Methods for rapidly identifying additives in mixtures using artificial intelligence
Patent Information
- Application Number
- TW114110243
- Authority / Receiving Office
- TW · TW
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-03-19
- Publication Date
- 2026-10-01
- Estimated Expiration
- 2045-03-18
AI Technical Summary
Traditional Raman spectroscopy struggles with accurately distinguishing individual components in complex mixtures due to overlapping absorption peaks, requiring pretreatment and often failing to achieve reliable identification.
A method utilizing artificial intelligence to analyze Raman spectra through a trained target feature identification module, involving data cleaning and a machine learning model to identify additives in mixtures, including interpolation, baseline correction, data smoothing, and intensity normalization, followed by classification using algorithms like Random Forest or CNN.
Significantly shortens analysis time and enhances identification efficiency and accuracy, enabling rapid and reliable detection of additives in complex mixtures.
Smart Images
Figure TWG2TA001077571_001 
Figure TWG2TA001077571_002 
Figure TWG2TA001077571_003
Abstract
Description
[Technical Field]
[0001] This invention relates to methods for identifying substances, and more particularly to a method for rapidly identifying additives in a mixture using artificial intelligence. [Previous Technology]
[0002] In traditional techniques, Raman spectroscopy is widely used for compound detection. Simply put, Raman spectroscopy provides information about chemical structures, which helps in identification. However, for complex mixtures, the absorption peaks in the Raman spectra of different components may overlap, causing interference and making it difficult to directly and accurately distinguish the individual components from the spectra. This interference poses a challenge to Raman spectroscopy in the analysis of mixtures.
[0003] To address this issue, pretreatment of the mixture is typically required to remove or reduce certain components in order to improve detection results. However, even after pretreatment, if the mixture contains a large number of components, the measured Raman spectrum may still be subject to significant interference, potentially failing to achieve the desired identification effect. Therefore, traditional techniques still have limitations when detecting mixtures. [Summary of the Invention]
[0004] Therefore, the main objective of this invention is to provide a method for rapidly identifying additives in a mixture using artificial intelligence, which can significantly shorten the analysis time, improve the identification efficiency, and ensure the reliability and accuracy of the detection results. It is applicable to fields such as chemistry, food, and medicine, especially in situations where it is necessary to quickly screen or detect whether a mixture contains a specific additive.
[0005] Therefore, in order to achieve the above-mentioned objective, the method for rapidly identifying additives in a mixture using artificial intelligence provided by the present invention performs the following steps using a computer or a computing program: using a target feature identification module established by a trained artificial intelligence technology to analyze the Raman spectrum of a substance to be analyzed, and using the artificial intelligence technology to determine whether the substance to be analyzed contains one or more identification features included in the target feature identification module, wherein the identification features are the spectral characteristics of the Raman spectra of a majority of different target substances included in the target feature identification module within their respective predetermined wavenumber ranges.
[0006] In one embodiment, the training system of the target feature recognition module includes the following steps: collecting Raman spectra of a plurality of target objects; extracting Raman spectral recognition features of the plurality of target objects respectively, and using them as training data for training.
[0007] In one embodiment, the Raman spectra of the targets are first processed by a data cleaning procedure before their respective identification features are extracted.
[0008] In one embodiment, the data cleaning procedure includes an interpolation processing step, a baseline correction processing step, a data smoothing processing step, and an intensity value normalization processing step, which respectively process the Raman spectra of each target object.
[0009] In one embodiment, the interpolation process is to interpolate the Raman spectra of each target object to complete the missing data in each standard Raman spectrum and / or the missing data in the Raman spectrum to be analyzed.
[0010] In one embodiment, the baseline correction process is performed by using the arPLS (asymmetric reweighted penalized Least Square) algorithm to perform baseline correction on the Raman spectra of each target object, so as to eliminate background noise or systematic bias in the Raman spectra of each target object.
[0011] In one embodiment, the data smoothing step involves using a filter to smooth the Raman spectra of each target object separately. The filter is processed using the following formula:
[0012] (I+λD TD)z=y; where I is the identity matrix, λ is the smoothing parameter, D is the difference matrix, z is the fitted smoothing sequence, and y is the input sequence.
[0013] In one embodiment, λ=1 can be used to smooth the noise in each of the standard Raman spectra and / or smooth the noise in the Raman spectrum to be analyzed.
[0014] In one embodiment, the intensity value normalization step normalizes the intensity range of each spectral feature so that its intensity range is between 0 and 1, and is processed by the following formula:
[0015] ; where y is the sequence to be normalized, yi is the i-th component in y, and is the i-th component after normalization.
[0016] In one embodiment, the artificial intelligence technology is a predetermined machine learning model.
[0017] In one embodiment, the machine learning model is a classification model constructed by one of the following algorithms: Random Forest, Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM).
[0018] In one embodiment, the identification features of each target object are labeled with an identification message, and each identification message is a two-dimensional barcode tag, a QR code tag, or a matrix barcode tag.
Implementation Method
[0019] First, the terms mentioned in this specification must be explained as follows:
[0020] The term 'machine learning (ML)' is a type of artificial intelligence (AI), which focuses on building systems that can learn or improve their performance based on the data used. In other words, artificial intelligence is a broad term that refers to systems and machines that can simulate human intelligence. The difference between machine learning and artificial intelligence is that although all machine learning belongs to artificial intelligence, not all artificial intelligence is machine learning.
[0021] The term 'interpolation' is used to estimate the values of missing or unknown parts of data, making predictions based on known data points. Common interpolation methods include linear interpolation, Lagrange interpolation, spline interpolation, etc., which can be used to fill in missing values in data, so that the data can be presented more completely.
[0022] The term 'baseline correction processing' is used to eliminate the effects of background noise or systematic bias in the data to ensure the authenticity of the measurement data.
[0023] The term 'data smoothing' refers to obtaining a smoother data curve by reducing fluctuations or noise in the data. Common methods include moving average, low-pass filtering, Savitzky-Golay filtering, etc., which can help reduce the impact of random fluctuations or high-frequency noise, making the data more stable.
[0024] The term 'intensity value standardization' refers to the process of adjusting data to the same range by standardizing data, since different measurement conditions may result in different ranges of intensity (or value), thus avoiding deviations caused by different measurement conditions and enabling more direct comparisons between data.
[0025] Next, in the first embodiment of the present invention, a method for rapidly identifying additives in a mixture using artificial intelligence is provided, comprising the following steps:
[0026] A target feature identification module established by a trained artificial intelligence technology is used to analyze the Raman spectrum of a substance to be analyzed, and the artificial intelligence technology determines whether the substance to be analyzed contains one or more identification features included in the target feature identification module, wherein the identification features are the spectral features of the Raman spectra of a majority of different targets included in the target feature identification module within their respective predetermined wavenumber ranges.
[0027] Wherein, the precipitate is a mixture. Each of the target substances can be a different pesticide, such as Thiram, Ferbam, Metiram, Meneb, Mancozeb, Paraquat, Parathion methyl, Carbaryl, Sudan Red, etc. For example, when the target substance is Sudan Red, the predetermined wavenumber range is between 1000 cm⁻¹ and 1599 cm⁻¹, and the predetermined wavenumber ranges of other target substances are also readily available from textbooks, encyclopedias, or related materials. Furthermore, each of the target substances is not limited to the examples mentioned above, and can also be any designated compound, such as food additives or toxins in food safety, or bacteria, viruses, or proteins in biomedicine.
[0028] The specific details of the method are explained below.
[0029] Furthermore, before their respective identification features are extracted, the Raman spectra of these targets undergo a data cleaning process to ensure the accuracy of the Raman spectra of each target. As shown in Figure 1, the data cleaning process includes an interpolation step, a baseline correction step, a data smoothing step, and an intensity value normalization step, which process the Raman spectra of each target respectively. These steps are executed in the aforementioned order, but are not limited to this; any of these steps may be appropriately reduced, or other data cleaning steps may be added. Among them:
[0030] The interpolation process involves interpolating the Raman spectra of each target object to complete the missing data in each standard Raman spectrum and / or the missing data in the Raman spectrum to be analyzed.
[0031] The baseline correction process uses the arPLS (asymmetric reweighted penalized Least Square) algorithm to perform baseline correction on the Raman spectra of each target object, so as to eliminate background noise or systematic bias in each standard Raman spectrum and / or eliminate background noise or systematic bias in the Raman spectrum to be analyzed.
[0032] The data smoothing process involves using a filter to smooth the Raman spectra of each target object. In this example, a Whittaker filter is used, and the following formula is applied:
[0033] (I+λD TD)z=y;
[0034] Where I is the identity matrix, λ is the smoothing parameter, D is the difference matrix, z is the fitted smoothing sequence, and y is the input sequence. In this example, λ=1, which can be used to smooth the noise in each standard Raman spectrum and / or smooth the noise in the Raman spectrum to be analyzed.
[0035] The intensity value standardization process standardizes the intensity range of the Raman spectrum of each target object, making its intensity range between 0 and 1, and is processed by the following formula:
[0036] ;
[0037] where y is the sequence to be normalized, yi is the i-th component in y, and is the i-th component after normalization.
[0038] For example, Figure 4 shows the standard Raman spectra after normalization.
[0039] In other words, the data cleaning process will process the Raman spectra of each target object into a converted data according to the interpolation processing step, the baseline correction processing step, the data smoothing processing step and the intensity value standardization processing step. The converted data does not contain the original data in the Raman spectrum of the target object, and only uses the converted data for subsequent operations such as feature extraction, training, analysis and identification.
[0040] The artificial intelligence technology is a predetermined machine learning model, which is a classification model constructed by one of the following algorithms: Random Forest, Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), or other general algorithms. For example, a 1D-CNN with 5 binary classifiers can be used as the machine learning model.
[0041] Furthermore, the training system of the target feature recognition module includes the following steps:
[0042] Collect Raman spectra of multiple target objects, wherein the collected Raman spectra are processed by the aforementioned data cleaning procedure.
[0043] Next, the Raman spectral recognition features of these complex target objects are extracted respectively, and used as training data for the training.
[0044] The data cleaning procedure and the extraction of the identification features can be performed only in the Raman spectra of these multiple targets within their respective predetermined wavenumber ranges, as training data, wherein the predetermined wavenumber ranges of different targets can be different.
[0045] Furthermore, the artificial intelligence technology, the target feature recognition module, and the machine learning model are pre-installed software or programs on a computing device, and the software or programs are executed through the processor (e.g., CPU, GPU, etc.) of the computing device. The computing device can be, but is not limited to, a computer, a screen, a tablet computer, a mobile phone, etc., and can use a display pre-installed on the computing device to display relevant data, information, or messages. The database is a storage medium built into the computing device, such as phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), flash memory disk, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., used to store the aforementioned data, software, or programs. For example, the Raman spectra of these targets can be read from a medical database of a medical institution and stored in the database. In addition, a handheld Raman spectrometer can be used to measure each target and the object to be analyzed, so as to obtain the Raman spectrum of each target and the Raman spectrum of the object to be analyzed, and then store them in the database.
[0046] Based on the above description, the simplified implementation process of the method provided by the present invention is as follows: preprocessing the Raman spectrum of the target object (i.e., data cleaning), collecting the Raman spectrum of the target object after preprocessing, extracting identification features, training, and analyzing the target feature identification module, so as to quickly screen or detect whether the mixture contains specific additives.
[0047] Furthermore, the main difference between the second embodiment and the first embodiment of the present invention is that the identification features of each target object are labeled with an identification message, thereby helping the target feature identification module to distinguish different target objects and perform accurate analysis and matching. The identification message may be, but is not limited to, a 2D barcode tag (as shown in Figure 3), a QR code tag, a matrix barcode tag (as shown in Figure 2), or other coded graphics.
[0048] The present invention has been described in detail above by way of various examples. Any simple modifications or changes made by those skilled in the art to the embodiments in the specification without departing from the spirit of the present invention shall be covered by the scope of the patent application in this case. [Simplified Explanation of the Diagram]
[0049] Figure 1 is a flowchart of data cleaning in the first embodiment of the present invention. Figure 2 is a schematic diagram of the identification information marked in the second embodiment of the present invention. Figure 3 is a schematic diagram of the Raman spectra of different targets and the identification information marked thereon.
Claims
1. A method for rapidly identifying additives in a mixture using artificial intelligence, comprising a computer or a computing program performing the following steps: analyzing the Raman spectrum of a analyte using a target feature identification module established by a trained artificial intelligence technique, and determining by the artificial intelligence technique whether the analyte contains one or more identification features included in the target feature identification module, wherein: These identification features are the spectral characteristics of the Raman spectra of most different targets included in the target feature identification module within their respective predetermined wavenumber ranges; the training system of the target feature identification module includes the following steps: collecting the Raman spectra of a complex number of targets; processing the Raman spectra of these targets with a data cleaning procedure, which includes a data smoothing step or an intensity value normalization step; the data smoothing step uses a filter to smooth the Raman spectra of each target separately, and processes them using the following formula: (I+λDT D)z=y, where I is the identity matrix, λ is the smoothing parameter, D is the difference matrix, z is the fitted smoothing sequence, and y is the input sequence; the intensity value normalization step normalizes the intensity range of the Raman spectrum of each target so that its intensity range is between 0 and 1, and processes it using the following formula: , where y is the sequence to be normalized, yi is the i-th component in y, and (y_i)̂ is the normalized i-th component. Each component extracts the Raman spectral identification features of the complex target objects, and uses them as training data for the training.
2. The method for rapidly identifying additives in a mixture using artificial intelligence as described in claim 1, wherein, The data cleaning process includes an interpolation step, a baseline correction step, a data smoothing step, and an intensity value normalization step, which process the Raman spectra of each target object respectively.
3. The method for rapidly identifying additives in a mixture using artificial intelligence as described in claim 2, wherein, The interpolation process involves interpolating the Raman spectra of each target object to fill in missing data in the standard Raman spectra and / or the Raman spectra to be analyzed.
4. The method for rapidly identifying additives in a mixture using artificial intelligence as described in claim 2, wherein, The baseline correction process uses the arPLS (asymmetric reweighted penalized Least Square) algorithm to perform baseline correction on the Raman spectra of each target object, so as to eliminate background noise or systematic bias in each standard Raman spectrum and / or eliminate background noise or systematic bias in the Raman spectrum to be analyzed.
5. A method for rapidly identifying additives in a mixture using artificial intelligence as described in any of claims 1 to 4, wherein, λ=1.
6. A method for rapidly identifying additives in a mixture using artificial intelligence as described in any of claims 1-4, wherein, This artificial intelligence technology is a pre-defined machine learning model.
7. The method for rapidly identifying additives in a mixture using artificial intelligence as described in claim 6, wherein, This machine learning model is a classification model constructed using one of the following algorithms: Random Forest, Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), or Long Short-Term Memory (LSTM).
8. A method for rapidly identifying additives in a mixture using artificial intelligence as described in any of claims 1-4, wherein, Each of the identification features is labeled with an identification message, which is a 2D barcode tag, a QR code tag, or a matrix barcode tag.