a gas sample is introduced into the sampling chamber, so that at least two of the sensing units in the array, including metal oxide, electrochemical, photoionization, colorimetric, and non-dispersive infrared sensors, are simultaneously exposed under the same gas sample conditions. The acquisition controller distributes the same clock and sampling trigger to each channel, and assigns a unified time mark to each sampling, thereby forming the array's original response sequence for the same time and the same gas sample. The array contains at least two different gas sensing units with different mechanisms. Specifically, it combines sensing units based on complementary mechanisms such as resistance change, Faraday current, photoionization current, visible or near-infrared absorption, and infrared absorption within the same sampling cavity to cover responses to different chemical reactions and optical absorption paths and provide distinguishable signal sources. The specific method for implementing temperature modulation and recording temperature sequences for the metal oxide is as follows: a segmented or continuous temperature trajectory is applied to the heater of the metal oxide sensing unit. The trajectory consists of a starting temperature, an ending temperature, a change mode, and a duration. The temperature sequence is recorded synchronously under this trajectory, and the corresponding response is obtained, so that the adsorption and reaction processes at different temperature ranges are explicitly encoded as auxiliary inputs for subsequent steps.

4. The artificial intelligence olfactory method based on heterogeneous sensor array according to claim 2, characterized in that: The concentration reference value is obtained by reading the intensity values ​​of the reference channel and the measurement channel within a unified time window in the non-dispersive infrared optical path, calculating the intensity ratio of the two channels, and mapping them according to the optical path parameters and device calibration coefficients to generate a concentration reference value sequence marked with time.

5. The artificial intelligence olfactory method based on heterogeneous sensor array according to claim 2, characterized in that: The temperature modulation segmentation ratio is specifically defined as follows: based on the temperature sequence, the standardized time-series data is divided into multiple temperature segments according to preset temperature boundaries. The mean or final value of the response of the metal oxide channel is extracted in each temperature segment, and a segmentation ratio set is formed by the ratio of the responses of adjacent temperature segments or preset combination temperature segments. The segmentation ratio is divided into multiple temperature segments with the starting temperature, ending temperature and change mode in the temperature sequence as boundaries. The segmentation ratio is recorded synchronously with a unified time stamp. This is used to distinguish the response modes corresponding to different adsorption and reaction barriers in complex mixed gas scenarios, so that the temperature-related information of the same channel is explicitly encoded by the segmentation ratio and can be compared with other mechanism channels. The fusion feature vector is a single input structure formed by classifying and splicing the features according to the mechanism under a unified scale. It includes the temperature modulation segment ratio and steady-state amplitude of the metal oxide channel, the steady-state amplitude and recovery time of the electrochemical channel, the response peak and recovery slope of the photoionization channel, the color change of the colorimetric channel, and the concentration reference value of the non-dispersive infrared channel. The features of each channel are composed in a fixed positional order and are labeled with the channel. The concentration reference value is embedded in the same vector as a quantitative anchor to support the subsequent classification and concentration estimation in the fusion model. This provides comparable and verifiable cross-mechanism input in business scenarios with complex mixed gases and low-concentration targets.

6. The artificial intelligence olfactory method based on heterogeneous sensor array according to claim 2, characterized in that: The specific steps for performing abnormal channel elimination and weight reduction based on channel health indicators are as follows: within a unified time window, calculate the noise level, baseline drift, and saturation signs for each channel to obtain a health score. Compare the health score with a preset elimination threshold and a weight reduction threshold. Channels below the elimination threshold are marked as abnormal and removed from the fused feature vector. Channels between the two thresholds are retained but assigned a reduced channel weight. Channels above the weight reduction threshold are assigned a standard weight. The channel weight corresponds to the feature segment of the corresponding channel in the fused feature vector and is synchronously saved as a channel weight vector. A weighted fusion strategy is adopted to weight and combine feature segments in the channel dimension: when weighted voting is used, the channel weight vector is normalized and then the same type of feature fields are weighted and aggregated to form a channel aggregation field. When attention weighting is used, the channel correlation between the channel weight vector and the fused feature vector is used as the input of the attention score. The channel attention weight is calculated and the feature segments are weighted in the channel dimension, giving priority to strengthening channel segments with high consistency with the concentration reference value.

7. The artificial intelligence olfactory method based on heterogeneous sensor array according to claim 2, characterized in that: The fusion model is a joint model that simultaneously completes category determination and concentration estimation in a single inference, comprising three parts: a shared input mapping layer, a discriminative branch, and a quantitative branch. The shared input mapping layer receives the cleaned fusion features and maintains the channel identifiers and field order in consistency. The discrimination branch uses the odor category set as the output space to classify the purified fusion features; The quantitative branch uses the concentration reference value as the quantitative anchor point to establish a mapping relationship between the purified fusion characteristics and the concentration of the target component, and outputs the concentration of at least one target component; The discrimination branch and the quantitative branch are linked through consistency constraints within the same inference process. The consistency constraints are based on the concentration reference value and the channel weight vector to check the compatibility of the outputs of the two branches. The confidence level is generated by combining the verification results of the consistency constraints with the matching degree of the channel weights, and is output simultaneously with the category and concentration to form the initial results of the category and concentration, which are used to provide usable judgment and quantitative basis in business scenarios of complex gas mixtures and low-concentration targets.

8. The artificial intelligence olfactory method based on heterogeneous sensor array according to claim 2, characterized in that: The consistency correction combining drift compensation and temperature and humidity covariates is as follows: taking the initial results of category and concentration as input, simultaneously reading the temperature and humidity covariates and concentration reference values, and aligning the time stamps of the three within a unified time window; The sensitivity correction coefficients for each channel were obtained by looking up tables based on the temperature and humidity covariates, and the zero-point and proportional correction parameters of drift compensation were applied to correct the contributions of each channel in the initial results of category and concentration. Using the concentration reference value as the quantitative benchmark, the corrected concentration is scaled and the quantitative contribution of each channel is weighted and merged according to the channel weight vector, so that the quantitative result is consistent with the concentration reference value within the preset consistency threshold. For category determination, if the revised quantitative value is out of bounds or inconsistent with the category label, the category is adjusted or marked as requiring review according to the consistency rule, and the confidence level of the determination is updated simultaneously.

Images