Adaptive calibration control method and system for gas detector

Abstract

The present application relates to the field of gas detector safety monitoring and control, and provides a gas detector adaptive calibration control method and system, the method comprising: obtaining high-frequency original electric signal time sequence and collecting external environment parameters, combining physical degradation model to calculate sensor equivalent aging time to build environment-internal state vector, and generating coupling time sequence characteristic vector by aggregating dynamic characteristic sub-vector; a double-path diagnosis unit extracts transient morphological fingerprint vector and context state vector, and outputs the final state vector through feedforward neural network fusion; an interference source attribution network generates an interference source weight vector, and generates a dynamic calibration vector through an attribution gate-expert hybrid model; a transient nonlinear calibration function is constructed to process high-frequency original electric signals, real-time temperature and humidity to output corrected concentration, and determine alarm or normal state according to weight threshold value. The present application combines physical-oriented dynamic characteristic extraction and model attribution diagnosis to realize transient and long-term double-scale decoupling and adaptive calibration control.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and control of gas detectors, and particularly to an adaptive calibration control method and system for gas detectors. Background Technology

[0002] With the widespread application of gas detection systems in critical fields such as petrochemicals, real-time high-precision decoupling and adaptive calibration of sensor signals have become core technologies for ensuring safe operation in the field. However, in complex industrial environments, sensor signals are highly coupled with factors such as target gas, multi-source interference, temperature and humidity drift, and sensor aging. Traditional calibration methods cannot effectively solve the two major safety problems caused by this: false alarms and missed alarms. Existing models are limited to post-event correction and hysteresis compensation for drift, and cannot achieve real-time decoupling of coupling interference at the moment of signal acquisition. It is urgent to solve the problems of insufficient utilization of the dynamic transient characteristics of the signal and the lag in adaptive calibration response in order to achieve intelligent upgrade of the system.

[0003] Chinese patent application CN115712267A discloses a host control method and host based on gas detectors. The method includes: based on a received self-test command, sequentially checking connection addresses and obtaining gas concentration values ​​collected by the gas detectors corresponding to those addresses; if the gas concentration value exceeds a preset threshold, controlling a speaker to sound an alarm and cyclically displaying the alarm address; and based on a received mute command, controlling the speaker to stop until the next alarm occurs. When the gas concentration values ​​of multiple gas detectors exceed the preset threshold, the speaker sounds an alarm, facilitating timely handling by staff. Furthermore, when the speaker sounds an alarm, a received mute command can be used to control the speaker to shut down, stopping the alarm. This helps staff understand the situation on-site and turn off the speaker, reducing the impact of continuous alarms on subsequent handling.

[0004] However, current technology still faces many challenges. When gas detectors are used for monitoring in petrochemical or refining sites, the rapid response process caused by target gas leaks is often accompanied by high concentrations of volatile organic compounds and other background interfering gases. Traditional gas detectors cannot effectively separate the coupling effects of these interfering substances on the sensor's original electrical signal when monitoring target gas concentrations, causing the measurement results to deviate from the true response of the target gas. When the concentration of background interfering gases changes significantly in an instant, it can cause high-frequency spikes in the original signal, leading to cross-sensitivity responses. Traditional systems often fail to identify these dynamic transient characteristics in time, resulting in misjudgments of target gas leaks, false alarms, and disruption of normal production processes. More seriously, when suppressive gases are present on-site and cause nonlinear suppression of the activity of electrochemical sensors, the sensor's sensitivity to the actual leaking target gas will drop sharply. If the system fails to detect this sensitivity decay caused by signal suppression in time, it will be unable to accurately measure the true concentration of the target gas, potentially leading to missed alarms and causing personnel to miss valuable intervention time. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides an adaptive calibration control method for a gas detector, the specific technical solution of which is as follows:

[0006] The sampling frequency is executed to obtain the high-frequency raw electrical signal timing, and the external environmental parameters are collected simultaneously. The equivalent aging time of the sensor is calculated in real time by combining the physical degradation model to construct the environment-internal state vector. The vector is then aggregated with the extracted dynamic feature sub-vector to construct the coupled timing feature vector.

[0007] The dual-path diagnostic unit is executed in parallel. The dual-path diagnostic unit includes a fast path and a slow path. The transient morphological fingerprint vector is extracted from the fast path and the context state vector is output from the slow path. The transient morphological fingerprint vector and the context state vector are fused using a feedforward neural network to output the final state vector.

[0008] The interference source attribution network is used to diagnose the final state vector to generate the interference source weight vector, and the attribution gating-expert hybrid model is used to generate the expert calibration vector. The expert calibration vector is dynamically weighted and converged based on the interference source weight vector to generate the dynamic calibration vector.

[0009] The physical model skeleton is dynamically parameterized using dynamic calibration vectors to construct an instantaneous nonlinear calibration function. The instantaneous nonlinear calibration function is called to process the high-frequency raw electrical signal, real-time temperature, and real-time humidity to output the corrected concentration. Attribution diagnosis is performed on the interference source weight vector in parallel to determine whether it is greater than a preset threshold. If any weight exceeds its preset threshold, the diagnostic alarm signal is set to the corresponding alarm state. If none of them exceed the threshold, the system is set to normal operation state.

[0010] Furthermore, the method for constructing the coupled temporal feature vector includes:

[0011] The control command sampling unit converts the comprehensive physicochemical stimuli sensed by the sensor in real time into high-frequency original electrical signal timing that retains the transient dynamic characteristics of the interference source and the target gas at the sampling frequency.

[0012] External environmental parameters, including real-time temperature and real-time humidity, are collected. Based on the historical high-frequency raw electrical signal timing, historical temperature, and historical humidity, the equivalent aging time of the sensor is calculated in real time through physical degradation model integration. The real-time temperature, real-time humidity, and sensor equivalent aging time are then constructed into an environment-internal state vector.

[0013] Based on the timing definition of the high-frequency original electrical signal, a sliding time window is defined to extract dynamic feature sub-vectors that characterize transient dynamic properties. The dynamic feature sub-vectors and the environment-internal state vector are then aggregated to output a coupled timing feature vector.

[0014] Furthermore, the method for calculating the equivalent aging time of the sensor through real-time integration using a physical degradation model includes: using a preset nonlinear degradation stress function to perform real-time integration calculations on the gas exposure surrogate value, historical temperature, and historical humidity, so as to diagnose and quantify the sensor equivalent aging time that reflects the cumulative physical loss of the sensor.

[0015] The gas exposure proxy value is calculated and stored by the AI ​​program controller based on the historical high-frequency raw electrical signal timing sequence at historical moments.

[0016] Furthermore, the method for extracting the dynamic feature sub-vector includes: calculating the first derivative of the signal within a sliding time window to quantify the rate of change of the timing of the high-frequency original electrical signal, and calculating the moving average of the signal within the sliding time window to provide a dynamic baseline reference for the timing of the high-frequency original electrical signal, and aggregating the first derivative of the signal, the moving average of the signal, and the timing of the high-frequency original electrical signal to construct the dynamic feature sub-vector.

[0017] Furthermore, the method for outputting the final state vector includes:

[0018] Will A short-time diagnostic sequence is constructed from dynamic feature sub-vectors. The short-time diagnostic sequence is analyzed using a physical-guided morphological diagnostic tool to identify the signal morphological fingerprint in real time and output a transient morphological fingerprint vector.

[0019] Long Short-Term Memory (LSTM) networks are used to temporally encode the environment-internal state vector to diagnose and model the long-term dependencies between temperature and humidity drift and sensor aging, and output the context state vector.

[0020] The transient morphological fingerprint vector and the context state vector are aggregated to form a diagnostic-state fusion vector. The diagnostic-state fusion vector is then subjected to nonlinear mapping and dimensionality compression using a feedforward neural network to output the final state vector.

[0021] Furthermore, the method for outputting the transient morphological fingerprint vector includes:

[0022] The driving morphological diagnostic tool performs diagnostic logic on the first derivative and moving average of the signal contained in the short-time diagnostic sequence.

[0023] The diagnostic logic includes: when the first derivative of the signal exhibits alternating positive and negative fluctuations or spikes, the output is a high-frequency interference pattern; when the first derivative of the signal exhibits a continuous positive gradient and the moving average of the signal continues to rise, the output is a target response pattern; when the first derivative of the signal exhibits a continuous negative gradient, the output is a suppression pattern; when both the first derivative of the signal and the moving average of the signal are stable, the output is a baseline pattern.

[0024] The diagnostic results of the diagnostic logic are quantified, and a probability distribution vector consisting of the confidence probabilities of the high-frequency interference pattern, target response pattern, suppression pattern and baseline pattern is generated and output as a transient morphological fingerprint vector.

[0025] Furthermore, the method for generating the dynamic calibration vector includes:

[0026] The interference source attribution network is used to perform interference source attribution decoupling on the final state vector to diagnose and quantify the contribution of each physical cause, and to generate and output the interference source weight vector; the interference source weight vector includes target gas response weight, environmental drift weight, cross-sensitivity source weight, transient suppression source weight, and permanent aging weight.

[0027] By using an attribution-gated-expert hybrid model, the final state vector is fed into multiple expert subnetworks in parallel to generate the corresponding expert calibration vector. The interference source weight vector is used as a dynamic gating weight to dynamically weight and converge the expert calibration vector to generate a dynamic calibration vector.

[0028] Furthermore, the method for generating the expert calibration vector includes: feeding the final state vector into multiple expert sub-networks in parallel, so that each expert sub-network independently diagnoses a physical cause in the final state vector and outputs the corresponding expert calibration vector based on its diagnosis result; the physical cause includes target gas, environmental drift, cross sensitivity, transient suppression source and permanent aging.

[0029] Furthermore, the method for outputting the corrected concentration includes:

[0030] The dynamic calibration vector is used as an instantaneous parameter set to perform dynamic parameterization on the fixed nonlinear calibration model skeleton in order to instantiate and output the instantaneous nonlinear calibration function.

[0031] The instantaneous nonlinear calibration function is invoked to process the timing of the high-frequency raw electrical signal, real-time temperature, and real-time humidity. The total additive interference and dynamic sensitivity, parameterized by the instantaneous nonlinear calibration function, are calculated in parallel. The corrected concentration is output by subtracting the total additive interference from the timing of the high-frequency raw electrical signal and dividing by the dynamic sensitivity.

[0032] An adaptive calibration control system for a gas detector, which is used to implement the aforementioned adaptive calibration control method for a gas detector, includes a coupling feature construction module, a diagnostic decoupling module, an attribution calibration module, and a calibration diagnosis module.

[0033] The coupling feature construction module is used to execute sampling frequency to obtain high-frequency raw electrical signal timing, and simultaneously collect external environmental parameters and combine them with the physical degradation model to calculate the sensor equivalent aging time in real time to construct an environment-internal state vector. It then performs vector aggregation with the extracted dynamic feature sub-vectors to construct a coupling timing feature vector.

[0034] The diagnostic decoupling module is used to execute the dual-path diagnostic unit in parallel. The dual-path diagnostic unit includes a fast path and a slow path. It extracts the transient morphological fingerprint vector in the fast path and outputs the context state vector in the slow path. It uses a feedforward neural network to fuse the transient morphological fingerprint vector and the context state vector to output the final state vector.

[0035] The attribution calibration module is used to generate an interference source weight vector by diagnosing the final state vector using the interference source attribution network, and to generate an expert calibration vector using the attribution gating-expert hybrid model. Based on the interference source weight vector, the expert calibration vector is dynamically weighted and converged to generate a dynamic calibration vector.

[0036] The calibration and diagnostic module is used to dynamically parameterize the physical model skeleton using dynamic calibration vectors to construct an instantaneous nonlinear calibration function, and to call the instantaneous nonlinear calibration function to process high-frequency raw electrical signals, real-time temperature, and real-time humidity to output corrected concentrations. In parallel, it performs attribution diagnosis on the interference source weight vectors to determine whether they are greater than a preset threshold. If any weight exceeds its preset threshold, the diagnostic alarm signal is set to the corresponding alarm state. If none of them exceed the threshold, the system is set to normal operating state.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention employs a high-frequency sampling strategy for diagnostic purposes and combines it with a sliding window to extract dynamic features. This fully preserves the transient dynamic features lost by traditional low-frequency sampling due to signal averaging or aliasing, thus solving the problem of false alarms caused by the inability to distinguish between rapid spike responses caused by interference sources and gradual rise responses caused by target gas.

[0039] This invention solves the real-time decoupling problem caused by the inability of traditional single models to simultaneously handle interference from two different time scales, namely fast spikes and slow drift, by constructing a dual-path diagnostic unit.

[0040] This invention constructs an attribution-gated-expert hybrid model, which uses an interference source attribution network to quantify the real-time contribution of each physical interference. This contribution is then used as a dynamic gating weight to adaptively weight the calibration suggestions of multiple expert sub-networks. This solves the problem that traditional single models cannot interpretably decouple and dynamically correct multi-source, coupled interference at the moment of data acquisition.

[0041] This invention applies dynamic calibration vectors to a fixed physical model skeleton by parallel execution of programmed nonlinear calibration control and accompanying diagnostics to achieve additive interference cancellation and multiplicative sensitivity correction, avoiding the problem of inaccurate correction concentration caused by signal coupling and nonlinear suppression; simultaneously, multi-level threshold diagnostic judgment is performed based on the interference source weight vector, realizing a technological leap from passive calibration to proactive health management. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the principle of the adaptive calibration control method for gas detectors of the present invention.

[0044] Figure 2 This is a functional block diagram of the gas detector adaptive calibration control system of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1:

[0047] Please see Figure 1 As shown, this embodiment provides an adaptive calibration control method for a gas detector, including:

[0048] Step S1000, execute sampling frequency Obtain the timing sequence of the high-frequency raw electrical signal Simultaneously, external environmental parameters are collected, and the equivalent aging time of the sensor is calculated in real time by combining the physical degradation model. To construct the environment-internal state vector , and the extracted dynamic feature sub-vectors Perform vector aggregation to construct coupled temporal feature vectors .

[0049] Specifically, this step aims to resolve the mapping relationship between the sensor's original electrical signal and the actual target gas concentration, which has been affected by various interferences, temperature and humidity drift, and sensor aging in the real physical world, i.e., a combination of physicochemical stimuli. The fundamental challenge of disruption. The aim of this step is to proactively extract and construct a coupled temporal feature vector from this highly coupled original physicochemical stimulus, which can characterize the dynamic transient properties of the signal and the physical health status of the sensor. This transforms a non-decoupled data source into a decoupled feature set, providing a unique data foundation with rich information for subsequent steps.

[0050] Further, step S1000 includes:

[0051] Step S1100: Control command sampling unit at sampling frequency The sensor senses the comprehensive physicochemical stimuli in real time. The timing sequence is converted into a high-frequency original electrical signal that fully preserves the dynamic transient characteristics of the interference source and the target gas. .

[0052] Specifically, this step aims to perform diagnostic data acquisition tasks, where the AI ​​program controller of the gas detector system actively controls its internal sampling unit to highly couple the comprehensive physicochemical stimuli of the sensor response. The timing sequence is converted into a digital high-frequency raw electrical signal without distortion. This ensures that the data acquired by the AI ​​program controller in subsequent steps fully retains the transient dynamic characteristics used to diagnose interference sources and target gases, rather than the steady-state values ​​obtained by traditional low-frequency sampling, which have been smoothed and lost information, thus laying a data foundation for subsequent steps.

[0053] The AI ​​program controller is the control core of the entire gas detector system, used to execute all algorithm steps. It is typically a microprocessor or application-specific integrated circuit.

[0054] In the specific implementation process, the sensor configured in this step serves as a physical interface, exposed to the field environment, and continuously senses the comprehensive physicochemical stimuli that mix the target gas, interfering gases, and environmental factors. This sensor will integrate physicochemical stimuli. It is converted in real time into a continuous, uncalibrated analog electrical signal.

[0055] The AI ​​program controller actively instructs its internal sampling units, such as high-precision analog-to-digital converters (ADCs), to execute a diagnostic data acquisition strategy. This acquisition strategy is characterized by its sampling frequency. The setting is not based on the known purpose of high-fidelity signal waveform reproduction, but on the purpose of ensuring the capture of specific diagnostics of transient interference events.

[0056] Therefore, the sampling unit operates at a sampling frequency much higher than that required for conventional data recording. This routine data recording typically only concerns steady-state results, for example, setting... That is, the sampling period The analog electrical signal is sampled and quantized at high density to obtain digitized high-frequency raw electrical signal time-series data. .

[0057] For example, in petrochemical sites, interference from volatile organic compounds (VOCs), which act as a source of cross-sensitivity, can potentially damage the timing data of high-frequency raw electrical signals. This generates rapid spikes or peaks, i.e., high-frequency dynamic characteristics, such as a signal spike to 50mV within 0.1 seconds and then disappearing immediately after 0.1 seconds; while the target gas, hydrogen sulfide... The response is a relatively gradual rise, exhibiting low-to-mid-frequency dynamic characteristics. For example, the signal slowly rises to a peak of 5mV within 2 seconds. Traditional low-frequency sampling has a sampling period of 1s, which is much longer than the 0.1s duration of the interference event. This leads to the loss of two key pieces of information: first, sampling aliasing, where rapid spikes are missed or misrepresented because they are not sampled in time; and second, signal averaging, where the sampler integrates or averages the signal within this period, resulting in the rapid peak value of volatile organic compounds (VOCs), such as 50mV at 0.1s, and the slow, low rise of VOCs, such as 5mV at 2s, being mathematically averaged into a potentially identical average value, such as 5mV. This information loss irreversibly confuses two physically distinct responses into a single signal rise data point. The controller then cannot distinguish whether this signal rise originates from the target gas hydrogen sulfide or VOCs, potentially leading to false alarms. The high-frequency sampling strategy in this step fully preserves this dynamic variability used for diagnosis, providing a data foundation for dynamic feature extraction in subsequent steps.

[0058] Step S1200: Collect data including real-time temperature. and real-time humidity The external environmental parameters, and based on the historical high-frequency raw electrical signal timing. Historical temperature and historical humidity The equivalent aging time of the sensor is calculated in real time by integrating the physical degradation model. The real-time temperature Real-time humidity and sensor equivalent aging time Construct as an environment-internal state vector .

[0059] Specifically, this step aims to address the technical problem of inaccurate output signals from gas detectors caused by temperature and humidity drift and sensor aging. It proposes an active compensation mechanism based on internal state modeling, which collects external environmental parameters in real time, including real-time temperature. and real-time humidity And in conjunction with step S1100, historical moments are collected. Historical high-frequency raw electrical signal timing It is used to diagnose and quantify an internal state variable that cannot be directly measured but is crucial for decoupling, namely the sensor's equivalent aging time. The final output is the environment-internal state vector. This provides crucial contextual information for subsequent steps to distinguish between permanent aging and transient inhibition.

[0060] In the specific implementation process, the timing sequence of the high-frequency raw electrical signal acquired in step S1100... Simultaneously, the AI ​​program controller executes this step in parallel to acquire auxiliary parameters. The core innovation of this step lies in transforming the sensor's aging time from a passive, cumulative power-on time decoupled from physical reality into an actively calculated internal state variable that reflects real physical wear and tear.

[0061] The AI ​​programmable controller obtains the real-time temperature inside the detector through auxiliary sensors within its sensor array, such as thermistors and humidity sensors. and real-time humidity Meanwhile, the AI ​​program controller diagnoses and calculates the cumulative equivalent operating time of the sensors based on historical data in its non-volatile memory. That is, the cumulative stress factor.

[0062] The real-time temperature For the current moment The environmental parameter values ​​around the sensor are collected in real time, in degrees Celsius (°C) or Kelvin (K). These values ​​are used to compensate for the drift caused by temperature changes in the sensor's instantaneous response and are also used as historical temperature data. It is stored to calculate long-term environmental stress. Similarly, the real-time humidity... For the current moment The environmental parameter values ​​around the sensor are collected in real time, and the unit is relative humidity percentage (%RH). These values ​​are used to compensate for the drift caused by changes in humidity to the sensor's instantaneous response, and are also used as historical humidity data. It is stored to calculate long-term environmental stress.

[0063] The cumulative stress factor It's not a simple timer, but rather an AI program controller that calculates the result in real-time using a preset physical degradation model in the background, for example, at a low frequency or during idle periods. The specific process formula is as follows:

[0064] ;

[0065] in, Indicates the current moment The sensor equivalent aging time is an internal state variable of the AI ​​program controller, used to quantify the cumulative physical wear and tear of the sensor and assess its true health status. Indicates the current moment, used to refer to the sensor's equivalent aging time. The immediacy of this state variable; The total power-on operating time of the sensor, i.e., the calendar age, is the upper limit of the definite integral, and its value is equal to the current time. , used to define the total time span for calculating cumulative stress; The integral variable represents a historical moment, and its value ranges from 0 to the total energized operating time. The continuous time is used to traverse the entire operating history of the sensor, that is, the total power-on operating time from the factory time 0 to the current time. ; Representing an infinitesimal time increment, it is the differential element of the integral and is used to represent the historical moment. Perform continuous summation; The nonlinear degradation stress function, stored in the memory of the AI ​​program controller, is a physical degradation model used to calculate at any historical moment. The instantaneous degradation rate; Indicating a historical moment The gas exposure proxy value is used to quantify the sensor at that historical moment. The environmental stress it withstands is determined by the AI ​​program controller based on historical data collected in step S1100. Historical high-frequency raw electrical signal timing Calculated and stored; and Representing historical moments Historical temperature and historical humidity are used to quantify the sensor at that moment. Environmental stresses it withstands.

[0066] Finally, the AI ​​program controller will use the obtained real-time temperature data. Real-time humidity and sensor equivalent aging time Together, they constitute and output the environment-internal state vector. .

[0067] Step S1300, based on the timing of the high-frequency original electrical signal Define sliding time window Extract dynamic feature sub-vectors that characterize transient dynamic properties. The dynamic feature subvector and environment-internal state vector Perform vector aggregation to output coupled temporal feature vectors. .

[0068] Specifically, this step aims to utilize the timing data of the high-frequency raw electrical signal acquired in step S1100. and the environment-internal state vector output by step S1200 Actively construct a single feature vector that integrates transient dynamic characteristics and internal state context, i.e., a coupled temporal feature vector. .

[0069] In the specific implementation process, this step is the convergence point of steps S1100 and S1200. In this step, the AI ​​program controller is at the current moment... Actively execute a multi-stage feature extraction and state fusion process.

[0070] To quantify transient dynamic characteristics, the AI ​​program controller is based on the timing of the acquired high-frequency raw electrical signals. Define a time Ending, size is Sliding time window .in, It is the AI ​​program controller that obtains the timing sequence of the high-frequency raw electrical signal. An ordered subset of data extracted from the data, whose values ​​are a set containing... A vector or list of the latest temporal sampling points is used to provide a local, temporal data context for subsequent dynamic feature calculations; This represents the size or length of the sliding window; its value is a preset hyperparameter used to define the sliding time window. The total number of sampling points included, and determines the time scale for calculating transient dynamic features; Indicating a historical moment The acquired high-frequency raw electrical signal values.

[0071] Based on the defined sliding time window Extract a dynamic feature vector .in, Indicates the current moment Dynamic feature vectors are used to digitally characterize the timing of high-frequency raw electrical signals. In the sliding time window Transient dynamic characteristics within; Indicates the current moment Based on sliding time window The first derivative of the signal is a scalar, which is expressed by the formula... Calculations show that The sampling period is used to quantify how fast the signal changes. Indicates the current moment Based on sliding time window The signal moving average value is a scalar value, determined by the formula... Calculations show that the timing of smoothing high-frequency raw electrical signals is achieved. Sampling may introduce transient noise and affect the timing of the original high-frequency electrical signal. The instantaneous amplitude provides a dynamic short-term baseline reference.

[0072] To achieve the fusion of dynamic characteristics and internal states, the AI ​​program controller uses the extracted dynamic feature sub-vectors as described above. and environment-internal state vector Perform vector aggregation and output an enhanced coupled temporal feature vector. .in, This represents the coupled temporal feature vector, which is a multi-dimensional feature vector used to aggregate and provide the AI ​​program controller at time step. Simultaneously diagnoses all dynamic and state characteristics required for cross sensitivity and nonlinear suppression or aging; This represents a vector aggregation operation.

[0073] Step S2000: The dual-path diagnostic unit is executed in parallel. The dual-path diagnostic unit includes a fast path and a slow path. The transient morphological fingerprint vector is extracted from the fast path. and output the context state vector on the slow path. Using feedforward neural networks to fuse transient morphological fingerprint vectors and context state vector Output the final state vector .

[0074] Specifically, this step aims to construct a dual-path diagnostic unit, which explicitly utilizes the coupled temporal feature vector output from step S1300 through the fast path in step S2100. The dynamic components in the process perform real-time morphological diagnosis to identify and suppress cross-sensitivity responses caused by interfering gases; in parallel, the coupled temporal feature vectors are utilized through the slow path in step S2200. The state components are subjected to long-term context encoding to distinguish between permanent sensor aging and temporary nonlinear suppression caused by target gas activation. Finally, the diagnostic results from the two different time scales are fused in real time to output the final state vector. This provides a decoupled high-order state basis for subsequent adaptive calibration steps.

[0075] Further, step S2000 includes:

[0076] Step S2100, will Dynamic feature vectors Constructed as a short time-series diagnostic sequence The short-time diagnostic sequence was analyzed using a physics-guided morphological diagnostic tool. Real-time identification of signal morphological fingerprints and output of transient morphological fingerprint vectors. .

[0077] Specifically, this step aims to execute the fast path task in the dual-path diagnostic unit, and to process the coupled temporal feature vector output from step S1300. Dynamic feature vectors in Perform real-time morphological diagnostics and output transient morphological fingerprint vectors. This step does not concern itself with the long-term drift trend, but focuses on quantifying the instantaneous shape characteristics of the current signal to identify whether the response originates from a cross-sensitivity source or target gas activation, providing crucial diagnostic basis for real-time false alarm suppression in subsequent steps.

[0078] In the specific implementation process, in order to capture the temporal evolution characteristics of the form, the AI ​​program controller buffers a sequence of data generated by the most recent... Dynamic feature vectors Constructed short-time diagnostic sequence .in, Indicates at time Ending, length is Short time-series diagnostic sequences are used to analyze dynamic feature vectors. The system dynamically constructs an ordered tensor or list to provide a temporal context for subsequent morphological diagnostics, enabling it to identify and diagnose the morphology or fingerprint of a signal, rather than just processing a single instant. Represents short-time diagnostic sequences The length of this parameter is a preset hyperparameter used to determine how long the subsequent morphological diagnostic tool should look back at the instantaneous history to diagnose the morphology. For example, if the sampling period... A short-time diagnostic sequence of 0.1s length If the value is 100, then the time scale for diagnosis is 10 seconds; Indicating a historical moment The dynamic feature sub-vectors are short-time diagnostic sequences. The first element is used to define the time start boundary of this short time-series diagnostic sequence.

[0079] This short-time diagnostic sequence The input is fed into a one-dimensional convolutional neural network unit. The innovation of this unit lies in its non-blind feature extractor nature, rather than a physically-guided morphological diagnostic tool. The coupled temporal feature vector is based on step S1300. This morphological diagnostic tool does not require learning the underlying physical concepts from scratch; instead, it is empowered to analyze the first derivative of the extracted signal. Signal moving average Key diagnostic features, etc.

[0080] The morphological diagnostic tool is trained to identify typical signal morphological fingerprints in real time:

[0081] Fingerprint 1, high-frequency interference pattern. When the first derivative of the signal... When alternating positive and negative swings or spike features appear, the diagnostic output is "spike fingerprint", a pattern usually caused by cross-sensitivity sources.

[0082] Fingerprint 2, target response pattern. When the first derivative of the signal... It exhibits a continuous positive gradient and a moving average of the signal. When the temperature continues to rise, the diagnostic output is a "response curve fingerprint," a pattern typically triggered by the target gas activation.

[0083] Fingerprint 3, Suppression Pattern. When the first derivative of the signal... When a continuous negative gradient or smooth decay characteristics are observed, the diagnostic output is "suppressed fingerprint," a pattern typically caused by suppressing gas.

[0084] Fingerprint four, baseline morphology. When the first derivative of the signal... Stable and signal moving average When the change is subtle, the diagnostic tool outputs a "baseline fingerprint".

[0085] Finally, the transient morphological fingerprint vector is output. .in, Indicates the current moment The transient morphological fingerprint vector, whose value is a probability distribution vector, that is, a multidimensional real number vector whose sum of all components is 1, is used to quantify the high-level diagnostic conclusion of the AI ​​controller on the current signal shape. The high-frequency interference pattern, i.e., the confidence probability of the spike fingerprint, is represented by the output of the morphological diagnostic instrument and is used to quantify the current short-time diagnostic sequence. The morphology was diagnosed as a spike fingerprint with high credibility; This represents the target response morphology output by the morphological diagnostic tool, i.e., the confidence probability of the response curve fingerprint, used to measure the current short-time-series diagnostic sequence. The morphology is used to diagnose the reliability of response curve fingerprints; The suppression morphology, represented by the output of the morphological diagnostic tool, is the confidence probability of the suppressed fingerprint and is used to quantify the current short-time diagnostic sequence. The morphology was diagnosed as suppressing the reliability of fingerprints; The baseline morphology, or baseline fingerprint, output by the morphological diagnostic tool is used to quantify the current short-time-series diagnostic sequence. The morphology is diagnosed as the baseline morphology, which is the confidence level of the signal being in a stable, baseline state without significant events.

[0086] Step S2200: Utilize a Long Short-Term Memory network to process the environment-internal state vector. Temporal encoding is performed to diagnose and model the long-term dependencies between temperature and humidity drift and sensor aging, outputting a context state vector. .

[0087] Specifically, this step aims to execute the slow path task in the dual-path diagnostic unit, in parallel with the fast path task in step S2100, on the coupled temporal feature vector output from step S1300. Environment-internal state vector Perform timing encoding and output the context state vector. This step does not concern itself with the instantaneous morphological characteristics of the current signal, but only captures the long-term dependencies of slowly changing characteristics, such as temperature and humidity drift and internal state evolution, such as sensor aging, providing crucial contextual information for subsequent steps to distinguish between permanent aging and transient suppression in real time.

[0088] In the specific implementation process, in parallel with the fast path of step S2100, the AI ​​program controller will couple the temporal feature vector. Environment-internal state vector The input is fed into a Long Short-Term Memory (LSTM) network for time-series encoding. The LSTM network is characterized by its specialization and diagnostic capabilities; it does not analyze signal morphology but specializes in capturing the long-term dependencies and drift trends of these slowly changing features over time. Based on the key diagnostic feature provided in step S1200, namely the sensor's equivalent aging time... This Long Short-Term Memory (LSTM) network is internally configured to model the sensor's health state (aging context) and baseline drift (environmental context) with high accuracy. The specific state update function is defined as follows:

[0089] ;

[0090] in, Indicates the current moment The context state vector, which takes the value of a high-dimensional real number vector, is used to encode and encapsulate the diagnostic conclusions of all slowly changing long-term historical contexts, such as temperature and humidity drift and physical aging. The state update function of a Long Short-Term Memory (LSTM) network is a type of Recurrent Neural Network (RNN) model used to receive the current input environment-internal state vector. and the previous moment context state vector , used to capture long-term dependencies; This indicates that the Long Short-Term Memory (LSTM) network was at the previous time step. The context state vector, whose values ​​are encoded until... All long-term history up to this point, such as cumulative drift and aging trends, are used to define the time frame. The calculations provide historical memory.

[0091] Step S2300: Aggregate transient morphological fingerprint vectors and context state vector A diagnostic-state fusion vector is formed, and a feedforward neural network is used to perform nonlinear mapping and dimensionality compression on the diagnostic-state fusion vector to output the final state vector. .

[0092] Specifically, this step aims to achieve the same result at the same time. The transient morphological fingerprint vector output by the fast path in step S2100 is... and the context state vector output by the slow path in step S2200 Perform nonlinear fusion to output the final state vector. To encapsulate the AI ​​controller for the current moment The highest-level diagnostic decision after signal decoupling, and the final state vector It is the sole basis for subsequent steps.

[0093] In the specific implementation process, this step is the convergence node of the fast path and slow path in the dual-path diagnostic unit, and the AI ​​program controller is at the same time. Parallel acquisition of fast morphological diagnostics, i.e., transient morphological fingerprint vectors and slow context diagnostics, i.e., context state vector The data is then aggregated in a multi-scale feature space to form a diagnostic-state fusion vector, enabling a joint representation of transient and steady-state information. This diagnostic-state fusion vector is then input to a diagnostic fusion layer, which preferably employs a feed-forward network (FFN) structure to achieve nonlinear mapping and dimensionality compression, outputting the final state vector. .

[0094] Step S3000: Diagnose the final state vector using the interference source attribution network. To generate interference source weight vector And an expert calibration vector is generated using an attribution-gated-expert hybrid model, based on the interference source weight vector. Dynamic weighted aggregation is performed on the expert calibration vectors to generate dynamic calibration vectors. .

[0095] Specifically, this step aims to use the final state vector output in step S2300 as a basis for... Interference source attribution decoupling is performed, and finally, a dynamic calibration vector is generated in real time. The dynamic calibration vector Regarding the current moment The interference components are identified and compensated, thereby fundamentally solving the problem of real-time decoupling.

[0096] Further, step S3000 includes:

[0097] Step S3100: Use the interference source attribution network to process the final state vector. Perform interference source attribution decoupling to diagnose and quantify the contribution of each physical cause, and generate and output the interference source weight vector. The interference source weight vector Including target gas response weights Environmental drift weights Cross-sensitivity source weights Instantaneous suppression source weights and permanent aging weight .

[0098] Specifically, this step aims to respond to the final state vector output by step S2300. It performs interference source attribution decoupling to diagnose and quantify which type of physical root cause is causing the interference, thereby outputting an interference source weight vector. This provides an explainable basis for attribution in subsequent steps.

[0099] In the specific implementation process, the AI ​​program controller will finalize the state vector. The input is fed into the Interference Source Attribution Network (ISAN), which can be a lightweight feedforward neural network (FFN) or a dedicated attention structure with a multi-head mechanism. The ISAN is trained from the final state vector. Potential disturbance features are identified, and the final state vector is transformed using a nonlinear transformation function. The original attribution scores are mapped to the original scores, then processed by the Softmax normalization function, and finally output as a vector of interference source weights with a sum of 1. .in, Indicates the current moment The interference source weight vector is a A probability distribution vector, where the sum of all components is 1, used to explicitly quantize the final state vector. The relative contributions of various physical origins; Represents the interference source weight vector The total number of components is equal to the total number of categories of physical causes that the AI ​​controller predefines and requires explicit decoupling; Indicates the current moment The target gas response weight is used to quantify the contribution of the target gas response; Indicates the current moment The environmental drift weight is used to quantify the contribution of drift caused by instantaneous environmental stresses such as temperature and humidity. Indicates the current moment The cross-sensitivity source weights are used to quantify the contribution of interfering gases to false alarm sources; Indicates the current moment The instantaneous suppression source weight is used to quantify the contribution of transient pollution sources; Indicates the current moment Permanent aging weights are used to quantify the equivalent aging time of the sensor. The contribution of the diagnosed permanent damage.

[0100] Step S3200: Using the attribution gating-expert hybrid model, the final state vector is... Multiple expert subnetworks are fed in parallel to generate corresponding expert calibration vectors, and the interference source weight vectors are utilized. As a dynamic gating weight, the expert calibration vector is dynamically weighted and aggregated to generate a dynamic calibration vector. .

[0101] Specifically, this step aims to respond to the final state vector output by step S2300. The interference source weight vector output in step S3100 Dynamic calibration vectors are generated in real time using an attribution-gated-expert hybrid model. It can identify and compensate for interference components in sensor signals at the moment of acquisition, thereby achieving real-time decoupling and dynamic correction of the signal layer.

[0102] In the specific implementation, the architecture of this step is designed as an Attribution-Gated Mixture of Experts (AG-MoE). This AG-MoE model comprises two sequential stages: an expert diagnosis stage and an attribution gating stage. The specific process is as follows:

[0103] The first stage is the expert diagnosis stage. In this stage, the final state vector is processed in parallel. Feed in separately A network of experts. Each expert subnetwork is a dedicated AI inverse model, and each expert subnetwork is trained independently to diagnose the final state vector. This process involves identifying specific physical causes, such as cross-sensitivity or permanent aging, and outputting corresponding expert calibration vectors. The specific formula is as follows:

[0104] ; ;

[0105] ; ;

[0106] ;

[0107] in, , , , as well as These represent the nonlinear transformation functions of the expert subnetworks for "target gas", "environmental drift", "cross sensitivity", "transient suppression", and "permanent aging", respectively. , , , as well as They represent the current time. The target gas expert calibration vector, environmental drift expert calibration vector, cross sensitivity expert calibration vector, transient suppression expert calibration vector, and permanent aging expert calibration vector.

[0108] The second stage is the attribution gating stage. This stage uses the interference source weight vector... As a gating weight, for all The outputs of each expert subnetwork are dynamically weighted and converged. The real-time impact of various interference sources is explicitly modeled using attribution gating units, and this model is used as a weighting coefficient to dynamically sum the calibration vectors of each expert, generating the final dynamic calibration vector. .

[0109] Step S4000, using dynamic calibration vector Dynamically parameterize the physical model skeleton to construct instantaneous nonlinear calibration functions Call the instantaneous nonlinear calibration function Processing high-frequency raw electrical signals Real-time temperature and real-time humidity Output corrected concentration Parallel processing of the interference source weight vector Perform attribution diagnosis and determine if any weight exceeds a preset threshold. If any weight exceeds its preset threshold, a diagnostic alarm signal is set. The corresponding alarm status is set. If none of them exceed the limit, the system is set to normal operating status.

[0110] Specifically, this step serves a dual purpose: firstly, to utilize the dynamic calibration vector generated in step S3200. Dynamically construct an instantaneous nonlinear calibration function And immediately apply it to the timing of the high-frequency raw electrical signal acquired in step S11000. And in step S1200, the real-time temperature inside the detector is obtained. and real-time humidity To output the decoupled corrected concentration Secondly, the interference source weight vector output in step S3100 is used. Implement accompanying diagnostics and output diagnostic alarm signals. This will enable a technological leap from automatic calibration to intelligent health management.

[0111] Further, step S4000 includes:

[0112] Step S4100, dynamically calibrate the vector As an instantaneous parameter set, it is used for the fixed nonlinear calibration model skeleton. Perform dynamic parameterization to instantiate and output the instantaneous nonlinear calibration function. .

[0113] Specifically, this step is designed to respond to the dynamic calibration vector generated in step S3200. The AI ​​controller at the current moment Dynamically parameterize a fixed physical model skeleton to construct the physical model at that moment. The only instantaneous nonlinear calibration function .

[0114] In the specific implementation process, this step is performed at each sampling time. Receives dynamically generated calibration vectors in real time This dynamic calibration vector Encapsulate the final data of all collaborative steps from step S1000 to step S3000.

[0115] Receive the dynamic calibration vector Injected into a nonlinear calibration model skeleton embedded in nonvolatile memory The calibration model skeleton is a pre-defined, physically meaningful nonlinear function, in which... It is a dynamically configurable set of parameters. Dynamic calibration vector. This step serves as the skeleton of the nonlinear calibration model. At any moment instantaneous parameter set .in, Indicates the current moment The instantaneous parameter set is obtained through dynamic calibration vectors. Instantiate the nonlinear calibration model skeleton Configurable parameter set The resulting intermediate variables have the same structure and dimensions as the configurable parameter set. Consistent; Represents parameterized actions, used to represent instantaneous parameter sets. The value is dynamically calibrated by the vector The determined parameters enable real-time generation of parameters for nonlinear calibration of the model skeleton. Dynamic configuration. Through the above parameterized actions, at time... Instantiate and output a parameterized, instantaneous nonlinear calibration function. .

[0116] Step S4200: Call the instantaneous nonlinear calibration function. Timing of high-frequency raw electrical signals Real-time temperature and real-time humidity Processing is performed, and the instantaneous nonlinear calibration function is calculated in parallel. Total additive disturbances with internal parameterization and dynamic sensitivity By analyzing the timing of the high-frequency raw electrical signal Subtract the total additive interference And divided by dynamic sensitivity Output corrected concentration .

[0117] Specifically, this step aims to invoke the instantaneous nonlinear calibration function constructed in step S4100. And immediately apply it to the original physical signal that is input in real time in steps S1100 and S1200, that is, the timing sequence of the high-frequency original electrical signal. Real-time temperature and real-time humidity This is to perform real-time decoupling and programmed nonlinear calibration control, ultimately outputting the decoupled corrected concentration. .

[0118] In the specific implementation process, this step calls the same time. The instantaneous nonlinear calibration function constructed by step S4100 Simultaneously, the high-frequency raw electrical signal transmitted in real time in step S1100 is acquired. And the real-time temperature input in step S1200 and real-time humidity .

[0119] To perform the computation that fuses physical information and AI, this step executes the instantaneous nonlinear calibration function called above. This instantaneous nonlinear calibration function In a physical sense, the calibration process is decoupled into two parallel stages: additive interference cancellation and multiplicative sensitivity correction, as follows:

[0120] The first stage is the additive interference cancellation stage. This step first calculates the total additive interference. The total additive interference It is composed of instantaneous nonlinear calibration function In the process, the dynamic calibration vector generated by step S3200 Real-time parameterized drift function and cross-interference function It is formed by stacking.

[0121] The drift function Get real-time temperature and real-time humidity As an input variable, the temperature drift coefficient is obtained. and humidity drift coefficient As an instantaneous parameter, it is used to calculate and compensate for environmental drift. The temperature drift coefficient is mentioned here. and humidity drift coefficient It is a dynamic calibration vector In response to the interference source weight vector in step S3100 Environmental drift weights The environmental drift weight The final state vector output in response to step S2300 The final state vector In response to the coupled temporal feature vector in step S2200 Real-time temperature and real-time humidity Long-term monitoring.

[0122] Similarly, the cross-interference function Obtain the cross sensitivity correction coefficient As an instantaneous parameter, it is used to calculate and counteract cross-sensitivity interference. The cross-sensitivity correction coefficient is... It is a dynamic calibration vector In response to the interference source weight vector in step S3100 Cross-sensitivity source weights The cross-sensitivity source weight In response to the real-time morphological diagnosis of the high-frequency interference pattern in step S2100, this diagnosis relies on the first derivative of the signal extracted in real time in step S1300. .

[0123] Ultimately, the AI ​​controller is based on the drift function. and cross-interference function Calculate and obtain the total additive disturbance. .

[0124] The second stage is the multiplicative sensitivity correction stage. Dynamic sensitivity is calculated in parallel. This dynamic sensitivity It is composed of instantaneous nonlinear calibration function In, by dynamic calibration vector Real-time parameterized suppression function and aging function For a solidified reference sensitivity It is obtained by nonlinear correction.

[0125] The suppression function Using instantaneous suppression correction factor This is to correct for transient suppression. The transient suppression correction coefficient is... It is a dynamic calibration vector In response to the interference source weight vector in step S3100 Instantaneous suppression source weights The instantaneous suppression source weight In response to the real-time morphological diagnosis of the suppression pattern in step S2100, the diagnosis relies on the first derivative of the signal extracted in real time in step S1300. .

[0126] The aging function Use permanent aging factor This is to correct for permanent aging. The permanent aging coefficient is... It is a dynamic calibration vector In response to the interference source weight vector in step S3100 Permanent aging weight in This permanent aging weight In response to the context state vector in step S2200 Long-term monitoring, which relies on the sensor equivalent aging time extracted in real time in step S1200. .

[0127] Ultimately, the AI ​​controller is based on the inhibition function. and aging function Calculate and obtain dynamic sensitivity .

[0128] Ultimately, the AI ​​controller uses high-frequency raw electrical signals... Subtract total additive interference and divided by dynamic sensitivity Calculate and output the final corrected concentration. .

[0129] Step S4300, adjust the interference source weight vector Perform attribution diagnosis and determine the interference source weight vector based on preset accompanying diagnostic control logic. Cross-sensitivity source weights Instantaneous suppression source weights Or permanent aging weight If any weight exceeds its preset threshold, a diagnostic alarm signal is set. If none of the alarms exceed the specified threshold, the status will be set to "System is running normally".

[0130] Specifically, this step aims to adjust the interference source weight vector output in step S3100. Perform continuous monitoring and attribution diagnostics to identify the corrected concentration output from step S4200. The underlying physical cause outputs diagnostic alarm signals that can be read by users or upper-level systems. This enables a shift from passive calibration to proactive health management, allowing for targeted alerts to be issued at the early stages of abnormal trends.

[0131] In the specific implementation process, the AI ​​program controller adjusts the interference source weight vector. Each component in the process undergoes real-time logical judgment, and accompanying diagnostic control logic is executed. Prior to this, this step pre-obtains preset threshold parameters for judgment, including a cross-sensitivity threshold. Instantaneous inhibition threshold Permanent aging threshold The preset threshold can be determined based on the upper limit of the 95% confidence interval of the statistical distribution of historical cross-interference data of sensors of this type in the gas detector, or preset as a fixed empirical value according to the sensor's factory calibration data. The specific execution logic of the accompanying diagnostic control logic is as follows:

[0132] First, cross-sensitivity monitoring. If the interference source weight vector... Cross-sensitivity source weights Greater than the preset cross sensitivity threshold The system will then programmatically set diagnostic alarm signals. The first alarm state indicates that the sensor response is affected by background gas interference, which may trigger a false alarm.

[0133] Second, transient suppression monitoring. If the cross-sensitivity source weights... Not exceeding the cross sensitivity threshold Then the AI ​​controller will further determine the weight vector of the interference source. Instantaneous suppression source weights Is it greater than the preset instantaneous suppression threshold? If the conditions are met, the system will programmatically set the diagnostic alarm signal. This is the second alarm state, indicating that the sensor may be temporarily electrochemically inhibited by interfering gases such as sulfur dioxide (SO2).

[0134] Third, permanent monitoring. If neither of the aforementioned two types of interference is triggered, the AI ​​controller determines the weight vector of the interference source. Permanent aging weight Is it greater than the permanent aging threshold? If the permanent aging threshold is exceeded The system will then programmatically set diagnostic alarm signals. This is the third alarm state, indicating that the sensor sensitivity is declining or the electrodes are aging.

[0135] Fourth, normal operation judgment. If none of the aforementioned three types of interference are triggered, the system will programmatically set diagnostic alarm signals. The system is in a "normal operating state".

[0136] Current moment Diagnostic alarm signals Its value can be a status code or a text string, and it is output through the display interface, remote monitoring interface, or host computer system to intuitively report the current status of the sensor and the correction concentration. This provides a clear basis for intervention by maintenance personnel or system scheduling layer, enabling intelligent health management and interpretability.

[0137] Example 2:

[0138] This embodiment, based on Embodiment 1, provides an adaptive calibration control system for a gas detector, such as... Figure 2 As shown, the system includes a coupling feature construction module, a diagnostic decoupling module, an attribution calibration module, and a calibration diagnosis module;

[0139] The coupling feature construction module is used to execute the sampling frequency. Obtain the timing sequence of the high-frequency raw electrical signal Simultaneously, external environmental parameters are collected, and the equivalent aging time of the sensor is calculated in real time by combining the physical degradation model. To construct the environment-internal state vector , and the extracted dynamic feature sub-vectors Perform vector aggregation to construct coupled temporal feature vectors .

[0140] The diagnostic decoupling module is used to execute the dual-path diagnostic unit in parallel. The dual-path diagnostic unit includes a fast path and a slow path, and extracts transient morphological fingerprint vectors on the fast path. and output the context state vector on the slow path. Using feedforward neural networks to fuse transient morphological fingerprint vectors and context state vector Output the final state vector .

[0141] The attribution calibration module is used to diagnose the final state vector using the interference source attribution network. To generate interference source weight vector And an expert calibration vector is generated using an attribution-gated-expert hybrid model, based on the interference source weight vector. Dynamic weighted aggregation is performed on the expert calibration vectors to generate dynamic calibration vectors. .

[0142] The calibration and diagnostic module is used to utilize dynamic calibration vectors. Dynamically parameterize the physical model skeleton to construct instantaneous nonlinear calibration functions Call the instantaneous nonlinear calibration function Processing high-frequency raw electrical signals Real-time temperature and real-time humidity Output corrected concentration Parallel processing of the interference source weight vector Perform attribution diagnosis and determine if any weight exceeds a preset threshold. If any weight exceeds its preset threshold, a diagnostic alarm signal is set. The corresponding alarm status is set. If none of them exceed the limit, the system is set to normal operating status.

[0143] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0144] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive calibration control method for a gas detector, characterized in that, include: The sampling frequency is executed to obtain the high-frequency raw electrical signal timing, and the external environmental parameters are collected simultaneously. The equivalent aging time of the sensor is calculated in real time by combining the physical degradation model to construct the environment-internal state vector. The vector is then aggregated with the extracted dynamic feature sub-vector to construct the coupled timing feature vector. The dual-path diagnostic unit is executed in parallel. The dual-path diagnostic unit includes a fast path and a slow path. The transient morphological fingerprint vector is extracted from the fast path and the context state vector is output from the slow path. The transient morphological fingerprint vector and the context state vector are fused using a feedforward neural network to output the final state vector. The interference source attribution network is used to diagnose the final state vector to generate the interference source weight vector, and the attribution gating-expert hybrid model is used to generate the expert calibration vector. The expert calibration vector is dynamically weighted and converged based on the interference source weight vector to generate the dynamic calibration vector. The physical model skeleton is dynamically parameterized using dynamic calibration vectors to construct an instantaneous nonlinear calibration function. The instantaneous nonlinear calibration function is called to process the high-frequency raw electrical signal, real-time temperature, and real-time humidity to output the corrected concentration. Attribution diagnosis is performed on the interference source weight vector in parallel to determine whether it is greater than a preset threshold. If any weight exceeds its preset threshold, the diagnostic alarm signal is set to the corresponding alarm state. If none of them exceed the threshold, the system is set to normal operation.

2. The adaptive calibration control method for a gas detector according to claim 1, characterized in that, The method for constructing the coupled temporal feature vector includes: The control command sampling unit converts the comprehensive physicochemical stimuli sensed by the sensor in real time into high-frequency original electrical signal timing that retains the transient dynamic characteristics of the interference source and the target gas at the sampling frequency. External environmental parameters, including real-time temperature and real-time humidity, are collected. Based on the historical high-frequency raw electrical signal timing, historical temperature, and historical humidity, the equivalent aging time of the sensor is calculated in real time through physical degradation model integration. The real-time temperature, real-time humidity, and sensor equivalent aging time are then constructed into an environment-internal state vector. Based on the timing definition of the high-frequency original electrical signal, a sliding time window is defined to extract dynamic feature sub-vectors that characterize transient dynamic properties. The dynamic feature sub-vectors and the environment-internal state vector are then aggregated to output a coupled timing feature vector.

3. The adaptive calibration control method for a gas detector according to claim 2, characterized in that, The method for calculating the equivalent aging time of a sensor in real time by integrating a physical degradation model includes: using a preset nonlinear degradation stress function to perform real-time integration calculations on the gas exposure proxy value, historical temperature, and historical humidity, so as to diagnose and quantify the sensor equivalent aging time that reflects the cumulative physical loss of the sensor. The gas exposure proxy value is calculated and stored by the AI ​​program controller based on the historical high-frequency raw electrical signal timing at historical moments.

4. The adaptive calibration control method for a gas detector according to claim 2, characterized in that, The method for extracting the dynamic feature sub-vector includes: calculating the first derivative of the signal within a sliding time window, calculating the moving average of the signal within the sliding time window, and aggregating the first derivative of the signal, the moving average of the signal, and the time sequence of the original high-frequency electrical signal to construct the dynamic feature sub-vector.

5. The adaptive calibration control method for a gas detector according to claim 1, characterized in that, The method for outputting the final state vector includes: Will A short-time diagnostic sequence is constructed from dynamic feature sub-vectors. The short-time diagnostic sequence is analyzed using a physical-guided morphological diagnostic tool to identify the signal morphological fingerprint in real time and output a transient morphological fingerprint vector. Long Short-Term Memory (LSTM) networks are used to temporally encode the environment-internal state vector to diagnose and model the dependence of temperature and humidity drift and sensor aging, and output the context state vector. The transient morphological fingerprint vector and the context state vector are aggregated to form a diagnostic-state fusion vector. The diagnostic-state fusion vector is then subjected to nonlinear mapping and dimensionality compression using a feedforward neural network to output the final state vector.

6. The adaptive calibration control method for a gas detector according to claim 5, characterized in that, The method for outputting the transient morphological fingerprint vector includes: The driving morphological diagnostic tool performs diagnostic logic on the first derivative and moving average of the signal contained in the short-time diagnostic sequence. The diagnostic logic includes: when the first derivative of the signal exhibits alternating positive and negative fluctuations or spikes, the output is a high-frequency interference pattern; when the first derivative of the signal exhibits a continuous positive gradient and the moving average of the signal continues to rise, the output is a target response pattern; when the first derivative of the signal exhibits a continuous negative gradient, the output is a suppression pattern; when both the first derivative of the signal and the moving average of the signal are stable, the output is a baseline pattern. The diagnostic results of the diagnostic logic are quantified, and a probability distribution vector consisting of the confidence probabilities of the high-frequency interference pattern, target response pattern, suppression pattern and baseline pattern is generated and output as a transient morphological fingerprint vector.

7. The adaptive calibration control method for a gas detector according to claim 1, characterized in that, The method for generating the dynamic calibration vector includes: The interference source attribution network is used to perform interference source attribution decoupling on the final state vector to diagnose and quantify the contribution of each physical cause, and to generate and output the interference source weight vector; the interference source weight vector includes target gas response weight, environmental drift weight, cross-sensitivity source weight, transient suppression source weight, and permanent aging weight. By using an attribution-gated-expert hybrid model, the final state vector is fed into multiple expert subnetworks in parallel to generate the corresponding expert calibration vector. The interference source weight vector is used as a dynamic gating weight to dynamically weight and converge the expert calibration vector to generate a dynamic calibration vector.

8. The gas detector adaptive calibration control method according to claim 7, characterized in that, The method for generating the expert calibration vector includes: feeding the final state vector into multiple expert sub-networks in parallel, so that each expert sub-network independently diagnoses one physical cause in the final state vector and outputs the corresponding expert calibration vector based on its diagnosis result; the physical causes include target gas, environmental drift, cross sensitivity, transient suppression source and permanent aging.

9. The adaptive calibration control method for a gas detector according to claim 1, characterized in that, The method for outputting the corrected concentration includes: The dynamic calibration vector is used as an instantaneous parameter set to perform dynamic parameterization on the fixed nonlinear calibration model skeleton in order to instantiate and output the instantaneous nonlinear calibration function. The instantaneous nonlinear calibration function is invoked to process the timing of the high-frequency raw electrical signal, real-time temperature, and real-time humidity. The total additive interference and dynamic sensitivity, parameterized by the instantaneous nonlinear calibration function, are calculated in parallel. The corrected concentration is output by subtracting the total additive interference from the timing of the high-frequency raw electrical signal and dividing by the dynamic sensitivity.

10. A gas detector adaptive calibration control system, used to implement the gas detector adaptive calibration control method according to any one of claims 1-9, characterized in that, The system includes a coupling feature construction module, a diagnostic decoupling module, an attribution calibration module, and a calibration diagnosis module. The coupling feature construction module is used to execute sampling frequency to obtain high-frequency raw electrical signal timing, and simultaneously collect external environmental parameters and combine them with the physical degradation model to calculate the sensor equivalent aging time in real time to construct an environment-internal state vector. It then performs vector aggregation with the extracted dynamic feature sub-vectors to construct a coupling timing feature vector. The diagnostic decoupling module is used to execute the dual-path diagnostic unit in parallel. The dual-path diagnostic unit includes a fast path and a slow path. It extracts the transient morphological fingerprint vector in the fast path and outputs the context state vector in the slow path. It uses a feedforward neural network to fuse the transient morphological fingerprint vector and the context state vector to output the final state vector. The attribution calibration module is used to generate an interference source weight vector by diagnosing the final state vector using the interference source attribution network, and to generate an expert calibration vector using the attribution gating-expert hybrid model. Based on the interference source weight vector, the expert calibration vector is dynamically weighted and converged to generate a dynamic calibration vector. The calibration and diagnostic module is used to dynamically parameterize the physical model skeleton using dynamic calibration vectors to construct an instantaneous nonlinear calibration function, and to call the instantaneous nonlinear calibration function to process high-frequency raw electrical signals, real-time temperature, and real-time humidity to output corrected concentrations. In parallel, it performs attribution diagnosis on the interference source weight vectors to determine whether they are greater than a preset threshold. If any weight exceeds its preset threshold, the diagnostic alarm signal is set to the corresponding alarm state. If none of them exceed the threshold, the system is set to normal operating state.

Images