Essential oil raw material active component content detection system based on intelligent sensor

Abstract

The invention discloses an essential oil raw material active component content detection system based on an intelligent sensor, and relates to the field of essential oil raw material active component content detection.The essential oil raw material active component content detection system comprises an equipment acquisition module used for acquiring chemical signals of volatile components in an essential oil sample and drawing a comprehensive signal time sequence map; the spectrum analysis module is used for automatically determining characteristic peak positions, peak widths and peak heights of different volatile components displayed on the comprehensive signal time sequence spectrum; the feature extraction module is used for extracting a multi-dimensional feature vector from the sensor signal in each time slice; the model construction module is used for constructing a double-branch attention quantification network and outputting the percentage content of the active ingredients of the essential oil raw materials; and the model training and updating module is used for training and updating the double-branch attention quantization network. The method has the advantages that through collaborative innovation of dynamic headspace sampling and the double-branch attention network, rapid lossless separation and high-precision quantitative detection of the active components of the essential oil are realized.

Description

Technical Field

[0001] This invention relates to the field of detecting the content of active ingredients in essential oil raw materials, and more specifically to a system for detecting the content of active ingredients in essential oil raw materials based on intelligent sensors. Background Technology

[0002] In the fields of cosmetics, food, medicine, and fragrance, plant essential oils occupy an indispensable position due to their rich bioactive components and unique aromas. Taking the cosmetics industry as an example, many high-end skincare products incorporate plant essential oils to enhance their skincare efficacy. Rose essential oil, for instance, is rich in various vitamins, minerals, and antioxidants, effectively moisturizing the skin and improving skin tone, making it highly popular among consumers. In the pharmaceutical field, tea tree essential oil, with its powerful antibacterial and anti-inflammatory properties, is widely used in wound care and the treatment of skin inflammation. Accurate analysis of the content of active ingredients in essential oils is crucial to ensuring their quality and efficacy. On the one hand, different levels of active ingredients directly determine the effects of essential oils in various applications. For example, in fragrance blending, precise ingredient content ensures the stability and uniqueness of the aroma; in pharmaceutical applications, accurate active ingredient content is the foundation for ensuring therapeutic effects and safety. On the other hand, as consumers increasingly demand higher product quality and safety, essential oil manufacturers need to meet market demands and enhance product competitiveness through precise ingredient analysis.

[0003] However, traditional methods for essential oil component analysis commonly use gas chromatography-mass spectrometry (GC-MS). While it can separate and identify multiple components in essential oils, the analytical process usually involves complex sample pretreatment, long chromatographic separation time, and is destructive to the sample, failing to meet the requirements for on-site, rapid, and non-destructive detection. Furthermore, in the complex mixture of essential oils, all sensors respond to multiple components simultaneously, and the degree of response varies. This results in the final signal received by the sensor becoming a "chaotic signal soup" whose source is difficult to distinguish, making it difficult to cope with the problem of cross-interference between components in complex essential oil systems. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a detection system for the content of active ingredients in essential oil raw materials based on intelligent sensors. This technical solution solves the problems mentioned in the background technology, such as the destructive nature of the sample and the inability to meet the requirements of on-site, rapid, and non-destructive detection. Furthermore, all sensors respond to multiple components simultaneously, and the degree of response varies. This results in the final signal received by the sensor becoming a "chaotic signal soup" from which it is difficult to directly distinguish the source, making it difficult to cope with the cross-interference problem of components in complex essential oil systems.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A system for detecting the content of active ingredients in essential oil raw materials based on intelligent sensors, comprising:

[0007] The device acquisition module is used to set up a data acquisition device group, acquire chemical signals of volatile components in essential oil samples, and plot a comprehensive signal time series spectrum.

[0008] The spectrum analysis module is used to automatically determine the position, peak width, and peak height of the characteristic peaks of different volatile components on the integrated signal time series spectrum based on the integrated signal time series spectrum.

[0009] The feature extraction module is used to divide the entire time series signal into several time slices corresponding to different volatile components according to the start and end positions of the feature peaks, and extract multi-dimensional feature vectors from the sensor signals in each time slice.

[0010] The model building module is used to construct a dual-branch attention quantization network with multi-dimensional feature vectors as input, and output the percentage content of active ingredients in essential oil raw materials.

[0011] The model training and update module is used to train and update the dual-branch attention quantization network based on the GC-MS detection results as a reference standard and historical detection data.

[0012] Preferably, the step of setting up a data acquisition device group to acquire the chemical signals of volatile components in the essential oil sample and plotting a comprehensive signal time series spectrum specifically includes:

[0013] A multi-channel dynamic headspace sampler was set up to perform programmed temperature-controlled heating and micro-pulse carrier gas purging on essential oil samples sealed in sample bottles to obtain time-series release data of different volatile components.

[0014] A biomimetic sensor array is connected in series at the gas outlet of the multi-channel dynamic headspace sampler via a constant-temperature transmission pipeline.

[0015] A set of multi-dimensional electrical parameter time-series signals are collected and output in real time through a biomimetic sensor array, wherein the electrical parameters include, but are not limited to, resistance value, frequency drift value, and conductivity change value.

[0016] Based on the time series signals of this set of multidimensional electrical parameters, a comprehensive signal time series spectrum for characterizing the changes in total volatile organic compounds in the headspace is generated by principal component analysis.

[0017] In this graph, the horizontal axis represents time, and the vertical axis represents the response value based on the fused signal strength of the sensor array.

[0018] Preferably, the step of automatically determining the characteristic peak positions, peak widths, and peak heights of different volatile components on the integrated signal time series spectrum specifically includes:

[0019] The time series spectrum of the integrated signal is smoothed and filtered to suppress noise, and the baseline of the spectrum is fitted and subtracted using a piecewise linear fitting method to obtain the baseline-corrected time series spectrum.

[0020] Calculate the first derivative spectrum of the time series spectrum after baseline correction, and locate all zero-crossing points that change from positive to negative in the first derivative spectrum. These zero-crossing points are the positions of the characteristic peaks, and the corresponding time points are recorded as peak times.

[0021] Using the peak time as a reference, trace back to the last positive derivative point in the first derivative plot, and record it as the starting point of the peak; trace back to the first negative derivative point in the first derivative plot, and record it as the ending point of the peak.

[0022] The peak width of the characteristic peak is calculated by subtracting the starting point from the ending point.

[0023] On the time series spectrum after baseline correction, the fused signal intensity response value corresponding to the peak time is read, and the average value of the response value at the start and end points of the peak is read as the local baseline.

[0024] The peak height of the characteristic peak is obtained by subtracting the local baseline from the fused signal intensity response value at the peak time.

[0025] The peak position, peak width, and peak height of each characteristic peak are used as the core quantification parameters of the volatile component release curve. These parameters are then associated with and bound to the raw response data of all sensors during the time period from the start point to the end point, thereby completing the quantification and data extraction of different volatile components in the time dimension.

[0026] Preferably, the step of dividing the entire time-series signal into several time slices corresponding to different volatile components based on the start and end positions of the characteristic peaks, and extracting multi-dimensional feature vectors from the sensor signals within each time slice specifically includes:

[0027] Based on the start and end points corresponding to each characteristic peak of the integrated signal time series spectrum, different volatile components are quantified into a series of independent time slices in the time dimension.

[0028] Within this time slice, for each sensor, calculate the maximum response value of that sensor within this time slice;

[0029] Within this time slice, for each sensor, calculate the average response slope of that sensor within this time slice;

[0030] Within this time slice, for each sensor, calculate the area of ​​integration of the sensor's response curve within this time slice;

[0031] The time required for the sensor's response value to recover to the baseline level after the sample stimulus is removed is denoted as the recovery time.

[0032] The maximum response value, average response slope, area integral of the response curve, and recovery time feature extracted by each sensor within this time slice are combined to form a local feature subvector.

[0033] From the comprehensive signal time sequence spectrum, the peak height and peak width corresponding to each time slice are directly extracted to form a global peak feature vector;

[0034] The local feature vectors of all sensors are concatenated with the global peak feature vector to form a multidimensional feature vector.

[0035] Preferably, the model building module specifically includes:

[0036] The network building unit is used in a dual-branch attention quantization network, which includes a component-aware branch, an interference modeling branch, and a fusion regression layer.

[0037] The component sensing branch unit is used to automatically find and weight the sensor features most relevant to the target active ingredient by utilizing an attention mechanism;

[0038] An interference modeling branch unit is used to simultaneously analyze all sensor signals to learn and compensate for cross-interference caused by other components to the detection of the target component;

[0039] The fusion regression layer unit is used to connect with the component perception branch unit and the interference modeling branch unit to fuse the outputs of the two branches. The output of the fusion regression layer is the percentage content of the active ingredients of the essential oil raw material.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention proposes a system for detecting the content of active ingredients in essential oil raw materials based on intelligent sensors. This system acquires chemical signals of volatile components in essential oil samples using a multi-channel dynamic headspace sampler and a biomimetic sensor array. A comprehensive signal time-series spectrum is generated based on principal component analysis (PCA). Automated spectrum analysis determines the characteristic peak parameters of different volatile components, achieving precise temporal separation of each component. Furthermore, time slices are divided based on characteristic peaks, and multidimensional feature vectors containing sensor dynamic parameter features are extracted. A dual-branch attention quantization network is constructed. The component perception branch automatically focuses on the sensor features most relevant to the target active ingredient using an attention mechanism. The interference modeling branch learns and quantifies cross-interference caused by coexisting components through a multilayer perceptron. Finally, interference compensation calculations are performed through a fusion regression layer. This system outputs the precise percentage content of active ingredients in essential oil raw materials. Simultaneously, using GC-MS detection results as a reference standard, and combining historical detection data, the network undergoes supervised training and incremental updates. This approach achieves high-precision and rapid quantitative detection of specific active ingredients in complex essential oil systems, overcoming the limitations of traditional methods that rely on lengthy chromatographic separations, sample destruction, and inability to be implemented on-site. Furthermore, through a dual mechanism of physical time-series separation and algorithmic interference compensation, it effectively solves the problem of signal cross-interference in complex mixtures, significantly improving the quantitative reliability of target components under near-co-eluent conditions. Moreover, the system continuously adapts to new sample characteristics through a continuous learning mechanism, achieving self-evolution of detection accuracy, providing an intelligent and iterative solution for the quality control of essential oil raw materials. Attached Figure Description

[0042] Figure 1 This is a structural block diagram of an essential oil raw material active ingredient content detection system based on a smart sensor according to the present invention;

[0043] Figure 2 The present invention collects chemical signals of volatile components in essential oil samples and plots a flowchart of the integrated signal time series spectrum.

[0044] Figure 3 A flowchart illustrating the automatic determination of the characteristic peak positions, peak widths, and peak heights of different volatile components on the integrated signal time-series spectrum according to the present invention;

[0045] Figure 4 The flowchart of the present invention is as follows: dividing the entire time series signal into several time slices corresponding to different volatile components, and extracting multidimensional feature vectors from the sensor signals in each time slice. Detailed Implementation

[0046] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0047] Reference Figure 1 As shown, a system for detecting the content of active ingredients in essential oil raw materials based on intelligent sensors includes:

[0048] The device acquisition module is used to set up a data acquisition device group, acquire chemical signals of volatile components in essential oil samples, and plot a comprehensive signal time series spectrum.

[0049] The spectrum analysis module is used to automatically determine the position, peak width, and peak height of the characteristic peaks of different volatile components on the integrated signal time series spectrum based on the integrated signal time series spectrum.

[0050] The feature extraction module is used to divide the entire time series signal into several time slices corresponding to different volatile components according to the start and end positions of the feature peaks, and extract multi-dimensional feature vectors from the sensor signals in each time slice.

[0051] The model building module is used to construct a dual-branch attention quantization network with multi-dimensional feature vectors as input, and output the percentage content of active ingredients in essential oil raw materials.

[0052] The model training and update module is used to train and update the dual-branch attention quantization network based on the GC-MS detection results as a reference standard and historical detection data.

[0053] Reference Figure 1 As shown, the model building module specifically includes:

[0054] The network building unit is used in a dual-branch attention quantization network, which includes a component-aware branch, an interference modeling branch, and a fusion regression layer.

[0055] The component sensing branch unit is used to automatically find and weight the sensor features most relevant to the target active ingredient by utilizing an attention mechanism;

[0056] An interference modeling branch unit is used to simultaneously analyze all sensor signals to learn and compensate for cross-interference caused by other components to the detection of the target component;

[0057] The fusion regression layer unit is used to connect with the component perception branch unit and the interference modeling branch unit to fuse the outputs of the two branches. The output of the fusion regression layer is the percentage content of the active ingredients of the essential oil raw material.

[0058] This solution involves acquiring chemical signals of volatile components in essential oil samples using a multi-channel dynamic headspace sampler and a biomimetic sensor array. A comprehensive signal time-series spectrum is generated based on principal component analysis. Automated spectrum analysis determines the characteristic peak parameters of different volatile components, achieving precise temporal separation of each component. Furthermore, time slices are divided based on characteristic peaks, and multi-dimensional feature vectors containing sensor dynamic parameters are extracted. A dual-branch attention quantification network is constructed. The component perception branch automatically focuses on the sensor features most relevant to the target active ingredient using an attention mechanism, while the interference modeling branch learns and quantifies cross-interference caused by coexisting components through a multilayer perceptron. Finally, interference compensation calculations are performed through a fusion regression layer, outputting the precise percentage content of the active ingredients in the essential oil raw material. Simultaneously, GC-MS detection results are used as a control standard, and historical detection data are combined to supervise training and incremental updates of the network. Through the synergistic innovation of dynamic headspace sampling and the dual-branch attention network, rapid, non-destructive separation and high-precision quantitative detection of essential oil active ingredients are achieved.

[0059] Reference Figure 2 As shown, the specific steps of collecting chemical signals of volatile components in essential oil samples and plotting a comprehensive signal time series spectrum include:

[0060] A multi-channel dynamic headspace sampler was set up to perform programmed temperature-controlled heating and micro-pulse carrier gas purging on essential oil samples sealed in sample bottles to obtain time-series release data of different volatile components.

[0061] A biomimetic sensor array is connected in series at the gas outlet of the multi-channel dynamic headspace sampler via a constant-temperature transmission pipeline.

[0062] A set of multi-dimensional electrical parameter time-series signals are collected and output in real time through a biomimetic sensor array, wherein the electrical parameters include, but are not limited to, resistance value, frequency drift value, and conductivity change value.

[0063] Based on the time series signals of this set of multidimensional electrical parameters, a comprehensive signal time series spectrum for characterizing the changes in total volatile organic compounds in the headspace is generated by principal component analysis.

[0064] In this graph, the horizontal axis represents time, and the vertical axis represents the response value based on the fused signal strength of the sensor array.

[0065] This can be explained by the fact that, compared to traditional static headspace or single-sensor detection methods, this scheme achieves a technological upgrade from component aliasing detection to temporal separation and analysis by employing a series architecture of a multi-channel dynamic headspace sampler and a biomimetic sensor array. Through the synergistic effect of programmed temperature control and micro-pulse carrier gas, the multi-channel dynamic headspace sampler utilizes the different desorption rates of different volatile components under specific temperatures and airflow conditions due to differences in boiling points and polarities, thereby achieving their separation in the time series and effectively avoiding signal interference and masking effects between components. The subsequently connected biomimetic sensor array, with its diverse functionalized interfaces, further enhances the temporal separation of components. Parallel capture and characteristic response of monomeric components after sequential separation significantly improves the detection sensitivity and selectivity of trace active ingredients in complex matrices. Among these, a multi-channel dynamic headspace sampler is a sample pretreatment device capable of parallel, dynamic, and controllable extraction and preliminary separation of volatile components in complex samples. By setting up a multi-channel dynamic headspace sampler, the essential oil sample sealed in the sample bottle is subjected to programmed temperature-controlled heating (linear heating from the initial temperature T1 to the final temperature T2 at a rate R) and micro-pulse carrier gas purging (the flow rate fluctuates periodically in a sinusoidal or rectangular wave mode within the range V1). This utilizes the specific temperature and gas response of different volatile components. The difference in desorption rates along the flow path enables separation over time. Secondly, the biomimetic sensor array is composed of a combination of surface-functionalized metal oxide semiconductor sensors, a quartz microbalance, and conductive polymer sensors. This array converts the chemical adsorption signal of a specific component into a quantifiable electrical parameter time-series signal, thereby enabling the identification and semi-quantitative analysis of specific active ingredients. The specific parameter ranges for the programmed temperature control and carrier gas flow rate are optimized based on the boiling point distribution and diffusion characteristics of typical plant essential oil volatile components, aiming to achieve effective separation of most target components. The temperature setting of the isothermal transport pipeline is designed to... The boiling point is higher than that of the high-boiling-point components, ensuring that all separated components can be transmitted to the sensor array in gaseous form without loss. The response value of the fused signal intensity is obtained by using a principal component analysis algorithm to fuse the multi-dimensional, standardized electrical parameter time-series signals (including resistance value, frequency drift value, and conductivity change value) output by the biomimetic sensor array. This score, as a dimensionless comprehensive index, quantifies the overall intensity and chemical activity of volatile components flowing through the sensor array at a specific time point, thereby unifying the heterogeneous signals of multiple sensors into a single variable time-series signal that can be used to plot release curves.

[0066] Reference Figure 3 As shown, the automatic determination of the characteristic peak positions, peak widths, and peak heights of different volatile components on the integrated signal time series spectrum specifically includes:

[0067] The time series spectrum of the integrated signal is smoothed and filtered to suppress noise, and the baseline of the spectrum is fitted and subtracted using a piecewise linear fitting method to obtain the baseline-corrected time series spectrum.

[0068] Calculate the first derivative spectrum of the time series spectrum after baseline correction, and locate all zero-crossing points that change from positive to negative in the first derivative spectrum. These zero-crossing points are the positions of the characteristic peaks, and the corresponding time points are recorded as peak times.

[0069] Using the peak time as a reference, trace back to the last positive derivative point in the first derivative plot, and record it as the starting point of the peak; trace back to the first negative derivative point in the first derivative plot, and record it as the ending point of the peak.

[0070] The peak width of the characteristic peak is calculated by subtracting the starting point from the ending point.

[0071] On the time series spectrum after baseline correction, the fused signal intensity response value corresponding to the peak time is read, and the average value of the response value at the start and end points of the peak is read as the local baseline.

[0072] The peak height of the characteristic peak is obtained by subtracting the local baseline from the fused signal intensity response value at the peak time.

[0073] The peak position, peak width, and peak height of each characteristic peak are used as the core quantification parameters of the volatile component release curve. These parameters are then associated with and bound to the raw response data of all sensors during the time period from the start point to the end point, thereby completing the quantification and data extraction of different volatile components in the time dimension.

[0074] It can be explained that the response value of the fused signal intensity reflected in the integrated signal time series spectrum is the first principal component score obtained by fusing the multi-dimensional, standardized electrical parameter time series signals (including resistance, frequency drift, and conductivity changes) output by the biomimetic sensor array using a principal component analysis algorithm. This score, as a dimensionless comprehensive index, quantifies the overall intensity and chemical activity of volatile components flowing through the sensor array at a specific time point and intuitively shows the separation trend of different components on the time axis. However, in order to achieve unmanned and standardized extraction of quantitative information of each component from this macroscopic overall trend, this solution uses a complete automated algorithm to mathematically interpret the integrated signal time series spectrum, automatically converting the intuitive peak shape into calculable structured parameters. This automated process replaces the reliance on pre-set, reusable digital signal processing steps (such as smoothing filtering, baseline correction, and derivative calculation) in traditional methods. Subjective judgment based on human experience ensures the objective consistency and high throughput of characteristic peak position, peak width, and peak height parameters. The automatically identified characteristic peak position reflects the elution time of the component in the separated sequence, a key indicator related to compound properties. The automatically calculated characteristic peak width reflects the diffusion degree of the component during separation, related to separation efficiency and component purity. The automatically measured characteristic peak height reflects the relative concentration or absolute content of the component in the headspace. This automatic peak boundary determination method based on derivative changes is more adaptable to complex situations with different peak shapes and baseline drift than the traditional fixed threshold method. In principle, it ensures the accuracy and robustness of the quantitative interpretation process without human intervention, laying a solid foundation for building an end-to-end intelligent analysis system. Specifically, a peak height threshold of three times the baseline standard deviation is set; peaks below this threshold are identified as false peaks and removed. For overlapping peaks, a Gaussian multi-peak fitting algorithm is used to split them before extracting parameters.

[0075] Reference Figure 4 As shown, dividing the entire time-series signal into several time slices corresponding to different volatile components, and extracting multi-dimensional feature vectors from the sensor signals within each time slice specifically includes:

[0076] Based on the start and end points corresponding to each characteristic peak of the integrated signal time series spectrum, different volatile components are quantified into a series of independent time slices in the time dimension.

[0077] Within this time slice, for each sensor, calculate the maximum response value of that sensor within this time slice;

[0078] Within this time slice, for each sensor, calculate the average response slope of that sensor within this time slice;

[0079] Within this time slice, for each sensor, calculate the area of ​​integration of the sensor's response curve within this time slice;

[0080] The time required for the sensor's response value to recover to the baseline level after the sample stimulus is removed is denoted as the recovery time.

[0081] The maximum response value, average response slope, area integral of the response curve, and recovery time feature extracted by each sensor within this time slice are combined to form a local feature subvector.

[0082] From the comprehensive signal time sequence spectrum, the peak height and peak width corresponding to each time slice are directly extracted to form a global peak feature vector;

[0083] The local feature vectors of all sensors are concatenated with the global peak feature vector to form a multidimensional feature vector.

[0084] This approach divides the entire time-series signal into several time slices corresponding to different volatile components by synthesizing the signal time-series spectrum, and obtains the corresponding peak width and peak height parameters. The purpose is to use the synthesized signal time-series spectrum as a "consensus view" of all sensor information, effectively overcoming the problems of missing peaks, false peaks, or inconsistent peak boundary determination caused by selectivity differences, noise interference, or response drift of a single sensor. This ensures the objectivity, completeness, and high reliability of the time-slice division. Based on this unified time benchmark, feature extraction is then performed by tracing back to the original response curves of each sensor, fully utilizing the retained, high-fidelity detailed information (such as kinetic processes and specific features). The multidimensional feature vector constructed by the biomimetic sensor possesses both the robustness of the fused signal and the richness of the original signal, providing a reliable data foundation for high-precision quantitative analysis in subsequent artificial intelligence models. Specifically, the maximum response value, derived from the time-series signals of the original electrical parameters (such as resistance, frequency drift, and conductivity changes) collected by each biomimetic sensor, refers to the largest value among all response values ​​of the sensor within that time slice. This value reflects the maximum signal intensity that the component can elicit. The average response slope refers to the slope of the sensor's response curve within that time slice, calculated from the rising edge (from the starting point to the...). The slope of the peak is the arithmetic mean of the slopes of all adjacent data points, reflecting the rate at which the component is captured by the sensor. The area under the integral of the response curve refers to the area under the curve calculated by numerically integrating (e.g., using the trapezoidal method) the response curve of the sensor within the time slice, reflecting the total effect of the component. In particular, due to the differences in molecular structure (polarity, functional groups, molecular weight) of different active components, their adsorption-desorption rates differ from those of the sensor's sensitive material. Components with strong binding forces (such as highly polar alcohols) result in slow desorption and longer recovery times; while components with weak binding forces (such as nonpolar alkenes) desorb rapidly and have shorter recovery times. By extracting the recovery time feature, additional dynamic basis can be provided for the component sensing branch of the subsequent dual-branch network to distinguish different component types, enhancing the model's ability to identify specific components in complex mixtures. Specifically, the time required for the sensor to recover to the baseline level after the removal of sample stimulation refers to the time required for the response curve of each sensor to fall from the peak to the baseline level after reaching the maximum value (peak) within the time slice. The baseline level is obtained by exposing the bionic sensor array to pure carrier gas for at least 60 seconds at the beginning of each detection cycle and taking the average value of the sensor output signal during this period as the baseline level.The response value refers to the time-series signal of the original electrical parameters acquired by the bionic sensor, and the response curve refers to the curve plotted with time as the horizontal axis and the original response value as the vertical axis. Specifically, to ensure the system's robustness under extreme conditions, such as when there are severely overlapping peaks in the integrated signal time-series spectrum that prevent automatic analysis of independent characteristic peaks, the system will activate a backup signal segmentation strategy. This strategy uses a fixed-time-window sliding method or a signal segmentation algorithm based on change points to divide the time-series signal into continuous time slices. It should be noted that the time slices generated in this mode may chemically correspond to the mixed release stage of multiple volatile components, rather than a single pure component. The system will specially label the multidimensional feature vectors extracted from such slices, and the subsequent dual-branch attention quantization network will identify these labels. In this case, the interference modeling branch will play a more crucial role, dedicated to decoupling the target component information from strong cross-interference signals, thereby ensuring that the system can still provide reliable quantitative estimates under non-ideal separation conditions.

[0085] The process of constructing a dual-branch attention quantization network using multi-dimensional feature vectors as input to output the percentage content of active ingredients in essential oil raw materials specifically includes:

[0086] The dual-branch attention quantization network includes a component-aware branch, an interference modeling branch, and a fusion regression layer;

[0087] The component sensing branch utilizes an attention mechanism to automatically find and weight the sensor features most relevant to the target active ingredient;

[0088] The interference modeling branch simultaneously analyzes all sensor signals to learn and compensate for cross-interference caused by other components in the detection of the target component;

[0089] The fusion regression layer is connected to the component perception branch unit and the interference modeling branch unit, and the outputs of the two branches are fused. The output of the fusion regression layer is the percentage content of the active ingredients of the essential oil raw material.

[0090] This can be explained by the fact that by constructing a dual-branch attention quantification network, an intelligent and precise mapping from feature data to component content is achieved. This is the core technical guarantee for the high-precision quantitative analysis of this solution. The network effectively solves the key technical challenge of component cross-interference in complex essential oil systems (in complex mixtures like essential oils, all sensors simultaneously produce varying responses to multiple components, ultimately resulting in a "chaotic signal soup" from which the sensor signals cannot be directly distinguished) through a dual-path collaborative mechanism of "component perception" and "interference modeling." Specifically, the component perception branch automatically focuses on the feature signals most relevant to the target component using an attention mechanism, much like an experienced analyst... Experts precisely pinpoint key evidence; the interference modeling branch simultaneously analyzes full-spectrum features, dynamically learns and quantifies background interference from coexisting components, and forms an effective compensation signal; finally, the measurement bias is corrected by integrating the outputs of the two branches through a fusion regression layer. This intelligent processing flow of "feature screening - interference compensation - fusion decision" upgrades the traditional single anti-interference strategy that relies on physical separation to a dual guarantee system of "physical separation and algorithm compensation," significantly improving the quantitative detection accuracy and reliability of target active ingredients in complex matrices. This allows the system to maintain stable analytical performance even under extreme conditions of near-complete co-eluenting, making it a decisive technical link for achieving rapid and accurate quality control of essential oil raw materials.

[0091] The component sensing branch unit specifically includes:

[0092] The set of output multidimensional feature vectors is denoted as: ,in, For the first There are features, where m is the total number of features;

[0093] Using a learnable query vector q, calculate the compatibility score between each feature and the query vector. : ,in, This is a compatibility calculation function;

[0094] The compatibility score is normalized using the softmax function to obtain the attention weights for each feature. : ;

[0095] The output multidimensional feature vector F is compared with its corresponding attention weight A. We perform a weighted summation to obtain the weighted feature representation Z. ;

[0096] The obtained weighted feature representation Z will be used as the output of the component-aware branch and fed into the fusion regression layer.

[0097] This can be explained by the fact that the component-aware branch is a key part of the dual-branch attention quantization network responsible for accurately identifying the features related to the target active ingredient. Its core lies in the clever use of the attention mechanism. When faced with multi-dimensional feature vector input, this branch acts like a highly experienced analytical expert, quickly and accurately finding the parts most relevant to the target active ingredient among numerous complex sensor features. Taking the analysis of linalool content in lavender essential oil as an example, after a large amount of feature data collected by the sensors is input into the component-aware branch, the attention mechanism begins to function. It calculates the degree of correlation between each sensor feature and the target active ingredient (linalool), assigning corresponding weights to different features. Features that are of significant indicative importance, such as the integrated area of ​​a specific metal oxide semiconductor sensor over a certain time slice, or the maximum response value of a quartz microbalance sensor, are given high weight, much like an expert identifying key evidence among numerous clues. Features unrelated to or weakly related to linalool, such as weak response signals of other minor components or noise signals from environmental interference, are given lower weight or even ignored. Through this weighted processing, the component sensing branch can focus its attention highly on sensor features closely related to the target active ingredient, thereby extracting the most valuable information and laying a solid foundation for subsequent accurate analysis of the target active ingredient's content. For compatibility calculation functions, mechanisms such as dot product, scaled dot product, or additive attention can be selected.

[0098] The interference modeling branch unit specifically includes:

[0099] The multidimensional feature vector set F is taken as input and nonlinearly transformed through a multilayer perceptron consisting of fully connected layers.

[0100] The last fully connected layer of the multilayer perceptron outputs a vector with the same dimension as the weighted feature representation Z obtained from the component perception branch, which is the interference vector I.

[0101] The interference vector I and the weighted feature representation Z are input into the fusion regression layer, and interference compensation calculation is performed for fusion.

[0102] This can be explained by the fact that the interference modeling branch plays a crucial role in the two-branch attention quantization network, primarily responsible for solving the intractable cross-interference problem in complex essential oil systems. In the complex mixture system of essential oils, each component may produce a unique response to the sensor, and these responses are intertwined and mutually influential, forming a complex interference network. For example, in lavender essential oil, in addition to the main active ingredients such as linalool and linalyl acetate, it also contains various other volatile compounds such as terpenes and oxides. When using a sensor to detect the linalool content, the presence of linalyl acetate and other components (such as β-caryophyllene) will cause significant cross-interference to the sensor signal, making it extremely difficult to accurately resolve the true content of linalool from the original signal alone. The interference modeling branch learns and compensates for these cross-interferences by simultaneously analyzing all sensor signals and deeply exploring their potential patterns. During the analysis process, this branch... Multilayer perceptrons quantify the interactions and interference relationships between different components. Through continuous learning and training, they can accurately identify the degree and pattern of interference caused by each component to the detection of the target component. For example, they can effectively quantify the superposition or masking effect of linalyl acetate on the linalool response signal on a specific sensor, whether it enhances or weakens the signal, and the specific magnitude of the change. Based on this, the interference modeling branch generates a corresponding compensation vector according to the learned interference information. When detecting linalool content, it compensates for the interfered parts in the original features according to the learned interference patterns (such as interference from linalyl acetate and β-caryophyllene), thereby eliminating or reducing the influence of cross-interference. This makes the final signal used for content regression purer and more accurate, effectively improving the reliability and accuracy of quantitative analysis of the target active ingredient. The fusion regression layer represents the interference-compensated features... The input is fed into a final fully connected regression layer (which may contain a linear activation function) to map the high-dimensional feature to the percentage content of the target active ingredient;

[0103] The interference compensation calculation is expressed as follows:

[0104]

[0105] In the formula, This is the feature representation after interference compensation. This is a learnable scaling factor used to control the strength of interference compensation.

[0106] The specific steps of using GC-MS detection results as a reference standard and training and updating the dual-branch attention quantization network based on historical detection data include:

[0107] Collect N essential oil raw material samples with broad representativeness. The representativeness refers to the fact that the samples cover different origins, different harvesting periods, different processing batches, and the target active ingredient content is distributed in a gradient.

[0108] GC-MS was used to standardize the detection of each sample, determine the percentage content of the target active ingredient in the sample, and use the result as the standard true value label for the sample.

[0109] The sample is detected using a dual-branch attention quantization network to obtain its multi-dimensional feature vector, thus forming the sensor digital fingerprint of the sample.

[0110] Using the sensor's digital fingerprint as the input feature and its corresponding GC-MS standard value as the training target, the dual-branch attention quantization network is trained under supervision.

[0111] During training, the mean squared error loss function between the network predictions and the GC-MS standard values ​​is minimized through the backpropagation algorithm until the model converges, thus obtaining the initial two-branch attention quantization network model.

[0112] During routine testing, the system automatically stores test results and sample information in the cloud platform database.

[0113] M samples with update value are periodically selected by authorized personnel. The criteria for value judgment include: samples with low confidence in model prediction, samples with chemical composition spectra that exceed the range of the initial training set, or samples that are questionable according to user feedback.

[0114] The selected M samples were sent to GC-MS for standard testing to obtain their latest standard true value labels;

[0115] An incremental training set is constructed by combining the sensor digital fingerprints of M samples with their latest GC-MS standard value labels;

[0116] Based on the incremental training set, the initial dual-branch attention quantization network model is incrementally trained to update the network weight parameters.

[0117] This solution establishes a reliable "supervised learning" mechanism for the intelligent sensor system by using GC-MS, the internationally recognized "gold standard" for chemical composition analysis, as the ultimate reference. In the initial training phase, the system learns from samples containing a large number of GC-MS detection results, much like an apprentice practicing repeatedly under the guidance of an expert. This establishes a high-precision mapping model from the sensor's digital fingerprint to the actual component content. Through a continuous update mechanism, the system can continuously self-calibrate and optimize using historical detection data. This not only combats the long-term drift problem of the sensor itself but also allows the system to gradually learn and adapt to new essential oil varieties or chemical spectra, enabling its analytical capabilities to maintain generalization ability when facing new data distributions. This ability continuously strengthens over time, ultimately forming an essential oil raw material active ingredient content detection system that becomes more accurate with use. This provides the core technological guarantee for the system to maintain high reliability and high precision in complex practical applications. The incremental training can be performed using full-parameter fine-tuning, partial layer freeze fine-tuning, or a strategy based on elastic weight consolidation.

[0118] In summary, the advantages of this invention are: through the synergistic innovation of dynamic headspace sampling and bi-branch attention network, rapid and non-destructive separation and high-precision quantitative detection of essential oil active ingredients are achieved.

[0119] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A system for detecting the content of active ingredients in essential oil raw materials based on intelligent sensors, characterized in that, include: The device acquisition module is used to set up a data acquisition device group, acquire chemical signals of volatile components in essential oil samples, and plot a comprehensive signal time series spectrum. The spectrum analysis module is used to automatically determine the position, peak width, and peak height of the characteristic peaks of different volatile components on the integrated signal time series spectrum based on the integrated signal time series spectrum. The feature extraction module is used to divide the entire time series signal into several time slices corresponding to different volatile components according to the start and end positions of the feature peaks, and extract multi-dimensional feature vectors from the sensor signals in each time slice. The model building module is used to construct a dual-branch attention quantization network with multi-dimensional feature vectors as input, and output the percentage content of active ingredients in essential oil raw materials. The model training and update module is used to train and update the dual-branch attention quantization network based on the GC-MS detection results as a reference standard and historical detection data.

2. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 1, characterized in that, The specific steps of setting up the data acquisition equipment group to collect the chemical signals of volatile components in the essential oil sample and plotting the comprehensive signal time series spectrum include: A multi-channel dynamic headspace sampler was set up to perform programmed temperature-controlled heating and micro-pulse carrier gas purging on essential oil samples sealed in sample bottles to obtain time-series release data of different volatile components. A biomimetic sensor array is connected in series at the gas outlet of the multi-channel dynamic headspace sampler via a constant-temperature transmission pipeline. A set of multi-dimensional electrical parameter time-series signals are collected and output in real time through a biomimetic sensor array, wherein the electrical parameters include, but are not limited to, resistance value, frequency drift value, and conductivity change value. Based on the time series signals of this set of multidimensional electrical parameters, a comprehensive signal time series spectrum for characterizing the changes in total volatile organic compounds in the headspace is generated by principal component analysis. In this graph, the horizontal axis represents time, and the vertical axis represents the response value based on the fused signal strength of the sensor array.

3. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 2, characterized in that, The automatic determination of the characteristic peak positions, peak widths, and peak heights of different volatile components on the integrated signal time series spectrum specifically includes: The time series spectrum of the integrated signal is smoothed and filtered to suppress noise, and the baseline of the spectrum is fitted and subtracted using a piecewise linear fitting method to obtain the baseline-corrected time series spectrum. Calculate the first derivative spectrum of the time series spectrum after baseline correction, and locate all zero-crossing points that change from positive to negative in the first derivative spectrum. These zero-crossing points are the positions of the characteristic peaks, and the corresponding time points are recorded as peak times. Using the peak time as a reference, trace back to the last positive derivative point in the first derivative plot, and record it as the starting point of the peak; trace back to the first negative derivative point in the first derivative plot, and record it as the ending point of the peak. The peak width of the characteristic peak is calculated by subtracting the starting point from the ending point. On the time series spectrum after baseline correction, the fused signal intensity response value corresponding to the peak time is read, and the average value of the response value at the start and end points of the peak is read as the local baseline. The peak height of the characteristic peak is obtained by subtracting the local baseline from the fused signal intensity response value at the peak time. The peak position, peak width, and peak height of each characteristic peak are used as the core quantification parameters of the volatile component release curve. These parameters are then associated with and bound to the raw response data of all sensors during the time period from the start point to the end point, thereby completing the quantification and data extraction of different volatile components in the time dimension.

4. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 3, characterized in that, The step of dividing the entire time-series signal into several time slices corresponding to different volatile components based on the start and end positions of the characteristic peaks, and extracting multi-dimensional feature vectors from the sensor signals within each time slice specifically includes: Based on the start and end points corresponding to each characteristic peak of the integrated signal time series spectrum, different volatile components are quantified into a series of independent time slices in the time dimension. Within this time slice, for each sensor, calculate the maximum response value of that sensor within this time slice; Within this time slice, for each sensor, calculate the average response slope of that sensor within this time slice; Within this time slice, for each sensor, calculate the area of ​​integration of the sensor's response curve within this time slice; The time required for the sensor's response value to recover to the baseline level after the sample stimulus is removed is denoted as the recovery time. The maximum response value, average response slope, area integral of the response curve, and recovery time feature extracted by each sensor within this time slice are combined to form a local feature subvector. From the comprehensive signal time sequence spectrum, the peak height and peak width corresponding to each time slice are directly extracted to form a global peak feature vector; The local feature vectors of all sensors are concatenated with the global peak feature vector to form a multidimensional feature vector.

5. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 4, characterized in that, The model building module specifically includes: The network building unit is used in a dual-branch attention quantization network, which includes a component-aware branch, an interference modeling branch, and a fusion regression layer. The component sensing branch unit is used to automatically find and weight the sensor features most relevant to the target active ingredient by utilizing an attention mechanism; An interference modeling branch unit is used to simultaneously analyze all sensor signals to learn and compensate for cross-interference caused by other components to the detection of the target component; The fusion regression layer unit is used to connect with the component perception branch unit and the interference modeling branch unit to fuse the outputs of the two branches. The output of the fusion regression layer is the percentage content of the active ingredients of the essential oil raw material.

6. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 5, characterized in that, The component sensing branch unit specifically includes: The set of output multidimensional feature vectors is denoted as: ,in, For the first There are features, where m is the total number of features; Using a learnable query vector q, calculate the compatibility score between each feature and the query vector. : ,in, This is a compatibility calculation function; The compatibility score is normalized using the softmax function to obtain the attention weights for each feature. : ; The output multidimensional feature vector F is compared with its corresponding attention weight A. We perform a weighted summation to obtain the weighted feature representation Z. ; The obtained weighted feature representation Z will be used as the output of the component-aware branch and fed into the fusion regression layer.

7. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 6, characterized in that, The interference modeling branch unit specifically includes: The multidimensional feature vector set F is taken as input and nonlinearly transformed through a multilayer perceptron consisting of fully connected layers. The last fully connected layer of the multilayer perceptron outputs a vector with the same dimension as the weighted feature representation Z obtained from the component perception branch, which is the interference vector I. The interference vector I and the weighted feature representation Z are input into the fusion regression layer, and interference compensation calculation is performed for fusion.

8. The essential oil raw material active ingredient content detection system based on intelligent sensors according to claim 7, characterized in that, The specific steps of using GC-MS detection results as a reference standard and training and updating the dual-branch attention quantization network based on historical detection data include: Collect N essential oil raw material samples with broad representativeness. The representativeness refers to the fact that the samples cover different origins, different harvesting periods, different processing batches, and the target active ingredient content is distributed in a gradient. GC-MS was used to standardize the detection of each sample, determine the percentage content of the target active ingredient in the sample, and use the result as the standard true value label for the sample. The sample is detected using a dual-branch attention quantization network to obtain its multi-dimensional feature vector, thus forming the sensor digital fingerprint of the sample. Using the sensor's digital fingerprint as the input feature and its corresponding GC-MS standard value as the training target, the dual-branch attention quantization network is trained under supervision. During training, the mean squared error loss function between the network predictions and the GC-MS standard values ​​is minimized through the backpropagation algorithm until the model converges, thus obtaining the initial two-branch attention quantization network model. During routine testing, the system automatically stores test results and sample information in the cloud platform database. M samples with update value are periodically selected by authorized personnel. The criteria for value judgment include: samples with low confidence in model prediction, samples with chemical composition spectra that exceed the range of the initial training set, or samples that are questionable according to user feedback. The selected M samples were sent to GC-MS for standard testing to obtain their latest standard true value labels; An incremental training set is constructed by combining the sensor digital fingerprints of M samples with their latest GC-MS standard value labels; Based on the incremental training set, the initial dual-branch attention quantization network model is incrementally trained to update the network weight parameters.

Images