Gas sensing-based online monitoring method for waste gas concentration in lithium carbonate production

Abstract

The application discloses a lithium carbonate production waste gas concentration online monitoring method based on gas sensing, which comprises the following steps: collecting multi-source coupled monitoring data and preprocessing to form a baseline response sequence; sequentially applying a three-state perturbation mode, collecting steady-state and recovery stage signals, and forming three groups of perturbation state response sequences; constructing a perturbation difference curve based on the baseline difference to generate a three-state dynamic trajectory tensor; inputting an improved InceptionTime model, multi-scale convolution gate fusion, and frequency domain enhancement to output a dynamic fingerprint vector; constructing a perturbation state coupled parameter space, applying a non-negative monotonic conservation constraint, and solving a concentration candidate set; calculating the conservation residual of the candidate solution, screening the optimal concentration solution, and outputting the real-time concentration and change rate. The application realizes long-term stable and high-precision online monitoring of waste gas concentration under complex working conditions of lithium carbonate production by constructing a three-state perturbation response mechanism and combining a dynamic fingerprint constraint inversion method.

Description

Technical Field

[0001] This invention relates to the field of online gas composition analysis technology, and in particular to a method for online monitoring of the concentration of lithium carbonate production waste gas based on gas sensing. Background Technology

[0002] The lithium carbonate production process generates industrial waste gas containing various gaseous components during roasting, acidification, lithium precipitation, and drying. Common components include acidic gases, nitrogen-containing gases, and carbon dioxide, accompanied by high humidity and particulate matter. To meet environmental emission standards and production process control requirements, online gas sensors are typically installed at the emission pipelines to continuously monitor the concentration of target gases using electrochemical or infrared sensors. Current technologies often employ a single sensor response signal combined with a temperature and humidity compensation model or empirical calibration curve to correct for concentration variations, thereby obtaining real-time concentration data.

[0003] However, in the actual operation of lithium carbonate production, the exhaust gas components are complexly coupled, humidity and dust fluctuate dramatically, and different gases exhibit differences in diffusion rates and adsorption hysteresis within the sampling link and sensing chamber, resulting in significant hysteresis and non-stationary changes in the sensing response curve. Existing technologies typically rely on fixed compensation coefficients or periodic calibration with standard gases for correction, which struggles to accurately reflect true concentration changes under highly non-stationary operating conditions. Furthermore, frequent calibration with standard gases increases operating costs and affects the continuity of online monitoring.

[0004] Existing technologies are mostly based on response models established under a single operating condition for concentration inversion. They lack systematic utilization of the dynamic characteristics of the disturbance loading and recovery process and fail to fully explore the inherent coupling relationship between gas responses under different disturbance conditions. This leads to problems such as measurement drift accumulation, unstable concentration inversion, and difficulty in distinguishing short-term anomalies in multi-gas strongly coupled environments.

[0005] Therefore, how to provide an online monitoring method for the concentration of lithium carbonate production exhaust gas based on gas sensing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an online monitoring method for the concentration of waste gas from lithium carbonate production based on gas sensing, in order to address the shortcomings of the prior art.

[0007] The online monitoring method for lithium carbonate production waste gas concentration based on gas sensing according to embodiments of the present invention includes: Sampling devices and gas sensing units are installed in the exhaust gas pipeline of lithium carbonate production to collect multi-source coupled monitoring data. The multi-source coupled monitoring data is preprocessed to form a baseline response sequence. Microscale electrothermal gradient perturbation, local microfluidic field perturbation, and microelectric polarization perturbation are sequentially applied in the sampling microcavity according to a preset time sequence. Gas sensing response signals are collected in the steady-state and recovery phases of each perturbation mode to form three sets of perturbation response sequences. Using the baseline response sequence as a reference, the three sets of perturbation dynamic response sequences are subjected to baseline difference processing to construct perturbation baseline difference curves and perturbation-recovery cyclic splicing processing to form hysteresis response curves. The hysteresis area, phase lag and amplitude change of each hysteresis response curve are calculated to generate a three-state dynamic trajectory tensor. The three-state dynamic trajectory tensor is input into the improved InceptionTime model, which includes multi-scale dilated convolution branches, hysteresis gated fusion branches, and phase-locked frequency domain enhancement branches. Multi-timescale features are extracted and gated fusion and frequency domain enhancement processing is performed to output the dynamic fingerprint vector. The dynamic fingerprint feature vector and three sets of perturbation response sequences are assembled into a three-state response tensor. The perturbation dynamic coupling parameter space is constructed and non-negativity constraints, monotonically increasing constraints and diffusion-adsorption energy conservation constraints are applied. The concentration vector candidate set is obtained by solving the problem. Select the unique concentration solution that satisfies the disturbance response conservation relationship from the candidate set of concentration vectors as the real-time true concentration value of the target gas, and output the real-time true concentration value and the rate of change.

[0008] Optionally, the multi-source coupled monitoring data includes the electrical signal response value of the target gas, temperature parameters, relative humidity parameters, gas flow rate parameters, and dust concentration parameters within the sampling chamber.

[0009] Optionally, the preprocessing of the multi-source coupled monitoring data to form a baseline response sequence includes: Multi-source coupled monitoring data are time-aligned according to a unified sampling period, and data from different sampling channels are interpolated to complete the data and construct a synchronous monitoring data sequence. Outlier identification and removal are performed on the synchronous monitoring data sequence. A sliding window statistical method is used to correct the mutation data that exceeds the set threshold, resulting in an outlier-free data sequence. Without applying a perturbation mode, the outlier-removed data sequence is filtered for steady-state intervals and averaged to form a baseline response sequence.

[0010] Optionally, the formation of three sets of perturbation dynamic response sequences includes: Set a disturbance period and divide each disturbance period into a disturbance stabilization phase and a recovery phase after disturbance removal, so as to keep the sampling cavity in a baseline state without applying any disturbance. A microscale electrothermal gradient perturbation is applied during the perturbation period. A controllable temperature gradient is formed in the sampling cavity by controlling the micro heating element. Gas sensing response signals are continuously collected during the perturbation stabilization phase. After the electrothermal perturbation is removed, response signals are collected again during the recovery phase. The response signals of the perturbation stabilization phase and the recovery phase are combined in chronological order to form the first set of perturbation response sequences. After completing the microscale electrothermal gradient perturbation, a local microflow field perturbation is applied. The gas flow state in the sampling cavity is changed by adjusting the perturbation structure. Gas sensing response signals are collected during the perturbation stabilization phase. After the flow field perturbation is removed, response signals are collected again during the recovery phase. The response signals of the two phases are combined in chronological order to form the second set of perturbation response sequences. After completing the local microfluidic field perturbation, a micro-electric field polarization perturbation is applied. An electric field environment is formed in the sampling cavity by controlling the electrode array. Gas sensing response signals are collected during the perturbation stabilization phase. After the electric field perturbation is removed, response signals are collected again during the recovery phase. The response signals of the two phases are combined in chronological order to form a third set of perturbation response sequences. The first set of disturbance dynamic response sequences, the second set of disturbance dynamic response sequences, and the third set of disturbance dynamic response sequences are aligned and stored according to a unified time index to form three sets of disturbance dynamic response sequences.

[0011] Optionally, the step of calculating the hysteresis area, phase lag, and amplitude change of each hysteresis response curve to generate a three-state dynamic trajectory tensor includes: Using the baseline response sequence as a reference, the three sets of disturbance dynamic response sequences are aligned with the baseline response sequence according to a unified time index. When the time indices are inconsistent, linear interpolation is performed on the baseline response sequence. The baseline response value is obtained at the corresponding time index and the difference is calculated with the disturbance dynamic response value to obtain three sets of baseline difference sequences. Within each disturbance cycle, the disturbance start time and disturbance removal time are recorded based on the control signal applied during the disturbance. The disturbance loading segment and disturbance recovery segment are determined in the corresponding reference differential sequence. The loading peak point is determined in the disturbance loading segment, and the recovery peak point is determined in the disturbance recovery segment. The disturbance start time, disturbance removal time, loading peak point, and recovery peak point are used as the anchor point set for cyclic splicing. Phase normalization resampling is performed on the perturbation loading segment and the perturbation recovery segment respectively. The perturbation loading segment is resampled into a loading sequence with a preset number of points, and the perturbation recovery segment is resampled into a recovery sequence with the same number of points. The loading sequence and the recovery sequence are paired point by point according to the same sequence number to construct the phase pairing sequence of the perturbation reference difference curve. Based on the anchor point set, the phase pairing sequence is cyclically spliced. The end of the loaded sequence is connected to the beginning of the recovery sequence according to the anchor point correspondence. The end of the recovery sequence is closedly connected to the beginning of the loaded sequence. At each connection position, a three-point weighted average with fixed weights is used for continuous processing to form the hysteresis response curve corresponding to the disturbance type. For each hysteresis response curve, calculate the hysteresis area, phase lag, and amplitude change, where: The hysteresis area is calculated by summing the differences between the loaded sequence and the recovery sequence at the same index using a trapezoidal summation method. Phase lag is the time difference between the time corresponding to the loading peak point and the time corresponding to the recovery peak point; The magnitude change is the difference between the maximum and minimum values ​​of the benchmark difference sequence; The hysteresis area, phase lag, and amplitude change corresponding to each disturbance type are stacked in order of disturbance type to generate a three-state dynamic trajectory tensor.

[0012] Optionally, the step of extracting multi-timescale features and performing gated fusion and frequency domain enhancement processing to output a dynamic fingerprint vector includes: The channel ordering and amplitude normalization of the three-state dynamic trajectory tensor are performed. The characteristic channels corresponding to the hysteresis area, phase hysteresis, and amplitude change are arranged in order of perturbation type to obtain the input tensor. The input tensor is fed into the multi-scale dilated convolution branch of the improved InceptionTime model in parallel. The multi-scale dilated convolution branch uses multiple sets of one-dimensional depthwise separable convolutions with different kernel lengths and different dilation rates to extract features from the input tensor. After each set of convolutions, channel mixing is performed by pointwise convolution to output a multi-timescale local feature set. The set of local features at multiple time scales is input into the hysteresis-gated fusion branch. The hysteresis-gated fusion branch generates gating weights based on the boundary positions of the disturbance stability phase and the recovery phase, as well as the positions of the loading peak point and the recovery peak point. Cross-gated weighted fusion is performed between features at different time scales and between features of different disturbance types, and the disturbance dynamic coupling fusion feature is output. The input tensor is input into the phase-locked frequency domain enhancement branch. The phase-locked frequency domain enhancement branch performs synchronous correlation accumulation on the input according to the perturbation phase-locked signal to extract the response component that is in phase with the perturbation time sequence. After frequency domain trend enhancement processing is performed on the response component, a frequency domain residual feature is formed. The frequency domain residual feature is superimposed with the perturbation dynamic coupling fusion feature to obtain the time-frequency joint enhancement feature. Global pooling and linear mapping are performed on the time-frequency joint enhancement features to output the dynamic fingerprint vector.

[0013] Optionally, the solution yields a candidate set of concentration vectors, including: The dynamic fingerprint vector and three sets of perturbation response sequences are obtained. The dynamic fingerprint vector is allocated according to the time window consistent with the perturbation response sequence. It is then combined with the three sets of perturbation response sequences within the corresponding time window according to the perturbation type dimension, time dimension, and feature dimension to form a three-state response tensor. The perturbation dynamic coupling parameter space is constructed based on the three-state response tensor, specifically as follows: Three sets of perturbation sub-tensors are extracted from the three-state response tensor using the perturbation type as the index. Principal component decomposition is then performed on each perturbation sub-tensor to obtain the perturbation principal feature basis. The angle between the corresponding subspaces of any two sets of principal eigenbases is calculated based on the three sets of perturbation principal eigenbases, and the perturbation coupling degree matrix is ​​formed. Based on the perturbation coupling degree matrix, select the principal feature basis vectors with an angle less than a preset threshold to form a shared subspace, and select the principal feature basis vectors with an angle greater than a preset threshold as perturbation-specific subspaces. Combine the shared subspace and each perturbation-specific subspace in a block diagonal manner to obtain the perturbation dynamic coupling parameter space. Set up concentration vector solution variables in the perturbation dynamic coupling parameter space, and apply non-negativity constraints and monotonically increasing constraints to the concentration vector solution variables; In the perturbation dynamic coupling parameter space, a diffusion-adsorption energy conservation constraint is applied so that the response energy composed of diffusion response contribution and adsorption hysteresis contribution in the same perturbation period satisfies the consistency condition among the three perturbation types. Under the conditions of satisfying non-negativity constraint, monotonically increasing constraint and diffusion-adsorption energy conservation constraint, the solution variables of the concentration vector are iteratively updated. Write the concentration vector obtained from each iteration into the concentration vector candidate set, and use the difference between the concentration vectors obtained from two adjacent iterations being less than a preset convergence threshold as the iteration termination condition, and output the concentration vector candidate set.

[0014] Optionally, the output of real-time concentration values ​​and rate of change includes: Obtain a candidate set of concentration vectors, and number and store each candidate concentration vector according to the time index. At the same time, obtain the hysteresis area, phase hysteresis and amplitude change data in the corresponding time window of the three-state dynamic trajectory tensor, as well as the dynamic fingerprint vector. The weighting coefficients for each disturbance type are calculated based on the dynamic fingerprint vector. The weighting coefficients are then applied to the hysteresis area, phase hysteresis, and amplitude change of the corresponding disturbance type to obtain the comprehensive response for each disturbance type. For each candidate concentration vector, the corresponding disturbance response conservation deviation value is calculated. The disturbance response conservation deviation value is the absolute difference between the sum of the comprehensive response values ​​of the three disturbance types and the response value calculated by the candidate concentration vector according to the conservation mapping relationship. The perturbation response conservation deviation values ​​of all candidate concentration vectors are compared, and the candidate concentration vector with the smallest conservation deviation value that is less than the preset deviation threshold is selected as the real-time true concentration vector of the target gas. The concentration change rate is calculated based on the real-time true concentration vector of the target gas in the current time window and the previous time window, and the real-time true concentration vector and the corresponding concentration change rate are output.

[0015] The beneficial effects of this invention are: This invention introduces three controllable perturbation mechanisms—electrothermal gradient perturbation, microfluidic field perturbation, and microelectric polarization perturbation—within the sampling microcavity. This enables the gas to develop repeatable dynamic response characteristics during perturbation loading and recovery. By constructing a perturbation reference differential curve and a hysteresis response curve using a baseline response sequence as a reference, the traditional single static response measurement is transformed into a multidimensional dynamic response measurement. Compared to existing technologies that rely solely on a single sensor signal and empirical compensation model, this invention mitigates the impact of humidity fluctuations, dust deposition, and sampling link adsorption hysteresis on the measurement results, improving signal stability and anti-interference capabilities under complex operating conditions.

[0016] This invention constructs a three-state dynamic trajectory tensor and uses an improved feature extraction model to obtain dynamic fingerprint vectors. It applies nonnegativity, monotonicity, and diffusion-adsorption energy conservation constraints in the perturbed dynamic coupling parameter space to perform concentration inversion. This transforms the concentration solution process from a single empirical regression to a multi-condition screening process with physical consistency constraints. It can effectively suppress model drift and parameter mismatch problems in strongly nonstationary and multi-gas coupling environments, avoid the impact of frequent standard gas calibration on the continuous operation of the system, and improve the accuracy and reliability of concentration inversion.

[0017] This invention introduces a conservation residual screening mechanism into the concentration candidate set to ensure that the final output concentration solution satisfies the disturbance response conservation relationship, thus making the measurement results self-consistent and verifiable. Compared with existing technologies, this invention can operate stably for a long time in the lithium carbonate production process, improving concentration measurement accuracy while maintaining continuous online monitoring, reducing maintenance costs, and enhancing the system's adaptability to sudden changes in operating conditions, demonstrating significant engineering application value. Attached Figure Description

[0018] 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: Figure 1 This is a flowchart of the online monitoring method for lithium carbonate production waste gas concentration based on gas sensing proposed in this invention; Figure 2This is a schematic diagram of the improved InceptionTime model structure of the online monitoring method for lithium carbonate production waste gas concentration based on gas sensing proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 and Figure 2 A method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing includes: Sampling devices and gas sensing units are installed in the exhaust gas pipeline of lithium carbonate production to collect multi-source coupled monitoring data. The multi-source coupled monitoring data is preprocessed to form a baseline response sequence. Microscale electrothermal gradient perturbation, local microfluidic field perturbation, and microelectric polarization perturbation are sequentially applied in the sampling microcavity according to a preset time sequence. Gas sensing response signals are collected in the steady-state and recovery phases of each perturbation mode to form three sets of perturbation response sequences. Using the baseline response sequence as a reference, the three sets of disturbance dynamic response sequences are subjected to baseline difference processing to construct disturbance baseline difference curves and perform disturbance-recovery cyclic splicing processing to form hysteresis response curves. The hysteresis area, phase lag and amplitude change of each hysteresis response curve are calculated to generate a three-state dynamic trajectory tensor. The three-state dynamic trajectory tensor is input into the improved InceptionTime model, which includes multi-scale dilated convolution branches, hysteresis gated fusion branches, and phase-locked frequency domain enhancement branches. Multi-timescale features are extracted and gated fusion and frequency domain enhancement processing is performed to output the dynamic fingerprint vector. The dynamic fingerprint feature vector and three sets of perturbation response sequences are assembled into a three-state response tensor. The perturbation dynamic coupling parameter space is constructed and non-negativity constraints, monotonically increasing constraints and diffusion-adsorption energy conservation constraints are applied. The concentration vector candidate set is obtained by solving the problem. Select the unique concentration solution that satisfies the disturbance response conservation relationship from the candidate set of concentration vectors as the real-time true concentration value of the target gas, and output the real-time true concentration value and the rate of change.

[0021] In this invention, the three active perturbations, the three-state dynamic trajectory tensor, the improved InceptionTime model, the perturbation dynamic coupling parameter space, and the multiple constraints are not independent algorithm stacks, but are connected in the order of mass transfer perturbation, sensing response, feature characterization, concentration inversion, and conservation verification. Microscale electrothermal gradient perturbation changes the diffusion rate and adsorption / desorption rate of the target gas in the sampling microcavity, local microfluidic field perturbation changes the boundary layer thickness and gas renewal rate, and microelectric field polarization perturbation changes the migration and adsorption orientation of polar gas molecules and charged particles on the sensing surface. All three perturbations act on the mass transfer response caused by the same target gas concentration, but through different channels. Therefore, their response difference curves have common concentration components and perturbation-specific components.

[0022] The baseline difference curve is used to remove the environmental baseline. The hysteresis area corresponds to the cumulative response difference during the perturbation loading and recovery phases. The phase hysteresis corresponds to the time delay of diffusion and adsorption hysteresis. The amplitude change corresponds to the change in sensing response intensity caused by the change in target gas concentration. These three indicators are stacked according to the perturbation type to form a three-state dynamic trajectory tensor. The improved InceptionTime model is used as a temporal feature extractor to extract multi-scale hysteresis features from this tensor and generate a dynamic fingerprint vector. The dynamic fingerprint vector is used to determine the weight of different perturbation types on concentration inversion. The perturbation dynamic coupling parameter space is constructed by three sets of principal feature bases of perturbation responses. The subspace angle is used to measure the coupling degree between common concentration components and perturbation-specific components. Non-negativity, monotonically increasing, and diffusion-adsorption energy conservation constraints are used to screen out candidate concentration solutions that do not conform to the physical mass transfer relationship.

[0023] In this embodiment, the multi-source coupled monitoring data includes the electrical signal response value of the target gas, the temperature parameters, relative humidity parameters, gas flow rate parameters, and dust concentration parameters within the sampling chamber.

[0024] In this embodiment, the multi-source coupled monitoring data is preprocessed to form a baseline response sequence, including: Multi-source coupled monitoring data are time-aligned according to a unified sampling period, and data from different sampling channels are interpolated to complete the data and construct a synchronous monitoring data sequence. Outlier identification and removal are performed on the synchronous monitoring data sequence. A sliding window statistical method is used to correct the mutation data that exceeds the set threshold, resulting in an outlier-free data sequence. Without applying a perturbation mode, the outlier-removed data sequence is filtered for steady-state intervals and averaged to form a baseline response sequence.

[0025] In practice, the unified sampling frequency of multi-source coupled monitoring data is set to 1Hz to 5Hz, with 2Hz used in the embodiment; the sliding window length is set to 20 to 60 sampling points, with 40 sampling points used in the embodiment; the outlier identification threshold is set to the amplitude of the change between adjacent sampling points exceeding 3 times the standard deviation of the same window, and the allowable deviation range of the benchmark response sequence is set to ±2% of the benchmark mean.

[0026] In this embodiment, three sets of perturbation dynamic response sequences are formed, including: A perturbation period is set, and each perturbation period is divided into a perturbation stabilization phase and a recovery phase after perturbation removal. The sampling cavity is kept in a baseline state without any perturbation. The perturbation period is 20 to 40 seconds, and each perturbation period is divided into a perturbation stabilization phase and a recovery phase after perturbation removal, as follows: In practice, each perturbation cycle is set to 30 seconds, with a 12-second perturbation loading phase, an 8-second perturbation stabilization phase, and a 10-second recovery phase after perturbation removal. This can be adjusted to between 20 and 40 seconds depending on the sampling chamber volume and the target gas diffusion time. A baseline recovery interval of at least 5 seconds is set between the three perturbations to ensure that the residual response of the previous perturbation falls back to within the allowable deviation range of the baseline response sequence before initiating the next perturbation.

[0027] Apply the corresponding disturbance control signal at the start of the disturbance period, record the start time of disturbance loading, and determine that the disturbance stabilization stage has been entered when the disturbance control quantity reaches the preset stable amplitude and is maintained for the preset duration. During the stabilization phase of the disturbance, the disturbance control quantity is kept constant, and the gas sensor response signal is continuously collected. When the preset stabilization phase duration is reached, the disturbance control signal is removed, and the disturbance removal time is recorded. After the disturbance control signal is removed, the gas sensor response signal is continuously collected. When the response signal falls back to the allowable deviation range of the reference response sequence or reaches the upper limit of the preset recovery time, the recovery phase is determined to be over, and the phase division of a disturbance cycle is completed. A microscale electrothermal gradient perturbation is applied during the perturbation period. A controllable temperature gradient is formed in the sampling cavity by controlling the micro heating element. Gas sensing response signals are continuously collected during the perturbation stabilization phase. After the electrothermal perturbation is removed, response signals are collected again during the recovery phase. The response signals of the perturbation stabilization phase and the recovery phase are combined in chronological order to form the first set of perturbation response sequences. The implementation parameters for the microscale electrothermal gradient perturbation are as follows: the micro heating element is arranged on the side wall or bottom wall of the sampling microcavity, and the temperature rise is controlled between 3°C and 8°C, with 5°C used in the embodiment; the heating rate is controlled between 0.5°C / s and 1.5°C / s, with 1°C / s used in the embodiment; the temperature of the micro heating element is driven by closed-loop temperature control to ensure that the overall temperature of the sampled gas does not exceed the rated operating temperature of the sensor, thus avoiding direct high-temperature impact on the sensing unit.

[0028] After completing the microscale electrothermal gradient perturbation, a local microflow field perturbation is applied. The gas flow state in the sampling cavity is changed by adjusting the perturbation structure. Gas sensing response signals are collected during the stabilization phase of the perturbation. After the flow field perturbation is removed, response signals are collected again during the recovery phase. The response signals from the two phases are combined in chronological order to form a second set of perturbation response sequences. The perturbation structure is a controllable perturbation component set downstream of the sampling microcavity inlet. The perturbation component includes a micro stepper motor and perturbation blades connected to its output shaft. The perturbation blades have an arc-shaped plate structure and the rotation axis is perpendicular to the mainstream gas flow direction. When the perturbation is applied, the micro stepper motor is driven to rotate the perturbation blades to a preset angle, forming a local velocity gradient and vortex region on both sides of the blades, thus changing the gas flow state in the sampling cavity. When the perturbation is removed, the perturbation blades are reset to the initial position, and the gas flow returns to the baseline state. The implementation parameters for local micro-flow field disturbance are as follows: the rotation angle of the disturbance blade is 10° to 25°, and 15° is used in the embodiment; the micro-flow field pulse velocity increment is 0.05m / s to 0.25m / s, and 0.12m / s is used in the embodiment. The micro stepper motor is arranged in the dustproof sealed cavity outside the sampling microcavity, and drives the disturbance blade to move through the corrosion-resistant rotating shaft or magnetic coupling structure, so as to avoid the driving components being directly exposed to the high temperature, high humidity and high dust exhaust gas environment.

[0029] When the dust concentration at the site is high, the local micro-flow field disturbance can also be replaced by a bypass pulse airflow formed by a corrosion-resistant miniature solenoid valve. The replacement method still aims to change the boundary layer thickness and gas renewal rate of the sampling microcavity, without changing the acquisition logic of the disturbance dynamic response sequence.

[0030] After completing the local microfluidic field perturbation, a micro-electric field polarization perturbation is applied. An electric field environment is formed in the sampling cavity by controlling the electrode array. Gas sensing response signals are collected during the perturbation stabilization phase. After the electric field perturbation is removed, response signals are collected again during the recovery phase. The response signals of the two phases are combined in chronological order to form a third set of perturbation response sequences. The implementation parameters for micro-electric field polarization perturbation are as follows: the electrode array uses inert metal or corrosion-resistant coated electrodes, and the electrode spacing is 2mm to 6mm, with 4mm used in the embodiment; a DC or low-frequency pulse voltage of 3V to 12V is applied, with 8V used in the embodiment, corresponding to an electric field strength of 0.75kV / m to 3kV / m; an insulating protective layer and a current-limiting circuit are set on the outside of the electrodes, and the perturbation current is limited to below the milliampere level, in order to form a controllable micro-electric field environment rather than ionized gas.

[0031] The first set of disturbance dynamic response sequences, the second set of disturbance dynamic response sequences, and the third set of disturbance dynamic response sequences are aligned and stored according to a unified time index to form three sets of disturbance dynamic response sequences.

[0032] The sampling microcavity of this invention is connected to the discharge pipeline through a heated sampling tube, a primary dust filter, and a condensate discharge structure. The active perturbation component and sensing unit are set inside the constant temperature and dustproof sampling box, and the deep learning inference hardware is set inside the field control cabinet or edge industrial control computer. It receives feature data and returns concentration results through wired communication. Therefore, the algorithm inference hardware is not directly in a high temperature, high humidity, and high dust exhaust gas environment.

[0033] In this embodiment, the hysteresis area, phase lag, and amplitude change of each hysteresis response curve are calculated to generate a three-state dynamic trajectory tensor, including: Using the baseline response sequence as a reference, the three sets of disturbance dynamic response sequences are aligned with the baseline response sequence according to a unified time index. When the time indices are inconsistent, linear interpolation is performed on the baseline response sequence. The baseline response value is obtained at the corresponding time index and the difference is calculated with the disturbance dynamic response value to obtain three sets of baseline difference sequences. Within each disturbance cycle, the disturbance start time and disturbance removal time are recorded based on the applied control signal. The disturbance loading segment and disturbance recovery segment are determined in the corresponding reference differential sequence. The loading peak point is determined in the disturbance loading segment, and the recovery peak point is determined in the disturbance recovery segment. The disturbance start time, disturbance removal time, loading peak point, and recovery peak point are used as a cyclically spliced ​​set of anchor points. Specifically, the loading peak point is determined within the disturbance loading segment as follows: In the corresponding baseline difference sequence, the response data within the loading segment is continuously scanned starting from the initial time of the disturbance. First, candidate intervals with response amplitudes reaching more than 95% of the maximum response amplitude of the loading segment are selected. Within the candidate intervals, the change in the response slope is calculated. When the response slope gradually decreases from a positive value and approaches zero for the first time, and the absolute value remains less than the preset slope threshold for at least three consecutive sampling points, the response value corresponding to the time point is determined as the loading peak point. If there are multiple time points within the loading segment that meet the conditions, the time point with the largest response amplitude is selected as the final loading peak point. The recovery peak point is determined in the disturbance recovery segment, specifically as follows: In the corresponding baseline difference sequence, the response data within the recovery segment is continuously scanned starting from the time the disturbance is removed. First, the extreme direction of the response value within the recovery segment is determined, and candidate intervals with response amplitudes reaching more than 95% of the extreme amplitude of the recovery segment are selected. Within the candidate intervals, the change in response slope is calculated. When the response slope gradually increases from a negative value and approaches zero for the first time, and the absolute value remains less than the preset slope threshold for at least three consecutive sampling points, the response value corresponding to the time point is determined as the recovery peak point. If there are multiple time points within the recovery segment that meet the conditions, the time point with the largest absolute value of the amplitude is selected as the final recovery peak point. Phase normalization resampling is performed on the perturbation loading segment and the perturbation recovery segment respectively. The perturbation loading segment is resampled into a loading sequence with a preset number of points, and the perturbation recovery segment is resampled into a recovery sequence with the same number of points. The loading sequence and the recovery sequence are paired point by point according to the same sequence number to construct the phase pairing sequence of the perturbation reference difference curve. Based on the anchor point set, the phase pairing sequence is cyclically spliced. The end of the loaded sequence is connected to the beginning of the recovery sequence according to the anchor point correspondence. The end of the recovery sequence is closedly connected to the beginning of the loaded sequence. At each connection position, a three-point weighted average with fixed weights is used for continuous processing to form the hysteresis response curve corresponding to the disturbance type. For each hysteresis response curve, calculate the hysteresis area, phase lag, and amplitude change, where: The hysteresis area is calculated by summing the differences between the loaded sequence and the recovery sequence at the same index using a trapezoidal summation method. Phase lag is the time difference between the time corresponding to the loading peak point and the time corresponding to the recovery peak point; The magnitude change is the difference between the maximum and minimum values ​​of the benchmark difference sequence; The hysteresis area, phase lag, and amplitude change corresponding to each disturbance type are stacked in order of disturbance type to generate a three-state dynamic trajectory tensor.

[0034] The dimensions of the three-state dynamic trajectory tensor are set according to the perturbation type, hysteresis feature, and time window. In this example, the perturbation type dimension is 3, the hysteresis feature dimension is 3, and the number of time windows is 32; the number of phase-normalized resampling points is 64. The hysteresis area of ​​the electrothermal perturbation channel mainly reflects the changes in thermally induced diffusion and adsorption / desorption, the hysteresis area of ​​the microfluidic field channel mainly reflects the changes in boundary layer mass transfer, and the phase hysteresis of the microelectric field channel mainly reflects the changes in polar molecule migration and surface adsorption orientation. All three reflect the same concentration change, forming a physical response structure at the tensor level that can be learned by the model.

[0035] In this embodiment, multi-timescale features are extracted and subjected to gated fusion and frequency domain enhancement processing to output a dynamic fingerprint vector, including: The channel ordering and amplitude normalization of the three-state dynamic trajectory tensor are performed. The characteristic channels corresponding to the hysteresis area, phase hysteresis, and amplitude change are arranged in order of perturbation type to obtain the input tensor. The input tensor is fed into the multi-scale dilated convolution branch of the improved InceptionTime model in parallel. The multi-scale dilated convolution branch uses multiple sets of one-dimensional depthwise separable convolutions with different kernel lengths and dilation rates to extract features from the input tensor. After each set of convolutions, channel mixing is performed by pointwise convolution to output a multi-timescale local feature set. The set of local features at multiple time scales is input into the hysteresis-gated fusion branch. The hysteresis-gated fusion branch generates gating weights based on the boundary positions between the disturbance stability phase and the recovery phase, as well as the positions of the loading peak point and the recovery peak point. Cross-gated weighted fusion is performed between features at different time scales and between features of different disturbance types, and the output disturbance dynamic coupling fusion features are output. The input tensor is input into the phase-locked frequency domain enhancement branch. The phase-locked frequency domain enhancement branch performs synchronous correlation accumulation on the input according to the perturbation phase-locked signal to extract the response component in phase with the perturbation time sequence. After frequency domain trend enhancement processing on the response component, frequency domain residual features are formed. The frequency domain residual features are superimposed with the perturbation dynamic coupling fusion features to obtain the time-frequency joint enhancement features. The perturbation phase-locked signal is generated by the application time sequence of microscale electrothermal gradient perturbation, local microfluidic field perturbation and microelectric polarization perturbation. Global pooling and linear mapping are performed on the time-frequency joint enhancement features to output the dynamic fingerprint vector.

[0036] The key structural parameters of the improved InceptionTime model are as follows: the multi-scale dilated convolution branch includes three sets of one-dimensional depthwise separable convolutions with kernel sizes of 3, 5, and 9, and dilation rates of 1, 2, and 4, respectively. Each set has 32 output channels, and the pointwise convolution kernel size is 1. The hysteresis-gated fusion branch has a gating vector dimension of 32. The phase-locked frequency domain enhancement branch uses phase-locked markers synchronized with three perturbation control signals to perform synchronous correlation accumulation on the in-phase response components. After global pooling, a 32-dimensional dynamic fingerprint vector is output through linear mapping.

[0037] The model training details are as follows: a training set is constructed using calibrated samples and online collected samples, and the training set, validation set, and test set are divided into three groups according to a ratio of 70%, 15%, and 15%, respectively; the number of training rounds is 120, the batch size is 32, the initial learning rate is 0.001, the optimizer is Adam, and the loss function is composed of a weighted average of the concentration prediction mean square error, the disturbance response conservation residual, and a monotonicity penalty term; training is stopped when the validation set loss does not decrease for 15 consecutive rounds, and the model parameters with the lowest validation set loss are retained.

[0038] The input for the inference stage is a tri-state dynamic trajectory tensor consisting of three disturbance types, three types of hysteresis features, and 32 time windows. The number of parameters for a single inference is controlled within the range that an industrial edge control computer can handle, and the inference time is less than one disturbance cycle. It can be deployed in a field control cabinet outside the sampling box.

[0039] In this embodiment, the candidate set of concentration vectors obtained by solving includes: The dynamic fingerprint vector and three sets of perturbation response sequences are obtained. The dynamic fingerprint vector is allocated according to the time window consistent with the perturbation response sequence. It is then combined with the three sets of perturbation response sequences within the corresponding time window according to the perturbation type dimension, time dimension, and feature dimension to form a three-state response tensor. The perturbation dynamic coupling parameter space is constructed based on the three-state response tensor, specifically as follows: Three sets of perturbation sub-tensors are extracted from the three-state response tensor using the perturbation type as the index. Principal component decomposition is then performed on each perturbation sub-tensor to obtain the perturbation principal feature basis. The angle between the corresponding subspaces of any two sets of principal eigenbases is calculated based on the three sets of perturbation principal eigenbases, and the perturbation coupling degree matrix is ​​formed. Based on the perturbation coupling degree matrix, select the principal feature basis vectors with an angle less than a preset threshold to form a shared subspace, and select the principal feature basis vectors with an angle greater than a preset threshold as perturbation-specific subspaces. Combine the shared subspace and each perturbation-specific subspace in a block diagonal manner to obtain the perturbation dynamic coupling parameter space. The preset threshold for the subspace angle can be set to 10° to 20°, and in this embodiment it is set to 15°. When the subspace angle between any two sets of principal feature bases is less than 15°, it is determined that they contain common concentration driving components and enter the shared subspace. When the angle is greater than 15°, it is determined that they mainly characterize the specific response of the corresponding perturbation type and enter the perturbation specific subspace.

[0040] Within the perturbation dynamic coupling parameter space, concentration vector solution variables are set, and non-negativity and monotonically increasing constraints are applied to these variables, where: Nonnegativity constraint refers to limiting each target gas concentration component in the candidate concentration vector so that the value of each target gas concentration component is not less than zero, and truncating each target gas concentration component to zero or back to the previous valid positive value when a negative value appears during the iterative update process. The monotonically increasing constraint means that within the same perturbation period, when the intensity of the corresponding perturbation response increases with time, the concentration component of the candidate concentration vector must not decrease within the corresponding time window. When the concentration component is detected to be inconsistent with the direction of response change, the concentration component is corrected according to the direction of perturbation response change so that the direction of concentration change is consistent with the direction of response intensity change. Diffusion-adsorption energy conservation constraints are applied within the perturbation dynamic coupling parameter space to ensure that the response energy, composed of diffusion response contribution and adsorption hysteresis contribution, satisfies the consistency condition among the three perturbation types within the same perturbation period. The concentration vector solution variables are iteratively updated under the conditions of non-negativity constraints, monotonically increasing constraints, and diffusion-adsorption energy conservation constraints. Specifically, the diffusion-adsorption energy conservation constraints are as follows: within each perturbation period, the diffusion response contribution is calculated for the baseline difference sequence during the perturbation loading stage, obtained by accumulating the change in the response curve during the loading stage relative to the baseline response sequence. The adsorption hysteresis contribution is calculated for the baseline difference sequence during the perturbation recovery stage, obtained by accumulating the decrease in the response curve during the recovery stage relative to the loading peak. The sum of the diffusion response contribution and adsorption hysteresis contribution corresponding to the three perturbation types is calculated respectively, and the difference between the results obtained from the three perturbation types is limited to not exceeding a preset allowable deviation range. When the difference between any two perturbation types exceeds the allowable deviation range, the concentration vector solution variables are adjusted and recalculated until the allowable deviation range is met. Write the concentration vector obtained from each iteration into the concentration vector candidate set, and use the difference between the concentration vectors obtained from two adjacent iterations being less than a preset convergence threshold as the iteration termination condition, and output the concentration vector candidate set.

[0041] Iterative updates are performed using the projection gradient method, with a maximum of 80 iterations. The preset convergence threshold for concentration vector differences is set to 0.0001 to 0.0005, and 0.0002 is used in this example. The normalization allowable deviation for the diffusion adsorption energy conservation constraint is set to 0.085, and candidate concentration vectors exceeding this threshold are not included in the final screening.

[0042] In this embodiment, the real-time actual concentration value and rate of change are output, including: Obtain a candidate set of concentration vectors, and number and store each candidate concentration vector according to the time index. At the same time, obtain the hysteresis area, phase hysteresis and amplitude change data in the corresponding time window of the three-state dynamic trajectory tensor, as well as the dynamic fingerprint vector. The weighting coefficients for each disturbance type are calculated based on the dynamic fingerprint vector. These weighting coefficients are then applied to the hysteresis area, phase hysteresis, and amplitude change of the corresponding disturbance type to obtain the comprehensive response for each disturbance type. Specifically, the weighting coefficients for each disturbance type are calculated based on the dynamic fingerprint vector as follows: The dynamic fingerprint vector is normalized so that all components of the dynamic fingerprint vector are in a uniform amplitude range. The feature sub-segments in the dynamic fingerprint vector are pre-divided according to the type of disturbance, and the feature sub-segments corresponding to electrothermal gradient disturbance, microfluidic field disturbance and microelectric polarization disturbance are extracted respectively. For each type of disturbance, the feature sub-segment is weighted and summed. The weighted sum of each disturbance type is then scaled proportionally to obtain the weight coefficient of each disturbance type. For each candidate concentration vector, the corresponding perturbation response conservation deviation is calculated. This deviation is the absolute difference between the sum of the combined responses of the three perturbation types and the response calculated from the candidate concentration vector according to the conservation mapping relationship, where the conservation mapping relationship is: Based on calibration sample data, a proportional correspondence rule between concentration and comprehensive response is established. Calibration sample data is obtained by introducing standard gas of known concentration into a sampling microcavity and recording the comprehensive response under three disturbance types. The standard gas concentration is set in stages according to the concentration range of the target gas in the production waste gas. Multiple sets of disturbance response data are repeatedly collected at each concentration level. With the standard gas concentration as the independent variable and the comprehensive response of the three disturbance types as the dependent variable, the response proportional coefficient corresponding to each target gas is obtained by multiple linear regression method to establish the proportional correspondence between concentration and comprehensive response. In actual calculation, each target gas concentration component in the candidate concentration vector is multiplied by the corresponding response proportional coefficient. All product results are accumulated to obtain the theoretical comprehensive response corresponding to the candidate concentration vector. Then, the difference between the theoretical comprehensive response and the sum of the comprehensive response of the three disturbance types is calculated and the absolute value is taken to obtain the disturbance response conservation deviation value. The perturbation response conservation deviation values ​​of all candidate concentration vectors are compared, and the candidate concentration vector with the smallest conservation deviation value that is less than the preset deviation threshold is selected as the real-time true concentration vector of the target gas. In this embodiment, the disturbance response conservation deviation threshold is set to 0.085. If the conservation deviation values ​​of all candidate concentration vectors are greater than this threshold, the current time window is marked as an abnormal disturbance window and the three-state micro-perturbation acquisition is restarted. If multiple candidate concentration vectors meet this threshold, the candidate concentration vector with the smallest conservation deviation value and continuous with the concentration change rate of the previous time window is selected as the real-time true concentration vector.

[0043] The concentration change rate is calculated based on the real-time true concentration vector of the target gas in the current time window and the previous time window, and the real-time true concentration vector and the corresponding concentration change rate are output.

[0044] Example 1: To verify the feasibility of this invention in practice, it was applied to an industrial-grade lithium carbonate production line in a lithium carbonate production enterprise. This production line includes roasting, acidification, lithium precipitation, and drying sections, with exhaust gases collected from each section before entering a main emission pipeline. The exhaust gas mainly contains carbon dioxide, hydrogen chloride, and a small amount of ammonia, accompanied by high humidity and dust particles. Long-term on-site operation data shows that the exhaust gas humidity fluctuates between 40% and 88%, and the dust concentration fluctuates between 20 mg / m³ and 150 mg / m³. The exhaust gas velocity changes significantly when the production load is adjusted, exhibiting typical strong non-stationary characteristics. Traditional online monitoring systems use electrochemical sensors combined with temperature and humidity compensation models for concentration correction, but under high humidity or sudden load changes, they suffer from response lag and baseline drift, typically requiring weekly calibration with standard gas to control errors.

[0045] The online monitoring system for lithium carbonate production exhaust gas concentration based on gas sensing, as described in this invention, is installed on the main exhaust duct of the production line. The system first continuously collects multi-source coupled monitoring data, including the target gas electrical signal response value, temperature, humidity, flow rate, and dust concentration parameters. The data is then time-aligned and outlier-removed to form a baseline response sequence. During normal production, the system sequentially applies microscale electrothermal gradient perturbations, local microfluidic field perturbations, and microelectric polarization perturbations according to a preset time sequence. Gas sensing response signals are collected during the loading and recovery phases of each perturbation, forming three sets of perturbation response sequences.

[0046] The specific engineering parameters in this embodiment are as follows: a uniform sampling frequency of 2Hz, a sliding window length of 40 sampling points, a perturbation period of 30 seconds (12 seconds for the loading phase, 8 seconds for the stabilization phase, and 10 seconds for the recovery phase); an electrothermal temperature rise of 5℃ and a heating rate of 1℃ / s; a microflow field pulse velocity increment of 0.12m / s; a perturbation blade angle of 15°; an electrode spacing of 4mm; an electrode voltage of 8V; 64 resampling points for the hysteresis curve; and 32 tensor windows for the three-state dynamic trajectory. The improved InceptionTime model uses convolution kernel sizes of 3, 5, and 9, expansion rates of 1, 2, and 4, 120 training epochs, a batch size of 32, a learning rate of 0.001, a subspace angle threshold of 15°, a conserved residual threshold of 0.085, and a concentration vector convergence threshold of 0.0002.

[0047] Subsequently, using the baseline response sequence as a reference, the three sets of perturbation dynamic response sequences were differentially processed to construct the perturbation baseline difference curve. The loading and recovery phases were then cyclically concatenated in chronological order to form a hysteresis response curve. The hysteresis area, phase lag, and amplitude change were calculated to generate a three-state dynamic trajectory tensor. This three-state dynamic trajectory tensor was input into the multi-scale convolution branch, hysteresis-gated fusion branch, and phase-locked frequency domain enhancement branch of the improved InceptionTime model to extract the dynamic fingerprint vector. Based on the dynamic fingerprint vector, a perturbation dynamic coupling parameter space was constructed. Non-negativity constraints, monotonically increasing constraints, and diffusion-adsorption energy conservation constraints were applied within this space to obtain a candidate set of concentration vectors. The real-time actual concentration values ​​and their rates of change were obtained through conservation residual screening.

[0048] The system ran continuously for 21 days without manual calibration; a baseline setting was performed only once during the initial installation phase. To verify measurement accuracy, comparative tests were conducted every three days at the same sampling point using a portable laboratory analyzer. Hydrogen chloride was selected as the comparison object, and the following data were obtained.

[0049] Table 1 Comparison of Online Monitoring Data for Hydrogen Chloride

[0050] As can be seen from the data in Table 1 above, among the seven selected sampling time points, using the laboratory analysis values ​​as a reference standard, the measurement results of the system of this invention maintained a high degree of consistency with the laboratory data, with a small range of error fluctuation. The relative error at each time point was controlled between 1% and 2%, with a minimum of 0.96% and a maximum of 1.71%, and an average relative error of approximately 1.33%, demonstrating good measurement stability and repeatability. Especially when the laboratory concentration values ​​were high on May 10th and May 13th, the system of this invention was still able to maintain a low error level, indicating that it still has good linear response capability under conditions of large concentration fluctuations.

[0051] In contrast, the values ​​measured by the traditional system were significantly higher than the laboratory analysis values ​​at all sampling points, exhibiting a systematic positive deviation, with the deviation ranging from approximately 3.2 mg / m³ to 4.7 mg / m³. This phenomenon indicates that the traditional system is affected by factors such as humidity fluctuations, dust deposition, and sampling link lag during actual operation, resulting in a combination of baseline rise and response lag. Particularly at the time points of May 4th and May 13th, the differences between the traditional system and laboratory values ​​reached 4.4 mg / m³ and 4.7 mg / m³, respectively, reflecting a further amplification of the error in the medium-to-high concentration range.

[0052] In summary, the system of this invention can maintain a measurement error of less than 2% during continuous operation without manual calibration, and the error distribution at each time point is uniform with no obvious accumulation trend. This indicates that by constructing hysteresis characteristics through three-state perturbation response and combining dynamic fingerprint and conservation constraints for concentration screening, it can effectively suppress measurement deviations caused by environmental interference and diffusion adsorption hysteresis, and achieve stable and high-precision online monitoring of lithium carbonate production exhaust gas concentration.

[0053] 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. A method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing, characterized in that, include: Sampling devices and gas sensing units are installed in the exhaust gas pipeline of lithium carbonate production to collect multi-source coupled monitoring data. The multi-source coupled monitoring data is preprocessed to form a baseline response sequence. Microscale electrothermal gradient perturbation, local microfluidic field perturbation, and microelectric polarization perturbation are sequentially applied in the sampling microcavity according to a preset time sequence. Gas sensing response signals are collected in the steady-state and recovery phases of each perturbation mode to form three sets of perturbation response sequences. Using the baseline response sequence as a reference, the three sets of disturbance dynamic response sequences are subjected to baseline difference processing to construct disturbance baseline difference curves and perform disturbance-recovery cyclic splicing processing to form hysteresis response curves. The hysteresis area, phase lag and amplitude change of each hysteresis response curve are calculated to generate a three-state dynamic trajectory tensor. The three-state dynamic trajectory tensor is input into the improved InceptionTime model, which includes multi-scale dilated convolution branches, hysteresis gated fusion branches, and phase-locked frequency domain enhancement branches. Multi-timescale features are extracted and gated fusion and frequency domain enhancement processing is performed to output the dynamic fingerprint vector. The dynamic fingerprint vector and three sets of perturbation response sequences are assembled into a three-state response tensor. The perturbation dynamic coupling parameter space is constructed and non-negativity constraints, monotonically increasing constraints, and diffusion-adsorption energy conservation constraints are applied. The candidate set of concentration vectors is obtained by solving the problem. Select the unique concentration solution that satisfies the disturbance response conservation relationship from the candidate set of concentration vectors as the real-time true concentration value of the target gas, and output the real-time true concentration value and the rate of change. The output includes real-time concentration values ​​and rates of change, including: Obtain a candidate set of concentration vectors, and number and store each candidate concentration vector according to the time index. At the same time, obtain the hysteresis area, phase hysteresis and amplitude change data in the corresponding time window of the three-state dynamic trajectory tensor, as well as the dynamic fingerprint vector. The weighting coefficients for each disturbance type are calculated based on the dynamic fingerprint vector. The weighting coefficients are then applied to the hysteresis area, phase hysteresis, and amplitude change of the corresponding disturbance type to obtain the comprehensive response for each disturbance type. For each candidate concentration vector, the corresponding disturbance response conservation deviation value is calculated. The disturbance response conservation deviation value is the absolute difference between the sum of the comprehensive response values ​​of the three disturbance types and the response value calculated by the candidate concentration vector according to the conservation mapping relationship. The perturbation response conservation deviation values ​​of all candidate concentration vectors are compared, and the candidate concentration vector with the smallest conservation deviation value that is less than the preset deviation threshold is selected as the real-time true concentration vector of the target gas. The concentration change rate is calculated based on the real-time true concentration vector of the target gas in the current time window and the previous time window, and the real-time true concentration vector and the corresponding concentration change rate are output.

2. The method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing according to claim 1, characterized in that, The multi-source coupled monitoring data includes the electrical signal response value of the target gas, temperature parameters, relative humidity parameters, gas flow rate parameters, and dust concentration parameters within the sampling chamber.

3. The method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing according to claim 1, characterized in that, The preprocessing of multi-source coupled monitoring data to form a baseline response sequence includes: Multi-source coupled monitoring data are time-aligned according to a unified sampling period, and data from different sampling channels are interpolated to complete the data and construct a synchronous monitoring data sequence. Outlier identification and removal are performed on the synchronous monitoring data sequence. A sliding window statistical method is used to correct the mutation data that exceeds the set threshold, resulting in an outlier-free data sequence. Without applying a perturbation mode, the outlier-removed data sequence is filtered for steady-state intervals and averaged to form a baseline response sequence.

4. The method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing according to claim 1, characterized in that, The formation of three sets of perturbation dynamic response sequences includes: Set a disturbance period and divide each disturbance period into a disturbance stabilization phase and a recovery phase after disturbance removal, so as to keep the sampling cavity in a baseline state without applying any disturbance. A microscale electrothermal gradient perturbation is applied during the perturbation period. A controllable temperature gradient is formed in the sampling cavity by controlling the micro heating element. Gas sensing response signals are continuously collected during the perturbation stabilization phase. After the electrothermal perturbation is removed, response signals are collected again during the recovery phase. The response signals of the perturbation stabilization phase and the recovery phase are combined in chronological order to form the first set of perturbation response sequences. After completing the microscale electrothermal gradient perturbation, a local microflow field perturbation is applied. The gas flow state in the sampling cavity is changed by adjusting the perturbation structure. Gas sensing response signals are collected during the perturbation stabilization phase. After the flow field perturbation is removed, response signals are collected again during the recovery phase. The response signals of the two phases are combined in chronological order to form the second set of perturbation response sequences. After completing the local microfluidic field perturbation, a micro-electric field polarization perturbation is applied. An electric field environment is formed in the sampling cavity by controlling the electrode array. Gas sensing response signals are collected during the perturbation stabilization phase. After the electric field perturbation is removed, response signals are collected again during the recovery phase. The response signals of the two phases are combined in chronological order to form a third set of perturbation response sequences. The first set of disturbance dynamic response sequences, the second set of disturbance dynamic response sequences, and the third set of disturbance dynamic response sequences are aligned and stored according to a unified time index to form three sets of disturbance dynamic response sequences.

5. The method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing according to claim 1, characterized in that, The calculation of the hysteresis area, phase lag, and amplitude change of each hysteresis response curve to generate a three-state dynamic trajectory tensor includes: Using the baseline response sequence as a reference, the three sets of disturbance dynamic response sequences are aligned with the baseline response sequence according to a unified time index. When the time indices are inconsistent, linear interpolation is performed on the baseline response sequence. The baseline response value is obtained at the corresponding time index and the difference is calculated with the disturbance dynamic response value to obtain three sets of baseline difference sequences. Within each disturbance cycle, the disturbance start time and disturbance removal time are recorded based on the control signal applied during the disturbance. The disturbance loading segment and disturbance recovery segment are determined in the corresponding reference differential sequence. The loading peak point is determined in the disturbance loading segment, and the recovery peak point is determined in the disturbance recovery segment. The disturbance start time, disturbance removal time, loading peak point, and recovery peak point are used as the anchor point set for cyclic splicing. Phase normalization resampling is performed on the perturbation loading segment and the perturbation recovery segment respectively. The perturbation loading segment is resampled into a loading sequence with a preset number of points, and the perturbation recovery segment is resampled into a recovery sequence with the same number of points. The loading sequence and the recovery sequence are paired point by point according to the same sequence number to construct the phase pairing sequence of the perturbation reference difference curve. Based on the anchor point set, the phase pairing sequence is cyclically spliced. The end of the loaded sequence is connected to the beginning of the recovery sequence according to the anchor point correspondence. The end of the recovery sequence is closedly connected to the beginning of the loaded sequence. At each connection position, a three-point weighted average with fixed weights is used for continuous processing to form the hysteresis response curve corresponding to the disturbance type. For each hysteresis response curve, calculate the hysteresis area, phase lag, and amplitude change, where: The hysteresis area is calculated by summing the differences between the loaded sequence and the recovery sequence at the same index using a trapezoidal summation method. Phase lag is the time difference between the time corresponding to the loading peak point and the time corresponding to the recovery peak point; The magnitude change is the difference between the maximum and minimum values ​​of the benchmark difference sequence; The hysteresis area, phase lag, and amplitude change corresponding to each disturbance type are stacked in order of disturbance type to generate a three-state dynamic trajectory tensor.

6. The method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing according to claim 1, characterized in that, The process of extracting multi-timescale features, performing gated fusion and frequency domain enhancement, and outputting a dynamic fingerprint vector includes: The channel ordering and amplitude normalization of the three-state dynamic trajectory tensor are performed. The characteristic channels corresponding to the hysteresis area, phase hysteresis, and amplitude change are arranged in order of perturbation type to obtain the input tensor. The input tensor is fed into the multi-scale dilated convolution branch of the improved InceptionTime model in parallel. The multi-scale dilated convolution branch uses multiple sets of one-dimensional depthwise separable convolutions with different kernel lengths and different dilation rates to extract features from the input tensor. After each set of convolutions, channel mixing is performed by pointwise convolution to output a multi-timescale local feature set. The set of local features at multiple time scales is input into the hysteresis-gated fusion branch. The hysteresis-gated fusion branch generates gating weights based on the boundary positions of the disturbance stability phase and the recovery phase, as well as the positions of the loading peak point and the recovery peak point. Cross-gated weighted fusion is performed between features at different time scales and between features of different disturbance types, and the disturbance dynamic coupling fusion feature is output. The input tensor is input into the phase-locked frequency domain enhancement branch. The phase-locked frequency domain enhancement branch performs synchronous correlation accumulation on the input according to the perturbation phase-locked signal to extract the response component that is in phase with the perturbation time sequence. After frequency domain trend enhancement processing is performed on the response component, a frequency domain residual feature is formed. The frequency domain residual feature is superimposed with the perturbation dynamic coupling fusion feature to obtain the time-frequency joint enhancement feature. Global pooling and linear mapping are performed on the time-frequency joint enhancement features to output the dynamic fingerprint vector.

7. The method for online monitoring of lithium carbonate production waste gas concentration based on gas sensing according to claim 1, characterized in that, The solution yields a candidate set of concentration vectors, including: The dynamic fingerprint vector and three sets of perturbation response sequences are obtained. The dynamic fingerprint vector is allocated according to the time window consistent with the perturbation response sequence. It is then combined with the three sets of perturbation response sequences within the corresponding time window according to the perturbation type dimension, time dimension, and feature dimension to form a three-state response tensor. The perturbation dynamic coupling parameter space is constructed based on the three-state response tensor, specifically as follows: Three sets of perturbation sub-tensors are extracted from the three-state response tensor using the perturbation type as the index. Principal component decomposition is then performed on each perturbation sub-tensor to obtain the perturbation principal feature basis. The angle between the corresponding subspaces of any two sets of principal eigenbases is calculated based on the three sets of perturbation principal eigenbases, and the perturbation coupling degree matrix is ​​formed. Based on the perturbation coupling degree matrix, select the principal feature basis vectors with an angle less than a preset threshold to form a shared subspace, and select the principal feature basis vectors with an angle greater than a preset threshold as perturbation-specific subspaces. Combine the shared subspace and each perturbation-specific subspace in a block diagonal manner to obtain the perturbation dynamic coupling parameter space. Set up concentration vector solution variables in the perturbation dynamic coupling parameter space, and apply non-negativity constraints and monotonically increasing constraints to the concentration vector solution variables; In the perturbation dynamic coupling parameter space, a diffusion-adsorption energy conservation constraint is applied so that the response energy composed of diffusion response contribution and adsorption hysteresis contribution in the same perturbation period satisfies the consistency condition among the three perturbation types. Under the conditions of satisfying non-negativity constraint, monotonically increasing constraint and diffusion-adsorption energy conservation constraint, the solution variables of the concentration vector are iteratively updated. Write the concentration vector obtained from each iteration into the concentration vector candidate set, and use the difference between the concentration vectors obtained from two adjacent iterations being less than a preset convergence threshold as the iteration termination condition, and output the concentration vector candidate set.

Images