A method for detecting the rapid annealing effect of metallic materials based on acoustic response

Abstract

This invention relates to the field of metal material heat treatment quality inspection technology, and particularly to a method for rapid annealing effect detection of metal materials based on acoustic response, comprising: Step 1: Acoustic excitation and synchronous acquisition; Step 2: Multi-physics coupling compensation: Based on temperature distribution data and the temperature dependence of material acoustic properties, the phase compensation amount of the vibration signal is calculated; the environmental vibration spectrum is obtained through a production line mechanical vibration monitoring device, and the excitation response component is separated from the original vibration signal using adaptive filtering; Step 3: Annealing sensitive feature extraction; Step 4: Dynamic quality mapping and control: The real-time feature vector is input into a pre-trained machine learning model, and the recrystallization ratio and grain size are output; when the output parameters exceed the process threshold, the annealing furnace control parameters are adjusted. Through non-destructive and precise acoustic monitoring technology, the annealing effect of metal materials can be monitored efficiently and accurately in high-speed, high-noise, and strong-interference industrial environments.

Description

Technical Field

[0001] This invention relates to the field of heat treatment quality testing technology for metallic materials, and in particular to a method for testing the rapid annealing effect of metallic materials based on acoustic response. Background Technology

[0002] The inspection of annealing effects in metallic materials is a core aspect of heat treatment quality control, and its accuracy directly impacts the material's mechanical properties, formability, and service reliability. In continuous annealing production lines (such as strip steel and copper wire rolling lines), it is necessary to capture the evolution of microstructure (such as recrystallization ratio and grain size) in real time under industrial conditions involving high-speed operation (linear speed ≥3m / s), strong vibration interference, and temperature gradients (±50℃). However, existing inspection technologies have significant limitations:

[0003] 1. Technical bottlenecks of existing detection methods:

[0004] (1) Destructive metallographic inspection: Grain size is observed under a microscope after cutting a sample, polishing and etching it. This method requires at least 24 hours to complete a single test and cannot be used for real-time control of the production line. The sampling coverage is less than 1%, which makes it impossible to reflect the uniformity of the strip along its entire length, and the destructive sampling results in material waste.

[0005] (2) Indirect physical quantity detection method:

[0006] Resistance method: It relies on the correlation between resistance value and structural defects, but is significantly affected by fluctuations in contact pressure and temperature drift. For example, in the inspection of copper wire annealing, the probe pressure needs to be adjusted multiple times to obtain the "lowest resistance value", the measurement error exceeds 15%, and it cannot distinguish between recrystallization and dislocation density changes.

[0007] Hardness method: Surface hardness cannot characterize the gradient of deep tissue. Pipe inspection requires slicing and testing point by point, which takes more than 45 minutes per piece. Temperature fluctuations of ±50℃ can cause hardness values ​​to drift by up to 10%.

[0008] (3) Advanced physical field analysis method

[0009] X-ray diffraction: The equipment is expensive and requires a vacuum environment, and the detection speed is less than 0.5m / s, making it unsuitable for high-speed production lines;

[0010] Ultrasonic phased array: In annealed coarse-grained materials, the sound attenuation is severe, the signal-to-noise ratio is less than 6dB, and the defect detection rate is less than 60%.

[0011] Therefore, there is an urgent need for a rapid annealing effect detection method for metallic materials based on acoustic response to solve the above problems. Summary of the Invention

[0012] To achieve the above objectives, this invention provides a method for detecting the rapid annealing effect of metallic materials based on acoustic response, comprising:

[0013] Step 1: Acoustic excitation and synchronous acquisition:

[0014] A broadband linear frequency modulated acoustic signal is emitted to a moving metal strip through a piezoelectric ceramic transducer array, and its frequency band covers the resonant modes in the thickness direction of the material.

[0015] During the excitation, a non-contact vibration sensor is used to collect the vibration response signal of the strip surface, and an infrared thermometer is used to obtain the surface temperature distribution.

[0016] Step 2: Multiphysics Coupling Compensation

[0017] Based on temperature distribution data and the temperature dependence of material acoustic properties, the phase compensation amount of the vibration signal is calculated.

[0018] The environmental vibration spectrum is obtained by the production line mechanical vibration monitoring device, and the excitation response component is separated from the original vibration signal by adaptive filtering.

[0019] Step 3: Extraction of Annealing-Sensitive Features:

[0020] Time-frequency analysis was performed on the compensated signal, and the sub-band with the largest change in energy concentration was selected based on the Brillouin scattering characteristics of the metal lattice.

[0021] Extract the energy proportion parameter and signal attenuation time constant of the sub-band;

[0022] Step 4: Dynamic Quality Mapping and Control

[0023] The real-time feature vector is input into a pre-trained machine learning model, which outputs the recrystallization ratio and grain size.

[0024] When the output parameters exceed the process threshold, the control parameters of the annealing furnace are adjusted.

[0025] Preferably, the method for determining the frequency band in step 1 includes:

[0026] A pre-established mapping relationship library between material thickness and resonant frequency was created: samples of the same material with varying thickness gradients were prepared, and a frequency sweep excitation experiment was performed on each sample to record the peak distribution range of its resonant frequency.

[0027] Based on the measured thickness of the strip detected online, the corresponding lower and upper frequency limits are matched from the mapping database;

[0028] The sweep range of the linear frequency modulation signal is set to the matched frequency interval, and the sweep period is dynamically adjusted according to the speed of the strip movement to ensure that each detection area is fully excited.

[0029] Preferably, the calculation of the phase compensation amount in step 2 includes:

[0030] The temperature-affected zones are divided along the sound wave propagation path, and the length of each zone is determined based on the spatial density of infrared temperature measurement points.

[0031] Obtain the average temperature measurement value of each section, and calculate the sound wave propagation time offset of that section by combining the calibration curve of the material sound velocity with temperature.

[0032] The propagation time offsets of each segment are accumulated and converted into the phase compensation amount of the vibration signal;

[0033] The calibration curve was obtained through laboratory calibration: the sound wave propagation speed of the same material at different temperature points was measured under a controlled temperature environment, and the sound speed-temperature function relationship was fitted.

[0034] Preferably, the implementation of the adaptive filtering includes:

[0035] The vibration spectrum collected by the mechanical vibration monitoring device is used as the reference noise signal;

[0036] A finite impulse response filter is constructed, and the filter coefficients are dynamically adjusted through an iterative algorithm to maximize the similarity between the filter output signal and the reference noise signal.

[0037] The pure acoustic response component is obtained by subtracting the filter output signal from the original vibration signal.

[0038] The termination condition of the iterative algorithm is that the power spectral density of the residual signal drops below a set ratio.

[0039] Preferably, the method for selecting the sub-band includes:

[0040] Wavelet packet decomposition is performed on the compensated signal to obtain a set of sub-bands covering the entire frequency band;

[0041] Calculate the difference in energy concentration for each sub-band in the fully annealed and unannealed states;

[0042] Select the combination of continuous sub-bands with the largest difference in energy concentration as the characteristic band, and its total bandwidth shall not exceed a preset proportion of the full bandwidth;

[0043] The energy concentration is defined as the ratio of sub-band signal energy to full-band signal energy.

[0044] Preferably, the extraction of the decay time constant includes:

[0045] Envelope extraction is performed on the characteristic sub-frequency band signal to identify data segments in the signal attenuation phase;

[0046] Select data points in the attenuation data segment that are above the environmental noise threshold;

[0047] The selected data points are fitted into an exponential decay curve using a nonlinear fitting algorithm.

[0048] The time required for the amplitude to decay to a specific proportion of the initial value is calculated based on the fitted curve.

[0049] Preferably, the training method for the machine learning model includes:

[0050] A sample set covering different annealing degrees was prepared, and acoustic feature vectors and metallographic detection data were collected simultaneously for each sample;

[0051] A regression algorithm with adjustable kernel function is used for training, and the kernel function type is adaptively selected according to the number of samples and feature dimensions;

[0052] The model hyperparameters are optimized through cross-validation until the mean absolute error between the predicted value and the metallographic detection value is lower than the preset requirement.

[0053] Preferably, the preparation of the sample set includes:

[0054] Gradient annealing process is applied to the same batch of materials to make the recrystallization ratio cover the range of 10% to 100%;

[0055] The number of samples for each annealed state is determined based on the material's microstructure uniformity; for materials with poor microstructure uniformity, the number of samples is increased.

[0056] Acoustic feature acquisition simulates the production line environment, introducing background vibration noise of the same magnitude as that used in online detection.

[0057] Preferably, the adjustment of the control parameters for the trigger annealing furnace includes:

[0058] When the recrystallization ratio is lower than the target value, a temperature control command is generated, and the temperature rise is proportional to the difference between the target value and the measured value.

[0059] When the grain size exceeds the upper limit, a cooling control command is generated, and the temperature drop is proportional to the difference between the measured value and the upper limit.

[0060] The proportionality coefficient is determined through process experiments: the rate of change of parameters is measured under a fixed temperature adjustment, and the average response coefficient of multiple experiments is taken.

[0061] Preferably, it also includes an online model update step:

[0062] Metallographic examination of strip samples is conducted periodically, with the examination cycle set according to the stability of the material batch.

[0063] Add the acoustic feature vectors and metallographic data of the new samples to the training set;

[0064] When the number of new samples reaches a set proportion of the original dataset, the machine learning model is retrained.

[0065] After running the old and new models in parallel for a predetermined period, the model with the smaller prediction error is selected as the current model to be used.

[0066] The beneficial effects of this invention are:

[0067] 1. This invention overcomes the destructive nature of metallographic inspection through a real-time monitoring method based on acoustic response. By using a piezoelectric ceramic transducer array and a non-contact vibration sensor, the acoustic response signal of the metal strip can be captured in real time while it is moving at high speed, and its temperature distribution can be measured simultaneously, comprehensively monitoring the microstructure evolution during the annealing process. This method eliminates the need for sample cutting, avoiding material waste, and provides efficient and accurate real-time data through continuous online monitoring, greatly improving detection efficiency and the response speed of quality control.

[0068] 2. The detection method of this invention combines acoustic excitation with a non-contact vibration sensor, solving the problems of contact interference and temperature fluctuations. By understanding the relationship between the acoustic signal and the thermal properties of the material, this invention can accurately capture the microstructural changes of the material during annealing. Furthermore, through temperature compensation and adaptive filtering techniques, the influence of environmental interference can be effectively eliminated, ensuring the accuracy of the detection signal. This method not only solves the error problems of the resistance method and hardness method, but also reflects the internal microstructure of the material, achieving accurate characterization of deep structural gradients.

[0069] 3. This invention overcomes the shortcomings of advanced physical field analysis methods in high-speed production line applications by employing an acoustic response-based detection method. Through precise linear frequency modulated acoustic excitation signals and efficient signal acquisition technology, this invention enables real-time detection under high-speed operating conditions, meeting the speed requirements of continuous annealing production lines. Unlike ultrasonic phased arrays, the acoustic response detection of this invention is unaffected by material acoustic attenuation, effectively avoiding the problem of low signal-to-noise ratio. Simultaneously, by utilizing time-frequency analysis and attenuation time constant extraction methods, it can accurately capture microstructural changes in metallic materials, significantly improving detection accuracy and avoiding the limitations of traditional technologies. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0071] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0072] Figure 2 This is a flowchart illustrating the steps of the method of the present invention to trigger the adjustment of the control parameters of the annealing furnace;

[0073] Figure 3 This is a flowchart illustrating the steps of online model updating in the method of this invention. Detailed Implementation

[0074] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0075] Please see Figures 1-3 This invention provides a method for detecting the rapid annealing effect of metallic materials based on acoustic response. In step 1, a wideband linear frequency modulated acoustic signal is first emitted using a piezoelectric ceramic transducer array. The frequency band of this signal can cover the resonant modes of the metal strip in the thickness direction, ensuring that various microstructural changes generated in the metal material during annealing are excited. After penetrating the metal strip, these acoustic signals interact with its internal microstructure, changing its propagation characteristics. Simultaneously, a non-contact vibration sensor synchronously collects the vibration response signal on the strip surface. These signals reflect the changes in the internal lattice of the metal, enabling real-time capture of the material's microscopic evolution process. Furthermore, an infrared thermometer acquires the surface temperature distribution data of the material, providing temperature change information for subsequent analysis and ensuring the accuracy of the detection process.

[0076] In step 2, based on temperature distribution data, this method calculates the phase compensation amount of the vibration signal. Temperature has a significant impact on the acoustic properties of materials; therefore, compensation is performed using temperature data to avoid errors caused by temperature fluctuations. On the other hand, the environmental vibration spectrum is obtained through production line mechanical vibration monitoring, and adaptive filtering technology is used to separate the noise generated by mechanical vibration and excitation signals from the original vibration signal, ensuring that the extracted signal only represents the true changes within the material.

[0077] In step 3, after compensation processing, time-frequency analysis is performed on the vibration signal. Due to the Brillouin scattering characteristics of metallic materials, microstructural changes during annealing lead to significant changes in the energy concentration of different frequency components. By analyzing the energy distribution of different sub-bands, the frequency band with the largest change in energy concentration is selected, and relevant features are extracted based on this, including parameters such as the signal decay time constant and the energy proportion of the sub-band. These features can effectively reflect the grain evolution and recrystallization degree of metallic materials during annealing.

[0078] In step 4, the feature vectors extracted in real time are input into a pre-trained machine learning model to obtain key parameters such as recrystallization ratio and grain size. These parameters directly affect the mechanical properties and service reliability of the material. When the parameters monitored in real time exceed the threshold set by the process, control signals can be automatically triggered to adjust control parameters such as the temperature and heating time of the annealing furnace, thereby achieving precise control of the annealing process and ensuring that the material meets the expected quality standards.

[0079] The application of acoustic response technology enables real-time, non-destructive detection of the annealing effect of metal strips, while maintaining high precision under complex conditions such as high temperature, high speed, and high vibration. Simultaneously, the combination of multi-physics coupling compensation and time-frequency analysis effectively improves the accuracy and robustness of the detection. Furthermore, the application of machine learning models makes the control of the annealing process more intelligent, enabling dynamic adjustment of process parameters and improving production efficiency and product quality.

[0080] In one possible implementation, to accurately determine the frequency band of the excitation signal, a series of identical metal material samples with varying thickness gradients are first prepared. Frequency sweep excitation experiments are then performed on these samples, recording the peak resonant frequencies of each sample at different thicknesses. This process encompasses the changes in the acoustic properties of the metal material at different thicknesses and provides fundamental data for subsequent frequency matching. Through multiple experiments, the mapping relationship between material thickness and resonant frequency is obtained, forming a mapping relationship library.

[0081] In actual production, the thickness of the strip may vary, so it is necessary to monitor the strip thickness in real time and match the corresponding frequency range from the established mapping database. Specifically, based on the actual thickness of the strip, the mapping database provides corresponding lower and upper frequency limits to ensure that the frequency range of the excitation signal matches the resonant frequency of the strip, thereby ensuring that the sound wave can effectively excite the resonant modes of the material.

[0082] Once the frequency range is determined, the next step is to set the sweep range of the linear frequency modulated signal to ensure that the signal sweep covers the matched frequency band. During the movement of the strip, the sweep period needs to be dynamically adjusted to avoid signal omission or repeated excitation. This adjustment is based on the actual movement speed of the strip, ensuring that each detection area is fully excited and that no potential changes in the annealing effect are missed.

[0083] By establishing a mapping library between thickness and resonant frequency, an appropriate excitation frequency range can be automatically selected based on the actual thickness of the metal strip, greatly improving the accuracy and specificity of the detection signal. Simultaneously, by dynamically adjusting the sweep period, complete signal excitation can be ensured, resulting in more accurate acoustic response data and achieving high-precision annealing effect detection. Compared to traditional fixed-frequency excitation methods, this method can adapt to variations in strip thickness, improving the flexibility and adaptability of the entire detection process.

[0084] In one possible implementation, the sound wave propagation path is divided into several temperature-affected sections based on the distribution density of spatial temperature points acquired by the infrared thermometer. The length of each section is determined by the spacing between the infrared thermometer points, ensuring that temperature changes within each section can be accurately reflected. The principle for dividing the sections is to ensure both the spatial resolution of the temperature information and to avoid excessive computational complexity due to too many sections.

[0085] Within each section, measurements were collected at corresponding temperature points, and the average temperature was calculated. Using a pre-obtained material sound velocity-temperature calibration curve in the laboratory, the average temperature of that section was mapped to the sound wave propagation speed, thereby calculating the sound wave propagation time offset for that section. The calibration curve was obtained under controlled temperature conditions by measuring the sound wave propagation speed of the same metallic material at different temperatures and fitting it into a continuous sound velocity-temperature function relationship, achieving a precise correspondence between temperature and sound velocity.

[0086] The calculated sound wave propagation time offsets for each segment along the propagation path are summed to obtain the total propagation time offset. This time offset is then converted into a phase compensation value for the vibration signal, enabling the acquired vibration signal to eliminate the error caused by temperature-induced propagation velocity variations and accurately reflect the acoustic properties of the material itself.

[0087] The above steps effectively overcome the influence of temperature gradient on sound wave propagation during rapid annealing, enabling vibration signals to more accurately reflect the material state and improving the detection accuracy of annealing effects. Furthermore, this method eliminates the need for contact measurement, allowing for real-time detection on moving strips, adapting to the high-speed operating environment of production lines, and avoiding signal distortion caused by temperature changes in traditional methods, thereby enhancing the reliability and repeatability of the detection results.

[0088] In one possible implementation, firstly, vibration signals of the strip during the annealing process are collected using a mechanical vibration monitoring device. These vibration signals contain external noise components, such as equipment operation or environmental vibrations. The collected vibration spectra are then used as reference noise signals for adaptive adjustment of subsequent filtering.

[0089] To remove noise, a Finite Impulse Response (FIR) filter needs to be constructed. This filter's coefficients are adjusted using an iterative algorithm to optimize its filtering effect. In each iteration, the algorithm dynamically adjusts the filter coefficients based on the characteristics of the reference noise signal, maximizing the similarity between the filter's output signal and the reference noise signal. In other words, the filter continuously adjusts its response to filter out noise signals to the greatest extent possible.

[0090] After processing by the adaptive filter, the output signal represents the signal after noise has been removed from the original vibration signal. Next, the filter output signal is subtracted from the original vibration signal to obtain the pure acoustic response components. These pure acoustic signals reflect the actual acoustic properties of the metallic material and can be used to determine the annealing effect.

[0091] The iterative process of adaptive filtering stops when a specific condition is met. Specifically, the iterative algorithm stops when the power spectral density of the residual signal (i.e., the energy of the noise component) drops below a set percentage. This condition ensures the accuracy of the filtering process, so that the final output signal no longer contains significant noise components.

[0092] Adaptive filtering technology effectively separates noise from the pure acoustic response components, significantly improving signal clarity and accuracy. This method can accurately extract acoustic signals reflecting the annealing effect of materials even in high-noise environments during annealing. Furthermore, the dynamic adjustment of the iterative algorithm makes the filtering process more intelligent and flexible, adapting to different noise conditions and maintaining efficient and stable detection performance in variable production environments.

[0093] In one possible implementation, the phase-compensated signal is first subjected to wavelet packet decomposition. Wavelet packet decomposition is a multi-scale signal analysis method that decomposes the original signal to obtain a set of sub-bands within different frequency ranges. Each sub-band represents the energy information of a specific frequency band in the original signal, thus making the frequency domain characteristics of the signal clearer. Through decomposition, all sub-bands of the entire frequency band can be covered, thereby providing comprehensive frequency domain information for subsequent analysis.

[0094] After wavelet packet decomposition, for each sub-band, the energy concentration difference between the fully annealed and unannealed states needs to be calculated. Energy concentration refers to the ratio of the signal energy of a sub-band to the total signal energy of the entire frequency band, reflecting the relative concentration of signal energy in that band. By comparing the energy concentration of each sub-band under different annealing states, the sensitivity of that band to the annealing effect can be evaluated. Sub-bands with larger energy concentration differences generally better reflect changes in the annealing state of the material.

[0095] After calculating the energy concentration difference of all sub-bands, the combination of continuous sub-bands with the largest energy concentration difference is selected as the characteristic band. This characteristic band combination, as the signal component that best reflects the annealing effect of the metallic material, will be used for subsequent annealing effect analysis. To avoid over-subdivision, the total bandwidth of the selected characteristic band does not exceed a preset proportion of the full bandwidth. This ensures the simplicity of the analysis while avoiding noise interference from excessive bandwidth.

[0096] By employing wavelet packet decomposition and energy concentration difference analysis, frequency domain features related to the annealing state can be accurately captured. This method can identify frequency bands that exhibit significant differences under different annealing states. These frequency bands are most sensitive to the annealing process, effectively improving detection accuracy. Furthermore, the selected characteristic frequency bands possess strong indicative capabilities of the annealing effect, effectively removing noise interference, reducing the influence of irrelevant signals, and ensuring the reliability and accuracy of the detection results.

[0097] In one possible implementation, firstly, envelope extraction is performed on the selected characteristic sub-band signal. The envelope is the outer contour curve of the signal, reflecting the amplitude variation trend. During annealing, the acoustic response signal of the metallic material gradually attenuates over time; the envelope clearly depicts this attenuation process, helping to determine the attenuation stage of the signal.

[0098] By observing the envelope, the attenuation stages of the signal can be identified. These attenuation data segments typically start from the maximum amplitude and gradually decrease. The attenuation process is an important indicator for studying the annealing effect of materials, because the physical properties of the material change during annealing, leading to changes in the attenuation mode of the acoustic response signal.

[0099] After identifying the attenuated data segments, data points exceeding the environmental noise threshold are further filtered out. Environmental noise can interfere with signal analysis; therefore, selecting only signal data significantly above the noise level for analysis effectively avoids interference from invalid information and improves the accuracy of the analysis results.

[0100] A nonlinear fitting algorithm is used to fit the selected effective data points into an exponential decay curve. The exponential decay curve is described by the following mathematical expression:

[0101] ;

[0102] in, It is the amplitude at time t. It is the initial amplitude. This is the attenuation constant, representing the rate at which the signal decays. This attenuation constant can be accurately determined using a nonlinear fitting algorithm.

[0103] Based on the fitted exponential decay curve, calculate the time required for the amplitude to decay to a specific percentage of its initial value. For example, a specific percentage can be set (such as decaying to 50% of the initial value), and then the time required to reach that decay percentage can be obtained from the fitted curve. This time is the decay time constant, which reflects the rate of signal decay.

[0104] This method for extracting the decay time constant accurately reflects the acoustic response changes of metallic materials during rapid annealing. The decay time constant provides a quantitative indicator that clearly describes the annealing effect of the metallic material. Compared with traditional methods, this method can more accurately analyze the signal decay process, avoid interference from environmental noise, and improve detection accuracy. Furthermore, the application of a nonlinear fitting algorithm makes the calculation of decay time more precise, providing reliable data support for real-time monitoring and analysis.

[0105] In one possible implementation, firstly, a sample set covering different annealing degrees needs to be prepared. The acquisition of each sample should simultaneously record two types of data: acoustic feature vectors and metallographic data. The acoustic feature vectors reflect the acoustic response of the metallic material under different annealing states, while the metallographic data provides information on the microstructure of the metallic material. By simultaneously acquiring these two types of data, rich information can be provided to the machine learning model, enabling the model to extract features related to changes in the material's microstructure from the acoustic signals, thereby more accurately predicting the annealing effect.

[0106] Next, a regression algorithm with an adjustable kernel function is used to train the sample set. Regression is a typical supervised learning method that makes predictions by learning the mapping relationship between input features and output results between samples. Here, the input features are acoustic feature vectors, and the output is metallographic detection data. The regression model employs kernel function technology, an effective technique for mapping data to a high-dimensional space, which improves the model's ability to fit complex data relationships. The choice of kernel function type is dynamic, adaptively adjusting based on the number of samples and the feature dimension. This means that for different sample sets, the model will select the most suitable kernel function to ensure the best fit.

[0107] To optimize model performance, cross-validation is used for both training and validation. Cross-validation is a technique that divides the sample set into multiple subsets, using one subset as the validation set in turn, and the rest as the training set. Cross-validation reduces overfitting of the model to the training data and improves its generalization ability. Furthermore, during training, the model's hyperparameters (such as kernel parameters and regularization coefficients) are continuously adjusted through cross-validation until an optimal combination of hyperparameters is found. The goal of this step is to minimize the model's prediction error.

[0108] The ultimate goal of the training process is to ensure that the mean absolute error (MAE) between the model's predicted values ​​and the metallographic detection values ​​is below a preset requirement. By continuously optimizing the kernel function and hyperparameters, the model is ultimately able to provide accurate predictions on real data. This evaluation criterion guarantees the model's predictive ability and practicality, ensuring its effectiveness in real-world applications.

[0109] This machine learning training method efficiently establishes a precise mapping relationship between acoustic features and the annealing effect of metallic materials. Compared to traditional detection methods, acoustic response-based detection is not only highly efficient but also provides sensitive and accurate predictions under different annealing states. Through a regression algorithm with an adjustable kernel function, the model can adapt to the characteristics of different sample sets, ensuring its fitting ability on complex data. Simultaneously, cross-validation and hyperparameter optimization steps effectively avoid overfitting, improve the model's generalization ability, and ensure its stable and reliable prediction of the annealing effect of metallic materials in practical applications.

[0110] In one possible implementation, a gradient annealing process is first applied to the same batch of material. This means that during annealing, the material undergoes varying degrees of heating and cooling to induce different recrystallization ratios. Specifically, the annealing process is controlled within a range, gradually varying the recrystallization ratio from 10% to 100%. This process simulates the microstructural changes of the material under different annealing states, providing a rich sample for subsequent acoustic characterization and metallographic analysis. These different annealing states cover the entire process from mild annealing to complete recrystallization, ensuring that the model can learn the differences in acoustic response under different degrees of annealing.

[0111] During sample preparation, the number of samples in each annealing state is adjusted according to the material's microstructure homogeneity. Materials with poor microstructure homogeneity typically have more complex and non-uniform microstructures, which can affect the extraction of acoustic features. Therefore, to better capture these microscopic differences, the number of samples for such materials needs to be increased. By increasing the number of samples, especially in materials with non-uniform microstructures, the training set can be made more comprehensive, reducing bias caused by insufficient samples and thus improving the model's ability to learn different states of the material.

[0112] To ensure the data has good practical application value during acoustic feature acquisition, the acquisition process needs to simulate a production line environment. This means that background vibration noise comparable to that in the online detection process needs to be introduced when acquiring acoustic signals. This design makes the acoustic signals in the training set more closely resemble real-world application scenarios, as online detection is often accompanied by varying degrees of environmental noise. By adding background noise during sample acquisition, the model can learn how to extract effective features from noisy signals during training, thereby improving the model's robustness and reliability in real-world environments.

[0113] This sample set preparation method ensures the representativeness and diversity of the model training data. Gradient annealing covers the entire annealing process from low to high recrystallization, enabling the model to capture acoustic changes at different annealing degrees. Adjusting the sample size based on the material's microstructure homogeneity ensures sufficient samples are collected from materials with heterogeneous microstructures, thereby improving the model's adaptability to complex structures. Simulating the production line environment to collect acoustic features and adding background noise enhances the model's adaptability to real-world production environments, effectively improving its accuracy and robustness in practical applications. In summary, this method improves data quality, enhances the model's predictive ability, and ensures the feasibility and practical application value of the detection method.

[0114] In one possible implementation, the recrystallization ratio of the material is first monitored in real time. When the recrystallization ratio is found to be lower than the set target value, the system generates a temperature control command to accelerate the recrystallization process by increasing the temperature. The amount of temperature increase is proportional to the difference between the target value and the actual measured value. Specifically, the larger the difference between the target value and the measured value, the larger the temperature increase, thus prompting the recrystallization ratio to reach the target requirement as quickly as possible. This step precisely controls the recrystallization process during annealing by adjusting the rate of temperature change.

[0115] Simultaneously, the system monitors the grain size. If the measurement results show that the grain size has exceeded the predetermined upper limit, the system will generate a cooling control command to prevent the grain from continuing to grow. The amount of cooling is proportional to the difference between the measured grain size and the upper limit. That is, the larger the grain size, the greater the temperature reduction, thereby effectively preventing the grain from continuing to grow and ensuring that the grain size remains within the target range.

[0116] To ensure the accuracy and stability of temperature control, the selection of the proportionality coefficient is crucial. This coefficient is determined through process experiments. During the experiments, under a fixed temperature adjustment, the impact rate of temperature changes on various material parameters (such as recrystallization ratio, grain size, etc.) is measured. The average response coefficient is calculated based on the results of multiple experiments to determine the most suitable proportionality coefficient. This proportionality coefficient ensures that the temperature control response matches actual production needs, avoiding excessive or insufficient temperature changes.

[0117] This proportional control-based annealing furnace temperature adjustment method enables precise temperature control, ensuring dynamic adjustment of the recrystallization ratio and grain size during annealing. Firstly, temperature control effectively increases the recrystallization ratio, ensuring the material achieves the desired microstructure. Secondly, temperature control prevents excessive grain growth, guaranteeing grain size stability. This method not only optimizes the annealing process and improves the quality of the metallic material but also responds to changes in real time, automatically adjusting temperature control to improve production efficiency and material performance. Finally, the experimentally determined proportional coefficient allows for more precise temperature adjustment, avoiding quality fluctuations caused by over-adjustment and ensuring the reliability and consistency of the annealing effect.

[0118] In one possible implementation, to ensure the model can adapt to changes in material batches in real time, it is first necessary to periodically extract strip samples from the production line for metallographic examination. Metallographic examination provides information on the material's microstructure, such as recrystallization ratio and grain size, which is crucial for understanding the annealing effect. The examination cycle is determined based on the stability of the material batch. For materials with large batch fluctuations, the examination cycle will be shortened accordingly; while for materials with good batch stability, the examination cycle can be appropriately extended. This measure ensures the representativeness and timeliness of sampling, avoiding the impact of large fluctuations in material properties on the accuracy of the test results.

[0119] After each metallographic examination, the obtained metallographic data and corresponding acoustic feature vectors are added to the training set. This data fusion process allows the model to learn from a continuous stream of new samples, thereby capturing more patterns and feature changes in the samples. This step ensures real-time updates to the training set data, maintaining the model's dynamic adaptability.

[0120] When the number of new samples reaches a predetermined proportion of the original dataset, the system will trigger a retraining of the machine learning model. This proportion is typically set based on the data volume variation pattern and the model's accuracy requirements. By introducing new samples, the model can learn more data reflecting the current production status, further improving its predictive ability and accuracy for the annealing process of metal materials.

[0121] After the new model completes training, the old and new models will run in parallel for a predetermined period. This parallel running allows for comparison of their prediction performance under the same conditions, evaluating whether the new model outperforms the old one. This comparison step ensures the rationality of the model update and avoids errors caused by instability or overfitting in the new model.

[0122] After the old and new models have run in parallel, the system will select the model with the smaller prediction error as the current model. This selection process is based on the accuracy of the actual prediction results, ensuring that the final selected model can provide the best prediction performance and avoiding the impact on the detection accuracy of the annealing effect due to model obsolescence or reduced accuracy.

[0123] Regularly sampling and metallographic examination ensured the representativeness and timeliness of the test data, avoiding prediction bias caused by data lag. Secondly, the addition of new samples allowed the model to gradually adapt to changes in material batches, improving its adaptability to different batches. A proportionally controlled model update mechanism ensured the model remained in an optimal state, avoiding overtraining or undertraining. The parallel operation of old and new models effectively reduced the risks associated with model updates, making each update more reliable. This online update mechanism ensured the long-term stability and reliability of the annealing effect detection method.

[0124] The following examples will illustrate this in detail:

[0125] This embodiment uses the steel annealing process as an example to describe a method for rapid annealing effect detection of metallic materials based on acoustic response. This method combines metallographic examination data with acoustic characteristics, and uses a machine learning model to update and adjust the annealing process in real time, ensuring that the annealing effect of the material meets the predetermined standard.

[0126] In this embodiment of the invention, the metal material type is low-alloy high-strength steel (such as Q345 steel), which is commonly used in high-strength components in fields such as construction and bridges.

[0127] Annealing process: Annealing is performed using a medium-frequency electric furnace. When the steel strip passes through the annealing furnace, the temperature is controlled to rise from 300°C to 900°C, and then gradually cooled to room temperature.

[0128] Sensors are installed to collect acoustic response data as the steel strip passes through the annealing furnace. The acoustic signals collected by the sensors are converted into spectral data through Fourier transform to obtain frequency response characteristics. The sampling frequency range is set from 20Hz to 5kHz, with sampling once per second and a sampling duration of 10 seconds.

[0129] The acoustic signal was converted to a frequency domain signal using a Fast Fourier Transform (FFT), and the following features were extracted:

[0130] Peak frequency: The frequency at which the highest energy is present in the acoustic response.

[0131] Total Power: The power spectral area of ​​a signal, representing the total energy of the signal.

[0132] Spectral characteristics: The state of the material is further analyzed by calculating the morphological characteristics of the spectrum (such as average amplitude, bandwidth, etc.).

[0133] A strip sample is taken every 500 meters of the production line for metallographic inspection, measuring the grain size (D) and recrystallization ratio (R). The metallographic inspection uses standard metallographic sectioning methods and is performed using an optical microscope or a scanning electron microscope (SEM).

[0134] Assume there are currently 1000 samples, each containing acoustic features (such as peak frequency, total power, spectral characteristics, etc.) and metallographic data (such as grain size, recrystallization ratio).

[0135] The training set data includes a pair of acoustic features and metallographic data, with the target variables being grain size (D) and recrystallization ratio (R).

[0136] A Support Vector Machine (SVM) regression model was selected for training. The model's goal is to predict metallographic data based on acoustic features. The model training steps are as follows:

[0137] The acoustic features of the training set are normalized so that the mean of the data is 0 and the standard deviation is 1.

[0138] We used the scikit-learn library in Python to train a support vector machine regression model, specifically using the RBF kernel function, with the following training parameters:

[0139] C (regularization parameter): 10

[0140] γ (kernel function parameter): 0.01

[0141] Epsilon (tolerance): 0.1

[0142] After every 1000 meters of strip is produced, a sample is taken for metallographic examination to measure its grain size (D) and recrystallization ratio (R).

[0143] The newly added metallographic data and acoustic features (acquired in real time via acoustic sensors) are added to the training set. Assume that 10 data points are added each time.

[0144] When the number of new samples reaches 10% of the original training set (i.e., 100 new samples), the machine learning model is retrained.

[0145] When new data is added, the old and new models are run in parallel, their prediction errors on the new test set are compared, and the model with the smaller error is selected as the current model to use.

[0146] Mean squared error (MSE) was chosen as the evaluation criterion.

[0147] MSE of the old model: 0.022

[0148] The MSE of the new model is 0.018.

[0149] Based on the comparison results, the new model was chosen to continue using it because its prediction error was smaller.

[0150] Based on metallographic examination results and acoustic response data, the annealing furnace temperature can be adjusted in real time to ensure the material reaches the expected grain size and recrystallization ratio. The temperature adjustment formula is set as follows:

[0151] ;

[0152] in:

[0153] The temperature value that needs to be adjusted (unit: °C).

[0154] and The target grain size and recrystallization ratio are determined.

[0155] and This is the current metallographic data.

[0156] =1.5 and =2.0 is the adjustment coefficient, which is determined based on experimental data.

[0157] In a batch production process, this method was used for annealing, followed by metallographic examination. The annealing effect using this method was compared with that of the traditional annealing method.

[0158] Traditional annealing methods lack real-time acoustic response monitoring, rely on fixed temperature and time settings, resulting in poor grain size uniformity and significant fluctuations in recrystallization ratio.

[0159] The method of this invention involves adjusting the temperature of the annealing furnace in real time and adjusting the annealing process based on acoustic characteristics and metallographic data to ensure a more uniform annealing effect.

[0160] Traditional method: standard deviation of grain size uniformity is 12.5 μm, and recrystallization ratio fluctuation is 7%.

[0161] The method of this invention has a standard deviation of 5.8 μm for grain size uniformity and a recrystallization ratio fluctuation of 3%.

[0162] The comparison shows that the method of the present invention can significantly improve the uniformity of the annealing process and reduce the fluctuation of grain size, thereby improving the mechanical properties and reliability of steel.

[0163] This invention combines acoustic response with metallographic data, and employs a machine learning-based online update mechanism to monitor and adjust the annealing effect of metallic materials in real time. Comparative experiments demonstrate that this method effectively improves the control precision of the annealing process, reduces material property fluctuations during production, and enhances the quality and stability of the final product.

[0164] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0165] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting the rapid annealing effect of metallic materials based on acoustic response, characterized in that, include: Step 1: Acoustic excitation and synchronous acquisition: A broadband linear frequency modulated acoustic signal is emitted to a moving metal strip through a piezoelectric ceramic transducer array, and its frequency band covers the resonant modes in the thickness direction of the material. During the excitation, a non-contact vibration sensor is used to collect the vibration response signal of the strip surface, and an infrared thermometer is used to obtain the surface temperature distribution. Step 2: Multiphysics Coupling Compensation: Based on temperature distribution data and the temperature dependence of material acoustic properties, the phase compensation amount of the vibration signal is calculated. The environmental vibration spectrum is obtained by the production line mechanical vibration monitoring device, and the excitation response component is separated from the original vibration signal by adaptive filtering. Step 3: Extraction of Annealing-Sensitive Features: Time-frequency analysis was performed on the compensated signal, and the sub-band with the largest change in energy concentration was selected based on the Brillouin scattering characteristics of the metal lattice. Extract the energy proportion parameter and signal attenuation time constant of the sub-band; The method for selecting the sub-band includes: Wavelet packet decomposition is performed on the compensated signal to obtain a set of sub-bands covering the entire frequency band; Calculate the difference in energy concentration for each sub-band between the fully annealed and unannealed states; Select the combination of continuous sub-bands with the largest difference in energy concentration as the characteristic band, and its total bandwidth shall not exceed a preset proportion of the full bandwidth; The energy concentration is defined as the ratio of sub-band signal energy to full-band signal energy; Step 4: Dynamic Quality Mapping and Control Input the real-time feature vector into the pre-trained machine learning model, and output the recrystallization ratio and grain size; When the output parameters exceed the process threshold, the control parameters of the annealing furnace are adjusted.

2. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, The method for determining the frequency band in step 1 includes: A pre-established mapping relationship library between material thickness and resonant frequency was created: samples of the same material with varying thickness gradients were prepared, and a frequency sweep excitation experiment was performed on each sample to record the peak distribution range of its resonant frequency. Based on the measured thickness of the strip detected online, the corresponding lower and upper frequency limits are matched from the mapping relationship library; The sweep range of the linear frequency modulation signal is set to the matched frequency interval, and the sweep period is dynamically adjusted according to the speed of the strip movement to ensure that each detection area is fully excited.

3. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, The calculation of the phase compensation amount in step 2 includes: The temperature-affected zones are divided along the sound wave propagation path, and the length of each zone is determined based on the spatial density of infrared temperature measurement points. Obtain the average temperature measurement value of each section, and calculate the sound wave propagation time offset of that section by combining the calibration curve of the material sound velocity with temperature. The propagation time offsets of each segment are accumulated and converted into the phase compensation amount of the vibration signal; The calibration curve was obtained through laboratory calibration: the sound wave propagation speed of the same material at different temperature points was measured under a controlled temperature environment, and the sound speed-temperature function relationship was fitted.

4. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, The implementation of the adaptive filtering includes: The vibration spectrum collected by the mechanical vibration monitoring device is used as the reference noise signal; A finite impulse response filter is constructed, and the filter coefficients are dynamically adjusted through an iterative algorithm to maximize the similarity between the filter output signal and the reference noise signal. The pure acoustic response component is obtained by subtracting the filter output signal from the original vibration signal. The termination condition of the iterative algorithm is that the power spectral density of the residual signal drops below a set ratio.

5. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, The extraction of the decay time constant includes: Envelope extraction is performed on the characteristic sub-frequency band signal to identify data segments in the signal attenuation phase; Select data points in the attenuation data segment that are above the environmental noise threshold; The selected data points are fitted into an exponential decay curve using a nonlinear fitting algorithm. The time required for the amplitude to decay to a specific proportion of the initial value is calculated based on the fitted curve.

6. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, The training methods for the machine learning model include: A sample set covering different annealing degrees was prepared, and acoustic feature vectors and metallographic detection data were collected simultaneously for each sample; A regression algorithm with adjustable kernel function is used for training, and the kernel function type is adaptively selected according to the number of samples and feature dimensions; The model hyperparameters are optimized through cross-validation until the mean absolute error between the predicted value and the metallographic detection value is lower than the preset requirement.

7. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 6, characterized in that, The preparation of the sample set includes: Gradient annealing process is applied to the same batch of materials to make the recrystallization ratio cover the range of 10% to 100%; The number of samples for each annealed state is determined based on the material's microstructure uniformity; for materials with poor microstructure uniformity, the number of samples is increased. Acoustic feature acquisition simulates the production line environment, introducing background vibration noise of the same magnitude as that used in online detection.

8. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, The adjustment of the control parameters for the trigger annealing furnace includes: When the recrystallization ratio is lower than the target value, a temperature control command is generated, and the temperature rise is proportional to the difference between the target value and the measured value. When the grain size exceeds the upper limit, a cooling control command is generated, and the temperature drop is proportional to the difference between the measured value and the upper limit. The proportionality coefficient is determined through process experiments: the rate of change of parameters is measured under a fixed temperature adjustment, and the average response coefficient of multiple experiments is taken.

9. The method for detecting the rapid annealing effect of metallic materials based on acoustic response according to claim 1, characterized in that, It also includes the online model update step: Metallographic examination of strip samples is conducted periodically, with the examination cycle set according to the stability of the material batch. Add the acoustic feature vectors and metallographic data of the new samples to the training set; When the number of new samples reaches a set proportion of the original dataset, the machine learning model is retrained. After running the old and new models in parallel for a predetermined period, the model with the smaller prediction error is selected as the current model to be used.

Images