A Failure Diagnosis Method for Heat-Resistant Steel Based on Acoustic Energy Corrected Laser-Induced Breakdown Spectroscopy

Abstract

This invention discloses a failure diagnosis method for heat-resistant steel based on laser-induced breakdown spectroscopy (LIBS) with acoustic energy correction, belonging to the field of heat-resistant steel failure detection technology. The method first simultaneously acquires spectral and acoustic signals of heat-resistant steel samples using a LIBS experimental platform. The acoustic energy value obtained by integrating the main acoustic band is used to correct the original spectrum in real time, reducing matrix effects and plasma wave interference. Then, a multi-threshold segmentation combined with an average spectrum comparison mechanism is employed to reduce the dimensionality of the corrected high-dimensional spectral data. Finally, a radial basis function kernel support vector machine collaborative model driven by recursive feature elimination, linear discriminant analysis, and particle swarm optimization is constructed to achieve high-precision determination of the aging level of heat-resistant steel. This invention utilizes the complementary advantages of photoacoustic multimodal signals, possessing the characteristics of non-destructive testing, strong anti-interference ability, and good model generalization performance. It can be adapted to complex industrial scenarios, providing reliable technical support for the safe operation and maintenance of high-temperature critical components in industries such as power and machinery.

Description

Technical Field

[0001] This invention belongs to the field of heat-resistant steel failure detection technology, specifically relating to a method for diagnosing heat-resistant steel failure based on acoustic energy-corrected laser-induced breakdown spectroscopy. Background Technology

[0002] Heat-resistant steel is a core material for critical high-temperature components in power plants, such as boilers, superheaters, and steam pipes. Its performance directly determines whether the unit can operate safely and stably for a long time under harsh conditions such as high temperature and high pressure. However, the actual operating environment of the unit is complex and variable. Different parts of the components experience significant differences in temperature, stress, and media corrosion conditions, leading to uneven degradation of the material's microstructure and performance. This unevenness makes the remaining lifespan and condition distribution of the critical components of the entire unit extremely complex, increasing the difficulty of operation and maintenance and limiting the full realization of the unit's overall energy efficiency and operational potential. Therefore, accurate diagnosis of the entire life cycle status of in-service heat-resistant steel and identification of key weak points in performance degradation have become core requirements for preventing major accidents, optimizing operation and maintenance strategies, and ensuring the safe operation of the unit under harsh conditions.

[0003] Accurate assessment of the service condition of heat-resistant steel requires precise judgment of its microstructure and the evolution of its mechanical properties. Currently, the mainstream condition assessment methods in the industry are mainly divided into two categories: one is the mechanical property-based experimental method, which is based on classical creep and endurance strength theories and infers the remaining life of the material through accelerated creep experiments on samples; the other is the life prediction method based on microstructure analysis, which relies on metallography and physical metallurgy theories to determine the material's damage state by observing and quantifying microstructure parameters such as precipitates and dislocation structures. However, both methods have significant limitations: they are both destructive or micro-destructive testing methods, which not only have long testing cycles and low efficiency, but are also limited by sampling conditions, making large-scale field application difficult.

[0004] Although conventional non-destructive testing techniques such as ultrasonic testing and eddy current testing have been applied in engineering, their sensitivity to the microstructural evolution of materials in the early aging stages is generally insufficient, failing to capture subtle performance degradation signals and making it difficult to meet the needs of accurate diagnosis of the aging state of heat-resistant steel. Especially for units in operation, frequent load fluctuations caused by peak-shaving operations result in significant transient and alternating characteristics in the temperature and stress experienced by critical components, with the differences under different operating conditions further amplified. Given the need to consider the complex operating history of the unit throughout its entire life cycle, accurately characterizing the time-varying patterns of heat-resistant steel performance degradation has become a core issue that urgently needs to be addressed in its service status assessment and remaining life prediction.

[0005] Laser-induced breakdown spectroscopy (LIBS), as an emerging non-destructive analysis method, can directly correlate the aging state of materials by analyzing the elemental and spectral characteristics of the material surface, providing a new path to overcome the shortcomings of traditional methods. However, in complex industrial applications, the spectral signal of LIBS is easily affected by matrix effects, material surface conditions, and plasma fluctuations, making it difficult to meet the stability and accuracy requirements of the analysis results. To address this issue, the paper "Chen Kaiqing et al. Evaluation of Aging Grade of T91 Heat-resistant Steel Coupled with Laser-Induced Photoacoustic Multimodal Information [J]. Journal of Shanghai University of Science and Technology, 2025(4)" proposes a new method for aging grade evaluation based on LIBS photoacoustic multimodal information fusion. The method converts the acoustic time-domain data into acoustic spectrum data through fast Fourier transform, normalizes the spectral data and acoustic data, filters the fused data, and builds an SVM model through LDA dimensionality reduction, achieving a high classification accuracy.

[0006] However, the existing scheme still has two key shortcomings: First, its research object is limited to artificially prepared aging samples, and it does not carry out prediction and verification for actual service aging samples. The aging influencing factors of actual samples are more complex and the classification difficulty is much higher than that of artificial samples. Traditional SVM models are difficult to meet the requirements of strong classification performance and robustness in actual scenarios. Second, the scheme directly couples spectral data with acoustic spectrum data without optimizing for data volatility. As a result, there is still room for improvement in data stability, which cannot fully guarantee the diagnostic reliability in actual industrial scenarios. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a failure diagnosis method for heat-resistant steel based on laser-induced breakdown spectroscopy technology with acoustic energy correction. First, plasma spectra and acoustic signals are simultaneously acquired, and the original spectrum is corrected in real time using acoustic energy data to improve signal stability and reliability. Then, multiple threshold segmentation is used to reduce the dimensionality of the corrected spectrum, and a mean spectral comparison mechanism is combined to screen key features characterizing the aging state. Finally, a support vector machine (SVM) method integrating recursive feature elimination (RFE), particle swarm optimization (PSO), and linear discriminant analysis (LDA) is fused to achieve high-precision classification of the aging level of heat-resistant steel. This method fully leverages the complementary advantages of photoacoustic multimodal signals, constructing a complete technical system integrating spectral correction, feature selection, and intelligent diagnosis. It features strong anti-interference capabilities, high feature extraction efficiency, and good model generalization performance, enabling accurate state assessment of heat-resistant steel materials in complex industrial environments and providing core technical support for the safe operation and maintenance of high-temperature critical components in industries such as power and machinery.

[0008] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for diagnosing the failure of heat-resistant steel based on acoustic energy-corrected laser-induced breakdown spectroscopy, comprising the following steps:

[0009] S1. Build a laser-induced breakdown spectroscopy (LIBS) experimental platform and simultaneously acquire spectral and acoustic signals of heat-resistant steel samples;

[0010] S2. Calculate the acoustic energy value E of the sound wave signal in the main band. a Utilizing the sound energy value E a The spectral signal is corrected to suppress spectral interference caused by laser energy fluctuations;

[0011] S3. Perform multiple threshold segmentation on the corrected spectral data to screen the core characteristic wavelength points related to the aging level and achieve data dimensionality reduction.

[0012] S4. Input the dimensionality-reduced spectral data into the machine learning model. Through feature selection, secondary dimensionality reduction and parameter optimization, construct a classification model for the aging level of heat-resistant steel and output the failure diagnosis results of the heat-resistant steel sample.

[0013] Furthermore, the specific operation process of step S2 includes:

[0014] S2.1 Prepare artificial aging test specimens and obtain actual service test specimens. The artificial aging test specimens are obtained through variable stress high temperature aging tests and are marked with a clear aging level.

[0015] S2.2 Integrate the acquired acoustic signal within the preset main band and calculate the acoustic energy value E. a The formula is as follows:

[0016]

[0017] In the formula, x(t) is the acoustic signal in the form of a time function, and t1 and t2 are the start and end times of the main band, respectively;

[0018] S2.3, Utilizing the sound energy value E a The original spectral intensity I0 under the same laser pulse is corrected to obtain a stable corrected spectral intensity I. E The correction formula is:

[0019]

[0020] Among them, I E I0 represents the original spectral intensity, where I is the corrected spectral intensity.

[0021] Furthermore, the specific operation process of step S3 includes:

[0022] S3.1. For the corrected spectra of artificial samples with different aging levels, calculate the average value for each wavelength to obtain the average spectrum A;

[0023] S3.2. Take the quotient of the corrected spectrum of the actual sample with the average spectrum A wavelength by wavelength to generate dimensionless data D;

[0024] S3.3 Calculate the absolute difference ΔD between dimensionless data D and 1, select the first n wavelength points in ΔD that are closest to 1, and generate a binary array M, where n = 1500~2500;

[0025] S3.4 Summing multiple sets of binary arrays M yields a feature importance array S. The wavelength points corresponding to the first m maximum values ​​in S are selected to generate the final filtering array F, where m = 1500~2500. F is multiplied wavelength by wavelength with the corrected spectrum to achieve data dimensionality reduction.

[0026] Furthermore, the specific operation process of step S4 includes:

[0027] S4.1. Use the reduced-dimensional spectral data of artificial samples as the training set and label it with aging level, and use the reduced-dimensional spectral data of actual samples as the test set to achieve separation of training and test verification data.

[0028] S4.2 Filtering core features through recursive feature elimination (RFE): The initial number of features for recursive feature elimination directly follows the dimensions of the dataset after multi-threshold segmentation in step S3—this dimension has already completed high-dimensional noise removal, providing a high-quality foundation for feature selection; the selection step size is set to 50 to balance feature selection efficiency and the retention of core information; the upper and lower limits of parameters are determined based on the feature contribution law: setting the lower limit of features to 1000 can avoid premature removal of features with weak discriminative power but cumulative contribution due to an excessively low initial number of features, and prevent the feature space from being oversimplified and damaging the model performance; combined with the characteristics of the multi-band distribution of spectral features, the upper limit of features is set to 2000 dimensions—this upper limit can ensure that all potential discriminative features of each band are included in the search space, and can also avoid feature redundancy, achieving a balance between the comprehensiveness of feature selection and the stability of the model;

[0029] S4.3 Construct a radial basis function (RBF) kernel support vector machine (SVM) model optimized by particle swarm optimization (PSO) algorithm. Input the dimensionality-reduced feature data into the model and output the aging level of the actual sample, the overall classification accuracy of the model, and the misclassification rate.

[0030] Furthermore, a particle swarm optimization algorithm is used to jointly optimize the penalty parameter C and the kernel function parameter γ of the support vector machine, wherein the search range of the penalty parameter C is 0.1~100, and the search range of the kernel function parameter γ is 0.001~1; the parameters of the particle swarm optimization algorithm itself are set as follows: learning factor is 2.0~2.3, population size is 100~110, and maximum number of iterations is 100~120.

[0031] Furthermore, as a preferred embodiment, the parameters of the particle swarm optimization algorithm are set as follows: learning factor of 2, population size of 100, and maximum number of iterations of 100.

[0032] Furthermore, the heat-resistant steel is at least one of T91 steel, P92 steel, and TP347H steel.

[0033] The beneficial effects of this invention are as follows:

[0034] 1. Strong model generalization ability and adaptability to real industrial scenarios: This application builds a machine learning model based on artificial samples, and uses actual aging samples only as test sets to participate in model evaluation and not in training; this mode of training with artificial samples and verifying with actual samples effectively avoids overfitting of the model to specific samples, significantly enhances the model's adaptability to complex real-world scenarios, and solves the shortcomings of traditional methods that can only handle artificial samples and are difficult to extend to the field.

[0035] 2. This application innovatively extracts "acoustic energy features" with clear physical meaning from acoustic time-domain signals and uses them to correct LIBS spectral data. This correction method can specifically suppress interference factors such as matrix effects and plasma fluctuations, significantly reduce the volatility of spectral signals, and provide a stable and reliable data source for subsequent feature extraction and modeling. Compared with the traditional scheme of directly coupling photoacoustic data, the data stability is significantly improved.

[0036] 3. This application proposes a "multi-threshold segmentation" method, which performs preliminary dimensionality reduction and pre-screening on high-dimensional spectral data before feature screening. By constructing an "average spectral comparison mechanism", key wavelength points that can characterize the aging state are accurately screened out, and redundant noise features are eliminated. This not only compresses the data dimension of subsequent modeling from the original high dimension to a reasonable range, but also reduces the interference of invalid features on the model, while improving modeling efficiency and classification accuracy.

[0037] 4. This application addresses the significantly increased classification difficulty when "extending from artificial samples to actual samples" by introducing a particle swarm optimization (PSO) algorithm to globally optimize the penalty parameter C and kernel function parameter γ of the SVM. Combined with recursive feature elimination (RFE) and linear discriminant analysis (LDA), a collaborative modeling scheme of "RFE+LDA+PSO-SVM" is formed, ultimately achieving high-precision classification of the aging grades of actual heat-resistant steel. Compared to the traditional SVM model, the classification accuracy is improved by 10%-15%.

[0038] 5. This application fully utilizes the complementary advantages of spectral signals (reflecting changes in elemental composition) and acoustic signals (reflecting the physical state of materials) to construct a complete technical system for spectral correction, feature screening, and intelligent diagnosis. This system has the characteristics of strong anti-interference ability (acoustic correction), high feature extraction efficiency (multiple threshold segmentation + RFE), and good model generalization performance (actual sample verification). It can realize fully automatic, rapid (single diagnosis ≤30 seconds), and accurate assessment of the aging state of heat-resistant steel in complex industrial environments, providing reliable technical support for unit operation and maintenance decisions.

[0039] 6. The heat-resistant steel failure diagnosis method proposed in this application is based entirely on LIBS non-destructive testing technology. It does not require sampling or damage treatment of in-service heat-resistant steel components, and data acquisition and diagnosis can be completed directly on-site. This not only avoids the damage to the structural integrity of components caused by traditional destructive testing and reduces operation and maintenance risks, but also eliminates cumbersome steps such as sampling and laboratory analysis, significantly shortening the testing cycle and reducing operation and maintenance costs. It is more suitable for on-site testing needs in industrial scenarios such as power plants. Attached Figure Description

[0040] Figure 1 These are metallographic images of artificial samples with different aging levels;

[0041] Figure 2 This is a structural diagram of the experimental system for synchronous acquisition of photoacoustic signals in Example 1; wherein, 1-laser, 2-delay generator, 3-spectrometer, 4-light-collecting lens, 5-three-dimensional translation stage, 6-microphone, 7-reflecting lens, 8-focusing lens, 9-computer, 10-sample;

[0042] Figure 3 These are spectral signal diagrams of five artificial samples;

[0043] Figure 4 The acoustic time-domain signal diagram of artificial sample T34;

[0044] Figure 5 These are spectral signal diagrams of five actual samples;

[0045] Figure 6 A flowchart for aging grade diagnosis of heat-resistant steel based on acoustic energy correction and machine learning;

[0046] Figure 7 This is a comparison chart showing the effect of acoustic energy correction on improving the stability of aging characteristic data of heat-resistant steel. Detailed Implementation

[0047] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] Example 1

[0049] This embodiment discloses a failure diagnosis method for heat-resistant steel based on laser-induced breakdown spectroscopy (LIBS) technology with acoustic energy correction. The method simultaneously acquires the spectral signal (reflecting the core characteristics of aging) and acoustic signal (a stable calibration basis) of the heat-resistant steel through a laser-induced breakdown spectroscopy (LIBS) platform. First, the acoustic energy is used to correct the interfered spectral data, improving data reliability. Then, multiple threshold segmentation is used to filter useful features, reducing data dimensionality and improving efficiency. Finally, a support vector machine (SVM) model combining recursive feature elimination (RFE) and particle swarm optimization (PSO) is used to automatically output the aging level, achieving full automation of the "acquisition, correction, dimensionality reduction, and diagnosis" process.

[0050] The specific steps are as follows:

[0051] S1. Build a LIBS experimental platform, synchronously collect photoacoustic signals, and use sound wave energy to eliminate spectral interference;

[0052] S1.1 Setting up the LIBS experimental platform

[0053] T91, a novel ferritic heat-resistant steel commonly used in supercritical and ultra-supercritical boilers of thermal power plants, was selected as the test substrate. Artificially aged samples were prepared through variable stress high-temperature aging tests.

[0054] 1) Cut samples from the same batch of T91 steel and process them into two metal samples with completely identical shapes and specifications to ensure that the initial state of the samples is uniform and to eliminate the interference of processing differences on subsequent aging tests;

[0055] 2) Two metal samples were suspended in a vacuum muffle furnace and constant tensile stress was applied by suspending counterweights below the samples to simulate the tensile stress state in actual service. Accelerated aging tests were carried out using two different load-temperature combinations with the following parameters: ① Load 3500N, heating temperature 675℃; ② Load 2500N, heating temperature 700℃.

[0056] 3) The condition of the sample is monitored in real time during the test. When the sample with a heating temperature of 700℃ breaks under the maximum stress, the entire test is terminated. At this time, the uniform duration of the two sets of tests is 895.7 hours. This duration ensures that the aging effect is fully manifested and that the sample breaks as a clear endpoint, thus ensuring the objectivity of the test data.

[0057] 4) After the test, the two sets of tested samples and one set of original samples (as a reference) were refined. Two sub-samples were cut from different areas of each set of tested samples, resulting in four sub-samples from the two sets of tested samples. Combined with the original reference sample, a total of five samples were obtained. These samples underwent metallographic preparation processes such as grinding, polishing, and etching to obtain five artificial samples covering different aging levels. Subsequently, metallographic microanalysis was performed on all samples according to the Chinese power industry standard "DL / T 884-2019 Technical Guidelines for Metallographic Inspection and Evaluation of Thermal Power Plants". Metallographic images of different aging levels are shown below. Figure 1 As shown.

[0058] After undergoing high-temperature failure testing under alternating stress, five artificial samples were obtained, designated as T11, T12, T22, T32, and T34. The corresponding aging levels are shown in Table 1.

[0059] Table 1. Aging levels of artificial samples

[0060] label Aging level T34 1 T32 2 T22 3 T12 4 T11 5

[0061] Actual sample preparation: Cut five small samples of the same size from the high-temperature equipment on site, and polish the surface with sandpaper until smooth (to remove oxide scale and avoid contaminating the signal). The five actual samples are recorded as S31, S11, S41, S51, and S61.

[0062] Equipment parameter settings: This application uses an experimental system integrating optical and acoustic signal acquisition for synchronous acquisition of photoacoustic signals. The system structure is as follows: Figure 2 As shown, the system mainly consists of core components such as a laser 1, a delay generator 2, a spectrometer 3, a light-collecting lens 4, a three-dimensional translation stage 5, a microphone 6, a reflecting lens 7, a focusing lens 8, and a computer 9. This system uses an Nd:YAG Q-switched pulsed laser (Eazy, Quantel) as the plasma excitation source. The laser operates at a wavelength of 1064 nm, with a pulse width of approximately 6 ns, and a maximum single-pulse output energy of 300 mJ. After being guided by several laser reflecting lenses 7, the laser beam is focused by the focusing lens 8 onto the surface of the sample 10 to form ablation points. The sample 10 is fixed on the three-dimensional translation stage 5, and the laser's position on the sample 10 surface is precisely controlled by adjusting the position of the translation stage.

[0063] The spectral detection section employs a four-channel spectrometer (AvaSpec-2048, Avantes), with a wavelength coverage of 179-829 nm and a spectral resolution of approximately 0.8-0.12 nm, meeting the requirements for simultaneous monitoring of multi-element characteristic spectral lines. The light-collecting probe is arranged laterally at 45° and, in conjunction with the light-collecting lens 4, focuses the plasma emitted light, improving the light signal collection efficiency. Both the spectrometer 3 and the laser 1 are connected and triggered via a delay generator 2, enabling precise control of parameters such as laser pulse energy, repetition frequency, and spectral acquisition delay time.

[0064] Acoustic signal acquisition employed a condenser microphone (ATR2500X-USB, Audio-Technica) to record the acoustic signals generated by laser-induced plasma. This microphone has an effective frequency response range of 20Hz-15000Hz, a cardioid polar pattern, and some ability to suppress ambient noise; its 24-bit quantization accuracy and 192kHz sampling rate ensure high-fidelity digital acquisition of the acoustic signals. In the experiment, microphone 6 was placed horizontally approximately 1m from the surface of sample 10, with no obstructions between them to minimize attenuation and reflection interference during sound wave propagation. The computer 9 used Audacity software for real-time recording and storage of the acoustic signals.

[0065] To achieve a strict correspondence between the spectral signal and the acoustic signal on the time axis, a movable baffle was installed at the emission port of laser 1 as a synchronization reference.

[0066] At the start of the experiment, with the baffle raised, spectral and acoustic signals were recorded simultaneously for approximately 3-5 seconds. The baffle was then lowered to block laser propagation, stopping plasma excitation. The signal intensity recorded by the spectrometer 3 and the recording software rapidly decayed to near zero and remained so for about 3 seconds. The baffle was then removed again, allowing the laser beam to be refocused on the surface of sample 10 through the optical system, generating new plasma. The subsequently acquired spectral and acoustic data can then be considered as synchronization signals corresponding to the same laser ablation process. Through this process of "mechanical baffle-signal interruption-re-excitation," precise time synchronization and correspondence between the spectral and acoustic channels were achieved.

[0067] Figure 3 The spectral signal diagrams of five artificial samples are shown. Figure 4 The image shows the acoustic time-domain signal of artificial sample T34, with the main band being 0.3~0.45ms, which best reflects the plasma state. Figure 5 The images show the spectral signals of five actual samples.

[0068] S1.2 Calculate the acoustic energy of the main band of the sound wave: Figure 4 Taking the acoustic time-domain signal diagram of the artificial sample T34 as an example, we can analyze... Figure 4By integrating the signal in the main band of the mid-sound wave, the acoustic energy value E of the artificial sample T34 can be obtained. a The formula is as follows:

[0069]

[0070] In the formula, x(t) represents the acoustic signal output by the microphone, a function of time. t1 and t2 define the integration time window, which here represent the start and end times of the main waveband, respectively (previous experiments have verified that the acoustic energy and laser energy are most strongly correlated within this window, accurately reflecting laser fluctuations). The physical meaning of acoustic energy lies in the mechanical kinetic energy converted from the portion of internal energy that fails to be released in the form of light radiation during the cooling process of laser-induced plasma. This energy drives the expansion of the plasma plume and excites shock waves, ultimately being detected as an acoustic signal. Its intensity directly reflects the energy distribution ratio between the radiation channel and the mechanical kinetic energy channel; a lower spectral intensity indicates a higher acoustic energy intensity. Therefore, it can serve as a calibration scale for spectral correction. This step yields the acoustic energy values ​​for five artificial samples and five actual samples.

[0071] S1.3 Correcting spectral data with acoustic energy: Acoustic energy E calculated using S1.2 a The original spectral intensity I0 acquired under the same laser pulse is corrected to obtain a stable corrected spectrum I. E The formula is as follows:

[0072]

[0073] In the formula, I E I0 represents the spectral intensity after correction; I0 represents the original spectral intensity.

[0074] This step yields the corrected spectra I of five artificial samples and five actual samples. E .

[0075] S2. Multiple threshold segmentation for data dimensionality reduction

[0076] The corrected spectral data is high-dimensional data, which contains a lot of redundant noise unrelated to aging and abnormal data that deviates from aging characteristics. It is necessary to screen the core features through multiple threshold segmentation to achieve data dimensionality compression and provide accurate input for subsequent modeling.

[0077] S2.1 Calculate the average spectrum A (establish a benchmark and eliminate component interference): Calculate the average value of the corrected spectra of the five artificial samples with different aging levels from step S1, wavelength by wavelength, to obtain the average spectrum A. The average spectrum A serves as a spectral consistency reference and can be used to characterize the statistical central features of the training samples in the overall spectral morphology. By comparing the spectrum of a single sample with A, random interference introduced by local ablation instability, plasma fluctuations, and small component fluctuations can be effectively reduced, allowing subsequent feature screening to focus more on spectral features that are stable in most samples.

[0078] S2.2 Generation of dimensionless data and feature screening: The corrected spectrum of the actual sample is divided wavelength by wavelength by the average spectrum A to obtain dimensionless data D. D represents the relative intensity ratio of the actual spectrum to the reference spectrum. The degree to which its value deviates from 1 directly reflects the difference between the wavelength point and the reference state, providing a quantitative basis for subsequent feature screening.

[0079] Based on dimensionless data D, the following steps are used to accurately screen the core characteristics of aging:

[0080] ① Calculate the absolute difference: Calculate using the formula ΔD=|D-1|, which essentially quantifies the degree of deviation between each wavelength point of the actual sample and the reference spectrum. The smaller ΔD is, the higher the consistency and stability of the wavelength point relative to the reference spectrum under different aging states. It can effectively reduce the uncertainty introduced by composition fluctuations and measurement noise, and is therefore more suitable as the base band for subsequent aging feature modeling. The larger ΔD is, the more likely the wavelength point contains both aging differences and large inter-class fluctuations. Its stability is insufficient and it is easy to introduce modeling noise. Therefore, it is preferentially eliminated in the pre-screening stage.

[0081] ② Feature selection and binarization: Set the number of core features to be selected to n=2000 (through previous experiments, it was verified that when n=2000, it can retain the difference features of all aging levels and remove noise to the greatest extent). Sort the values ​​in each group ΔD from smallest to largest, and mark the wavelength points corresponding to the first 2000 minimum values ​​after sorting as 1 (representing the core aging feature points to be retained), and mark the remaining wavelength points as 0 (representing the irrelevant noise points to be removed), thus obtaining multiple sets of binarized arrays M.

[0082] S2.4 Obtain the final dimensionality reduction dataset: Sum the binarized array M of all actual samples to obtain the feature importance array S (the larger the value of a wavelength point in S, the more actual samples that wavelength is identified as a stable aging feature point, reflecting the higher the statistical reliability of it as a general aging feature); then select the wavelength points corresponding to the top 2000 maximum values ​​from S and mark them as 1, and mark the remaining wavelength points as 0 to generate the final screening array F; multiply F with the corrected spectrum (including the corrected spectrum of artificial samples and the corrected spectrum of actual samples) wavelength by wavelength:

[0083] If the value of a certain wavelength point in F is 1, the corrected spectral intensity of that wavelength point is retained as is (e.g., if the corrected intensity is 800 counts, the multiplication result is still 800 counts).

[0084] If the value of a certain wavelength point in F is 0, the corrected spectral intensity of that wavelength point becomes 0 (equivalent to removing the noise signal at that wavelength point).

[0085] Based on the above calculation method, a new training set (obtained by multiplying the corrected spectrum of the artificial sample by F, generating a new training set with known aging level labels for subsequent machine learning model training) and a test set (obtained by multiplying the corrected spectrum of the actual sample by F, generating a new test set with unknown aging level labels for verifying the diagnostic accuracy of the model) can be obtained, and the data dimension is significantly reduced from high dimension to 2000 dimensions.

[0086] S3. Machine Learning Model Construction and Aging Diagnosis: Based on a dimensionality-reduced dataset, a process of "feature selection - dimensionality reduction - parameter optimization - classification" is constructed to achieve automatic aging level diagnosis. For details, please refer to [link / reference]. Figure 6 .

[0087] S3.1 Data Calibration

[0088] Training set: Corrected spectral data of artificial samples, labeled according to aging level (level 1 data is labeled with "1", level 2 with "2", ..., level 5 with "5"), used by the model to learn the correspondence between features and levels;

[0089] Test set: Corrected spectral data of actual samples, unlabeled for the time being, used to verify the accuracy of the model.

[0090] S3.2 Recursive Feature Elimination (RFE) for core feature selection

[0091] Using a support vector machine (SVM) with a linear kernel function as the evaluator, iterative feature selection is performed on the training set data, with the following specific settings:

[0092] The initial number of features is set to 1000; the screening step size is set to 50 (50 "features with the lowest contribution" are removed in each iteration); the iteration termination condition is: when the number of retained features i is greater than 2000, the screening stops, and the core feature set that can characterize the aging difference is finally obtained.

[0093] S3.3 Linear Discriminant Analysis (LDA) Secondary Dimensionality Reduction: The feature set after RFE screening already possesses strong aging discrimination capabilities. However, to further improve the separability of the feature space and fully utilize class information, supervised linear transformation of the core features is required. By calculating the cumulative information contribution rate of each principal component after LDA dimensionality reduction, it is ultimately determined to reduce the spectral data to 4 dimensions. This operation maximizes the inter-class differences between different aging levels, minimizes intra-class fluctuations within the same level, and further compresses the data dimensionality, improving the computational efficiency and generalization ability of subsequent models.

[0094] S3.4 Particle Swarm Optimization (PSO)-Driven RBF Kernel SVM Modeling: The feature data after LDA quadratic dimensionality reduction is input into a Radial Basis Function (RBF) kernel support vector machine (SVM). Through kernel function adaptation and global parameter optimization, accurate classification of the aging level of heat-resistant steel is achieved. The specific process is as follows:

[0095] The reason for choosing the RBF kernel function is that the "feature-level" relationship of heat-resistant steel aging is non-linear (e.g., the feature of level 3 is not a simple superposition of "level 1 + level 2"). The RBF kernel can efficiently fit this kind of non-linear relationship, and the classification accuracy is 10%-15% higher than that of the linear kernel.

[0096] The key parameters of the SVM are globally optimized using the Particle Swarm Optimization (PSO) algorithm to avoid local optima problems caused by manual parameter tuning. The specific settings are as follows:

[0097] Parameters to be optimized: Penalty parameter C (search range 0.1~100, its core function is to balance the model's fit to the training samples and its generalization ability, avoiding overfitting (C too large) or underfitting (C too small)), kernel function parameter γ (search range 0.001~1, controls the fitting accuracy of the kernel function to nonlinear relationships, directly affecting the model's ability to capture the sample distribution in the feature space).

[0098] PSO configuration: learning factor set to 2.0 (balancing convergence speed and accuracy), population size 100 (parallel search of multiple sets of parameters to improve optimization efficiency), maximum number of iterations 100 (parallel convergence of parameters after iteration).

[0099] Besides PSO, simulated annealing can also be used as an alternative strategy for parameter optimization. This algorithm simulates the "annealing process" of a solid being heated to a high temperature and then slowly cooled, constructing an intelligent random search mechanism. In the parameter space search, it not only accepts "excellent solutions" that improve model performance, but also accepts "inferior solutions" with a certain probability of slightly worse performance, thus effectively escaping local optima. The optimal penalty parameter C and kernel width parameter γ can also be obtained through this algorithm, ultimately improving the model's generalization ability. Substituting the optimized parameters into an RBF kernel SVM for training can achieve the same modeling goal as PSO optimization, realizing accurate classification of the aging grades of heat-resistant steel.

[0100] S3.6 Model Diagnosis and Result Output: Based on the core data advantages after acoustic energy correction (such as...) Figure 7 As shown, the stability of the acoustically corrected data under different feature numbers is significantly better than that of the original data, effectively improving feature reliability. By inputting the test set data of the actual sample into the optimized RBF-SVM model, the following accurate diagnosis and output can be achieved:

[0101] The model automatically outputs "aging level" and "accuracy" to quantify the aging degree of each sample and the reliability of the judgment.

[0102] A single diagnosis takes ≤30 seconds, achieving fully automatic, fast and accurate judgment.

[0103] The aging level determination results of the actual samples obtained through the above model calculations are shown in Table 2. It should be noted that the artificial samples with aging levels "1" (brand new, unaged state) and "5" (tensile failure state) correspond to two extreme conditions in the aging process of heat-resistant steel. Their characteristic signals are easily affected by boundary effects, resulting in a slightly lower determination accuracy compared to intermediate aging levels. This phenomenon is consistent with the actual testing patterns under extreme conditions and is within a reasonable range. Ultimately, the overall classification accuracy of the five actual samples reached 94%, fully verifying the reliability and practicality of this method.

[0104] Table 2. Aging levels of actual samples

[0105] label Aging level accuracy Misclassification rate S31 1 84.17% 2:13.3%；4:0.83% S11 2 99.17% 1:0.83% S41 3 100% 0 S51 4 100% 0 S61 5 85% 2:15%

[0106] The actual sample aging level data output by this model was compared with the results of offline analysis (aging level evaluation of crystal phase diagram). The two were completely consistent, further confirming the accuracy and reliability of the "acoustic energy correction + machine learning" method disclosed in this application in the aging diagnosis of heat-resistant steel.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. However, the above description is merely a specific embodiment of the present invention, and the technical features of the present invention are not limited thereto. Any other embodiments derived by those skilled in the art without departing from the technical solution of the present invention should be covered within the scope of the present invention.

Claims

1. A method for diagnosing the failure of heat-resistant steel based on acoustic energy-corrected laser-induced breakdown spectroscopy, characterized in that, Includes the following steps: S1. Build a laser-induced breakdown spectroscopy experimental platform and simultaneously acquire spectral and acoustic signals of heat-resistant steel samples; S2. Calculate the acoustic energy value E of the sound wave signal in the main band. a Utilizing the sound energy value E a The spectral signal is corrected to eliminate spectral interference caused by laser energy fluctuations: S2.1 Prepare artificial aging test specimens and obtain actual service test specimens. The artificial aging test specimens are obtained through variable stress high temperature aging tests and are marked with a clear aging level. S2.2 Integrate the acquired acoustic signal within the preset main band and calculate the acoustic energy value E. a The formula is as follows: ； In the formula, x(t) is the acoustic signal in the form of a time function, and t1 and t2 are the start and end times of the main band, respectively; S2.3, Utilizing the sound energy value E a The original spectral intensity I0 under the same laser pulse is corrected to obtain a stable corrected spectral intensity I. E The correction formula is: ； Among them, I E I0 represents the original spectral intensity, where I is the corrected spectral intensity. S3. Perform multiple threshold segmentation on the corrected spectral data to screen core characteristic wavelengths related to aging levels, thereby achieving data dimensionality reduction: S3.

1. For the corrected spectra of artificial samples with different aging levels, calculate the average value for each wavelength to obtain the average spectrum A; S3.

2. Divide the corrected spectrum of the actual sample by wavelength and the average spectrum A to generate dimensionless data D. S3.3 Calculate the absolute difference ΔD between dimensionless data D and 1, select the first n wavelength points in ΔD that are closest to 1, and generate a binary array M, where n = 1500~2500; S3.4 Summing multiple sets of binary arrays M yields a feature importance array S. The wavelength points corresponding to the first m maximum values ​​in S are selected to generate the final filtering array F, where m = 1500~2500. F is multiplied wavelength by wavelength with the corrected spectrum to achieve data dimensionality reduction. S4. Input the dimensionality-reduced spectral data into the machine learning model. Through feature selection, secondary dimensionality reduction and parameter optimization, construct a classification model for the aging level of heat-resistant steel and output the failure diagnosis results of the heat-resistant steel sample.

2. The method for diagnosing the failure of heat-resistant steel based on acoustic energy-corrected laser-induced breakdown spectroscopy as described in claim 1, characterized in that, The specific operation process of step S4 includes: S4.

1. Use the reduced-dimensional spectral data of artificial samples as the training set and label it with aging level, and use the reduced-dimensional spectral data of actual samples as the test set to separate the training data from the test verification data. S4.2 Filtering core features through recursive feature elimination: The initial number of features for recursive feature elimination directly follows the dimensions of the dataset after the multiple threshold segmentation in step S3. The filtering step size is set to 50, and the upper and lower limits of the parameters are determined based on the feature contribution law: the lower limit of the feature is set to 1000; combined with the characteristics of the multi-band distribution of spectral features, the upper limit of the feature is set to 2000. S4.3 Construct a radial basis function kernel support vector machine model optimized by particle swarm optimization algorithm, input the dimensionality-reduced feature data into the model, and output the aging level of the actual sample, the overall classification accuracy of the model, and the misclassification rate.

3. The method for diagnosing the failure of heat-resistant steel based on acoustic energy-corrected laser-induced breakdown spectroscopy as described in claim 2, characterized in that, The particle swarm optimization algorithm is used to jointly optimize the penalty parameter C and the kernel function parameter γ of the support vector machine. The search range of the penalty parameter C is 0.1~100, and the search range of the kernel function parameter γ is 0.001~1. The parameters of the particle swarm optimization algorithm itself are set as follows: learning factor is 2.0~2.3, population size is 100~110, and maximum number of iterations is 100~120.

4. The method for diagnosing the failure of heat-resistant steel based on acoustic energy-corrected laser-induced breakdown spectroscopy as described in claim 3, characterized in that, The parameters of the particle swarm optimization algorithm are set as follows: learning factor is 2, population size is 100, and maximum number of iterations is 100.

5. The method for diagnosing the failure of heat-resistant steel based on acoustic energy-corrected laser-induced breakdown spectroscopy as described in claim 1, characterized in that, The heat-resistant steel is at least one of T91 steel, P92 steel, and TP347H steel.

Images