A method and system for graded screening of the contribution of a medicinal herb odor detection sensor array
By using baseline correction and channel contribution index set to define grade boundaries, an effective subset of sensors was selected, which solved the problem of inconsistent sensor channel contributions in the detection of medicinal herbs' odor and improved the stability and accuracy of the detection data.
Patent Information
- Application Number
- CN202611267442.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-10
AI Technical Summary
In existing medicinal herb odor detection technologies, the sensor channels contribute inconsistently to the target detection label, leading to decreased data stability during the detection process. Furthermore, the lack of an effective channel screening mechanism can easily result in problems such as low-contribution channels not being removed, redundant channels being repeatedly retained, and conditionally contributing channels being mistakenly deleted.
Effective response sequences are formed through baseline correction, channel contribution index sets are calculated and grade boundaries are defined, core retention, auxiliary retention, conditional retention and elimination channels are screened out, candidate subsets are constructed and recognition gain is verified, and only response data of effective sensor subsets are received for medicinal herb odor detection.
This improves the stability and relevance of the input data for medicinal herb odor detection, reduces interference from invalid channels, and ensures the accuracy and effectiveness of the sensor channel contribution level screening process.
Smart Images

Figure CN123069032A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medicinal herb odor detection technology, and more specifically, to a method and system for graded screening of the contribution of a medicinal herb odor detection sensor array. Background Technology
[0002] The contribution grading and screening method of sensor array for detecting the odor of medicinal materials involves the field of medicinal material odor detection technology. Specifically, it refers to the method of using sensor arrays to collect the volatile odor response values of medicinal materials and screening the effective channels for subsequent detection based on the contribution of each sensor channel to the target detection label.
[0003] Existing medicinal herb odor detection technologies can collect multi-channel response values corresponding to the volatile odors of medicinal herbs through devices such as electronic noses and gas-sensitive sensor arrays. By using baseline correction, feature extraction, and recognition models, the technologies can determine the target detection labels such as the type, origin, grade, and quality status of medicinal herbs, thereby converting the odor information of medicinal herbs into quantifiable sensor response data.
[0004] However, in detection scenarios where the odor components of medicinal materials are complex and there are significant differences between sample batches, the contributions of different sensor channels to the target detection label are not consistent. Some channels only produce weak responses close to background fluctuations, some channels have insufficient repeatability stability under the same target detection label, and some channels have highly correlated redundant responses with other channels. If the above channels are used together with the effective channels for medicinal material odor detection input, the detection process is easily affected by the differences in effective odors and the disturbances of ineffective channels at the same time, resulting in a decrease in the stability of the detection input data.
[0005] While existing technologies can perform routine feature extraction and model recognition on sensor response values, they typically lack a hierarchical evaluation chain for response effectiveness, label discrimination, repeatability stability, and redundancy interference. They also lack a processing mechanism to filter core retained channels, auxiliary retained channels, and conditionally retained channels into the effective sensor subset according to different rules. This can easily lead to problems such as low-contribution channels not being removed, redundant channels being repeatedly retained, and conditionally contributing channels being mistakenly deleted.
[0006] Therefore, a method for classifying and screening the contribution of a sensor array for detecting the odor of medicinal materials is needed. This method can form a reproducible set of channel contribution indicators and channel contribution level results based on the response values of labeled samples. It can also determine the effective sensor subset through candidate subset construction and gain verification, thereby reducing the interference of invalid sensor channels on the input data for detecting the odor of medicinal materials.
[0007] In view of this, this application proposes a method and system for graded screening of the contribution of a medicinal herb odor detection sensor array to solve the above problems. Summary of the Invention
[0008] To overcome the problem that existing technologies for detecting the aroma of medicinal materials lack a basis for classifying and screening sensor channel contributions, leading to low-contribution channels, channels with insufficient repeatability, and redundant channels easily participating in the detection input, this application aims to provide a method and system for classifying and screening the contribution of a sensor array for detecting the aroma of medicinal materials. Through continuous processing of effective response sequences, channel representative response values, channel contribution index sets, channel contribution level results, candidate subsets, and effective sensor subsets, the sensor channels participating in the detection input of the aroma of medicinal materials are transformed from all channel inputs to effective channel inputs selected according to contribution level and recognition gain.
[0009] To achieve the above objectives, this application provides the following technical solution:
[0010] A method for graded screening of the contribution of a medicinal herb odor detection sensor array includes:
[0011] Read the response values of labeled samples containing target detection tags, and perform baseline correction on the response values from entering the odor collection area to before the odor is cleared to form an effective response sequence;
[0012] Calculate the channel representative response value based on the effective response sequence, and obtain the channel contribution index set based on the channel representative response value;
[0013] Based on the channel contribution index set, the level boundary and redundancy high boundary are determined. By comparing the channel contribution index set with the level boundary, each sensor channel is divided into core retained channels, auxiliary retained channels, conditionally retained channels and eliminated channels, and the channel contribution level results are obtained.
[0014] Candidate subsets are constructed based on the channel contribution level results. Core reserved channels are added to the candidate subsets, and auxiliary reserved channels with supplementary label pairs and whose absolute correlation coefficient with any sensor channel in the candidate subset is less than or equal to the redundancy high boundary are added to the candidate subsets.
[0015] The recognition gain threshold is determined based on the recognition stability value perturbation of the eliminated channels. Condition-reserved channels whose recognition gain value is greater than or equal to the recognition gain threshold and whose absolute value of the correlation coefficient with any sensor channel in the subset is less than or equal to the redundancy high boundary are added to the candidate subset. The candidate subset after adding the condition-reserved channels is determined as the effective sensor subset.
[0016] When detecting the odor of medicinal materials, only the response data of the corresponding channels of the effective sensor subset are received to generate the input data for detecting the odor of medicinal materials.
[0017] Furthermore, methods for forming effective response sequences include:
[0018] For each labeled sample and each sensor channel, the median value of the response value of the labeled sample before entering the odor collection area is read as the baseline representative value. The baseline corrected response value is obtained by subtracting the baseline representative value from the response value of the labeled sample after entering the odor collection area until the odor is cleared. The baseline fluctuation limit is determined based on the fluctuation of the response value of the labeled sample before entering the odor collection area. When the absolute value of the baseline corrected response value of three consecutive sampling times is greater than the baseline fluctuation limit, the baseline corrected response value is intercepted from the first sampling time of the three consecutive sampling times to form an effective response sequence.
[0019] Furthermore, the methods for calculating the channel-represented response value and the degree of response validity include:
[0020] The maximum absolute value of the baseline correction response value at each sampling time in the effective response sequence is taken as the representative response value of the channel. The labeled samples whose representative response value of the channel is greater than or equal to the effective response threshold are counted as effective response samples. The ratio of the number of effective response samples to the number of labeled samples is taken as the response effectiveness of the corresponding sensor channel.
[0021] Furthermore, the methods for calculating the label differentiation degree include:
[0022] According to the target detection label, the channel representative response values under the same sensor channel are grouped. The median value of the channel representative response values in each group is taken as the tag response representative value. The average of the absolute values of the difference between any two tag response representative values is calculated as the tag separation value. The average of the absolute values of the difference between the channel representative response value in each group and the corresponding tag response representative value is calculated as the tag fluctuation value. The tag separation value is divided by the sum of the tag separation value and the tag fluctuation value to obtain the tag discrimination degree. When the sum of the tag separation value and the tag fluctuation value is zero, the tag discrimination degree is recorded as zero.
[0023] Furthermore, methods for calculating repeatability stability include:
[0024] According to the target detection label, the representative response values of the same sensor channel are grouped. The median value of the representative response values of the channel in each group is taken as the repeatability reference value. The average value of the absolute value of the difference between the representative response value of each channel in the group and the repeatability reference value is calculated as the repeatability deviation value of the group. The repeatability deviation value of the group is divided by the sum of the absolute value of the repeatability reference value and the minimum positive value to obtain the label repeatability deviation value. The minimum positive value is the minimum resolvable response value of the sensor channel. The average value of all label repeatability deviation values is taken as the channel repeatability deviation value. The maximum value between the result obtained by subtracting the channel repeatability deviation value from zero is taken as the repeatability stability.
[0025] Furthermore, the methods for calculating the degree of redundancy interference include:
[0026] Arrange the representative response values of the same sensor channel according to the order of the labeled samples to form the channel response vector of the sensor channel. Calculate the absolute value of the correlation coefficient between the channel response vector of the current sensor channel and the channel response vectors of other sensor channels, and take the largest absolute value of the correlation coefficient as the degree of redundancy interference of the current sensor channel.
[0027] Furthermore, before the auxiliary reserved channels are added to the candidate subset, they are arranged in descending order of label discrimination. When the label discrimination of two auxiliary reserved channels is the same, they are arranged in ascending order of redundancy interference. When the redundancy interference is still the same, they are arranged in ascending order of sensor channel number, and the auxiliary reserved channels are then judged in turn whether to be added to the candidate subset according to the arrangement results.
[0028] Furthermore, the methods for determining supplementary label pairs include:
[0029] Any two different target detection labels are combined into a target detection label pair. Based on the difference in the channel representative response values corresponding to the two target detection labels under the same sensor channel, it is determined whether the sensor channel can distinguish the target detection label pair. The target detection label pairs that the auxiliary reserved channel can distinguish and that are not yet covered by the candidate subset are determined as the supplementary label pairs of the auxiliary reserved channel.
[0030] Furthermore, the methods for identifying stable value perturbations and calculating gain values include:
[0031] The basic recognition stability value is calculated based on the channel representative response value and target detection label corresponding to the candidate subset; the removed channel is temporarily added to the candidate subset, the recognition stability value after addition is calculated, and the absolute value of the difference between the added recognition stability value and the basic recognition stability value is used as the recognition stability value perturbation. The recognition gain threshold is determined based on the recognition stability value perturbation; the condition-retained channel is temporarily added to the candidate subset, the temporary recognition stability value is calculated, and the difference between the temporary recognition stability value and the basic recognition stability value is used as the recognition gain value.
[0032] A contribution grading and screening system for a medicinal herb odor detection sensor array, comprising the following steps:
[0033] Response correction module: used to read the response values of labeled samples containing target detection tags, and to perform baseline correction on the response values from the time of entering the odor collection area until the odor is cleared, forming an effective response sequence;
[0034] Indicator Acquisition Module: Used to calculate the channel representative response value based on the effective response sequence, and to obtain the channel contribution indicator set based on the channel representative response value;
[0035] The classification module is used to determine the classification boundary and redundancy high boundary based on the channel contribution index set. It compares the channel contribution index set with the classification boundary to classify each sensor channel into core retained channels, auxiliary retained channels, conditionally retained channels, and eliminated channels, thus obtaining the channel contribution classification result.
[0036] Candidate building module: used to build candidate subsets based on channel contribution level results, add core reserved channels to candidate subsets, and add auxiliary reserved channels with supplementary label pairs and whose absolute correlation coefficient with any sensor channel in the candidate subset is less than or equal to the redundancy high boundary to candidate subsets;
[0037] Gain verification module: used to determine the recognition gain threshold based on the recognition stability value disturbance of the eliminated channel, add the condition-retained channels whose recognition gain value is greater than or equal to the recognition gain threshold and whose absolute value of the correlation coefficient with any sensor channel in the subset is less than or equal to the redundancy high boundary to the candidate subset, and determine the candidate subset after adding the condition-retained channels as the effective sensor subset;
[0038] Input generation module: When detecting the odor of medicinal materials, it only receives the response data of the corresponding channel of the valid sensor subset and generates the input data for detecting the odor of medicinal materials.
[0039] The technical effects and advantages of the contribution grading and screening method and system for detecting the odor of medicinal materials proposed in this application are as follows:
[0040] First, this application corrects the baseline of the labeled sample response values and forms an effective response sequence, so that the channel contribution evaluation is based on the effective response data from the time of entering the odor collection area to the time of odor removal, reducing the impact of background fluctuations and invalid responses on the representative response values of the channel, and providing a more stable input basis for the subsequent calculation of the channel contribution index set.
[0041] Second, this application obtains a set of channel contribution indicators based on the channel representative response value, and uses the effectiveness of the response, the degree of label differentiation, the degree of repeatability and the degree of redundancy interference as the evaluation criteria for sensor channels. This makes channel selection no longer dependent on a single response amplitude, but can simultaneously distinguish weak response channels, channels with insufficient label differentiation ability, channels with large repeatability deviation of the same label, and channels with strong redundancy interference.
[0042] Third, this application determines the level boundary and redundancy high boundary based on the channel contribution index set, and obtains the channel contribution level result accordingly, so that each sensor channel can be divided into core retained channel, auxiliary retained channel, condition retained channel and rejection channel, thereby setting a differentiated processing method for channels with different contribution levels and reducing the probability of invalid channels directly entering the medicinal odor detection input.
[0043] Fourth, this application directly adds core reserved channels when constructing candidate subsets, and only adds auxiliary reserved channels with supplementary label pairs and absolute correlation coefficients less than or equal to the redundancy high boundary to the subset. This makes the auxiliary reserved channels not only need to provide new target detection label pair discrimination capabilities, but also need to avoid forming excessive redundancy with the already selected channels, thereby improving the label coverage capability and channel combination effectiveness of the candidate subset.
[0044] Fifth, this application determines the identification gain threshold based on the identification stability value perturbation of the eliminated channel, and uses the identification gain value and the absolute value of the correlation coefficient to verify the conditionally retained channel, so that the conditionally retained channel can only enter the candidate subset when it can bring effective identification gain and does not exceed the redundancy limit, thereby reducing the risk of conditionally contributing channels being mistakenly deleted or mistakenly retained.
[0045] Sixth, this application only receives the response data of the corresponding channels of the effective sensor subset to generate the medicinal herb odor detection input data when detecting the odor of medicinal materials. This prevents low-response, low-discrimination, low-stability, and high-redundancy channels from participating in the final detection input, thereby improving the stability and specificity of the medicinal herb odor detection input data. Attached Figure Description
[0046] Figure 1 This is a flowchart of a method for graded screening of the contribution of a medicinal herb odor detection sensor array according to Embodiment 1 of this application;
[0047] Figure 2 This is a schematic diagram illustrating the contribution grading and screening of the medicinal herb odor detection sensor array in Embodiment 1 of this application;
[0048] Figure 3 This is a block diagram of a graded screening system for a medicinal herb odor detection sensor array, as shown in Embodiment 2 of this application. Detailed Implementation
[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0050] Example 1
[0051] See Figure 1-2 As shown, this embodiment provides a method for graded screening of the contribution of a medicinal herb odor detection sensor array, including:
[0052] The multi-channel response data of the labeled samples are read to form the effective response sequence of each medicinal herb odor sample under each sensor channel. The labeled sample refers to the medicinal herb odor sample with a known target detection label. The target detection label is a label field selected for this medicinal herb odor detection task. The label field is taken from one of the following: medicinal herb type, origin, grade, and quality status. The multi-channel response data includes sample number, target detection label, sensor channel number, sampling time, and response value. The same sensor channel number corresponds to an odor sensor channel in the sensor array. The effective response sequence serves as the basis for calculating the channel representative response value and the channel contribution index set.
[0053] When the number of target detection labels in the labeled samples is less than two, stop the channel contribution classification screening, and re-execute the method after supplementing the labeled samples with different target detection labels.
[0054] For each medicinal herb odor sample, the continuous sampling data before the medicinal herb odor sample enters the odor collection area is taken as the blank baseline segment. The blank baseline segment refers to the background response data formed by the sensor channel before contacting the current medicinal herb odor sample. It includes the sampling time and response value and is used to calculate the baseline representative value corresponding to the current medicinal herb odor sample. The continuous sampling data after the medicinal herb odor sample enters the odor collection area until the odor removal begins is taken as the injection response segment. The injection response segment refers to the response data formed after the sensor channel contacts the current medicinal herb odor sample and emits odor. When either the blank baseline segment or the injection response segment is not collected, the effective response sequence of the medicinal herb odor sample under the corresponding sensor channel is recorded as an empty sequence. The empty sequence refers to a valid response sequence that retains the sample number, target detection label, and sensor channel number, but does not generate the sampling time and baseline correction response value. When calculating the channel representative response value, the channel representative response value corresponding to the sample number and the sensor channel number is recorded as zero.
[0055] For each herbal odor sample and each sensor channel that is not recorded as an empty sequence, the median value of the response value within the blank baseline segment is calculated as the baseline representative value of the herbal odor sample under that sensor channel; the baseline representative value is subtracted from the response value at each sampling time within the injection response segment to obtain the baseline corrected response value at that sampling time; the baseline corrected response values of the same herbal odor sample and the same sensor channel are arranged in order from earliest to latest sampling time to obtain the baseline corrected response sequence.
[0056] In the baseline correction response sequence corresponding to the same herbal odor sample and the same sensor channel, three consecutive sampling times are searched sequentially from earliest to latest. The baseline fluctuation limit of the sensor channel is used to determine whether the baseline correction response value exceeds the background fluctuation range. When the absolute value of the baseline correction response value of the three consecutive sampling times is greater than the baseline fluctuation limit of the corresponding sensor channel, the first sampling time of the three consecutive sampling times is taken as the effective response start point, and the baseline correction response value between the effective response start point and the end of the injection response segment is extracted to form an effective response sequence. When no three consecutive sampling times that meet the above conditions are found, the effective response sequence of the herbal odor sample under the sensor channel is recorded as an empty sequence.
[0057] The baseline fluctuation limit refers to the response amplitude boundary used to distinguish between background fluctuations and the effective response of medicinal herb odor. It is obtained by reading the historical blank response data formed by the same sensor channel when no medicinal herb odor sample is placed, calculating the absolute difference between each response value in the historical blank response data and the median value of the historical blank response data, forming a blank fluctuation difference sample. When the number of blank fluctuation difference samples is greater than or equal to ten, the 95th percentile value of the blank fluctuation difference sample is taken as the baseline fluctuation limit of the sensor channel. When the number of blank fluctuation difference samples is less than ten, three times the minimum resolvable response value corresponding to the acquisition resolution of the sensor channel is taken as the baseline fluctuation limit.
[0058] The effective response sequence includes the sample number, target detection label, sensor channel number, sampling time, and baseline correction response value. It is used as input data in the process of generating the channel contribution index set to calculate the channel representative response value, response effectiveness, label discrimination, repeatability stability, and redundancy interference.
[0059] A set of channel contribution indicators is generated for each sensor channel based on the effective response sequence. The set of channel contribution indicators includes the effectiveness of the response, the degree of label differentiation, the degree of repeatability and stability, and the degree of redundancy interference, and serves as the basis for classifying the channel contribution level results.
[0060] The channel representative response value is determined based on the effective response sequence. The channel representative response value refers to the single-value response result of a medicinal herb odor sample participating in the contribution evaluation under a sensor channel. When the effective response sequence is an empty sequence, the channel representative response value of the medicinal herb odor sample under the sensor channel is recorded as zero. When the effective response sequence is not an empty sequence, the maximum value of the absolute value of the baseline corrected response value at each sampling time in the effective response sequence is taken as the channel representative response value of the medicinal herb odor sample under the sensor channel.
[0061] The effectiveness of the response is determined based on the representative response value of the same sensor channel across all labeled samples. The effective response threshold of the sensor channel is used as the judgment boundary. When the representative response value of the channel is greater than or equal to the effective response threshold, the corresponding medicinal herb odor sample is counted as an effective response sample of the sensor channel. The number of effective response samples is divided by the total number of labeled samples to obtain the effectiveness of the sensor channel's response. When the total number of labeled samples is zero, the channel screening is stopped, and the method is re-executed after adding labeled samples.
[0062] The effective response threshold refers to the boundary of the representative response value of the channel used to determine whether the sensor channel generates an effective response of medicinal herb odor. It is obtained by reading the historical blank response data and the representative response value of the channel under the labeled sample corresponding to the same sensor channel. First, the baseline fluctuation limit is determined according to the historical blank response data. Then, the representative response value of the channel under the labeled sample that is less than or equal to the baseline fluctuation limit is taken as a weak response sample, and the representative response value of the channel that is greater than the baseline fluctuation limit is taken as a candidate response sample. When the number of candidate response samples is greater than or equal to ten, the maximum value between the tenth percentile value of the candidate response sample and the baseline fluctuation limit is taken as the effective response threshold. When the number of candidate response samples is less than ten, the baseline fluctuation limit is taken as the effective response threshold.
[0063] In another possible implementation, the effective response threshold is determined based on the historical labeled detection records of the same sensor channel. The historical labeled detection records follow the data structure of the labeled samples and include historical discrimination results formed after the detection has been completed. When the historical discrimination results are consistent with the target detection label, the corresponding medicinal herb odor sample is regarded as the correctly identified sample. When the historical discrimination results are inconsistent with the target detection label, the corresponding medicinal herb odor sample is regarded as the incorrectly identified sample. Candidate thresholds are scanned from small to large according to the channel representative response values. The proportion of the number of channel representative response values greater than or equal to the candidate thresholds and belonging to the incorrectly identified samples is counted as the miscounting ratio. The smallest candidate threshold that makes the miscounting ratio less than or equal to the preset miscounting upper limit is selected as the effective response threshold. The preset miscounting upper limit is the maximum proportion of incorrect samples allowed to enter the effective response samples in the screening task. When the total number of incorrectly identified samples is zero, the maximum value between the 10th percentile value of the channel representative response value in the historical labeled detection records and the baseline fluctuation limit is taken as the effective response threshold.
[0064] The label discrimination level is used to represent the ability of the same sensor channel to distinguish the differences in the odor of medicinal materials between different target detection labels. The channel representative response values of the target detection labels under the same sensor channel are grouped according to their respective target detection labels. The median value of the channel representative response values within each label group is calculated as the label response representative value. The average of the absolute values of the differences between any two label response representative values is calculated as the label separation value. The average of the absolute values of the differences between each channel representative response value and the corresponding label response representative value within each label group is calculated as the label fluctuation value. The label discrimination level of the sensor channel is obtained by dividing the label separation value by the sum of the label separation value and the label fluctuation value. When the sum of the label separation value and the label fluctuation value is zero, the label discrimination level is recorded as zero.
[0065] The repeatability stability is determined based on the response deviation of different medicinal herb odor samples under the same target detection label. The representative response values of the same sensor channel are grouped according to the target detection label. The median value of the representative response values within each label group is calculated as the repeatability baseline value. The average of the absolute values of the differences between each channel representative response value and the repeatability baseline value within the label group is calculated as the intra-group repeatability deviation value for that label group. The intra-group repeatability deviation value is divided by the sum of the absolute value of the repeatability baseline value and the minimum positive value to obtain the label repeatability deviation value for that label group. The minimum positive value is taken as the minimum resolvable response value corresponding to the acquisition resolution of the sensor channel. The average of all label repeatability deviation values is calculated as the channel repeatability deviation value. When the channel repeatability deviation value is less than or equal to one, the repeatability stability is obtained by subtracting the channel repeatability deviation value from one. When the channel repeatability deviation value is greater than one, the repeatability stability is recorded as zero.
[0066] The degree of redundancy interference is determined based on the response correlation between different sensor channels. The representative response values of the same sensor channel under all labeled samples are sorted according to the sample number to form the channel response vector of that sensor channel. The absolute value of the correlation coefficient between the channel response vector of the current sensor channel and the channel response vector of each other sensor channel is calculated, and the maximum value of the absolute value of the correlation coefficient is taken as the degree of redundancy interference of the current sensor channel. When there is only one sensor channel in the sensor array, the degree of redundancy interference of that sensor channel is recorded as zero.
[0067] The rule for calculating the absolute value of the correlation coefficient is as follows: After aligning the channel response vector of the current sensor channel with the channel response vector of another sensor channel according to their common sample numbers, the Pearson correlation coefficient is used to calculate the correlation coefficient between the two, and the absolute value of this correlation coefficient is taken as the absolute value of the correlation coefficient. The calculation formula is as follows:
[0068] ;
[0069] in, Indicates sensor channel With sensor channel The correlation coefficient between them This represents the number of sample numbers shared by the two channel response vectors. Indicates sensor channel In the Each channel under a common sample number represents a response value. Indicates sensor channel In the Each channel under a common sample number represents a response value. Indicates sensor channel The channels under the common sample number represent the mean response value. Indicates sensor channel The channel representative under the common sample number is the mean response value; when the denominator of the formula is zero, the absolute value of the correlation coefficient between the two is recorded as zero; when all values in the channel response vector of the current sensor channel are the same, or when all values in the channel response vector of the other sensor channel used for comparison are the same, the absolute value of the correlation coefficient between the two is recorded as zero; when the sample numbers corresponding to the two channel response vectors are not completely consistent, only the sample numbers that the two channel response vectors share are used for calculation; when the number of shared sample numbers is less than three, the absolute value of the correlation coefficient between the two is recorded as zero.
[0070] The response effectiveness, label differentiation, repeatability stability, and redundancy interference in the channel contribution index set are all one-to-one with the sensor channel number and are used as input data in the process of determining the channel contribution level, to classify core retained channels, auxiliary retained channels, conditionally retained channels, and eliminated channels.
[0071] The channel contribution level result for each sensor channel is determined based on the channel contribution index set. The channel contribution level result includes one of the following level labels: core retained channel, auxiliary retained channel, conditionally retained channel, and eliminated channel. This label is used to distinguish the processing methods of different sensor channels in the effective sensor subset construction process.
[0072] The grade boundaries are determined based on the channel contribution index set. First, a response effectiveness sample is formed by combining the response effectiveness of all sensor channels, a label discrimination sample is formed by combining the label discrimination of all sensor channels, a repeatability stability sample is formed by combining the repeatability stability of all sensor channels, and a redundancy interference sample is formed by combining the redundancy interference of all sensor channels. The grade boundaries include high response contribution boundary, high label contribution boundary, high stability contribution boundary, minimum response retention boundary, minimum label retention boundary, minimum stability retention boundary, low redundancy boundary, and high redundancy boundary. Specifically, the high response contribution boundary is taken as the 70th percentile value of the response effectiveness sample, the high label contribution boundary is taken as the 70th percentile value of the label discrimination sample, and the high stability contribution boundary is taken as the 70th percentile value of the repeatability stability sample. The minimum retention boundary is taken as the 30th percentile of the response effectiveness sample, the minimum retention boundary of the label is taken as the 30th percentile of the label distinguishability sample, and the minimum retention boundary of stability is taken as the 30th percentile of the repeatability stability sample; the low redundancy boundary is taken as the 50th percentile of the redundancy interference sample, and the high redundancy boundary is taken as the 80th percentile of the redundancy interference sample; the percentile values are determined by sorting the corresponding samples from smallest to largest. When the percentile position is not an integer, it is obtained by linear interpolation of two adjacent sorted values; when the number of sensor channels is less than three, the high contribution boundary of the response, the high contribution boundary of the label, and the high contribution boundary of stability are taken as the maximum value of the corresponding samples, the minimum retention boundary of the response, the minimum retention boundary of the label, and the minimum retention boundary of stability are taken as the minimum value of the corresponding samples, and the low redundancy boundary and the high redundancy boundary are taken as the maximum value of the redundancy interference sample.
[0073] In another possible implementation, when there are sensor channels with different response dimensions and different ranges in the sensor array, the level boundaries are determined by grouping according to the response dimensions and ranges; when the number of sensor channels in the same group is greater than or equal to three, the response high contribution boundary, tag high contribution boundary, stability high contribution boundary, response minimum retention boundary, tag minimum retention boundary, stability minimum retention boundary, redundancy low boundary, and redundancy high boundary are calculated in that group; when the number of sensor channels in the same group is less than three, the level boundaries of all sensor channels in that group are uniformly determined.
[0074] The core reserved channels are determined based on the conditions of high contribution and low redundancy. When the response effectiveness of a certain sensor channel is greater than or equal to the high contribution boundary of the response, the tag discrimination is greater than or equal to the high contribution boundary of the tag, the repeatability stability is greater than or equal to the high contribution boundary of the stability, and the redundancy interference is less than or equal to the low redundancy boundary, the sensor channel is designated as the core reserved channel.
[0075] The auxiliary reserved channel is determined based on the minimum response, strong label discrimination, minimum stability, and redundancy constraints. For sensor channels that are not classified as core reserved channels, when the response effectiveness of a sensor channel is greater than or equal to the minimum response retention boundary, the label discrimination is greater than or equal to the label high contribution boundary, the repeatability stability is greater than or equal to the minimum stability retention boundary, and the redundancy interference is less than or equal to the redundancy high boundary, the sensor channel is classified as an auxiliary reserved channel.
[0076] Condition-reserved channels are determined based on minimum response, minimum label discrimination, and minimum stability conditions. For sensor channels that are not classified as core-reserved channels or auxiliary-reserved channels, when the response effectiveness of a sensor channel is greater than or equal to the minimum response retention boundary, the label discrimination is greater than or equal to the minimum label retention boundary, and the repeatability stability is greater than or equal to the minimum stability retention boundary, the sensor channel is classified as a condition-reserved channel.
[0077] The elimination channel is determined based on not entering the retention level and the high redundancy replacement condition. For sensor channels not classified as core retention channels, auxiliary retention channels, or conditional retention channels, they are classified as elimination channels. For sensor channels already classified as auxiliary or conditional retention channels, if their redundancy interference level is greater than the redundancy high boundary, and there is another sensor channel already classified as a core, auxiliary, or conditional retention channel, and the response effectiveness, tag discrimination, and repeatability of this sensor channel are greater than or equal to the response effectiveness, tag discrimination, and repeatability of this sensor channel, then this sensor channel is downgraded from an auxiliary or conditional retention channel to an elimination channel. When the response effectiveness, tag discrimination, and repeatability of two sensor channels are the same, the retention order is determined according to the redundancy interference level from small to large and the sensor channel number from small to large, and only the sensor channel with the later retention order is downgraded to an elimination channel.
[0078] When no sensor channel is designated as a core retention channel, and there are sensor channels that are not designated as rejection channels, the sensor channel with the highest label differentiation among the sensor channels that were not designated as rejection channels is selected as the core retention channel. When no sensor channel is designated as a core retention channel, and all sensor channels are designated as rejection channels, the sensor channel with the highest label differentiation among all sensor channels is selected as the core retention channel, and its original grade label is adjusted from rejection channel to core retention channel. When there are two or more sensor channels with the same label differentiation, the sensor channel with the highest repeatability is selected as the core retention channel. When the repeatability is still the same, the sensor channel with the smallest sensor channel number is selected as the core retention channel.
[0079] The channel contribution level results correspond one-to-one with the sensor channel number and are used as input data in the candidate subset construction process. Core retained channels directly enter the candidate subset, auxiliary retained channels participate in supplementary screening, conditionally retained channels enter the recognition gain verification, and eliminated channels do not enter the effective sensor subset.
[0080] Candidate subsets are constructed based on the channel contribution level results, channel contribution index set, and channel representative response value. The candidate subsets include core reserved channels and auxiliary reserved channels that meet the supplementary label differentiation condition and redundancy restriction condition, and serve as the basic subset for the verification of conditional reserved channels.
[0081] All core reserved channels are added to the candidate subset, and the covered label pairs in the candidate subset are initialized to label pairs that can be distinguished by the core reserved channels. The label pair refers to the combination formed by any two different target detection labels. The distinguishable label pairs are determined by the response difference between two target detection labels under the same sensor channel.
[0082] When determining which tag pairs a sensor channel can distinguish, for any sensor channel and any tag pair, read the channel representative response values corresponding to the two target detection tags in the tag pair, calculate the median value of the channel representative response values corresponding to the two target detection tags respectively, and obtain the two tag response representative values; calculate the absolute value of the difference between the two tag response representative values as the tag pair response difference; calculate the average of the absolute values of the differences between the channel representative response values and the corresponding tag response representative values within each of the two target detection tags, and take the average of the two averages as the tag pair fluctuation value; when the tag pair fluctuation value is greater than zero and the tag pair response difference is greater than or equal to twice the tag pair fluctuation value, the sensor channel is recorded as able to distinguish the tag pair; when the tag pair fluctuation value is zero and the tag pair response difference is greater than zero, the sensor channel is recorded as able to distinguish the tag pair; when the tag pair fluctuation value is greater than zero and the tag pair response difference is less than twice the tag pair fluctuation value, or when the tag pair fluctuation value is zero and the tag pair response difference is equal to zero, the sensor channel is recorded as unable to distinguish the tag pair.
[0083] When the number of target detection labels is greater than the preset number of labels, the number of misjudgments between different target detection labels in the historical labeling and detection records is read. The target detection label pairs composed of any two target detection labels are arranged in descending order of the number of misjudgments. The target detection label pairs with the highest number of misjudgments are judged first to see if they can be distinguished by the sensor channel. The preset number of labels is determined based on the upper limit of the number of target detection label pairs that can be processed in a single screening.
[0084] The auxiliary reserved channels are arranged from highest to lowest label distinguishability; when the label distinguishability of two auxiliary reserved channels is the same, they are arranged from lowest to highest redundancy interference; when the redundancy interference is still the same, they are arranged from lowest to highest sensor channel number.
[0085] The auxiliary reserved channels are processed sequentially according to the sorting results. Tag pairs that the current auxiliary reserved channel can distinguish and are not included in the already covered tag pairs are identified as supplementary tag pairs for the current auxiliary reserved channel. These supplementary tag pairs refer to newly covered target detection tag pairs relative to the candidate subset. When the number of supplementary tag pairs for the current auxiliary reserved channel is greater than or equal to one, and the absolute value of the correlation coefficient between the current auxiliary reserved channel and any sensor channel in the candidate subset is less than or equal to the redundancy high boundary, the current auxiliary reserved channel is added to the candidate subset, and the supplementary tag pairs of the current auxiliary reserved channel are added to the already covered tag pairs. When the number of supplementary tag pairs for the current auxiliary reserved channel is zero, or the absolute value of the correlation coefficient between the current auxiliary reserved channel and any sensor channel in the candidate subset is greater than the redundancy high boundary, the current auxiliary reserved channel is not added to the candidate subset.
[0086] The candidate subset serves as the basic subset in the output process of the effective sensor subset. It is used to calculate the basic recognition stability value and to verify whether the condition-preserved channel can be added.
[0087] Based on the candidate subset, condition-reserved channels, channel representative response values, and target detection labels, it is verified whether the condition-reserved channels meet the recognition gain conditions, thus forming an effective sensor subset and medicinal herb odor detection input data.
[0088] First, calculate the basic stable value of the candidate subset; after deduplicating the sample numbers in the multi-channel response data, form a sample number set, and arrange the labeled samples according to the order of their first appearance in the multi-channel response data, assigning verification serial numbers sequentially from the beginning; divide the labeled samples into five verification rounds, and in each verification round, the medicinal herb odor samples with the same remainder when the verification serial number is divided by five are used as verification samples, and the remaining medicinal herb odor samples are used as reference samples; when the number of labeled samples is less than five, a one-sample verification method is adopted, selecting one medicinal herb odor sample as the verification sample each time, and the remaining medicinal herb odor samples as reference samples; when the target detection label of the verification sample in any verification round does not have a corresponding medicinal herb odor sample in the reference samples, the Euclidean distance between the verification sample and the label center vector corresponding to the target detection label is not calculated, and the verification sample is not included in the total number of verification samples.
[0089] The calculation method for the basic recognition stability value is as follows: In each verification round, the reference samples are grouped according to the target detection label. The median value of the representative response value of each sensor channel in the candidate subset under the same target detection label is calculated to form the label center vector of that target detection label. The representative response value of each sensor channel in the candidate subset of the verification sample is read to form the verification sample response vector. The Euclidean distance between the verification sample response vector and each label center vector is calculated using the following formula:
[0090] ;
[0091] in, Indicates verification sample With target detection labels The Euclidean distance between the corresponding label center vectors This indicates the number of sensor channels in the candidate subset. Indicates verification sample In the candidate subset Each sensor channel represents a response value. Indicates object detection label The label center vector in the candidate subset is the first The center value under each sensor channel; when the number of sensor channels in the candidate subset is zero, the Euclidean distance is not calculated, and the basic recognition stability value is recorded as zero; the target detection label corresponding to the label center vector with the smallest Euclidean distance is used as the discriminant label of the verification sample; when there are at least two label center vectors with the same Euclidean distance, the target detection label that appears first in the labeled samples is selected as the discriminant label; when the discriminant label of the verification sample is consistent with the target detection label of the verification sample, the verification sample is regarded as the correctly identified sample; the number of correctly identified samples in all verification rounds is divided by the total number of verification samples to obtain the basic recognition stability value of the candidate subset; when the total number of verification samples is zero, the basic recognition stability value is recorded as zero.
[0092] In another possible implementation, when the labeled samples have been pre-divided into reference samples and verification samples, the basic identification stability value is calculated using a fixed verification set method. The reference samples are used to form label center vectors according to the target detection labels, and the verification samples are used to calculate the Euclidean distance between the verification sample response vector and each label center vector. The basic identification stability value is calculated according to the proportion of the number of verification samples whose discriminant labels are consistent with the target detection labels to the total number of verification samples. When the total number of verification samples is zero, the basic identification stability value is recorded as zero.
[0093] The recognition gain threshold is determined based on the perturbation samples. Each eliminated channel is temporarily added to the candidate subset, and the recognition stability value after temporarily adding the eliminated channel is calculated according to the calculation method of the basic recognition stability value. The absolute value of the difference between the recognition stability value after temporarily adding the eliminated channel and the basic recognition stability value is calculated to form a perturbation sample. When the number of perturbation samples is greater than or equal to three, the 95th percentile value of the perturbation sample is taken as the recognition gain threshold. When the number of perturbation samples is less than three, the recognition gain threshold is set to 0.02. The 0.02 is the default minimum recognition gain threshold, which is based on the minimum effective stability improvement corresponding to the recognition stability value range of 0 to 1.
[0094] The condition-reserved channels are arranged from highest to lowest label discrimination; when two condition-reserved channels have the same label discrimination, they are arranged from highest to lowest repeatability; when the repeatability is still the same, they are arranged from lowest to highest sensor channel number.
[0095] The condition-reserved channels are processed sequentially according to the sorting results. The current condition-reserved channel is temporarily added to the candidate subset. The identification stability value of the temporary subset is calculated according to the calculation method of the basic identification stability value. The identification gain value of the current condition-reserved channel is obtained by subtracting the basic identification stability value of the current candidate subset from the identification stability value of the temporary subset. When the identification gain value is greater than or equal to the identification gain threshold, and the absolute value of the correlation coefficient between the current condition-reserved channel and any sensor channel in the candidate subset is less than or equal to the redundancy high boundary, the current condition-reserved channel is added to the candidate subset, and the basic identification stability value of the candidate subset is updated with the identification stability value of the temporary subset. When the identification gain value is less than the identification gain threshold, or the absolute value of the correlation coefficient between the current condition-reserved channel and any sensor channel in the candidate subset is greater than the redundancy high boundary, the current condition-reserved channel is not added to the candidate subset.
[0096] After completing all condition-reserved channel verifications, the current candidate subset is determined as the effective sensor subset. During the detection of medicinal herb odor, only the response data of the medicinal herb odor sample to be detected under the sensor channel corresponding to the effective sensor subset is received. The effective response sequence of the medicinal herb odor sample to be detected is extracted according to the effective response sequence acquisition process, and the channel representative response value of the medicinal herb odor sample to be detected is determined according to the channel representative response value calculation rules in the channel contribution index set generation process, forming the medicinal herb odor detection input data. The medicinal herb odor detection input data includes the effective response sequence and channel representative response value corresponding to each sensor channel of the medicinal herb odor sample to be detected in the effective sensor subset. The effective response sequence corresponding to the medicinal herb odor sample to be detected includes the sample number, sensor channel number, sampling time, and baseline correction response value, but does not contain the target detection label, and is used as the input for the medicinal herb odor detection result generation process.
[0097] The effective sensor subset limits the range of sensor channels involved in the input of medicinal herb odor detection. The input data for medicinal herb odor detection is used to detect the type, origin, grade or quality status of medicinal herbs, so that sensor channels with low response, low discrimination, low stability or high redundancy do not participate in the final medicinal herb odor detection input.
[0098] Example 2
[0099] See Figure 3 As shown, this embodiment provides a contribution grading and screening system for medicinal herb odor detection sensor arrays. The method for implementing the contribution grading and screening method for medicinal herb odor detection sensor arrays includes:
[0100] Response correction module: used to read the response values of labeled samples containing target detection tags, and to perform baseline correction on the response values from the time of entering the odor collection area until the odor is cleared, forming an effective response sequence;
[0101] Indicator Acquisition Module: Used to calculate the channel representative response value based on the effective response sequence, and to obtain the channel contribution indicator set based on the channel representative response value;
[0102] The classification module is used to determine the classification boundary and redundancy high boundary based on the channel contribution index set. It compares the channel contribution index set with the classification boundary to classify each sensor channel into core retained channels, auxiliary retained channels, conditionally retained channels, and eliminated channels, thus obtaining the channel contribution classification result.
[0103] Candidate building module: used to build candidate subsets based on channel contribution level results, add core reserved channels to candidate subsets, and add auxiliary reserved channels with supplementary label pairs and whose absolute correlation coefficient with any sensor channel in the candidate subset is less than or equal to the redundancy high boundary to candidate subsets;
[0104] Gain verification module: used to determine the recognition gain threshold based on the recognition stability value disturbance of the eliminated channel, add the condition-retained channels whose recognition gain value is greater than or equal to the recognition gain threshold and whose absolute value of the correlation coefficient with any sensor channel in the subset is less than or equal to the redundancy high boundary to the candidate subset, and determine the candidate subset after adding the condition-retained channels as the effective sensor subset;
[0105] Input generation module: When used for detecting the odor of medicinal materials, it only receives the response data of the corresponding channel of the valid sensor subset and generates the input data for detecting the odor of medicinal materials.
[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0107] In conclusion, the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for graded screening of the contribution of a medicinal herb odor detection sensor array, characterized in that, include: Read the response values of labeled samples containing target detection tags, and perform baseline correction on the response values from entering the odor collection area to before the odor is cleared to form an effective response sequence; Calculate the channel representative response value based on the effective response sequence, and obtain the channel contribution index set based on the channel representative response value; Based on the channel contribution index set, the level boundary and redundancy high boundary are determined. By comparing the channel contribution index set with the level boundary, each sensor channel is divided into core retained channels, auxiliary retained channels, conditionally retained channels and eliminated channels, and the channel contribution level results are obtained. Candidate subsets are constructed based on the channel contribution level results. Core reserved channels are added to the candidate subsets, and auxiliary reserved channels with supplementary label pairs and whose absolute correlation coefficient with any sensor channel in the candidate subset is less than or equal to the redundancy high boundary are added to the candidate subsets. The recognition gain threshold is determined based on the recognition stability value perturbation of the eliminated channels. Condition-reserved channels whose recognition gain value is greater than or equal to the recognition gain threshold and whose absolute value of the correlation coefficient with any sensor channel in the subset is less than or equal to the redundancy high boundary are added to the candidate subset. The candidate subset after adding the condition-reserved channels is determined as the effective sensor subset. When detecting the odor of medicinal materials, only the response data of the corresponding channels of the effective sensor subset are received to generate the input data for detecting the odor of medicinal materials.
2. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, Methods for forming effective response sequences include: For each labeled sample and each sensor channel, the median value of the response value of the labeled sample before entering the odor collection area is read as the baseline representative value. The baseline corrected response value is obtained by subtracting the baseline representative value from the response value of the labeled sample after entering the odor collection area until the odor is cleared. The baseline fluctuation limit is determined based on the fluctuation of the response value of the labeled sample before entering the odor collection area. When the absolute value of the baseline corrected response value of three consecutive sampling times is greater than the baseline fluctuation limit, the baseline corrected response value is intercepted from the first sampling time of the three consecutive sampling times to form an effective response sequence.
3. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, The methods for calculating the channel-representing response value and the degree of response validity include: The maximum absolute value of the baseline correction response value at each sampling time in the effective response sequence is taken as the representative response value of the channel. The labeled samples whose representative response value of the channel is greater than or equal to the effective response threshold are counted as effective response samples. The ratio of the number of effective response samples to the number of labeled samples is taken as the response effectiveness of the corresponding sensor channel.
4. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, Methods for calculating the distinguishability of labels include: According to the target detection label, the channel representative response values under the same sensor channel are grouped. The median value of the channel representative response values in each group is taken as the tag response representative value. The average of the absolute values of the difference between any two tag response representative values is calculated as the tag separation value. The average of the absolute values of the difference between the channel representative response value in each group and the corresponding tag response representative value is calculated as the tag fluctuation value. The tag separation value is divided by the sum of the tag separation value and the tag fluctuation value to obtain the tag discrimination degree. When the sum of the tag separation value and the tag fluctuation value is zero, the tag discrimination degree is recorded as zero.
5. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, Methods for calculating repeatability stability include: According to the target detection label, the representative response values of the same sensor channel are grouped. The median value of the representative response values of the channel in each group is taken as the repeatability reference value. The average value of the absolute value of the difference between the representative response value of each channel in the group and the repeatability reference value is calculated as the repeatability deviation value of the group. The repeatability deviation value of the group is divided by the sum of the absolute value of the repeatability reference value and the minimum positive value to obtain the label repeatability deviation value. The minimum positive value is the minimum resolvable response value of the sensor channel. The average value of all label repeatability deviation values is taken as the channel repeatability deviation value. The maximum value between the result obtained by subtracting the channel repeatability deviation value from zero is taken as the repeatability stability.
6. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, Methods for calculating the degree of redundancy interference include: Arrange the representative response values of the same sensor channel according to the order of the labeled samples to form the channel response vector of the sensor channel. Calculate the absolute value of the correlation coefficient between the channel response vector of the current sensor channel and the channel response vectors of other sensor channels, and take the largest absolute value of the correlation coefficient as the degree of redundancy interference of the current sensor channel.
7. The method for graded screening of the contribution of a medicinal herb odor detection sensor array according to claim 1, characterized in that, Before the auxiliary reserved channels are added to the candidate subset, they are arranged in descending order of label discrimination. When the label discrimination of two auxiliary reserved channels is the same, they are arranged in ascending order of redundancy interference. When the redundancy interference is still the same, they are arranged in ascending order of sensor channel number, and the auxiliary reserved channels are then judged to be added to the candidate subset according to the arrangement results.
8. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, The methods for determining supplementary label pairs include: Any two different target detection labels are combined into a target detection label pair. Based on the difference in the channel representative response values corresponding to the two target detection labels under the same sensor channel, it is determined whether the sensor channel can distinguish the target detection label pair. The target detection label pairs that the auxiliary reserved channel can distinguish and that are not yet covered by the candidate subset are determined as the supplementary label pairs of the auxiliary reserved channel.
9. The method for graded screening of contribution of medicinal herb odor detection sensor array according to claim 1, characterized in that, The methods for identifying stable disturbances and calculating gain values include: The basic recognition stability value is calculated based on the channel representative response value and target detection label corresponding to the candidate subset; the removed channel is temporarily added to the candidate subset, the recognition stability value after addition is calculated, and the absolute value of the difference between the added recognition stability value and the basic recognition stability value is used as the recognition stability value perturbation. The recognition gain threshold is determined based on the recognition stability value perturbation; the condition-retained channel is temporarily added to the candidate subset, the temporary recognition stability value is calculated, and the difference between the temporary recognition stability value and the basic recognition stability value is used as the recognition gain value.
10. A contribution grading and screening system for a medicinal herb odor detection sensor array, implementing the contribution grading and screening method for a medicinal herb odor detection sensor array as described in any one of claims 1-9, characterized in that, include: Response correction module: used to read the response values of labeled samples containing target detection tags, and to perform baseline correction on the response values from the time of entering the odor collection area until the odor is cleared, forming an effective response sequence; Indicator Acquisition Module: Used to calculate the channel representative response value based on the effective response sequence, and to obtain the channel contribution indicator set based on the channel representative response value; The classification module is used to determine the classification boundary and redundancy high boundary based on the channel contribution index set. It compares the channel contribution index set with the classification boundary to classify each sensor channel into core retained channels, auxiliary retained channels, conditionally retained channels, and eliminated channels, thus obtaining the channel contribution classification result. Candidate building module: used to build candidate subsets based on channel contribution level results, add core reserved channels to candidate subsets, and add auxiliary reserved channels with supplementary label pairs and whose absolute correlation coefficient with any sensor channel in the candidate subset is less than or equal to the redundancy high boundary to candidate subsets; Gain verification module: used to determine the recognition gain threshold based on the recognition stability value disturbance of the eliminated channel, add the condition-retained channels whose recognition gain value is greater than or equal to the recognition gain threshold and whose absolute value of the correlation coefficient with any sensor channel in the subset is less than or equal to the redundancy high boundary to the candidate subset, and determine the candidate subset after adding the condition-retained channels as the effective sensor subset; Input generation module: When detecting the odor of medicinal materials, it only receives the response data of the corresponding channel of the valid sensor subset and generates the input data for detecting the odor of medicinal materials.
