Ceramic product nondestructive testing method based on ensemble learning
Through the integrated learning method, non-destructive testing of ceramic products is used using sound signals, the problem of insufficient detection accuracy and efficiency in the automated production line of ceramic products is solved, and high-precision and high-efficiency ceramic product inspection is achieved.
Patent Information
- Application Number
- CN202510150876.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-11
- Publication Date
- 2025-04-25
AI Technical Summary
The prior art cannot meet the high efficiency and high precision non-destructive testing needs of ceramic product automation production lines, and traditional knocking methods and visual methods cannot effectively detect internal defects of ceramic products.
Using an integrated learning method, the sound signals are collected by tapping ceramic products, starting and end point detection, noise reduction processing and wavelet packet decomposition, defect characteristic indicators are constructed, and defect detection framework for integrated learning is established to realize non-destructive detection of ceramic products.
It improves the accuracy and efficiency of non-destructive testing of ceramic products, and is better than other machine learning methods based on decision trees and neural networks, which can effectively reduce misjudgment of manual detection and improve production efficiency and product quality.
Smart Images
Figure CN119881081A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a nondestructive testing method for ceramic products, in particular to a nondestructive testing method for ceramic products based on ensemble learning, and belongs to the technical field of structural nondestructive testing. Background Art
[0002] Ceramics are a widely used inorganic non-metallic material with excellent properties such as high temperature resistance, corrosion resistance and wear resistance. Ceramic products are usually made of natural or synthetic inorganic compounds through high-temperature sintering process. The preparation process is complicated and the processes are numerous. Even the finished products on the same production line have certain differences in quality and specifications. At the same time, the ceramic material itself has the characteristics of high brittleness and low toughness. Tiny internal defects will lead to stress concentration, which will quickly cause the destruction of ceramic products.
[0003] Therefore, in order to improve the integrity and uniformity of ceramic products, their quality inspection is very important.
[0004] In the quality inspection of traditional ceramic products, the percussion method or visual method is usually used to determine whether the ceramic products have defects. Obviously, these methods cannot meet the needs of automated production in the ceramic manufacturing industry. In order to improve the production efficiency and product quality of related industries, it is particularly important to study a high-efficiency and high-precision non-destructive testing method suitable for automated production lines of ceramic products. Summary of the invention
[0005] In order to solve the deficiencies of the prior art, the object of the present invention is to provide a nondestructive testing method for ceramic products based on ensemble learning.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] A nondestructive testing method for ceramic products based on ensemble learning comprises the following steps:
[0008] S1. Knock on any position of the ceramic product, collect sound signals from different directions, and detect the starting and ending points of the sound signals;
[0009] S2, performing noise reduction and normalization preprocessing on the sound slice signal after the start and end point detection;
[0010] S3. Decompose the preprocessed sound signal according to the characteristics of the sound signal, calculate the spectrum peaks and corresponding frequencies of each part of the signal after decomposition, and construct defect characteristic indicators;
[0011] S5. Use the defect feature index composition vector as the input of the defect detection framework to output the health status of the ceramic product.
[0012] The training of the above defect detection framework includes the following steps:
[0013] S1. Knock on any position of the ceramic product, collect sound signals from different directions, and detect the starting and ending points of the sound signals to obtain slicing signals;
[0014] S2, performing noise reduction and normalization preprocessing on the sound slice signal after the start and end point detection;
[0015] S3. Decompose the preprocessed sound signal according to the characteristics of the sound signal, calculate the spectrum peaks and corresponding frequencies of each part of the signal after decomposition, and construct defect characteristic indicators;
[0016] S4, repeating steps A1-A3, obtaining sound signals of defects of different positions and types of ceramic products, or adding noises of different degrees and types, and calculating corresponding defect characteristic indicators;
[0017] The defect feature index composition vector is taken as input, and the output is the health status of the ceramic product, and an integrated learning defect detection framework is constructed.
[0018] The above method uses an algorithm based on short-time average energy-zero-crossing rate to detect the start and end points of the sound signal, and the steps are:
[0019] A1. Set the high threshold H and the low threshold L. The high threshold intersects with the short-time energy curve at points H1 and H2, and the low threshold intersects with the short-time energy curve at points L1 and L2. If the valid signal is outside the H1-H2 segment, search in the direction of H1 to the left and H2 to the right until points L1 and L2 are found.
[0020] A2, short-time average zero-crossing rate, represents the cumulative number of times the sound signal crosses the time axis per unit time. Taking the short-time average zero-crossing rate as an indicator, follow the direction of L1 to the left and L2 to the right to find two intersection points T1 and T2 with the time axis. T1 is the start time of the original sound signal, and T2 is the end time of the original sound signal.
[0021] The above health conditions:
[0022] If it is defective, the output result is one-dimensional [0,1] data, 0 represents no defect, 1 represents defect, and the closer the output data is to 1, the higher the probability of defect;
[0023] If it is a defect type, the output result is data of several dimensions [0,1], where each dimension represents a certain type of defect, 0 represents the absence of such defect, and 1 represents the presence of such defect. The closer the output data is to 1, the higher the probability of containing such defect.
[0024] In the above method, the denoising is performed based on the ant colony algorithm-variational modulus decomposition and reconstruction method.
[0025] In the above method, the decomposition of the pre-processed sound signal is performed using a wavelet packet decomposition method.
[0026] The above method uses a Bayesian optimizer to optimize the ensemble method, base learner type, base learner number, and variable parameters of the base learner of the ensemble learning, and constructs a defect detection framework for the ensemble learning.
[0027] In the above method, the ensemble learning ensemble method includes the parallel method Bagging, the serial method Boosting, and the hybrid method Stacking.
[0028] The present invention is beneficial in that:
[0029] The present invention provides a nondestructive testing method for ceramic products based on ensemble learning, which has the following advantages:
[0030] 1. Using the ant colony algorithm-variational mode decomposition and reconstruction algorithm to reduce the noise of the knocking sound signal can significantly reduce the pollution of environmental noise, make the collected sound signal more accurate, and improve the accuracy of the non-destructive testing algorithm of ceramic products;
[0031] 2. The proposed nondestructive testing method for ceramic products based on ensemble learning has better detection accuracy than other machine learning nondestructive testing methods based on decision trees, neural networks, etc. while ensuring detection efficiency;
[0032] The nondestructive testing method of the present invention is an intelligent method that can use knocking sound information to directly carry out nondestructive testing on ceramic products on an automated production line, thereby reducing misjudgments in manual testing, improving production efficiency and product quality, and has strong practicality and wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is an instrument layout diagram for non-destructive testing of ceramic products of the present invention.
[0034] Figure 2 This is a flow chart of the signal denoising method based on ant colony algorithm-variational modulus decomposition and reconstruction.
[0035] Figure 3 Flowchart of the construction of a training dataset for ensemble learning.
[0036] Figure 4 This is a flowchart of non-destructive testing of ceramic products based on ensemble learning. DETAILED DESCRIPTION
[0037] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] A nondestructive testing method for ceramic products based on ensemble learning, the steps are:
[0039] S1. Use a steel pendulum ball suspended in the air to knock on the same position of the ceramic product. Place pickups at 2 to 5 fixed positions to collect the sound signals generated by the ceramic product being knocked. The detection equipment is arranged as follows: Figure 1 shown.
[0040] The start and end point detection of the original sound signal collected by the pickup is performed using an algorithm based on short-time average energy-zero crossing rate to accurately extract the start and end time of the knocking sound signal. Specifically:
[0041] A1: Set the high threshold H and the low threshold L. The high threshold and the short-time energy curve intersect at points H1 and H2, and the low threshold and the short-time energy curve intersect at points L1 and L2. If the valid signal is outside the H1-H2 segment, search in the direction of H1 to the left and H2 to the right until points L1 and L2 are found.
[0042] The expression of short-time average zero-crossing rate is
[0043] where x n (t) represents the sound signal,
[0044] A2: The short-time average zero-crossing rate represents the cumulative number of times the sound signal crosses the time axis in a unit time. Using the short-time average zero-crossing rate as an indicator, follow the L1 found in A1 to the left and L2 to the right to find two intersections with the time axis, T1 and T2; T1 is the start time of the original sound signal, and T2 is the end time of the original sound signal.
[0045] S2. For the sound slice signal after the start and end point detection, the ant colony algorithm-variational modulus decomposition and reconstruction method is used for noise reduction preprocessing. The flow chart is as follows: Figure 2 shown.
[0046] S2.1: Set the number of decomposition layers K and penalty parameter α;
[0047] S2.2: Initialize the number of layers k = 1, parameter n = 1;
[0048] S2.3: Initialize variables
[0049] S2.4: Calculate variables according to iterative formula
[0050] S2.5: Calculate variables according to iterative formula
[0051] S2.6: Calculate variables according to iterative formula
[0052] S2.7: For a given discrimination accuracy e, if the error Then n=n+1, return to step S2.4; otherwise, stop iteration and go to the next step;
[0053] S2.8: Calculate the signal residual after the k-th layer decomposition
[0054] S2.9: Let k = k + 1. If k ≤ K, then n = 1 and return to step S2.3; otherwise, stop the iteration and go to the next step.
[0055] S2.10: Finally, the original signal is decomposed into the superposition of k component signals, the noise component signal part is removed, the noise-reduced sound signal is reconstructed and normalized.
[0056] Preferably, the ant colony algorithm is used to optimize the decomposition layer number K and the penalty parameter α of the variational modulus decomposition, so that the noise signal component can be separated more accurately by using the variational modulus decomposition, and the steps are:
[0057] S2.11: Take the parameter combination [K, α] consisting of the decomposition layer number K and the penalty parameter α as the position of the population, and randomly generate a population of m parameter combinations as the initial position of the population;
[0058] S2.12: Perform variational modulus decomposition on the sound slice signal according to different parameter combinations [K, α] and calculate the fitness value of each position response where p j is the normalized value of the envelope signal obtained after the signal is Hilbert demodulated;
[0059] S2.13: Compare the size of the fitness value and update the individual local extreme value and the population global extreme value of [K, α];
[0060] S2.14: Using pheromones to update the formula τ ij =(1-ρ)τ ij +Δτ ij Update the position of the population, where ρ is the volatility coefficient of the pheromone, For path x n =i to y n = the amount of pheromone released on j, the constant Q is the amount of pheromone released each time, L n is the length of path n, δ(x n =i,y n =j) is the indicator function;
[0061] S2.15: loop iteration, go to step S2.12, until the number of iterations reaches the maximum set value G, then output the optimal [K, α] value for subsequent variational modulus decomposition.
[0062] Step S3: Set the striking point as h, and the sound signal collected by the pickup is pre-processed and converted into v hr (t) represents, r = 1, 2, ... rc is the pickup number, and the sound slice signal after noise reduction and normalization preprocessing is calculated using the wt layer wavelet packet decomposition to obtain 2 wt The decomposed signal components Perform Fourier transform on these signal components to obtain the corresponding spectrum and obtain its spectrum peak and the corresponding frequency According to the number r of the pickup, the defect characteristic index vector DI for nondestructive testing is formed as follows: DI = [D1, D2, ..., D rc ] T ,
[0063] The elements
[0064] Step S4: The defect characteristic index DI is used as the input of the integrated learning, and the health status of the ceramic product is used as the output of the integrated learning to establish the framework of the integrated learning.
[0065] Preferably, a Bayesian optimizer is used to perform parameter optimization on the ensemble method, base learner type, and base learner number of ensemble learning.
[0066] Preferably, the ensemble learning method can be a parallel method Bagging, a serial method Boosting, a hybrid method Stacking, etc., taking the Bagging method as an example:
[0067] S4.1: Extract training sets from the original sample set. In each round, use the Bootstraping method to extract n training samples from the original sample set. Perform k rounds of extraction in total to obtain k independent training samples.
[0068] S4.2: Each time a training set is used to obtain a corresponding training model, a total of k models are obtained from k training sets;
[0069] S4.3: Taking the classification of nondestructive testing of ceramic products as the topic, the k models obtained are voted to obtain the classification results.
[0070] Preferably, the base learner type of ensemble learning can be different machine learning algorithms such as decision tree, neural network, support vector machine, etc., taking CART classification decision tree as an example;
[0071] S4.4: Select an independent variable and a value that divides the data space into two parts, one part where all points satisfy the condition and the other part where all points do not satisfy the condition;
[0072] S4.5: For each partition point, calculate its Gini coefficient Among them, c represents the number of categories in the data set D, g i Represents the probability of the occurrence of the i-th category, and selects the point with the smallest Gini coefficient for division;
[0073] S4.6: Divide the entire data space into multiple subsets, each subset adopts the same division method until the stopping condition is met;
[0074] S4.7: For each leaf node, classification is performed using the combined strategy of the base learners.
[0075] Preferably, the combination strategy of the base learners selects a voting method suitable for classification.
[0076] Preferably, the integrated learning output represents the health status of the ceramic product. If only the ceramic product needs to be detected for defects, the output is a one-dimensional data label; if the defect type of the ceramic product needs to be distinguished, the output is a multi-dimensional data label representing various defect types.
[0077] A microphone is used to collect the original sound signals of defects in different positions and types of ceramic products, and noise of different degrees and types is added to expand the data set. The corresponding defect feature indicators are calculated to construct a training data set for ensemble learning.
[0078] Preferably, the training data set needs to include various defect types of ceramic products included in the integrated learning output vector in step S4. The construction process of the training data set is as follows: Figure 3 .
[0079] The tested and verified integrated learning architecture can be used for non-destructive testing of corresponding ceramic products, which serves as an integrated learning defect detection framework.
[0080] S5. For the ceramic product to be tested, the defect characteristic index of the knocking sound signal is obtained through steps S1 to S3, which is used as the input of the integrated learning defect detection framework that has been trained in step S4. The output is the health status of the ceramic product. The non-destructive testing process is as follows: Figure 4 shown.
[0081] In step S5, if only the ceramic product needs to be tested for defects, the output result is one-dimensional [0, 1] data, where output data 0 represents that the ceramic product has no defects, and output data 1 represents that the ceramic product has defects. The closer the output data is to 1, the higher the probability that the ceramic product has defects.
[0082] If it is necessary to detect the defect type of the ceramic product in step S5, the output result is data of several dimensions [0, 1], where each dimension represents a certain type of defect of the ceramic product. The output data 0 of this dimension represents that the ceramic product does not have this type of defect, and the output data 1 represents that the ceramic product contains this type of defect. The closer the output data is to 1, the higher the probability that the ceramic product contains this type of defect.
[0083] Example 1
[0084] A nondestructive testing experiment was conducted on ceramic bowls using the ceramic product nondestructive testing method based on ensemble learning described in the present invention.
[0085] The ceramic bowl used in this experiment was fired at 1300℃, with a diameter of 15.5cm and a height of 6cm. It was placed on a vibration isolation platform. A steel pendulum ball of fixed height was used to hit the ceramic bowl at a fixed position. The experimental arrangement is as follows: Figure 1 As shown in Figure 2, the health status of ceramic bowls can be divided into two types: intact structure and defective structure.
[0086] In this experiment, three mobile phones were arranged, located on both sides and above the ceramic bowl. They can collect the sound signals propagating in all directions of space after the ceramic bowl is hit. The sampling frequency is set to 48kHz.
[0087] Firstly, the start and end points of the knocking signal are detected using an algorithm based on short-time average energy-zero-crossing rate to obtain the slicing signal. Then, the slicing signal is denoised using a method based on ant colony algorithm-variational mode decomposition and reconstruction, and the preprocessed sound signal is obtained after normalization. Finally, the sound signal is decomposed using wavelet packets and the defect characteristic indicators are calculated.
[0088] The test found that when the wavelet packet is decomposed into 4 layers, the test results of ensemble learning can achieve the best accuracy and efficiency. Therefore, the input of ensemble learning, the defect feature index dimension is 2×3×2 4 , where 2 represents the spectrum peak and the corresponding frequency, 3 represents the number of pickups, and 2 4 Represents the number of sub-signals after wavelet packet decomposition. The output of ensemble learning is one-dimensional [0, 1] data. Output 0 means that the ceramic product has no defects, and output 1 means that the ceramic product has defects. The closer the output is to 1, the higher the probability that the ceramic product has defects.
[0089] Due to the characteristics of the ceramic bowl sound signal, the Bayesian optimizer is used to optimize the parameters of the ensemble learning. The selected ensemble method is Bagging, the base learner type is decision tree, the number of base learners is 121, the maximum number of splits is 2329, and the number of sampling variables is 2.
[0090] After two tests, 2760 sets of data were collected, 420 sets of data from the first test were used as the test set, and 2340 sets of data from the second test were used as the training set.
[0091] Finally, it can be obtained that the accuracy of the non-destructive detection of ceramic bowls based on ensemble learning of the present invention is as high as 99.8%. Compared with the accuracy of the detection method based on decision tree at 96.3% and the accuracy of the detection method based on neural network at 89.9%, the accuracy of the method proposed in the present invention is the best.
[0092] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form, and any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.
Claims
1. A nondestructive testing method for ceramic products based on ensemble learning, characterized in that: The following steps are involved: S1. Knock on any position of the ceramic product, collect sound signals from different directions, and detect the starting and ending points of the sound signals; S2, performing noise reduction and normalization preprocessing on the sound slice signal after the start and end point detection; S3. Decompose the preprocessed sound signal according to the characteristics of the sound signal, calculate the spectrum peaks and corresponding frequencies of each part of the signal after decomposition, and construct defect characteristic indicators; S5. Use the defect feature index composition vector as the input of the defect detection framework to output the health status of the ceramic product.
2. The method according to claim 1, characterized in that: The training of the defect detection framework includes the following steps: S1. Knock on any position of the ceramic product, collect sound signals from different directions, and detect the starting and ending points of the sound signals to obtain slicing signals; S2, performing noise reduction and normalization preprocessing on the sound slice signal after the start and end point detection; S3. Decompose the preprocessed sound signal according to the characteristics of the sound signal, calculate the spectrum peaks and corresponding frequencies of each part of the signal after decomposition, and construct defect characteristic indicators; S4, repeating steps A1-A3, obtaining sound signals of defects of different positions and types of ceramic products, or adding noises of different degrees and types, and calculating corresponding defect characteristic indicators; The defect feature index composition vector is taken as input, and the output is the health status of the ceramic product, and an integrated learning defect detection framework is constructed.
3. The method according to claim 1 or 2, characterized in that: The start and end point detection of the sound signal is performed using an algorithm based on short-time average energy-zero crossing rate. The steps are as follows: A1. Set the high threshold H and the low threshold L. The high threshold intersects with the short-time energy curve at points H1 and H2, and the low threshold intersects with the short-time energy curve at points L1 and L2. If the valid signal is outside the H1-H2 segment, search in the direction of H1 to the left and H2 to the right until points L1 and L2 are found. A2, short-time average zero-crossing rate, represents the cumulative number of times the sound signal crosses the time axis per unit time. Taking the short-time average zero-crossing rate as an indicator, follow the direction of L1 to the left and L2 to the right to find two intersection points T1 and T2 with the time axis. T1 is the start time of the original sound signal, and T2 is the end time of the original sound signal.
4. The method according to claim 1, characterized in that The health status: If it is defective, the output result is one-dimensional [0, 1] data, 0 represents no defect, 1 represents defect, and the closer the output data is to 1, the higher the probability of defect; If it is a defect type, the output result is data of several dimensions [0, 1], where each dimension represents a certain type of defect, 0 represents the absence of such defect, and 1 represents the presence of such defect. The closer the output data is to 1, the higher the probability of containing such defect.
5. The method according to claim 1 or 2, characterized in that: The denoising is performed based on an ant colony algorithm-variational modulus decomposition and reconstruction method.
6. The method according to claim 1 or 2, characterized in that: The decomposition of the preprocessed sound signal is performed using a wavelet packet decomposition method.
7. The method according to claim 2, characterized in that: The Bayesian optimizer is used to optimize the ensemble method, base learner type, number of base learners, and variable parameters of base learners of ensemble learning, and a defect detection framework of ensemble learning is constructed.
8. The method according to claim 2, characterized in that: The ensemble learning method includes the parallel method Bagging, the serial method Boosting, and the hybrid method Stacking.