A method and system for recognizing the processing stage of polygonum multiflorum thunb by multi-temperature zone electronic nose

Abstract

The application discloses a kind of Polygonum multiflorum nine steaming nine sun multi-temperature zone active sniffing resonance depth identification method and system.Polygonum multiflorum sample headspace gas is sent into electronic nose sensor array with independent heating channel, and multiple sensors are in low-temperature adsorption, medium-temperature reaction and high-temperature desorption state according to temperature excitation matrix, and the hardware execution log of synchronous acquisition electronic nose timing response, temperature and valve is collected;According to signal quality and hardware excitation, generate random resonance parameters, enhance temperature step edge, low-temperature weak response section or high-temperature recovery section;Again, the original response, resonance enhanced response and hardware excitation characterization are fused, and the steaming and sunning cycle category, continuous processing process value and confidence are output;When confidence is insufficient, the next temperature excitation sequence is automatically selected for reexamination.The application can improve the accuracy, robustness and interpretability of adjacent processing stage recognition of Polygonum multiflorum.

Description

Technical Field

[0001] This invention relates to the fields of electronic nose gas sensing and artificial intelligence recognition technology, and in particular to a method and system for identifying the nine-steaming and nine-drying processing stages of Polygonum multiflorum by actively sniffing volatile gases from Polygonum multiflorum samples through multi-sensor differential temperature control, and by co-regulating the sensor temperature excitation matrix with random resonance parameters and depth model uncertainty. Background Technology

[0002] Polygonum multiflorum is a commonly used Chinese medicinal herb. Its nine-steaming and nine-drying processing will cause continuous changes in volatile components, color and internal chemical components. These changes will affect the efficacy. Therefore, it is of practical significance to quickly distinguish between adjacent stages from the first to the ninth steaming.

[0003] In current production of Polygonum multiflorum, the significant price differences between different steaming and drying cycles lead to the market practice of passing off low-steamed and high-steamed varieties as high-steamed ones, and inferior products as genuine ones. This creates an urgent need for rapid and objective identification during the processing stage. However, current methods for distinguishing between different processing stages often rely on traditional methods such as manual experience and color observation, which are highly subjective and prone to large judgment errors, making it difficult to meet the rapid screening needs in the production and distribution process. Electronic noses, which utilize multi-channel gas sensor arrays to collect the volatile gas responses of samples, can form odor fingerprints in a short time, showing broad application prospects in areas such as tracing the origin of Chinese medicinal materials, processing status, and quality identification.

[0004] However, existing electronic nose systems mostly use a fixed sensor operating temperature, and the sampling hardware typically only performs the functions of air intake, air conduction, and detection. The sensor heating unit is also mostly understood as an auxiliary component to maintain the operating temperature, and is not used as a programmable active excitation source. For the nine-steaming and nine-drying processing of Polygonum multiflorum, the release ratio of volatile components in the sample changes continuously with the degree of processing at different stages, and the adsorption, surface reaction, and desorption rates of gas molecules on the surface of gas-sensitive materials such as metal oxides are closely related to the sensor operating temperature. Therefore, the electronic nose response obtained at a single temperature is often only a local result under a certain detection condition, and is prone to problems such as response saturation and insufficient sensitivity at low concentration stages.

[0005] Further research revealed that sensor temperature steps, temperature rise / fall ramps, and short-duration high-temperature pulses not only alter the response amplitude and rise slope of the same headspace gas on different sensor surfaces, but also change the matching relationship between weak signals, system noise, and external noise during stochastic resonance processing. Treating sensor temperature control as fixed hardware and independently setting stochastic resonance parameters may lead to a mismatch between the weak dynamic segments generated by temperature excitation and the stochastic resonance enhancement window, resulting in unstable signal enhancement and even introducing redundant fluctuations that reduce the identification stability of adjacent processing stages.

[0006] Furthermore, the most discriminative sensor temperature range differs across different processing stages and response intensities of Polygonum multiflorum. Some samples exhibit amplified slight differences in volatiles during the low-temperature adsorption phase, while others show more pronounced differences in response slope during the mid-temperature reaction phase. Still others require observation of hysteresis characteristics through high-temperature desorption. The classification confidence and uncertainty information output by the deep learning model can reflect the adequacy of the current temperature excitation matrix and the stochastic resonance enhancement strategy.

[0007] Therefore, an active sniffing electronic nose detection scheme is needed, so that the sampling hardware no longer performs fixed passive collection, but instead divides multiple gas sensors into a low-temperature adsorption group, a medium-temperature reaction group, and a high-temperature desorption group. Simultaneously, based on the dynamic response quality of the *Polygonum multiflorum* sample, the effect of stochastic resonance enhancement, and the uncertainty of the depth model, the target temperature, heating / cooling slope, and stochastic resonance parameters of each sensor are coordinated and adjusted. This creates a closed-loop linkage between multi-sensor differentiated temperature adjustment, nonlinear stochastic resonance enhancement, and the depth time-series recognition model, thereby more completely capturing the dynamic volatilization characteristics of *Polygonum multiflorum* during its nine-stage steaming and drying process, improving recognition accuracy, robustness, and interpretability. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-temperature zone electronic nose identification method and system for Polygonum multiflorum during the processing stage, thereby solving the problems of insufficient dynamic separability caused by fixed sensor operating temperature, unstable generalization of fixed random resonance parameters, and inability to actively re-examine low-confidence samples in the electronic nose detection of Polygonum multiflorum after nine steaming and nine drying. The invention provides a method and system that coordinates sensor temperature excitation, random resonance enhancement, and ordered depth identification.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The system of the present invention includes a multi-sensor differentiated temperature-controlled sampling excitation unit, an electronic nose sensor array, a temperature excitation matrix generation module, a sampling quality judgment module, a random resonance parameter adjustment module, a feature fusion module, a steam generation identification module, a re-inspection decision module, and a parameter update library.

[0011] The multi-sensor differentiated temperature-controlled sampling excitation unit controls the target temperature, heating / cooling slope, and high-temperature pulse width of each sensor in the electronic nose array, and simultaneously controls the clean air channel, sample gas channel, multi-way valve group, and flow controller, enabling the same sample of Polygonum multiflorum headspace gas to be detected in parallel at different sensor temperature zones. The electronic nose sensor array consists of multiple gas sensors with independent heating capabilities, used to collect the electronic nose response signals of Polygonum multiflorum volatile gases in the low-temperature adsorption zone, medium-temperature reaction zone, and high-temperature desorption zone. The temperature excitation matrix generation module generates the target temperature sequence of each sensor for each sampling time period and corrects the heating power based on the actual temperature feedback. The sampling quality judgment module evaluates the quality of the current sampling signal, including signal-to-noise ratio assessment, baseline drift detection, and signal quality vector input. The system comprises the following modules: a stochastic resonance parameter adjustment module, a feature fusion module, and a parameter update library. The stochastic resonance parameter adjustment module automatically adjusts the potential function parameters, noise intensity, and integration step size of the stochastic resonance branch based on the signal quality vector and temperature excitation execution parameters to enhance the weak dynamic signal components generated by differential temperature control. The feature fusion module fuses the original electronic nose response, the stochastic resonance-enhanced response, and the temperature excitation execution parameters to form a comprehensive odor feature. The steaming cycle identification module, connected to the feature fusion module, identifies the processing stage of the sample based on the fused features. The re-inspection decision module, connected to the steaming cycle identification module, evaluates the confidence level of the identification results and generates the next round of differential temperature control sampling excitation sequence or re-inspection instruction when the confidence level is below a threshold. The parameter update library stores the temperature excitation matrix, stochastic resonance parameters, model weights, and re-inspection results for subsequent optimization.

[0012] This invention also provides a multi-temperature zone electronic nose identification method for the processing stage of Polygonum multiflorum, comprising the following steps:

[0013] S1. Place the Polygonum multiflorum sample to be tested in the sample headspace chamber, and form a volatile gas headspace according to the preset mass, filling volume, and headspace equilibrium time. Record the sample number, batch, and evaporation cycle. Before entering the testing chamber, flush the sensor chamber with clean air and allow each sensor to complete baseline recovery at a preset cleaning temperature.

[0014] S2. Generate a multi-sensor differentiated temperature-controlled sampling excitation sequence based on the task to be detected, including the target temperature, temperature rise / fall slope, and high-temperature pulse width of each sensor in multiple time periods. Among them, some sensors are assigned to different temperature zones in the same time period, including low-temperature adsorption zone, medium-temperature reaction zone, and high-temperature desorption zone.

[0015] S3. The multiple gas sensors in the electronic nose sensor array are divided into three temperature control groups according to their spatial positions, and each group is controlled by an independent heating drive channel. The controller outputs PWM to each heating channel according to the temperature excitation matrix Theta(t), and corrects the heating power according to the temperature feedback closed loop.

[0016] S4. Acquire the baseline response of each temperature control group under the clean air channel. When the deviation between the actual temperature of each sensor and the target temperature is kept within the preset allowable range and the baseline drift is lower than the preset threshold, switch the multi-way valve group to allow the headspace gas of the Polygonum multiflorum sample to enter the sensor gas chamber with a preset carrier gas flow rate.

[0017] S5. Execute the temperature excitation matrix Theta(t) during the sample gas entering the sensor gas chamber: the low temperature group is used to prolong the adsorption time of gas molecules on the surface of the sensitive material, the medium temperature group is used to enhance the surface reaction difference, and the high temperature group is used to accelerate desorption recovery and form a recovery segment difference.

[0018] S6. Synchronously collect the multi-channel electronic nose timing response y(t), target temperature of each sensor, actual temperature of each sensor, heating power, valve status and timestamp to form a hardware execution log m(t); divide the response curve into segments according to the temperature switching time to obtain dynamic response data of the low temperature adsorption segment, the medium temperature reaction segment and the high temperature desorption segment.

[0019] S7. Calculate the signal quality vector based on the multi-channel timing response and hardware execution log. and hardware excitation vector The signal quality vector includes dynamic range, signal-to-noise ratio, baseline drift, etc.; the hardware excitation vector includes temperature excitation matrix number, target temperature zone, actual temperature zone, and temperature rise / fall slope.

[0020] S8. Transfer the signal quality vector and hardware excitation vector The input parameter generator outputs the potential function parameters, noise intensity, and integration step size of the stochastic resonance branch, so that the stochastic resonance processing window matches the temperature step edge, the low-temperature weak response section, or the high-temperature recovery section.

[0021] S9. Extract the original and resonant representations using the original response encoder and the stochastic resonant response encoder, respectively, and extract the temperature excitation representation using the hardware excitation embedding encoder. Based on the gating coefficient and attention weight, fuse the three representations to obtain the stage representation, and use the stage recognition model to output the category probability, continuous processing progress value, and confidence level for each steaming and drying round.

[0022] S10. When the confidence level meets the preset output conditions, output the evaporation cycle identification result; when the confidence level does not meet the preset output conditions, select the re-examination sequence of the next round of temperature zone rotation from the candidate temperature excitation library based on the model uncertainty, signal quality vector and temperature tracking error.

[0023] S11. Execute the next round of differentiated temperature-controlled sampling excitation sequence and re-acquire the electronic nose response until the output conditions are met. Write the confirmed samples, temperature excitation matrix, hardware execution log, random resonance parameters and recognition results into the parameter joint update library.

[0024] To further limit the control effect of temperature regulation on stochastic resonance processing, the stochastic resonance parameters are jointly determined by the sampling quality score and the temperature hardware excitation features. The temperature hardware excitation feature vector can be expressed as:

[0025]

[0026] in, This is the temperature hardware excitation feature vector corresponding to the nth sampling. This represents the average value of the target temperature or actual temperature of each sensor within the sampling window. This refers to the temperature span or temperature step amplitude between multiple temperature zones. This refers to the rate of temperature change or the intensity at the edge of a temperature step. , and These represent the residence times in the low-temperature adsorption zone, the medium-temperature reaction zone, and the high-temperature desorption zone, respectively. This represents the normalized temperature tracking error.

[0027] Based on the above temperature hardware excitation characteristics, the generation relationship of random resonance parameters can be expressed as:

[0028]

[0029] in, The parameter generation function is obtained through training, calibration, or table lookup, and ψ is its function parameters or model parameters; and These are the parameters of the stochastic resonance potential function. For input gain, For noise intensity, This is the integration step size. The stochastic resonance parameters are not set independently of the hardware, but rather depend on the current sampling quality. and temperature excitation state Synchronous changes enable the resonance enhancement to match the isothermal actions of low-temperature weak response, medium-temperature reaction changes, and high-temperature desorption recovery.

[0030] Gated fusion of three-branch representations can be expressed as:

[0031]

[0032] in, For the original response characterization, Characterization of stochastic resonance enhancement response, For hardware stimulus representation, , and These are the gating attention weights for the corresponding branches. This represents the stage representation after fusion. It shows that the deep model does not rely solely on any one signal, but rather adaptively allocates the contributions of the three types of representations based on sample quality and temperature excitation state.

[0033] Compared with existing technologies, this invention has the following advantages: This invention transforms the sensor heating unit from a fixed insulation component into an active sniffing excitation source. Through multi-sensor differentiated temperature control, it enables the same *Polygonum multiflorum* headspace gas to generate multi-view dynamic responses under low-temperature adsorption, medium-temperature reaction, and high-temperature desorption conditions. Compared to fixed-temperature sampling, it can obtain richer temperature-sensor-time odor fingerprints. Simultaneously, it generates random resonance parameters through the temperature excitation matrix, hardware execution log, and signal quality vector, instead of using fixed noise intensity. When adjacent evaporation cycles are difficult to distinguish at a fixed temperature, model uncertainty can be used to select the next cycle for low-temperature extension, medium-temperature encryption, and high-temperature pulse, transforming the detection device from a passive acquisition system into a model-driven active temperature sniffing system, improving the processing capability of boundary and anomalous samples. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the system structure of the present invention.

[0035] Figure 2 This is a flowchart of the multi-temperature zone electronic nose recognition method of the present invention.

[0036] Figure 3 This is a radar chart showing the maximum response values ​​of the six sensors for the nine steaming cycles of this invention.

[0037] Figure 4 This is a normalized average response heatmap of each steam-sensor in this invention.

[0038] Figure 5 This is a PCA distribution diagram of the dynamic features of the electronic nose of the present invention.

[0039] Figure 6 This is a diagram of the five-fold cross-validation identification confusion matrix of the present invention.

[0040] The attached figures include: 1-sample air inlet channel, 2-air inlet channel, 3-sampling pump, 4-cleaning pump, 5-multi-way valve group, 6-sensor, 7-temperature probe, 8-PTC heater, 9-sensor air chamber, 10-data acquisition card, 11-control module, 12-industrial computer. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Example 1: Active sniffing front-end sampling device system based on multi-sensor differentiated temperature control

[0043] like Figure 1 As shown, the active sniffing front-end sampling device in this embodiment includes a sample headspace chamber, a clean air channel 2, a sample gas channel 1, a multi-way valve group 5, a sensor gas chamber 9, an electronic nose sensor array 6, a PTC heater 8, a temperature probe 7, a control module 11, and a host computer 12. The control module 11 can be a microcontroller, PLC, or embedded controller, and the host computer 12 can be an industrial computer used to issue temperature excitation matrices, record hardware execution logs, and call back-end recognition models.

[0044] The sample headspace chamber is used to contain a preset mass of Polygonum multiflorum sample and form a volatile gas headspace. The outlet of the sample headspace chamber is connected to a multi-way valve assembly via a sample gas channel, and a clean air channel is connected to the multi-way valve assembly via an air filter. The multi-way valve assembly 5 has a clean air position and a sample gas position. The clean air position is used for sensor baseline cleaning and post-detection recovery, and the sample gas position is used to deliver Polygonum multiflorum headspace gas into the sensor gas chamber. Six gas sensors S1 to S6 are installed in the sensor gas chamber 9. Each gas sensor is a metal oxide semiconductor sensor, and each sensor is equipped with an independent heater 8. The sensor gas chamber 9 is also equipped with a gas chamber temperature probe 7 to record non-target disturbances and provide data for back-end model correction. The control module 11 uses the target temperature of each sensor as input and the actual temperature as feedback, and adjusts the heating power using a segmented proportional control method.

[0045] In one specific implementation, the six gas sensors are divided into a low-temperature adsorption group (G_L), a medium-temperature reaction group (G_M), and a high-temperature desorption group (G_H). The target temperature T_L for the low-temperature adsorption group is set in the range of 150℃ to 220℃, the target temperature T_M for the medium-temperature reaction group is set in the range of 230℃ to 300℃, and the target temperature T_H for the high-temperature desorption group is set in the range of 310℃ to 380℃. The low-temperature adsorption group is used to enhance the adsorption differences of weakly volatile substances on the surface of the sensitive material, the medium-temperature reaction group is used to enhance the differences in surface reaction rates, and the high-temperature desorption group is used to observe the differences in rapid response and desorption rates.

[0046] The control module 11 organizes the target temperatures of each sensor in each sampling time period into a temperature excitation matrix Theta, where the matrix rows correspond to the sensor numbers and the matrix columns correspond to the sampling time periods. For example, in the first time period, S1 to S2 can be in the low-temperature adsorption zone, S3 to S4 in the medium-temperature reaction zone, and S5 to S6 in the high-temperature desorption zone; in the second time period, S1 to S2 can be switched to the medium-temperature reaction zone, S3 to S4 to the high-temperature desorption zone, and S5 to S6 to the low-temperature adsorption zone; the third time period can continue to rotate, so that the same sample of Polygonum multiflorum headspace gas can be detected under multiple temperature zone combinations. Thus, the electronic nose output is no longer a response curve at a single fixed temperature, but a three-dimensional dynamic response that includes sensor number, temperature zone status, and time period information.

[0047] The sampling execution process in this embodiment is as follows:

[0048] A1. After powering on, the control module 11 reads the sensor number, heating channel number, temperature calibration parameters, and the temperature excitation matrix Theta selected for this test, and checks whether each heating channel, multi-way valve group 5, and air chamber temperature probe 7 are online.

[0049] A2. Switch the multi-way valve group 5 to the clean air position, so that clean air enters the sensor air chamber at the first cleaning flow rate, and at the same time raise each sensor to the cleaning temperature and maintain the preset cleaning time to remove the residue of the previous sample.

[0050] A3. After cleaning, lower each sensor to the initial target temperature corresponding to the first column of the temperature excitation matrix, and collect baseline responses under clean air conditions. When the deviation between the actual temperature and the target temperature remains within the allowable range and the baseline drift is less than the preset threshold, proceed to the sample detection stage.

[0051] A4. Place the pre-set mass of Polygonum multiflorum sample into the sample headspace chamber and seal it. After the pre-set headspace equilibration time, the control module 11 switches the multi-way valve group 5 to the sample gas position, so that the sample headspace gas enters the sensor gas chamber at the second sampling flow rate.

[0052] A5. While the sample gas enters the gas chamber 9, the control module 11 performs temperature adjustment column by column according to the temperature excitation matrix Theta: increasing the heating power for sensors that need to be heated and decreasing the heating power for sensors that need to be cooled, until the corresponding temperature zone is reached and maintained for a preset time.

[0053] A6. Within each temperature range, synchronously record the gas-sensitive response, resistance or voltage output, target temperature, actual temperature, heating power, temperature tracking error, valve status, and timestamp of each sensor.

[0054] A7. After sampling the current temperature segment is completed, the control module 11 switches to the next temperature segment according to the temperature excitation matrix. If the current temperature segment is the low-temperature adsorption segment, the focus is on recording the response rise slope and curve area; if the current temperature segment is the medium-temperature reaction segment, the focus is on recording the maximum response value, peak time, and cross-sensor response ratio; if the current temperature segment is the high-temperature desorption segment, the focus is on recording the recovery speed and residual offset.

[0055] A8. By moving the original low-temperature group to medium or high temperature, the original medium-temperature group to high or low temperature, and the original high-temperature group to low or medium temperature, the continuous response of the same sensor in different temperature zones can be obtained.

[0056] A9. After the sample testing is completed, the multi-way valve group 5 is switched to the clean air position. When all sensors return to the baseline allowable range, the front-end sampling ends.

[0057] A10 and control module 11 send the multi-channel electronic nose timing response y(t), temperature excitation matrix Theta(t), hardware execution log m(t), and sensor status flag generated in this test to host computer 12.

[0058] A11 and host computer 12 slice y(t) according to the temperature range boundary, generate low temperature adsorption characteristics, medium temperature reaction characteristics, high temperature recovery characteristics and cross-temperature range difference characteristics, and provide them to the sampling quality judgment module, the random resonance parameter adjustment module and the subsequent depth recognition model.

[0059] The cross-temperature range difference characteristics include the maximum response difference, response curve area difference, peak time difference, and response ratio of different sensor groups in the same temperature range for the same sensor in the low, medium, and high temperature ranges. Using this method, even if the response peak values ​​of two different processing stages of Polygonum multiflorum are similar at a fixed temperature, they can be distinguished by the kinetic differences in the low-temperature adsorption, medium-temperature reaction, and high-temperature desorption processes.

[0060] The multi-sensor differentiated temperature-controlled sampling excitation unit outputs a multi-channel electronic nose timing response y(t), a temperature excitation matrix Theta(t), and a hardware execution log m(t). m(t) includes the target temperature, actual temperature, heating power, temperature tracking error, and valve status for each sensor. This data, aligned with a unified timestamp, is then fed into the sampling quality judgment module, the random resonance parameter adjustment module, and the feature fusion module.

[0061] Example 2: Ordered Identification and Active Re-inspection of Polygonum multiflorum after Nine Steamings

[0062] This embodiment uses the multi-channel electronic nose timing response y(t), temperature excitation matrix Theta(t), and hardware execution log m(t) output from Embodiment 1 to sequentially identify the first to ninth steaming and drying cycles of Polygonum multiflorum. First, baseline subtraction is performed on y(t), followed by time alignment and channel normalization based on temperature switching moments. Then, the target temperature, actual temperature, heating power, temperature tracking error, and valve status of each sensor are extracted from m(t), enabling the model to simultaneously perceive odor response and multi-sensor differentiated temperature adjustment sampling actions.

[0063] For the nth sample, the hardware excitation vector This includes headspace temperature, target temperature of each sensor, actual temperature of each sensor, temperature step amplitude, heating / cooling slope, holding time, high-temperature pulse width, and sampling excitation sequence number e. Signal quality vector. This includes high-frequency perturbation ratio, baseline drift, dynamic range, response slope, peak hysteresis, recovery speed, saturation ratio, and temperature tracking error.

[0064]

[0065]

[0066] Parameter generator G receives and The output stochastic resonance branch parameters include potential function parameters a and b, input gain k, noise intensity D, integration step size dt, and gating coefficient g. For weak responses formed in the low-temperature adsorption section or short-term recovery differences appearing after a high-temperature pulse, the input gain can be increased, the potential barrier can be lowered, or the stochastic resonance processing window can be aligned to the temperature step edge. For samples with high noise, high humidity, large flow tracking error, large temperature tracking error, or significant baseline drift in the high-temperature section, the gating coefficient g can be reduced, allowing the original response branch or hardware excitation branch to occupy a higher weight.

[0067]

[0068]

[0069] The raw response encoder can employ one-dimensional convolutional networks, temporal convolutional networks, Transformer encoders, gated recurrent networks, or combinations thereof to extract the steady-state amplitude, rising edge, temperature-switching response segment, recovery segment, and cross-channel correlation of the raw electronic nose response. The stochastic resonance response encoder processes the enhanced response output by learnable stochastic resonance layers to extract weak dynamic features. The hardware excitation embedding encoder integrates the temperature excitation matrix Theta(t) and the hardware excitation vector... The encoding is used as a hardware representation to avoid the model misinterpreting the differences in response under different temperature conditions as differences in the sample itself, while preserving the contribution of different temperature zones to distinguishing adjacent evaporation cycles.

[0070] The three-branch representations are fused using a gated attention mechanism. The fusion weights are determined by the signal quality vector. Hardware excitation vectors Together with the intermediate uncertainty of the model, the high-quality samples mainly depend on the original response, while the weights of the stochastic resonance branch and the hardware excitation branch can be appropriately increased for boundary samples, weak response samples, samples with significant differences in low-temperature adsorption, or samples with significant high-temperature recovery lag.

[0071]

[0072] The stage identification model includes a nine-class classification output head, an ordinal output head, and a confidence output head. The nine-class classification output head outputs the probabilities of the first to the ninth steaming and drying rounds; the ordinal output head outputs the continuous processing progress value r; and the confidence output head calculates the final confidence score C based on the maximum class probability, the consistency between the classification result and the ordinal result, signal quality, and hardware tracking error. During the training phase, the model employs steaming round classification loss, ordinal regression loss, ordered trajectory loss across steaming rounds for the same sample or batch, stage consistency loss under different sampling stimuli for the same sample, and stochastic resonance-gated sparsity or stability regularization.

[0073]

[0074] When the confidence level C of the first sample is greater than or equal to the threshold tau, the system outputs the evaporation / drying cycle identification result; when C is less than the threshold tau, the active re-inspection control module... , The model uncertainty is used to select the next sampling excitation sequence from the candidate temperature excitation library, and the active sniffing front end of Example 1 is controlled to resample.

[0075] The algorithm was validated using a fixed-sample electronic nose dataset of nine steaming and nine drying cycles of Polygonum multiflorum. The dataset contained 270 records, covering the first to ninth steaming cycles, with 30 records per steaming cycle and 30 sample numbers. Using a five-fold outer-layer validation and inner-layer parameter tuning grouped by sample number, the optimal multi-domain feature model correctly identified 266 records, achieving an accuracy of 98.52%, a macro-average F1 score of 98.52%, and an ordinal mean absolute error of 0.033. These results demonstrate the feasibility of using the electronic nose's odor response to distinguish between the nine steaming cycles of Polygonum multiflorum.

[0076] On the fixed sampled data, 251 records were automatically received with a posterior confidence threshold of 0.60 and the accuracy of the received samples was 100%. This result indicates that the low-confidence sample set contains error risk and is suitable for triggering the next round of multi-sensor differentiated temperature control hardware sniffing in Example 1. Therefore, it illustrates the necessity of the model confidence re-verification mechanism. The effect of active sniffing temperature excitation can be verified by comparing and verifying schemes such as fixed temperature, unified temperature step, differentiated temperature zone parallel, temperature zone rotation, and high-temperature pulse re-verification in subsequent acquisition systems.

[0077] To further explain the source, composition, and verification results of the basic verification data for the above algorithm, this embodiment provides supplementary explanations in Tables 1 to 3.

[0078] Table 1. Composition of Experimental Data for Fixed-Sampling Electronic Nose

[0079] Steaming and drying cycles Number of records / entries Sample number involved Data Description Steaming 1st 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. Steaming 2nd time 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 3rd steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 4th steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 5th steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 6th steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 7th steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 8th steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. 9th steaming 30 30 sample numbers The electronic nose records are fixed and sampled during each evaporation and drying cycle. total 270 30 sample numbers Covering the first to ninth steaming

[0080] Table 1 shows that the experimental data are evenly distributed from the first to the ninth steaming. Each steaming and drying cycle contains 30 fixed sampling electronic nose records, totaling 270 records. This can avoid bias in the identification results caused by a significant increase or decrease in the number of samples in a certain steaming and drying stage.

[0081] Table 2. Validation Results of the Algorithm

[0082] project numerical values illustrate Total amount of verification data 270 items Covering the first to ninth steaming Record the number of times for each drying cycle 30 articles The number of steaming and drying cycles is the same for all nine cycles. Correctly identify the number of records 266 items The nine-category classification stage identifies the number of correctly recorded records. Number of misidentified records 4 Records that were not correctly categorized into the corresponding steaming and drying cycles out of 270 entries. Overall accuracy 98.52% 266 / 270×100% Macro average F1 98.52% Used to evaluate the overall identification effect of each evaporation / drying cycle category. Ordinal Mean Absolute Error 0.033 Used to evaluate the order deviation between predicted steam counts and actual steam counts. Verification method 50% off verification Five-fold outer layer verification was performed, grouped by sample number, and inner layer parameter tuning was conducted.

[0083] Table 2 shows that, out of 270 fixed-sample electronic nose records, the model correctly identified 266 records, with an overall accuracy and macro-average F1 of 98.52%. This indicates that the electronic nose response of Polygonum multiflorum at different steaming and drying stages is distinguishable, and provides an algorithmic basis for the subsequent introduction of multi-temperature zone active sniffing and stochastic resonance enhancement.

[0084] Table 3 Confidence Threshold Screening Results

[0085] project numerical values illustrate Confidence threshold τ 0.60 Used to determine whether to accept the results of this round of identification. Automatically receive record count 251 items, accounting for 92.96% Records with a confidence level not lower than the threshold Number of low-confidence re-examination records 19 items, accounting for 7.04% Records whose confidence level falls below the threshold and triggers the next round of review Automatic sample reception accuracy 100% No errors were detected in the automatic reception section. Purpose of re-inspection Trigger active sniffing Low-confidence records are used to select the next round of temperature excitation sequence.

[0086] Table 3 shows that the confidence threshold can distinguish between high-confidence identification results and low-confidence samples to be re-examined. When the confidence threshold is 0.60, 251 records can be directly output as identification results, and 19 low-confidence records enter the next round of multi-temperature zone active sniffing re-examination, thus forming a closed loop of hardware temperature excitation, random resonance parameter update and deep model uncertainty feedback.

[0087] Figure 3This is a radar chart showing the maximum response values ​​of the six sensor indicators for the nine steaming cycles of this invention. The chart illustrates the distribution of the maximum response values ​​of the six sensor indicators S1-S6 during the first to ninth steaming cycles. Each sub-chart corresponds to one steaming cycle, with each direction representing a different sensor indicator. The filled area reflects the relative magnitude of the maximum response values ​​of each sensor during that steaming cycle. This chart allows for a direct comparison of the changes in sensor response characteristics between different steaming cycles, thus characterizing the differences in sample state during the processing.

[0088] Figure 4 This is a heatmap showing the normalized average response of each steaming cycle and sensor. The graph first calculates the average response of each steaming cycle and each sensor at 600 time points and 30 samples, then normalizes it by sensor channel to observe the overall response strength changes across different steaming cycles on each sensor channel.

[0089] Figure 5 This is a PCA distribution map of the dynamic features of the electronic nose. Using the mean, dynamic range, response slope, and recovery difference extracted from each record as input, the high-dimensional electronic nose features are reduced to two principal component spaces to visually demonstrate the aggregation and separation of samples from different batches.

[0090] Figure 6 This figure illustrates the confusion matrix identified by five-fold cross-validation. The actual number of steaming cycles is plotted on the vertical axis, and the predicted number of steaming cycles is plotted on the horizontal axis. The diagonal lines represent the number of correctly identified samples, and the off-diagonal lines represent the number of misclassified samples. This demonstrates that the electronic nose data has a basis for identifying the processing stages of Polygonum multiflorum and shows the main confusion between adjacent steaming cycles.

[0091] The above Figures 3 to 6 In the data, the first 600 time points of each electronic nose record are uniformly extracted, corresponding to the timing response of the 6 sensor channels at a sampling frequency of 60 s and 10 Hz; each figure is used to supplement the explanation of experimental data format, response differences, feature separability and recognition results.

[0092] This invention can be used for rapid stage identification, batch quality monitoring, and production process sample tracking of Polygonum multiflorum during its nine-steaming and nine-drying processing. It can also be extended to other electronic nose detection scenarios for Chinese medicinal materials with continuous processing stages. Those skilled in the art can appropriately adjust the flow range, response target interval, model structure, and training parameters according to specific applications without departing from the concept of this invention; all such adjustments should be considered within the scope of protection of this invention.

Claims

1. A multi-temperature zone active sniffing resonance depth recognition method for Polygonum multiflorum (He Shou Wu) after nine steaming and nine drying processes, characterized in that... include: The Polygonum multiflorum sample to be tested is placed in the sample headspace chamber to form a volatile gas headspace, and clean air is used to restore the baseline of the electronic nose sensor air chamber and sensor array. A temperature excitation matrix Theta(t) is generated. Based on the temperature excitation matrix Theta(t), the target temperature, heating / cooling slope, and high-temperature pulse width of each gas sensor in the electronic nose sensor array are controlled during multiple sampling time periods, so that at least some gas sensors are in the low-temperature adsorption zone, the medium-temperature reaction zone, and the high-temperature desorption zone, respectively. The temperature excitation matrix Theta(t) is executed during the period when the volatile gas headspace enters the sensor gas chamber at a set carrier gas flow rate, and the multi-channel electronic nose timing response y(t) is collected. The target temperature, actual temperature, heating power, temperature tracking error, valve status, and timestamp of each gas sensor are recorded simultaneously to form a hardware execution log m(t). The multi-channel electronic nose timing response y(t) is segmented according to the temperature switching time. The signal quality vector and hardware excitation vector are calculated in combination with the hardware execution log m(t). Random resonance branch parameters are generated based on the signal quality vector and hardware excitation vector. The dynamic response corresponding to the temperature step edge, the low-temperature weak response segment, or the high-temperature recovery segment is enhanced through the random resonance branch. The original response representation, stochastic resonance enhanced response representation, and hardware excitation representation are fused using gated attention. The stage recognition model outputs the probability of the steaming and drying round category of Polygonum multiflorum, the value of continuous processing, and the confidence level. When the confidence level or the consistency between the steaming and drying round category probability and the value of continuous processing does not meet the preset output conditions, the next round of temperature excitation sequence is selected and resampled according to the model uncertainty, signal quality vector, and temperature tracking error. When the preset output conditions are met, the processing stage recognition result is output, and the sample, temperature excitation matrix, hardware execution log, stochastic resonance branch parameters, and recognition result are written into the parameter update library.

2. The method according to claim 1, characterized in that, The temperature excitation matrix Theta(t) has rows corresponding to gas sensor numbers and columns corresponding to sampling time periods. The matrix elements are the target temperature zone, target temperature, heating / cooling slope, holding time, or high-temperature pulse width of the corresponding gas sensor within the corresponding sampling time period. The target temperature of the low-temperature adsorption zone is 150°C to 220°C, the target temperature of the medium-temperature reaction zone is 230°C to 300°C, and the target temperature of the high-temperature desorption zone is 310°C to 380°C. Furthermore, the gas sensor temperature zones are rotated within different sampling time periods, so that the same gas sensor can obtain continuous responses of the same headspace gas of Polygonum multiflorum under different temperature zones.

3. The method according to claim 1, characterized in that, The active sniffing front end that executes the temperature excitation matrix Theta(t) includes a sample headspace chamber, a clean air channel, a sample gas channel, a multi-way valve group, a sensor gas chamber, an electronic nose sensor array, multiple independent heaters, a temperature probe, and a control module. The control module is used to control the multi-way valve group to switch between the clean air position and the sample gas position, and to perform closed-loop correction of the PWM of each independent heating drive channel based on the actual temperature obtained by the temperature feedback detection unit.

4. The method according to claim 1, characterized in that, The signal quality vector includes high-frequency perturbation ratio, baseline drift, dynamic range, response slope, peak hysteresis, recovery speed, saturation ratio, and temperature tracking error; the hardware excitation vector includes target temperature, actual temperature, heating power, temperature tracking error, and valve status for each sensor.

5. The method according to claim 1, characterized in that, The random resonance branch parameters include potential function parameters, input gain, noise intensity, integration step size, and gating coefficient. When the signal quality vector indicates a weak response in the low-temperature adsorption section or a short-term recovery difference after a high-temperature pulse, the input gain is increased, the potential barrier is lowered, or the random resonance processing window is aligned to the temperature step edge. When the signal quality vector indicates excessive noise, high humidity, large flow tracking error, large temperature tracking error, or significant baseline drift in the high-temperature section, the gating coefficient of the random resonance branch is reduced to mitigate the adverse effects of random resonance enhancement on the recognition results.

6. The method according to claim 1, characterized in that, The stage recognition model includes a raw response encoder, a stochastic resonance response encoder, a hardware excitation embedded encoder, a gated attention fusion layer, a nine-class output head, an ordinal output head, and a confidence output head. The raw response encoder is used to extract the steady-state amplitude, rising edge, temperature switching response segment, recovery segment, and cross-channel correlation of the raw electronic nose response. The stochastic resonance response encoder is used to extract the weak dynamic features after stochastic resonance enhancement. The hardware excitation embedded encoder is used to encode the temperature excitation matrix Theta(t) and the hardware execution log m(t). The confidence output head generates the final confidence based on the maximum class probability, the consistency between the nine-class output and the ordinal output, signal quality, and hardware tracking error.

7. A multi-temperature zone active sniffing resonance depth recognition system for Polygonum multiflorum (He Shou Wu) after nine steaming and nine drying processes, characterized in that... The system is used to perform the method according to any one of claims 1 to 6, the system comprising a multi-sensor differentiated temperature-controlled sampling and excitation unit, an electronic nose sensor array, a temperature excitation matrix generation module, a sampling quality judgment module, a stochastic resonance parameter adjustment module, a feature fusion module, a vaporization identification module, a re-inspection decision module, and a parameter update library; the multi-sensor differentiated temperature-controlled sampling and excitation unit is used to enable the same *Polygonum multiflorum* headspace gas to be detected in parallel in a low-temperature adsorption zone, a medium-temperature reaction zone, and a high-temperature desorption zone; the stochastic resonance parameter adjustment module is used to generate stochastic resonance branch parameters based on the signal quality vector and the hardware excitation vector; the feature fusion module is used to fuse the original response characterization, the stochastic resonance enhanced response characterization, and the hardware excitation characterization; the re-inspection decision module is used to generate the next round of temperature excitation sequence or re-inspection instruction when the identification confidence is insufficient; and the parameter update library is used to store samples, temperature excitation matrix, hardware execution log, stochastic resonance branch parameters, and identification results.

Images