Spectrum detection system and method applied to protective adhesive tape

Abstract

The invention discloses a spectrum detection system and method applied to a protective adhesive tape, and relates to the technical field of self-adhesive elastic bandage detection. A self-adaptive classification strategy is designed to automatically endow a reference sample with subclass labels, and an accurate infrared spectrum of the reference sample is obtained by adopting a character directional spectrum preparation method; a class average component proportion table and class standard spectrums are generated through a deviation correction method so as to preliminarily construct a standard comparison database for tasks to be detected, samples to be detected are divided into parts, infrared spectrums of single parts are compared with class average standard spectrums of subclass labels through a finished product quality inspection algorithm, efficient and accurate sample quality inspection is carried out, and the quality inspection accuracy of the samples is improved. A component analysis algorithm is adopted to intelligently judge whether a single component corresponds to an existing subclass label, two spectrum analysis methods are designed for the single component which cannot correspond to the single component so as to accurately predict a component proportion table of the single component, a standard comparison database is automatically updated, and automatic component analysis of an unknown to-be-detected sample is realized.

Description

Technical Field

[0001] The invention relates to the technical field of self-adhesive elastic bandage detection, and in particular to a spectrum detection system and method applied to protective tapes. Background Art

[0002] To ensure the safety and effectiveness of low-temperature adhesive self-adhesive elastic bandages, accurate testing of their performance is particularly important. Traditional testing methods usually require manual operation, which is cumbersome and time-consuming, increases production costs, and makes it difficult to achieve rapid testing. In addition, many traditional testing methods require destructive or invasive operations on samples, such as cutting, grinding, or dissolving, which wastes precious material resources and may cause irreversible effects on the overall performance of the bandage. Manual testing is subjective to the test results and depends on the experience and skill level of the tester, resulting in inconsistency and inaccuracy in quality control.

[0003] The existing Chinese patent application with publication number CN114112985A discloses a near-infrared spectrometer and a near-infrared online detection method, including a light source module, a spectrum collection module and a spectrum analysis module; the light source module uses near-infrared light to irradiate the sample to be tested; the spectrum collection module transmits the near-infrared light reflected by the sample to be tested to the spectrum analysis module, and the spectrum analysis module is used to analyze and process the near-infrared spectrum obtained by the infrared detection unit to complete the detection of the category or component of the sample to be tested, thereby improving the detection efficiency of the near-infrared spectrometer.

[0004] However, the existing technologies for testing samples require prior knowledge support. For example, the sample classification task can only divide the samples to be tested into pre-set categories. The component analysis task can only determine the proportion of each component under the premise of knowing the components contained in the samples to be tested. Spectral detection has limitations, and there is no effective database collection and update system to store the test results. All analysis and processing steps need to be repeated for test samples with the same components, which reduces the detection efficiency.

[0005] Therefore, it is particularly important to develop an efficient, accurate and non-destructive spectral detection system and method for quality inspection of low-temperature adhesive self-adhesive elastic bandages. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention proposes a spectral detection system and method for protective tape to perform efficient and accurate sample quality inspection and sample analysis.

[0007] The technical solution to achieve the purpose of the present invention is:

[0008] The spectral detection method applied to the protective tape includes the following specific steps:

[0009] Obtain a benchmark sample set and count the components involved in all benchmark samples therein, and re-sort them according to the dictionary order to construct a total component table;

[0010] Obtain all components and component ratios of a single benchmark sample, query the total component table to build the corresponding component ratio table, select the comparison sample based on the adaptive classification strategy and assign the benchmark sample a sub-category label;

[0011] Select subclass labels, select a spectrometry method based on the trait-oriented spectrometry strategy, construct a class spectrum set corresponding to the subclass labels, generate class standard spectra through the deviation correction method, and store subclass labels, class average component ratio tables, class standard spectra, class component ratio sets and class spectrum sets to construct a standard comparison database;

[0012] Obtain samples to be inspected and tasks to be done, identify the tasks to be done as sample quality inspection tasks or sample analysis tasks, and process the tasks to be done through a finished product quality inspection algorithm or a component analysis algorithm, respectively.

[0013] Furthermore, constructing the composition ratio table includes the following specific steps:

[0014] Get the Benchmark samples of Ingredients , synchronous acquisition The proportion of ingredients corresponding to , is the benchmark sample number, For the Benchmark samples The total number of ingredients;

[0015] generate All zero component ratio table of dimension , It is the dimension of the total component table;

[0016] Look up the total ingredient list to determine the Ingredients In the total composition table, the The proportion of ingredients Fill in the all-zero component ratio table The corresponding position until All component proportions are filled in the all-zero component proportion table , then the composition ratio table Build completed.

[0017] Furthermore, selecting comparison samples based on the adaptive classification strategy and assigning small class labels to the benchmark samples includes the following specific steps:

[0018] Get the Benchmark samples Ingredient ratio table and the existing Small category labels ,in, For the Small category labels;

[0019] Get the Small category labels Corresponding comparison component ratio table , Number the sub-category labels, perform unified position judgment, and determine the ingredient ratio table And comparison ingredient ratio table Whether the non-zero value positions in the are uniform;

[0020] If the non-zero value positions are not uniform, perform label exhaustion judgment to determine the sub-category label number Is it equal to the total number of existing sub-category labels? ;

[0021] If the sub-category label number Less than the total number of existing sub-category labels , then further select Small category labels Corresponding comparison component ratio table Continue with the unified judgment of position;

[0022] If the sub-category label number Equal to the total number of sub-category labels currently available , then the benchmark sample is assigned Small category label , and the benchmark sample Ingredient ratio table As a subcategory label Corresponding comparison component ratio table ;

[0023] If the non-zero value positions are uniform, obtain the composition ratio table And comparison ingredient ratio table The upper limit error is determined by performing proportional error judgment to determine whether the upper limit error is less than or equal to the error threshold;

[0024] If the upper limit error is less than or equal to the error threshold, the small class label Assigning a benchmark sample Otherwise, the label exhaustion judgment will continue, based on the small category label number Is it equal to the total number of existing sub-category labels? Decision to proceed with unified position judgment or assign benchmark sample Small category label , and the benchmark sample Ingredient ratio table As a subcategory label Corresponding comparison component ratio table ;

[0025] Execute sample exhaustion judgment and determine the benchmark sample number Is it less than the total number of benchmark samples in the benchmark sample set? , if less than Benchmark samples Continue to execute the adaptive classification strategy. If it is equal, stop executing the adaptive classification strategy.

[0026] Furthermore, the spectrum generation method is selected based on the trait-oriented spectrum generation strategy, and the class spectrum set corresponding to the subclass label is constructed, which includes the following specific steps:

[0027] Filter all small class labels from the benchmark sample set The benchmark sample set is constructed based on , class sample set The total number of class samples is ;

[0028] Judgement Class samples Whether the physical property of the powder is powder or not is used to select the spectrum preparation method, and the selected spectrum preparation method is used to obtain the class sample Infrared spectra of three different detection points, recorded as the first infrared spectrum , Second infrared spectrum and the third infrared spectrum ;

[0029] Calculate the first infrared spectrum , Second infrared spectrum and the third infrared spectrum Cosine similarity between pairs to obtain class samples The average similarity ;

[0030] If the average similarity Greater than or equal to the similarity threshold , then the first infrared spectrum , Second infrared spectrum and the third infrared spectrum The mean of the class sample Infrared spectrum ;

[0031] Get subcategory labels Corresponding Infrared spectra of samples of each class, construct sub-class labels The corresponding class spectrum set ,in, Small category label The corresponding Infrared spectra of samples of this type, .

[0032] Furthermore, generating a quasi-standard spectrum by the deviation correction method includes the following specific steps:

[0033] Get subcategory labels The corresponding class spectrum set , extract class sample set The component ratio table corresponding to all class samples in the class is used to construct the class component ratio set ;

[0034] Calculate the class component ratio set The average of all component ratio tables in the class is used to obtain the class average component ratio table ;

[0035] Calculate the class component ratio set The total difference of proportions corresponding to each component proportion table in the formula is used to construct the total difference set of proportions , the total difference of the proportion Take the negative and generate the class weight set through the Softmax function , and based on the class weight set Pair Spectral Set Perform weighted summation to generate a quasi-standard spectrum .

[0036] Furthermore, the finished product quality inspection algorithm is used to process the sample quality inspection task, including the following specific steps:

[0037] Get samples to be tested , assuming that the sample to be tested is Total includes Parts , and it is known Parts Corresponding to the sub-category labels already in the standard comparison database , is the total number of parts, where For the Parts The corresponding standard compares the existing small class labels in the database. ;

[0038] Lock part , based on the trait-directed profiling strategy to obtain the Infrared spectrum of components , is the part number;

[0039] Query the standard comparison database to obtain subcategory labels Corresponding standard spectrum ;

[0040] Execute similarity judgment and judge Infrared spectrum of components With class standard spectrum Cosine similarity of Is it greater than or equal to the similarity threshold? ;

[0041] If otherwise, determine the sample to be tested If yes, then execute the component exhaustive judgment to determine the component number Whether the total number of parts has been reached , if it has been reached, the sample to be tested is determined If qualified, otherwise, lock the part , get the Infrared spectrum of components To continue to perform similarity judgment.

[0042] Furthermore, the component analysis algorithm is used to process the sample analysis task, including the following specific steps:

[0043] Get samples to be tested , lock the part Based on the trait-directed profiling strategy, the Infrared spectrum of components ,

[0044] Perform adaptive acquisition, assuming that there are Small category labels , based on the existing Small category labels Whether to select to obtain a class standard spectrum or compare infrared spectrum for the missing state;

[0045] Perform similarity comparison and The standard spectra or comparison infrared spectra of each type are sequentially compared with the Infrared spectrum of components Calculate the cosine similarity to obtain the maximum cosine similarity and determine whether it is greater than or equal to the similarity threshold ;

[0046] If it is greater than or equal to the similarity threshold, it is assumed to correspond to the small class label , perform label status analysis and determine the sub-category label Is it missing? If not, extract the class average component ratio table. As a table of ingredient ratios Otherwise, the composition ratio table is obtained through spectral identification method , the ingredient ratio table and Infrared spectrum of components Insert sub-category labels separately The corresponding class component ratio set and class spectrum set;

[0047] Further execution of the quantity reaching judgment is performed to determine whether the total number of class samples reaches the quantity threshold. If it has reached the threshold, the class average component ratio table is generated through the deviation correction method. And standard spectrum , remove the subcategory label missing state;

[0048] If it is less than the similarity threshold, the component ratio table is obtained through spectral identification method , giving the part Small category label , the ingredient ratio table and Infrared spectrum of components As a comparison component ratio table and compare infrared spectra , creating new space in the standard comparison database and comparing the composition ratio table and compare infrared spectra Insert sub-category labels separately Set sub-class labels in the corresponding class component ratio set and class spectrum set is missing state;

[0049] Execute component exhaustive judgment to determine the component number Whether the total number of parts has been reached , if it has been reached, then the sample to be tested The analysis is completed, otherwise the first part , continue to perform adaptive acquisition and similarity comparison, and independently decide whether to perform label status analysis and quantity compliance judgment until the sample to be tested Analysis complete.

[0050] Furthermore, the spectral identification method includes the following specific steps:

[0051] Determine part The major category is cloth or rubber, get the Infrared spectrum of components , preprocess the first Infrared spectrum of components Generate High-quality infrared spectra of components ;

[0052] The PCA algorithm is used to extract High-quality infrared spectra of components In Main features to form a main feature sequence;

[0053] Based on part The corresponding pre-trained LS-SVM model is selected for the major categories, and the main feature sequence is input into the LS-SVM model to obtain the first part The component sequence includes part In Ingredients, The components are The main features correspond one to one, so it satisfies , the kernel function of the LS-SVM model uses the RBF function, which includes the variance and penalty factor Two main parameters;

[0054] Based on part The major categories of selection correspond to the pre-trained partial least squares model to process the main feature sequence and determine The main features correspond to The ratio of the ingredients to obtain the part Ingredient ratio table .

[0055] Furthermore, the LS-SVM model and the partial least squares model are pre-trained for fabric and rubber, respectively. Taking fabric as an example, the pre-training of the LS-SVM model and the partial least squares model includes the following specific steps:

[0056] Obtain the sub-class labels and class standard spectra corresponding to all cloth raw materials from the standard comparison database to construct a cloth raw material dataset for training the LS-SVM model;

[0057] Divide the cloth raw material data set into training set and validation set according to a certain ratio, and set the variance of the RBF function , penalty factor ;

[0058] The loss function of the LS-SVM model was selected as the root mean square error, and the stochastic gradient descent method was used for pre-training. The training rounds and initial learning rate were set to 100 and 0.01 respectively until the loss function of the LS-SVM model converged. The hyperparameters of the LS-SVM model were fixed and tested using the validation set;

[0059] Obtain the class average component ratio table and class standard spectrum corresponding to all fabric finished parts from the standard comparison database, generate the corresponding main feature sequence by PCA processing the class standard spectrum, and construct a fabric finished parts data set based on the main feature sequence and class average component ratio table to train the partial least squares model;

[0060] The evaluation criteria of the partial least squares model are set, including the correlation coefficient, cross-check standard deviation and prediction standard deviation. Fit a partial least squares model and pass the remaining 20 Conduct inspection.

[0061] Based on the further improvement of the above method, the improved spectral identification method adds a GAN network to realize the feature alignment of fabrics and rubbers, and uses residual convolutional networks, dimension adaptation networks and asynchronous convolutional networks instead of PCA to extract features in infrared spectra from multiple angles in depth. At the same time, the component prediction network is used instead of the partial least squares model to obtain the first part Ingredient ratio table .

[0062] Preferably, the improved spectral identification method comprises the following specific steps:

[0063] Get the Infrared spectrum of components , preprocess the first Infrared spectrum of components To generate the High-quality infrared spectra of components ;

[0064] The first High-quality infrared spectra of components Input into the GAN network to eliminate the characteristic differences between cloth and rubber and generate a unified infrared spectrum ;

[0065] Deep extraction of unified infrared spectra via residual convolutional networks The main features in the image are obtained, and the dimensions of the main features are adjusted through the dimension adaptation network to generate dimensionally appropriate features;

[0066] The asynchronous convolutional network processes the dimension-appropriate features simultaneously through three different convolutional blocks, outputs three convolutional features from different angles, and connects them to obtain multimodal features;

[0067] The multimodal features are input into the component prediction network, and the dimension reducer is used to reduce the multimodal features to , is the dimension of the total component table, and is adaptively predicted by the Softmax regressor The composition ratio of each component, output composition ratio table .

[0068] Specifically, the residual convolution network is improved based on the ResNet18 network. The convolution kernel is replaced by The convolution kernel is , and 4 residual blocks are set. The number of channels corresponding to the 4 residual blocks are , , and , the three convolution blocks in the asynchronous convolutional network are the first convolution block, the second convolution block and the third convolution block. The first convolution block adopts The convolution kernel is convolved and pooled, and the second convolution block uses The convolution kernel is used for convolution, and the third convolution block first uses Convolution kernel is used for convolution, and then Convolution is performed with the convolution kernel.

[0069] Furthermore, the improved spectral identification method includes GAN network, residual convolution network, asynchronous convolution network and component prediction network. The data set used for pre-training consists of class standard spectra and class average component ratio tables in the standard comparison database. The cross entropy is selected as the loss function. The Adam optimization algorithm is used for training. The training rounds and initial learning rate are set to 500 and 0.001 respectively. The hyperparameters are fixed after the training is completed.

[0070] The beneficial effect of the improved spectral identification method is that the feature alignment of fabric and rubber is achieved through the GAN network without separate training. At the same time, the residual convolutional network, dimension adaptation network and asynchronous convolutional network are used to obtain deep features in the infrared spectrum from multiple angles, so that feature extraction is no longer limited to the upper limit of the number of PCA. At the same time, the prediction accuracy of the composition ratio table is greatly improved by replacing the partial least squares model with a composition prediction network with better regression performance.

[0071] A spectrum detection system applied to protective tape, used to implement the above-mentioned spectrum detection method applied to protective tape, including a classification module, a collection module, a scanning module, a storage module and a task module;

[0072] The classification module generates a component ratio table of the benchmark sample, assigns a sub-class label to the benchmark sample based on an adaptive classification strategy, and generates a spectrum-deficient data packet;

[0073] The collection module obtains the missing spectrum data packet, initiates a scan request to obtain the infrared spectrum, and generates a full data packet to deliver to the storage module;

[0074] The scanning module responds to the scanning request, identifies the requester and the number, obtains the infrared spectrum based on the trait-oriented spectrum strategy and feeds back to the requester, the number includes the reference sample number and the component number;

[0075] The storage module receives the full data packet and stores it in the standard comparison database by classification, and controls the setting and cancellation of the missing status of the sub-category label;

[0076] The task module obtains and identifies pending tasks, and independently selects a finished product quality inspection algorithm or a component analysis algorithm to complete the pending tasks.

[0077] Further, the classification module includes a tabulation unit, a first main storage unit, a classification unit and a communication unit;

[0078] The first main memory unit stores the total component table for a long term;

[0079] The tabulation unit obtains the known component ratios of the total component table and the reference sample, generates an all-zero code and fills the known component ratios into positions corresponding to the all-zero code to generate a component ratio table of the reference sample;

[0080] The classification unit obtains the component ratio table of the reference sample, assigns a subclass label based on an adaptive classification strategy, and packages the subclass label, the reference sample number and the component ratio table to generate a spectrum-missing data packet;

[0081] The communication unit receives and temporarily stores the known component ratio of the reference sample, and transmits the spectrum-deficient data packet to the transfer module.

[0082] Further, the transfer module includes a supplementary unit and a transmission unit;

[0083] The supplement unit receives the missing spectrum data packet, extracts the reference sample number, generates a scanning request and sends it to the scanning module to obtain the infrared spectrum of the reference sample, and fills in the missing spectrum data packet to generate a full data packet;

[0084] The transmission unit forwards the full data packet to the storage module.

[0085] Further, the storage module includes a classification unit and a second main storage unit;

[0086] The classification unit receives the full data packet and extracts the sub-class label therefrom, determines whether it is a sub-class label already in the standard comparison database, if it is already in the standard comparison database, extracts the component ratio table and infrared spectrum in the full data packet and directly stores them in the standard comparison database, if not, marks and generates the comparison component ratio table and comparison infrared spectrum corresponding to the sub-class label, stores them in the second main storage unit and sets the sub-class label to a missing state;

[0087] The second main memory unit stores the standard comparison database, and determines in real time whether the total number of class samples of the small class labels in the missing state reaches the quantity threshold. If reached, the class average component ratio table and class standard spectrum are generated through the deviation correction method, and the missing state of the small class labels is cancelled.

[0088] Furthermore, the task module includes a judgment unit, a quality inspection unit, and an analysis unit;

[0089] The judgment unit receives and identifies the task to be inspected to select and activate the quality inspection unit or the analysis unit;

[0090] The quality inspection unit executes the finished product quality inspection algorithm to identify the total number of components and sub-category labels of the samples to be inspected. part , based on part number Generate a scan request to obtain the Infrared spectrum of components And determine whether the sample to be tested is qualified;

[0091] The analysis unit executes the component analysis algorithm to identify the total number of components of the sample to be tested. part , based on part number Generate a scan request to obtain the Infrared spectrum of components By comparing the existing class standard spectrum or infrared spectrum, decide whether to directly obtain the sub-class label and component ratio table from the standard request database If not, obtain the composition ratio table through spectral identification method , and give the part New subclass label, combined with new subclass label and part number , ingredient ratio table and Infrared spectrum of components Generates full data packets and forwards them to the storage module.

[0092] Compared with the prior art, the present invention has the following significant advantages:

[0093] 1. By constructing a component ratio table to quantify the benchmark sample, an adaptive classification strategy is designed to automatically assign sub-category labels to the benchmark sample. At the same time, based on the physical properties of the benchmark sample, the property-oriented spectrometry method is used to obtain the precise infrared spectrum of the benchmark sample, and the class-averaged component ratio table and class standard spectrum are generated through the deviation correction method to uniformly represent the sub-category label, thereby reducing the complexity of subsequent sample quality inspection or sample analysis tasks and automatically constructing a complete standard comparison database;

[0094] 2. Divide the samples to be inspected into components, design a quality inspection algorithm, and compare the infrared spectrum of a single component with the average standard spectrum of the corresponding sub-category label in the standard comparison database, so that efficient and accurate sample quality inspection can be carried out in batches;

[0095] 3. Divide the samples to be tested into components, design a component analysis algorithm, and intelligently determine whether a single component of the sample to be tested corresponds to an existing subcategory label in the standard database. If it cannot correspond to the existing subcategory label, design two spectral analysis methods to accurately predict the component ratio table of a single component, and automatically update the standard comparison database, thereby realizing automated component analysis of unknown samples to be tested. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] Figure 1 A flow chart of a spectral detection method applied to protective tape;

[0097] Figure 2 A flow chart of assigning sub-category labels to benchmark samples based on the adaptive classification strategy in the present invention;

[0098] Figure 3 A flow chart of obtaining a class spectrum set corresponding to a subclass label based on a trait-oriented spectrum generation strategy in the present invention;

[0099] Figure 4 This is a flow chart of the sample quality inspection task processing by the finished product quality inspection algorithm in the present invention;

[0100] Figure 5 This is a flow chart of the sample analysis task processed by the component analysis algorithm in the present invention. DETAILED DESCRIPTION

[0101] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0102] Example 1

[0103] like Figure 1 As shown, a specific embodiment of the present invention discloses a spectrum detection method applied to a protective tape, comprising the following specific steps:

[0104] Obtain a benchmark sample set, count the components involved in all benchmark samples in the benchmark sample set, reorder the components according to the dictionary order, and construct a total component table;

[0105] Obtain all components and component ratios of a single benchmark sample, query the total component table to build a component ratio table of a single benchmark sample, select comparison samples based on an adaptive classification strategy and assign sub-class labels to the benchmark samples;

[0106] Select subclass labels, select a spectrogram method based on the trait-oriented spectrogram strategy to construct a class spectrum set corresponding to the subclass labels, generate class standard spectra through the deviation correction method, and store subclass labels, class average component ratio tables, class standard spectra, class component ratio sets and class spectrum sets to construct a standard comparison database;

[0107] Obtain samples to be inspected and tasks to be done, identify the tasks to be done, and select a finished product quality inspection algorithm or a component analysis algorithm to process the tasks to be done.

[0108] Specifically, the benchmark sample set is composed of a large number of benchmark samples. The components and component proportions of the benchmark samples are known. The benchmark samples are divided into two categories: cloth and glue. From another perspective, the benchmark samples include finished parts and raw materials used to manufacture finished parts. Finished parts actually refer to the cloth and single-sided low-temperature glue that make up the low-temperature glue self-adhesive elastic bandage. There are many types of raw materials. Taking cloth as an example, they include cotton, polyester fiber and polyester fiber.

[0109] Furthermore, constructing a composition ratio table of a single benchmark sample includes the following specific steps:

[0110] Get the Benchmark samples of Ingredients , synchronous acquisition The proportion of ingredients corresponding to , is the benchmark sample number, , is the total number of benchmark samples, For the Benchmark samples The total number of components, and ;

[0111] Generate an all-zero component ratio table with the same dimensions as the total component table , assuming that components, then the zero component ratio table and the total component table are Dimensional, all zero component ratio table Middle Value , ;

[0112] Query the total ingredient table and get the Ingredients Position in the total composition table, assuming the Ingredients In the total ingredient list position, then the all-zero component ratio table No. The 0 at position 0 is updated to The proportion of ingredients ,Right now ;

[0113] Until The proportions of each component are filled in the zero component proportion table , then Benchmark samples Ingredient ratio table Build completed.

[0114] like Figure 2 As shown, further, selecting comparison samples based on the adaptive classification strategy and assigning small class labels to the benchmark samples includes the following specific steps:

[0115] Get the first Benchmark samples Corresponding ingredient ratio table , and get the current existing Small category labels , For the small class labels; the existing small class labels are the previous Small class labels assigned by benchmark samples;

[0116] Select Small category labels Corresponding comparison component ratio table , Number the subclass labels;

[0117] Execute the unified judgment of the position and judge the composition ratio table And comparison ingredient ratio table Whether the non-zero value positions in the are uniform;

[0118] If the non-zero value positions are not uniform, the composition ratio table And comparison ingredient ratio table The composition of the benchmark samples is inconsistent. Not a subcategory label , perform label exhaustion judgment and determine the sub-category label number Is it equal to the total number of existing sub-category labels? ;

[0119] If the sub-category label number Less than the total number of sub-category labels currently available , then further derive the Small category labels Corresponding comparison component ratio table Continue with the unified judgment of position;

[0120] If the sub-category label number Equal to the total number of sub-category labels currently available , existing sub-category labels None of them can correspond to the benchmark samples , given the benchmark sample New subcategory tags , and the benchmark sample Ingredient ratio table As a subcategory label Corresponding comparison component ratio table ;

[0121] If the non-zero value positions are uniform, the composition ratio table And comparison ingredient ratio table The components in the table are consistent, the upper limit error is obtained, and the proportion error judgment is performed to determine whether the upper limit error is less than or equal to the error threshold. Specifically, the upper limit error refers to the component proportion table. Subtract comparison ingredient ratio table And take the absolute value to get the difference component ratio table The maximum value in ;

[0122] If the upper limit error is less than or equal to the error threshold, the small class label Assigning a benchmark sample ;

[0123] If the upper limit error is greater than the error threshold, the component ratio table And comparison ingredient ratio table The ingredients in the samples are the same but the proportions vary greatly. Not a subcategory label , perform label exhaustive judgment, based on the small class label number Is it equal to the total number of sub-category labels currently available? Decision to proceed with unified position judgment or assign benchmark sample Small category label , and the ingredient ratio table As a comparison table of ingredients ;

[0124] Execute sample exhaustion judgment and determine the benchmark sample number Is it less than the total number of benchmark samples in the benchmark sample set? If it is less than, then further Benchmark samples Execute the adaptive classification strategy. If it is equal to , the adaptive classification strategy stops executing.

[0125] like Figure 3 As shown, further, selecting a spectrum making method based on a trait-oriented spectrum making strategy and constructing a class spectrum set corresponding to a small class label includes the following specific steps:

[0126] For Small category labels , filter out all small class labels from the benchmark sample set The benchmark sample set is constructed based on ,in, The sample set of class Middle Class samples, The sample set of class The total number of class samples;

[0127] Judgment class sample set Middle Class samples Whether the physical property of the sample is powder, if it is powder, the potassium bromide tablet method is selected as the spectrum preparation method, if it is not powder, the ATR-FTIR method is selected as the spectrum preparation method. Both the potassium bromide tablet method and the ATR-FTIR method use infrared light in the 4000-400 band for scanning;

[0128] Randomly select class samples The three different positions of the detector are taken as detection points, and the infrared spectra of the three detection points are obtained by the selected spectrum preparation method, which is recorded as the first infrared spectrum , Second infrared spectrum and the third infrared spectrum ;

[0129] Calculate the first infrared spectrum , Second infrared spectrum and the third infrared spectrum The cosine similarity between the two is denoted as , and ,in, The first infrared spectrum With the second infrared spectrum The cosine similarity between ;

[0130] Get class samples The average similarity :

[0131] ;

[0132] If the average similarity Greater than or equal to the similarity threshold , then the first infrared spectrum , Second infrared spectrum and the third infrared spectrum The mean of the class sample Infrared spectrum ;

[0133] Get subcategory labels Corresponding Infrared spectra of samples of each class, construct sub-class labels The corresponding class spectrum set .

[0134] Furthermore, generating a quasi-standard spectrum by the deviation correction method includes the following specific steps:

[0135] For small category labels , get the class spectrum set , and synchronously obtain the class sample set The component ratio table corresponding to all class samples in the class is used to construct the class component ratio set ,in, The sample set of class Middle The component ratio table corresponding to each class sample, , Ingredient ratio table Middle The proportion of ingredients;

[0136] Get the average component ratio table , average composition ratio table is the class component ratio set The mean of all composition ratio tables in , where The average composition ratio table Middle The average proportion of each class, and ;

[0137] For ingredient ratio table , calculate the composition ratio table The corresponding total difference in proportion , the total difference in proportion The actual composition ratio table Table of proportions relative to the class average The total degree of deviation;

[0138] Statistical component ratio set The total difference of proportions corresponding to all component proportion tables in the constructs the total difference set of proportions , the total difference of the proportion Take the negative and generate the class weight set through the Softmax function , and satisfies ,in, and The composition ratio table is The corresponding total difference of proportions and class weights. It is worth noting that the total difference of proportions is first calculated before entering the Softmax function The reason for taking a negative value is that when the Softmax function normalizes a set of data into a set of weights, the larger the data, the larger the corresponding weight. However, the present invention hopes to use the proportional total difference set The total difference of each proportion in the table represents the confidence of the corresponding benchmark sample. The larger the total difference of the proportion, the closer the component proportion table corresponding to the benchmark sample is to the class average component proportion table. The larger the deviation, the lower the confidence of the benchmark sample, and the smaller the corresponding weight assigned. Therefore, the total difference of the proportion is reduced by taking the negative operation. The total difference of the largest proportion becomes the smallest;

[0139] Based on the class weight set , using weighted summation to process the class spectrum set , generating a standard-like spectrum .

[0140] Specifically, the tasks to be done include sample quality inspection tasks and sample analysis tasks;

[0141] The sample quality inspection task means that if the sample to be inspected is known to be a finished product or raw material, and the corresponding sub-category labels of the finished product components or raw materials that make up the finished product are known at the same time, the sample to be inspected is identified through technical means to determine whether it is qualified. For example, the finished product or raw material is a low-temperature adhesive self-adhesive elastic bandage or cotton, and the finished product components that make up the low-temperature adhesive self-adhesive elastic bandage include known types of cloth and low-temperature glue. The sample quality inspection needs to identify whether the components of the cloth, low-temperature glue or cotton are qualified.

[0142] The sample analysis task means that it is known that the sample to be tested is a finished product or raw material, but the subcategory labels of the finished product components or raw materials that make up the finished product are not known in advance. It is necessary to use technical means to measure and analyze to obtain the component ratio table of the sample to be tested in order to determine whether the sample to be tested is assigned a subcategory label that already exists in the standard comparison database.

[0143] like Figure 4 As shown, further, the finished product quality inspection algorithm is used to process the sample quality inspection task, including the following specific steps:

[0144] Get samples to be tested Since the sample to be tested may be a finished product or raw material, it is assumed that the sample to be tested Total includes Parts ,like , the sample to be tested is a raw material or a single finished product component. , then the sample to be tested is The finished product consists of finished parts. Parts They correspond to the sub-category labels already in the standard comparison database, denoted as , is the total number of parts, For the Parts The corresponding standard compares the existing small class labels in the database. Take low temperature adhesive self-adhesive elastic bandage as an example. , the first component For fabric, corresponding to the sub-category label , the second component Low temperature glue for covering one side of the fabric, corresponding to the small category label ;

[0145] Lock part , based on the trait-directed profiling strategy, select the profiling method to obtain the first part No. Infrared spectrum of components , is the part number;

[0146] Query the standard comparison database to obtain subcategory labels Corresponding standard spectrum ;

[0147] Execute similarity judgment and calculate the Infrared spectrum of components With class standard spectrum Cosine similarity of , determine the cosine similarity Is it greater than or equal to the similarity threshold? ;

[0148] If otherwise, determine the sample to be tested If yes, then execute the component exhaustive judgment to determine the component number Whether the total number of parts has been reached , if it has been reached, the sample to be tested is determined If qualified, otherwise, lock the part , based on the trait-directed profiling strategy to obtain the Infrared spectroscopy , based on small class labels Get a standard spectrum , continue to perform similarity judgment and component exhaustion judgment.

[0149] like Figure 5 As shown, further, the component analysis algorithm is used to process the sample analysis task, including the following specific steps:

[0150] Get samples to be tested , lock the part Based on the trait-directed profiling strategy, the Infrared spectrum of components ;

[0151] Perform adaptive acquisition to lock in the standard comparison database Small category labels , is the total number of existing subclass labels in the standard comparison database, and determines whether the existing subclass labels are missing. If they are missing, obtain the comparison infrared spectrum corresponding to the subclass label. Otherwise, obtain the class standard spectrum corresponding to the subclass label. The missing state means that the class average component ratio table and class standard spectrum cannot be calculated temporarily due to insufficient sample quantity of the subclass label.

[0152] Perform similarity comparison and convert the total number to The class standard spectrum or comparison infrared spectrum is sequentially compared with the first Infrared spectrum of components Calculate the cosine similarity to obtain the maximum cosine similarity and determine whether the maximum cosine similarity is greater than or equal to the similarity threshold ;

[0153] If it is greater than or equal to the similarity threshold, it means that part Corresponding to the existing sub-category labels in the standard comparison database, it is assumed that the corresponding sub-category labels ;

[0154] Further perform label status analysis to determine sub-category labels Is it missing? If not, extract the subcategory label. Corresponding class average component ratio table , as the first part Ingredient ratio table Otherwise, the composition ratio table is obtained by spectral identification method. , the ingredient ratio table and Infrared spectrum of components Insert sub-category labels separately The corresponding class component ratio set and class spectrum set;

[0155] Further execute the quantity reaching standard judgment and determine the sub-category label Whether the total number of class samples reaches the quantity threshold, if it has reached it, the class average component ratio table is generated by the deviation correction method And standard spectrum , remove the subcategory label missing state;

[0156] If it is less than the similarity threshold, it means part Unable to match the sub-category labels in the standard comparison database, the first part Ingredient ratio table , giving the part New subcategory tags , will part Ingredient ratio table and Infrared spectrum of components As a comparison component ratio table and compare infrared spectra , in the standard comparison database, a small class label is stored , average composition ratio table , Standard spectrum , a new space of component ratio sets and spectrum sets, comparing the component ratio table and compare infrared spectra Insert sub-category labels separately The corresponding class component ratio set and class spectrum set, due to the class average component ratio table And standard spectrum Temporarily missing, set subcategory label is missing state;

[0157] Execute component exhaustive judgment to determine the component number Whether the total number of parts has been reached , if it has been reached, then the sample to be tested The analysis is completed, otherwise, the part , continue to perform adaptive acquisition, similarity comparison and component exhaustion judgment, and independently decide whether to perform label status analysis and quantity compliance judgment.

[0158] Furthermore, the spectral identification method includes the following specific steps:

[0159] Determine part The main categories, taking low temperature adhesive self-adhesive elastic bandage as an example, part The main categories are cloth or rubber. Before the component analysis, it is temporarily impossible to know the first part The specific proportion of ingredients cannot be assigned a sub-category label;

[0160] Get the Infrared spectrum of components , preprocess the first Infrared spectrum of components To eliminate interference factors such as noise and baseline drift, generate the High-quality infrared spectra of components ;

[0161] The PCA algorithm is used to extract High-quality infrared spectra of components The main feature sequence in is given by The main feature is essentially the Infrared spectrum of components The vibration absorption band caused by the specific functional group in the part ingredients;

[0162] Based on part The corresponding pre-trained LS-SVM model is selected for the major categories, and the main feature sequence is input into the LS-SVM model to obtain the first part The component sequence includes part Extracted by PCA algorithm The main features correspond to components, and , the kernel function of the LS-SVM model uses the RBF function, which includes the variance and penalty factor Two pre-set main parameters;

[0163] Based on part The major categories of selection correspond to the pre-trained partial least squares model to process the main feature sequence, and determine the The main features correspond to The composition ratio of each component is compared with the total composition table, and a zero composition ratio table of the same dimension is generated and the composition ratio is filled in the corresponding position to obtain the first part Ingredient ratio table .

[0164] Furthermore, there are two LS-SVM models and two partial least squares models, each applicable to fabric and rubber, with the same structure, which need to be pre-trained in advance. Taking fabric as an example, the pre-training of the LS-SVM model and partial least squares model applicable to fabric includes the following specific steps:

[0165] Since the LS-SVM model analyzes the main features and determines the components corresponding to the main features, and the infrared spectrum of the raw materials composed of a single component obtained by the property-oriented spectrum strategy contains only one main feature, which happens to achieve a one-to-one correspondence between the main feature and the component. Therefore, the sub-class labels and class standard spectra corresponding to all cloth raw materials are obtained from the standard comparison database to construct a cloth raw material data set, where the class standard spectrum is the input and the class label is the output;

[0166] The cloth raw material data set is divided into a training set and a validation set according to a certain ratio, and the main parameters of the RBF function are set, among which the variance , penalty factor ;

[0167] The loss function of the LS-SVM model is selected as the root mean square error to represent the gap between the predicted class label and the actual class label. The stochastic gradient descent method is used for pre-training, and the training rounds and initial learning rate are set to 100 and 0.01 respectively;

[0168] Until the loss function of the LS-SVM model converges, the hyperparameters of the LS-SVM model are fixed and tested using the validation set. When the average root mean square error of the validation set input into the LS-SVM model is less than or equal to the loss threshold, the LS-SVM model pre-training is completed;

[0169] Since the role of the partial least squares model is to determine the component ratio of the sample, it is necessary to select the class-average component ratio table and class-standard spectrum corresponding to all cloth finished parts from the standard comparison database, and generate the corresponding main feature sequence by PCA processing the class-standard spectrum. The main feature sequence is used as input and the class-average component ratio table is used as output to construct the cloth finished parts data set;

[0170] The essence of the partial least squares model is the partial least squares method. The evaluation criteria of the partial least squares method are set, including the correlation coefficient, the cross-check standard deviation and the prediction standard deviation. The 80% of the cloth finished parts data set is used. Fit the hyperparameters of the partial least squares model and pass the remaining 20 A test is performed. If the evaluation criteria are all less than or equal to the evaluation criterion threshold, the partial least squares model pre-training is completed.

[0171] Example 2

[0172] An improvement is made based on the spectral detection method for protective tape disclosed in Example 1. To reduce redundant description, this example only introduces the improved part, and the same part as in Example 1 is not introduced.

[0173] The spectral identification method proposed in Example 1 extracts the first High-quality infrared spectra of components of The number of main features extracted by the PCA algorithm There is a limit on the number of , part Total number of components Greater than the maximum quantity , PCA can only extract the part The greatest contribution The corresponding components The main features are: , and the part Remaining The components are treated as impurities and removed by PCA, so that the component ratio table obtained by the spectral identification method in Example 1 is The greatest contribution The proportion of the ingredients is too high and will be removed The proportion of the components was divided among the In addition, the LS-SVM model and the partial least squares model in Example 1 need to be trained for fabric and rubber respectively. Considering the excellent feature extraction performance of the convolutional neural network and the feature alignment function of the GAN network, a spectral identification method based on a neural network improvement is proposed to obtain a more accurate component ratio table. .

[0174] Preferably, the spectrum identification method based on neural network improvement includes the following specific steps:

[0175] Get the Infrared spectrum of components , preprocess the first Infrared spectrum of components To eliminate interference factors such as noise and baseline drift, generate the High-quality infrared spectra of components ;

[0176] The first High-quality infrared spectra of components Input the GAN network to eliminate the feature differences between cloth and rubber, achieve feature alignment, and generate a unified infrared spectrum ;

[0177] Deep extraction of unified infrared spectra via residual convolutional networks The main features in the image are then adjusted through a dimension adaptation network consisting of a linear layer, a batch normalization layer, and a ReLu activation function to generate dimensionally appropriate features for further input into the asynchronous convolutional network.

[0178] The asynchronous convolutional network processes the dimension-appropriate features simultaneously through three different convolutional blocks, further learns the dimension-appropriate features from different angles, and connects the convolutional features output by the three convolutional blocks to obtain multimodal features of the specified dimension.

[0179] The multimodal features are input into the component prediction network. First, a dimensionality reducer consisting of a combination of three linear layers and a batch normalization layer is used to reduce the multimodal features to , is the dimension of the total ingredient table, and then the Softmax regressor composed of a linear layer and a Softmax activation function is used to adaptively predict the proportion of each component in the total ingredient table, and output the Component ratio table .

[0180] Specifically, the residual convolutional network can effectively solve the problems of gradient diffusion and gradient explosion. It can be improved based on the existing ResNet18 network and The convolution kernel is replaced by The convolution kernel is used for one-dimensional convolution, and 4 residual blocks are set. The number of channels corresponding to the 4 residual blocks are , , and , when the unified infrared spectrum After inputting the residual convolutional network, the infrared spectrum is first unified through one-dimensional convolution learning The correlation between adjacent values ​​in the image is then analyzed through four residual blocks for deep feature mining, and the main features are obtained after pooling.

[0181] Specifically, the three convolution blocks in the asynchronous convolutional network are defined as the first convolution block, the second convolution block and the third convolution block. The first convolution block adopts The convolution kernel is convolved and pooled, and the second convolution block uses The convolution kernel is used for convolution, and the third convolution block first uses Convolution kernel is used for convolution, and then Convolution is performed with the convolution kernel.

[0182] Preferably, due to the introduction of the GAN network, the pre-training data sets of the GAN network, residual convolutional network, asynchronous convolutional network and component prediction network included in the spectral identification method proposed in this embodiment can be directly extracted from the standard comparison database, the pre-training data set includes class standard spectra and class average component ratio tables, the loss function selects cross entropy, and the Adam optimization algorithm is used for training. The training rounds and the initial learning rate are set to 500 and 0.001, respectively, and the hyperparameters are fixed after the training is completed.

[0183] The beneficial effect of the improved spectral identification method of this embodiment is that the feature alignment of fabric and rubber is achieved through the GAN network without separate training, and the residual convolution network, dimension adaptation network and asynchronous convolution network are used to obtain deep features in the infrared spectrum from multiple angles, so that feature extraction is no longer limited to the number upper limit of PCA. At the same time, a component prediction network with better regression performance was used to replace the partial least squares model, which greatly improved the prediction accuracy of the component ratio table.

[0184] Example 3

[0185] The present invention also discloses a spectrum detection system applied to a protective tape, which is used to implement the spectrum detection method applied to the protective tape proposed in Example 1 and Example 2, and includes a classification module, a collection module, a scanning module, a storage module and a task module;

[0186] The classification module generates a component ratio table of the benchmark sample, assigns a sub-class label to the benchmark sample based on an adaptive classification strategy, and generates a spectrum-deficient data packet;

[0187] The collection module obtains the missing spectrum data packet, initiates a scan request to obtain the infrared spectrum, and generates a full data packet to deliver to the storage module;

[0188] The scanning module responds to the scanning request, identifies the requester and the number, obtains the infrared spectrum based on the trait-oriented spectrum strategy and feeds back to the requester, the number includes the reference sample number and the component number;

[0189] The storage module receives the full data packet and stores it in the standard comparison database by classification, and controls the setting and cancellation of the missing status of the sub-category label;

[0190] The task module obtains and identifies pending tasks, and independently selects a finished product quality inspection algorithm or a component analysis algorithm to complete the pending tasks.

[0191] Further, the classification module includes a tabulation unit, a first main storage unit, a classification unit and a communication unit;

[0192] The first main memory unit stores the total component table for a long term;

[0193] Tabulation unit obtains the total composition table and the Benchmark samples The known component ratio is used to generate an all-zero code and fill the known component ratio into the corresponding position of the all-zero code to generate a reference sample Ingredient ratio table ;

[0194] Taxonomic unit acquisition Benchmark samples Ingredient ratio table , assign small class labels based on adaptive classification strategy , pack small category tags , Benchmark sample number And ingredient ratio table Generate spectrum-deficient data packets;

[0195] The communication unit receives and temporarily stores the Benchmark samples The known component ratio is transmitted to the transfer module and then the missing spectrum data packet is cleared. Benchmark samples of known ingredients.

[0196] Further, the transfer module includes a supplementary unit and a transmission unit;

[0197] The supplementary unit receives the missing spectrum data packet and extracts the reference sample number , generate a scan request and send it to the scanning module to obtain the benchmark sample Infrared spectrum , fill in the missing spectrum data packets to generate full data packets and deliver them to the transmission unit;

[0198] The transmission unit receives the full data packet and forwards it to the storage module.

[0199] Further, the storage module includes a classification unit and a second main storage unit;

[0200] The classification unit receives the full data packet, identifies the sub-class label, and determines whether it is an existing sub-class label in the standard comparison database. If it is, the component ratio table and infrared spectrum in the full data packet are extracted and directly stored in the class component ratio set and class spectrum set corresponding to the sub-class label in the standard comparison database. Otherwise, the component ratio table and infrared spectrum corresponding to the sub-class label are marked as the comparison component ratio table and comparison infrared spectrum corresponding to the sub-class label, and a new space is opened in the main memory unit for storage, and the sub-class label is set to be in a missing state.

[0201] The second main memory unit stores the standard comparison database for a long time and monitors the total number of class samples of small class labels that are in a missing state in real time. If the number threshold is reached, the class average component ratio table and class standard spectrum are generated through the deviation correction method, and the missing state of the small class label is cancelled.

[0202] Furthermore, the task module includes a judgment unit, a quality inspection unit, and an analysis unit;

[0203] The judgment unit receives and identifies the task to be inspected, and selects to activate the quality inspection unit or the analysis unit;

[0204] The quality inspection unit executes the finished product quality inspection algorithm to identify the total number of components in the sample to be inspected and subcategory labels , for the part , determine the part number To generate a scan request, get the Infrared spectrum of components To determine whether the sample to be tested is qualified;

[0205] The analysis unit executes the component analysis algorithm to identify the total number of components in the sample to be tested , for the part , determine the part number To generate a scan request, get the Infrared spectrum of components , compare with the existing quasi-standard spectra or comparison infrared spectra in the standard comparison database, and determine whether to directly call the standard comparison database to obtain the first part Sub-category labels and ingredient ratio tables If not, obtain the composition ratio table through spectral identification method , and give the part New subclass label, combined with new subclass label and part number , ingredient ratio table and Infrared spectrum of components Generates full data packets and forwards them to the storage module.

[0206] The present invention discloses a spectral detection system and method for protective tape, designs an adaptive classification strategy to automatically assign sub-category labels to reference samples, adopts a trait-oriented spectrometry method to obtain accurate infrared spectra of reference samples, generates a class-average component ratio table and a class-standard spectrum through a deviation correction method, and preliminarily constructs a standard comparison database for tasks to be inspected, divides the samples to be inspected into components, adopts a finished product quality inspection algorithm to compare the infrared spectrum of a single component with the class-average standard spectrum of the sub-category label, and performs efficient and accurate sample quality inspection, adopts a component analysis algorithm to intelligently determine whether a single component corresponds to an existing sub-category label, and designs two spectral analysis methods for single components that cannot be matched to accurately predict the component ratio table of the single component, and automatically updates the standard comparison database, thereby realizing automated component analysis of unknown samples to be inspected.

[0207] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A spectral detection method for protective tape, characterized in that: The specific steps include: Count and re-order the components involved in all benchmark samples to construct a total component table; Obtain the composition and composition ratio of a single benchmark sample, query the total composition table to build a composition ratio table, and assign sub-category labels to the benchmark sample based on an adaptive classification strategy; Based on the property-oriented spectrum generation strategy, a class spectrum set corresponding to the subclass label is constructed, and a class standard spectrum is generated by the deviation correction method. The subclass label, class average component ratio table, class standard spectrum, class component ratio set and class spectrum set are stored accordingly to construct a standard comparison database. Obtain samples to be inspected and tasks to be done, identify the tasks to be done as sample quality inspection tasks or sample analysis tasks, and complete the quality inspection or component analysis of the samples to be inspected through the finished product quality inspection algorithm or component analysis algorithm respectively.

2. The spectrum detection method for protective tape according to claim 1, characterized in that: The method of assigning small class labels to benchmark samples based on the adaptive classification strategy includes the following specific steps: Get the benchmark sample Ingredient ratio table and existing Small category labels; Get subcategory labels Corresponding comparison component ratio table , unified judgment of execution position, judgment of component ratio table And comparison ingredient ratio table Whether the non-zero value positions in the are uniform; If not unified, execute label exhaustion judgment based on sub-category label number Is it equal to the total number of existing sub-category labels? Decide on the comparison ingredient ratio table Make a unified position decision or assign a benchmark sample Small category label ; If unified, obtain the ingredient ratio table And comparison ingredient ratio table The upper limit error is used to perform proportional error judgment to determine whether the upper limit error is less than or equal to the error threshold. If so, the benchmark sample is assigned Small category label , otherwise, perform label exhaustive judgment to determine the proportion of the selected comparison ingredients Make a unified position decision or assign a benchmark sample Small category label ; Perform sample exhaustion judgment based on the benchmark sample number Is it less than the total number of benchmark samples in the benchmark sample set? Decide whether to target a benchmark sample Continue to execute the adaptive classification strategy or stop, For the A benchmark sample.

3. The spectrum detection method for protective tape according to claim 1, characterized in that: The method of constructing a class spectrum set corresponding to a small class label based on a trait-oriented spectrum generation strategy includes the following specific steps: Obtain all benchmark samples with single small class labels to construct a class sample set; Determine whether the physical property of a single type of sample is powder to select a spectrometry method, and use the selected spectrometry method to obtain a first infrared spectrum, a second infrared spectrum, and a third infrared spectrum of three different detection points of the single type of sample; Calculate the cosine similarity between the first infrared spectrum, the second infrared spectrum and the third infrared spectrum respectively to obtain the average similarity of samples of a single class; If the average similarity is greater than or equal to the similarity threshold, the average of the first infrared spectrum, the second infrared spectrum and the third infrared spectrum is taken as the infrared spectrum of the single-class sample; Obtain the infrared spectra of all single-class samples corresponding to a single subclass label and construct a class spectrum set corresponding to a single subclass label.

4. The spectrum detection method for protective tape according to claim 3, characterized in that: The deviation correction method comprises the following specific steps: Obtain a class spectrum set corresponding to a single small class label, extract the component ratio table corresponding to all class samples in the class sample set to construct a class component ratio set; Calculate the mean of all component proportion tables in the class component proportion set to obtain the class average component proportion table; The total difference of proportions corresponding to each component proportion table in the class component proportion set is calculated to construct a total difference set of proportions, the total difference set of proportions is negative and a class weight set is generated through the Softmax function, and the class standard spectrum is generated through weighted summation.

5. The spectrum detection method for protective tape according to claim 1, characterized in that: The component analysis algorithm is used to process sample analysis tasks, including the following specific steps: Lock samples for testing No. part And get the Infrared spectrum of components ; Perform adaptive acquisition based on existing If the sub-class label is missing, choose to obtain the corresponding class standard spectrum or compare the infrared spectrum; Perform similarity comparison, calculate all class standard spectra or compare infrared spectra with the first Infrared spectrum of components The cosine similarity of is obtained to obtain the maximum cosine similarity, and whether it is greater than or equal to the similarity threshold; If it is greater than or equal to the similarity threshold, perform label status analysis. If the corresponding sub-category label If it is not missing, extract the average component ratio table As a table of ingredient ratios , on the contrary, the composition ratio table is obtained by spectral identification method , the ingredient ratio table and Infrared spectrum of components Insert into the standard comparison database; Further execution of the quantity reaching judgment is performed to determine whether the total number of class samples reaches the quantity threshold. If it has reached the threshold, the class average component ratio table is generated through the deviation correction method. And standard spectrum , remove the subcategory label missing state; If it is less than the similarity threshold, the component ratio table is obtained through spectral identification method , giving the part Small category label , the ingredient ratio table and Infrared spectrum of components As a comparison component ratio table and compare infrared spectra Store in standard comparison database and set subcategory labels is missing state; Enforce part exhaustion decisions based on part number Whether the total number of parts has been reached Determine the sample to be tested Analysis completed or locked part Continue to analyze and obtain the ingredient ratio table Until the sample is tested Analysis complete.

6. The spectrum detection method for protective tape according to claim 5, characterized in that: The spectral identification method comprises the following specific steps: Determine the major category of a single component, obtain the infrared spectrum, and pre-process the infrared spectrum through SV filtering to generate a high-quality infrared spectrum; The main features in the high-quality infrared spectrum are extracted by PCA algorithm to form a main feature sequence; Based on the major categories of a single component, a LS-SVM model with a pre-trained kernel function of RBF function is selected, and the main feature sequence is input into the LS-SVM model to obtain the component sequence of the single component; The pre-trained partial least squares model corresponding to the major category selection of a single component is used to process the main feature sequence, and the component ratio of all components is determined to obtain the component ratio table of a single component.

7. The spectrum detection method for protective tape according to claim 6, characterized in that: The spectral identification method can be further improved, and the improved spectral identification method includes the following specific steps: Obtain the infrared spectrum of a single component and generate a high-quality infrared spectrum by preprocessing the infrared spectrum through SV filtering; The high-quality infrared spectra are input into the GAN network to eliminate the characteristic differences between cloth and rubber and generate a unified infrared spectrum; The main features in the unified infrared spectrum are deeply extracted through the residual convolutional network, and the main feature dimensions are adjusted through the dimension adaptation network to generate dimensionally adapted features; The asynchronous convolutional network processes the dimension-appropriate features simultaneously through three different convolutional blocks, outputs three convolutional features from different angles, and connects them to obtain multimodal features; The multimodal features are input into the component prediction network, the dimension reducer is used to reduce the dimension of the multimodal features, and the component ratio of a single component is adaptively predicted through the Softmax regressor, and the component ratio table is output.

8. The spectrum detection method for protective tape according to claim 1, characterized in that: The construction of the composition ratio table comprises the following specific steps: Obtain all components of a single benchmark sample and simultaneously obtain the corresponding component ratios of all components; Generate an all-zero component ratio table with the dimension being the total number of components; The total ingredient table is queried to determine the position of a single ingredient in the total ingredient table, and the ingredient ratio of the single ingredient is filled into the corresponding position of the all-zero ingredient ratio table, until all ingredient ratios are filled into the all-zero ingredient ratio table, and the ingredient ratio table is constructed.

9. The spectrum detection method applied to protective tape according to claim 1, characterized in that: The finished product quality inspection algorithm is used to process sample quality inspection tasks, including the following specific steps: Obtain all components of the sample to be inspected and the corresponding sub-category labels; Focus on a single component and obtain infrared spectra based on the property-oriented spectrometry strategy; Query the standard comparison database to obtain the class standard spectrum corresponding to the sub-class label of a single component; Execute similarity judgment to determine whether the cosine similarity between the infrared spectrum and the class standard spectrum is greater than or equal to a similarity threshold; If the answer is no, the sample to be tested is judged to be unqualified. If the answer is yes, the component exhaustion judgment is performed to determine whether the component number reaches the total number of components. If it has reached it, the sample to be tested is judged to be qualified. Otherwise, the infrared spectrum of the next single component is obtained to continue the similarity judgment.

10. A spectrum detection system for protective tape, used to implement the spectrum detection method for protective tape as claimed in any one of claims 1 to 9, characterized in that: It includes classification module, collection module, scanning module, storage module and task module; The classification module generates a component ratio table of the benchmark sample, assigns a sub-class label to the benchmark sample based on an adaptive classification strategy, and generates a spectrum-deficient data packet; The transfer module obtains the missing spectrum data packet, initiates a scan request to obtain the infrared spectrum, and generates a full data packet to deliver to the storage module; The scanning module responds to the scanning request, identifies the requester and the number, obtains the infrared spectrum based on the trait-oriented spectrum strategy and feeds back to the requester, the number includes the reference sample number and the component number; The storage module receives the full data packet and stores it in the standard comparison database by classification, and controls the setting and cancellation of the missing status of the sub-category label; The task module obtains and identifies pending tasks, and independently selects a finished product quality inspection algorithm or a component analysis algorithm to complete the pending tasks.

Images