Equipment quality detection method and system based on multi-source data

Abstract

The application discloses an equipment quality detection method and system based on multi-source data, and relates to the technical field of equipment nondestructive detection and quality evaluation. The method comprises the following steps: for the same detection area on the surface of equipment to be detected, time-domain vibration response signals and spectrum characteristic signals are synchronously executed; acoustic stiffness features of the time-domain vibration response signals are extracted to obtain a first feature vector set; surface stress features of the spectrum characteristic signals are extracted to obtain a second feature vector set; the first feature vector set and the second feature vector set are fused to obtain a cooperative feature set; and the cooperative feature set is input into a pre-constructed and trained deep belief network model to obtain a defect type, thereby solving the problems that in the existing equipment quality detection method, acoustic and spectrum detection means are independent of each other, information cannot be effectively cooperated and causally related.

Description

Technical Field

[0001] This invention relates to the field of equipment nondestructive testing and quality assessment technology, specifically to an equipment quality testing method and system based on multi-source data. Background Technology

[0002] As modern equipment continues to develop towards higher reliability and longer service life, microscopic defects and material degradation that may occur in its key components (such as engine blades, armor structures, and load-bearing frames) during manufacturing and service have become core factors affecting the safe operation of equipment. Currently, commonly used testing methods in the field of nondestructive testing mainly include the following categories: 1) Acoustic-based ultrasonic flaw detection technology: This technology detects internal volumetric defects such as cracks, porosity, and inclusions by emitting sound waves into the object under test and receiving the reflected echoes. This type of method has high sensitivity to deep structural anomalies, but its ability to assess surface conditions (such as corrosion and stress distribution) is severely insufficient. Furthermore, the detection results are significantly affected by human operating factors such as coupling state and probe frequency, resulting in poor consistency among different operators.

[0003] 2) Spectroscopic compositional analysis techniques: These techniques use laser excitation to generate characteristic spectra on the material surface, which are then used to identify the material grade, elemental content, and chemical information such as surface corrosion products. While these methods are exceptionally sensitive to surface composition, the extracted signals only reflect information at the micrometer-level depth of the material's surface. They cannot obtain mechanical integrity parameters such as stiffness distribution and elastic modulus changes within the structure, nor can they effectively respond to hidden defects (such as closed microcracks).

[0004] 3) Attempts have been made to simply stitch together ultrasonic and spectral data, but these methods mostly involve feature stitching plus a conventional classifier, lacking in-depth exploration of the physical coupling relationship between the acoustic and spectral domains. Specifically, this manifests in the following ways: the spatial resolution mismatch between the two data sources is not resolved, making it impossible to align data at the same location; a cross-domain causal relationship model between surface condition and internal structure is not established, resulting in the system being unable to determine whether surface corrosion and internal cracks are caused by the same physical cause or are independent damage events when they coexist, leading to a high false alarm rate and serious misclassification of defect types. Summary of the Invention

[0005] To address the problems in existing equipment quality inspection methods where acoustic and spectroscopic detection methods are independent, information cannot be effectively coordinated or correlated, leading to difficulties in distinguishing the physical causes of surface damage and internal defects, and hindering the accurate and comprehensive identification and judgment of surface and internal defects, this invention provides an equipment quality inspection method based on multi-source data. The method includes: S1. For the same detection area on the surface of the equipment under test, the synchronous control module simultaneously issues a start command to drive the acoustic detection unit and the spectral analysis unit to synchronously perform data acquisition actions. The acoustic detection unit transmits the original excitation signal to the detection area and receives the time-domain vibration response signal returned by the detection area. The spectral analysis unit excites the detection area and obtains the spectral characteristic signal of the surface material of the detection area under excited emission. S2. Extract the acoustic stiffness features of the time-domain vibration response signal to obtain the first feature vector set; S3. Extract the surface stress features of the spectral feature signal to obtain a second feature vector set; S4. Merge the first feature vector set and the second feature vector set to obtain a collaborative feature set; S5. Input the collaborative feature set into a pre-built and trained deep belief network model to obtain the defect type.

[0006] By driving the acoustic detection unit and the spectral analysis unit to simultaneously acquire data in the same detection area through a synchronous control module, strict consistency of the data source in the time reference is ensured for subsequent fusion analysis. Acoustic stiffness features from the acoustic time-domain signal and surface stress features from the spectral signal are extracted separately to form the first and second feature vector sets, realizing the separate quantitative description of the integrity of the internal structure of the equipment and the chemical and mechanical state of the surface material. By fusing to obtain a collaborative feature set, and then inputting the collaborative feature set into a deep belief network model for defect type identification, the technical problems of traditional single detection methods being one-sided in information and unable to take into account the comprehensive judgment of surface corrosion and internal cracks are solved. This realizes the integrated and automated comprehensive detection of surface defects and internal defects of the equipment, significantly improving the comprehensiveness of defect type identification and the accuracy of judgment.

[0007] Furthermore, the specific steps of S2 include: S201. Perform deconvolution processing on the time-domain vibration response signal to obtain an intrinsic transfer function sequence; S202. Obtain the theoretical propagation speed range of the original excitation signal in the material of the equipment under test, and obtain the time threshold of each type of wave arriving at the receiver based on the theoretical propagation speed range and the geometric thickness parameters of the detection area. The intrinsic transfer function sequence is segmented based on the time threshold and divided into three continuous and non-overlapping physical information windows on the time axis: the first time window, the second time window, and the third time window. S203. Compare the frequency domain energy distribution within each physical information window with the original excitation signal, calculate the dynamic percentage of the extra energy offset outside the excitation frequency band to the total response energy, obtain the frequency band energy leakage factor, and obtain the energy leakage gradient vector based on the frequency band energy leakage factor. S204. Extract the envelope decay curve of the vibration waveform within the second time window, perform linear trend analysis in the logarithmic domain on the envelope decay curve to obtain the dynamic damping ratio; compare the peak frequency offset of the inherent transfer function sequence with the peak frequency of the standard transfer function of the same type of health equipment to obtain the dynamic stiffness deviation value. S205. Extract the bottom attenuation coefficient of the third time window, and stack the energy leakage gradient vector, the dynamic damping ratio, the dynamic stiffness deviation value and the bottom attenuation coefficient vertically according to the same spatial coordinate to form a four-dimensional structural response original parameter matrix; perform extreme value normalization processing on the structural response original parameter matrix, expand the normalized matrix into a one-dimensional feature sequence according to the preset grid division rules, and bind it with the corresponding spatial coordinate label to form a first feature vector set.

[0008] By sequentially performing deconvolution on the time-domain vibration response signal to remove probe coupling interference and obtain the intrinsic transfer function sequence, and then dividing the intrinsic transfer function sequence into three physical information windows based on the sound velocity time window, the energy leakage gradient vector, dynamic damping ratio, dynamic stiffness deviation value, and bottom surface attenuation coefficient are extracted from the three information windows respectively. Finally, these four-dimensional features are stacked and normalized to form the first feature vector set. This achieves the decoupling of multi-dimensional structural response parameters that purely reflect the stiffness of the equipment body from the aliased original acoustic signal. Moreover, the four-dimensional features correspond to three different physical depth ranges: near-surface, deep, and full penetration, respectively, enabling subsequent defect localization to have depth resolution capability. This solves the problem of ambiguity in defect type caused by traditional ultrasonic testing relying only on a single image value or a single velocity index, and significantly improves the ability of acoustic features to distinguish different defect types such as cracks, porosity, and material hardening.

[0009] Furthermore, the specific steps of S201 include: A1. Taking the start time of the original excitation signal as zero point, scan the first arriving interface reflection wavefront in the time-domain vibration response signal, calculate the propagation time difference between the interface reflection wavefront and the start time, obtain the system delay compensation amount, and obtain the time-calibrated excitation-response data pair based on the system delay compensation amount. A2. Based on the excitation-response data pair, obtain the first wave interval that appears first on the time axis in the time-domain vibration response signal, perform morphological analysis on the waveform profile of the first wave interval, obtain the energy transmission loss coefficient and the waveform broadening distortion factor, and combine the energy transmission loss coefficient and the waveform broadening distortion factor in a time and space to construct a system interference feature template. A3. Based on the energy transmission loss coefficient, the waveform broadening distortion factor and the system delay compensation, the time-domain vibration response signal is corrected to obtain the initial residual signal, and the effective response wavelet parameter list is obtained based on the initial residual signal. A4. Based on the standard wall thickness of the equipment under test and the known sound velocity range, calculate the theoretical time window for the sound wave to travel through the wall thickness once round trip. Based on the theoretical time window, process the effective response wavelet parameter list to obtain the acoustic impedance relative change trend sequence. Normalize the amplitude of the acoustic impedance relative change trend sequence to generate the intrinsic transfer function sequence.

[0010] It achieves adaptive modeling and targeted removal of parasitic interference from non-equipment components such as probes, coupling agents, and cables in the detection link, solving the technical problem of incomparable acoustic features due to differences in coupling states at different detection points, and significantly improving the purity, stability, and cross-point repeatability of acoustic feature extraction.

[0011] Furthermore, in step A3, the specific steps for obtaining the effective response wavelet parameter list based on the initial residual signal include: a1. On the time axis of the initial residual signal, search for the position of the extreme point with the largest absolute value of amplitude, and take the extreme point as the position center to extract a local signal segment with the same duration as the original excitation waveform of the original excitation signal. a2. Perform shape similarity matching between the overall envelope shape of the local signal segment and the original excitation waveform to obtain a shape similarity value; calculate the amplitude compression ratio and arrival time delay increment of the local signal segment relative to the original excitation waveform; if the shape similarity value is higher than a preset first threshold, determine that the local signal segment is an effective structural response wavelet, and obtain effective parameters based on the amplitude compression ratio, the delay increment, and the phase polarity; a3. Obtain the parasitic oscillation component in the local signal segment that matches the shape of the system interference feature template, delete the parasitic oscillation component in the initial residual signal, and obtain a new residual signal; a4. Return to a1, update the initial residual signal to the new residual signal, and iteratively execute a1-a4 until the maximum absolute amplitude of the current residual signal drops below a preset percentage of the amplitude of the first wave of the time-domain vibration response signal, and determine that the iteration has converged; obtain the list of effective response wavelet parameters based on the effective parameters.

[0012] Instead of the traditional one-time global deconvolution in the frequency domain, the structure response wavelet is peeled off one by one in the time domain by matching pursuit. This allows the effective wavelet in the weak signal to be picked out one by one in the iteration process, avoiding the defect of strong reflection wave masking weak reflection wave. It solves the problem of insufficient extraction capability of traditional deconvolution for weak signals (such as small crack echoes) and significantly improves the detection sensitivity of early small defects.

[0013] Furthermore, the specific steps of a3 include: a301. Based on the delay increment, the system interference feature template is translated as a whole on the time axis so that the peak phase of the system interference feature template completely overlaps with the peak phase of the parasitic oscillation component in the local signal segment. A preset protection window is extended on each side of the translated system interference feature template. The template waveform within the protection window is subjected to edge shaving processing to generate interference stripping sub-wave. a302. Compare the amplitude compression ratio with the maximum absolute amplitude of the local signal segment to generate a projection weight coefficient. Multiply the projection weight coefficient by the interference stripping wavelet to generate an interference projection component. a303. The local signal segment is sampled point by point in the time domain and its morphological coherence is determined with the interference projection component to generate a purified local signal segment. a304. The purified local signal segments are backfilled into the global residual signal of the current round based on the original start and end positions of the local signal segments when they were extracted. The global residual signal after backfilling is then subjected to a full-time domain smooth transition process to generate an updated global residual signal. a305. Calculate the percentage decrease in the total energy of the updated global residual signal relative to the total energy before the start of this iteration. If the percentage decrease is greater than the preset significant change threshold, the stripping operation is deemed effective, and the updated global residual signal is taken as the new residual signal. If the percentage decrease is less than or equal to the significant change threshold, it is determined that there are no obvious parasitic interference components that can be stripped from the global residual signal, the current residual signal is marked as basically pure, a convergence command is output, and the iteration loop ends.

[0014] It achieves non-destructive separation of parasitic interference and effective signal in a mixed state in the time domain. The components of the effective signal that are out of phase or orthogonal to the interference are well preserved. It solves the technical problem that the effective signal, especially the weak reflection wave that is out of phase with the interference, is easily damaged when the matching component is directly subtracted. It significantly improves the fidelity of the effective signal in the interference removal process and ensures the integrity of the micro-defect reflection wave in the stripping process.

[0015] Furthermore, the specific steps of a303 include: For each sampling point, determine whether the amplitude direction of the local signal segment at that point is consistent with the amplitude direction of the interference projection component; if the two directions are consistent, it is determined that the energy part of the sampling point comes from the contribution of parasitic interference, and the absolute value of the amplitude of the interference projection component at that point is subtracted from the amplitude of the sampling point in the local signal segment. If the two directions are opposite, the signal shape at that sampling point is determined to be unrelated to interference, and the original value is maintained.

[0016] Using amplitude direction (i.e., phase) as the physical criterion for distinguishing between interference and valid signals, rather than relying on amplitude magnitude for fuzzy subtraction, this method achieves precise attenuation of co-frequency and in-phase superimposed interference and complete protection of quadrature / out-of-phase valid signals from a physical perspective. It solves the problem of valid signal distortion caused by the inability to distinguish phase in conventional filtering or subtraction operations, significantly improves the waveform fidelity of the purified signal, and provides a highly reliable time-domain waveform basis for the subsequent restoration of acoustic impedance trend sequences.

[0017] Furthermore, the specific steps of S3 include: S301. Remove the strong continuous background of the spectral feature signal to obtain an effective time frame. Divide the effective time frame according to the delay time after excitation to obtain a time segmented spectral data group. The time segmented spectral data group includes early transient data, mid-term steady-state data and late decay data. S302. Obtain high-energy state spectral lines and ground-state spectral lines with preset fixed wavelengths, and obtain effective time-series segmented spectral data based on the time-series segmented spectral data set, the high-energy state spectral lines and the ground-state spectral lines. S303. Obtain the matrix main element and key alloying element of the equipment under test, obtain the internal standard reference line of the matrix main element and the target characteristic spectral line of the key alloying element, obtain the integral intensity value of the target characteristic spectral line in the mid-term steady-state data of the effective time-series segmented spectral data group, obtain the ratio of the integral intensity value to the internal standard intensity value of the internal standard reference line, and obtain the relative intensity ratio. Obtain the theoretical standard composition table of the equipment to be tested, and obtain the theoretical standard relative ratio based on the theoretical standard composition table. For the key alloying elements, calculate the percentage fluctuation of the relative strength ratio relative to the theoretical standard relative ratio. Arrange the floating percentages in the order of fixed elements to obtain the stoichiometric deviation vector. S304. In the late decay data of the effective time-series segmented spectral data group, obtain the spectral peak of the corrosion-sensitive element, obtain the net integral intensity value of the spectral peak, compare the net integral intensity value with the preset corrosion warning intensity threshold, and obtain the absolute intensity level of the corrosion product. The net oxygen intensity value is extracted and arranged in chronological order to form an oxygen intensity-time decay sequence. The oxygen intensity-time decay sequence is then fitted with a linear trend to obtain an oxygen decay line. The slope direction and slope magnitude of the oxygen decay line are then calculated. The corrosion activity index is obtained based on the absolute strength level of the corrosion products, the tilt direction, and the tilt amplitude. Based on the absolute strength level of the corrosion products and the corrosion activity index, a surface corrosion state characteristic tuple is obtained; S305. In the mid-term steady-state data of the effective time-series segmented spectral data set, select the specific dislocation-sensitive spectral line that is most sensitive to lattice strain as the target spectral line. Obtain a reference spectral profile template corresponding to a healthy standard part of the same material and heat treatment state as the equipment under test. The reference spectral profile template includes the standard half-width and standard peak center position of the target spectral line under stress-free and distortion-free state. The target spectral line profile of the target spectral line is completely superimposed on the reference spectral line profile template at the same scale. Obtain the half-width at half-maximum (WHM) value and the peak center wavenumber position of the target spectral line, calculate the percentage increase in WHM value relative to the standard WHM value, and calculate the offset of the peak center wavenumber position relative to the standard peak center position. The percentage of broadening increment is mapped to a dimensionless lattice distortion tensor factor, and the peak center wavenumber position is converted into a residual stress pointing vector, which includes stress direction components and stress amplitude components. S306. Multiply the time-series segmented spectral data group marked as low confidence by the confidence attenuation weighting coefficient to obtain the attenuated segmented spectral data group. Based on the attenuated segmented spectral data group and the time-series segmented spectral data group marked as high confidence, obtain the cleaned data. The stoichiometric deviation vector, surface corrosion state feature tuple, lattice distortion tensor factor, and residual stress pointing vector are vertically stacked according to a preset dimensional order to obtain the original feature row vector of the material state. The original feature vectors of the material state are divided according to the grid division rules, normalized, expanded into a one-dimensional sequence, and bound to the corresponding spatial coordinate labels to form the second feature vector set.

[0018] This method extracts four independent dimensions of physicochemical information from a single spectral signal: material composition deviation, corrosion activity, lattice distortion, and residual stress. It also achieves spatial alignment with the acoustic grid, solving the technical problems of traditional spectral analysis, which is limited to element content determination, has a single information dimension, and cannot be spatially registered with acoustic data. This significantly expands the information contribution dimensions of spectral data in equipment quality assessment.

[0019] Furthermore, in step S302, the specific steps for obtaining a valid time-series segmented spectral data set include: In the early transient data, the intermediate steady-state data, and the late decay data, the intensity ratio of the high-energy spectral line and the ground-state spectral line in each frame is obtained. With time as the horizontal axis and the intensity ratio as the vertical axis, the convergence trajectory curve is obtained, and the convergence time, convergence speed, and stability value of the convergence trajectory curve are obtained. If the convergence time is within a preset convergence time range, the stable value is within a preset stable range, and the convergence speed is within a preset variation range, then the time-series segmented spectral data set is determined to be valid and marked with high confidence. If the convergence time is not within the preset convergence time range, but the convergence speed is within the preset variation range, then it is determined that there are abnormal attachments in the detection area, and a low confidence level is marked. If the convergence time is not within the preset convergence time range and the convergence speed is not within the preset variation range, then the time-series segmented spectral data group is determined to be invalid and the data group is discarded. If the convergence time is within a preset convergence time range, the convergence speed is within a preset variation range, but the stable value is not within the preset stable range, then the material of the equipment under test is deemed questionable.

[0020] By utilizing the physical evolution law of the spectral signal itself, data quality can be self-graded. Without the need for external standard samples, the reliability of each excitation data can be determined in real time under on-site testing conditions. This solves the blind spot problem in on-site testing where unreliable spectral data due to surface oil, oxide layer or roughness differences cannot be effectively identified and processed, and significantly improves the on-site adaptability and data reliability of spectral detection.

[0021] Furthermore, the specific steps of S4 include: The data of each fine grid node in the second feature vector set is assigned to the nearest coarse grid node in the first feature vector set according to the spatial coordinate label, and a weighted average is performed on all spectral data belonging to the same coarse grid. Feature vectors belonging to the same coarse grid are concatenated vertically to form a joint feature vector; For each grid node, the response sensitivity of acoustic features to spectral features and the modulation response of spectral features to acoustic features are calculated, and the cross-domain response gradient and modulation coefficients are obtained respectively. By stacking the joint feature vector, cross-domain response gradient, and modulation coefficient according to spatial coordinate labels in three dimensions, a collaborative feature set is obtained.

[0022] It realizes the quantification of physical causal relationship between the acoustic domain and the spectral domain, so that the fusion result not only carries the information of the two types of data themselves, but also carries the causal relationship criteria of whether the surface corrosion is caused by internal defects or whether the internal cracks have affected the surface. It solves the technical problem of traditional fusion methods that only perform feature splicing and cannot distinguish between independent damage and related damage, and significantly improves the interpretability of defect type determination and the anti-false alarm capability under complex working conditions.

[0023] The present invention also provides an equipment quality inspection system based on multi-source data, the system comprising: Data Unit: Used to simultaneously issue start commands based on the synchronous control module for the same detection area on the surface of the equipment under test, driving the acoustic detection unit and the spectral analysis unit to synchronously perform data acquisition actions. The acoustic detection unit emits the original excitation signal to the detection area and receives the time-domain vibration response signal returned by the detection area. The spectral analysis unit excites the detection area and obtains the spectral characteristic signal of the surface material of the detection area under excited emission. First feature unit: used to extract the acoustic stiffness features of the time-domain vibration response signal to obtain a first feature vector set; Second feature unit: used to extract surface stress features from the spectral feature signal to obtain a second feature vector set; Fusion unit: used to fuse the first feature vector set and the second feature vector set to obtain a collaborative feature set; Result unit: used to input the collaborative feature set into a pre-built and trained deep belief network model to obtain the defect type.

[0024] The principle and beneficial effects of this system are similar to those of this method, and will not be elaborated further on this system.

[0025] One or more technical solutions provided by this invention have at least the following technical effects or advantages: 1. This invention simultaneously collects acoustic vibration response signals and spectral feature signals from the same detection area, and extracts multi-dimensional features covering the internal structural stiffness and the chemical and mechanical state of the surface material, thereby achieving comprehensive information coverage of the equipment from the surface to the interior, and from composition to mechanics. This fundamentally overcomes the limitations of single detection methods, which provide only partial information and cannot simultaneously identify surface and internal defects, and significantly improves the comprehensiveness of defect identification.

[0026] 2. This invention constructs a system interference feature template based on the first-wave morphology deconstruction and combines iterative stripping with a matching pursuit strategy to achieve non-destructive directional separation of parasitic interference in the detection link from the structural response of the equipment body in the time domain. This ensures that the final output intrinsic transfer function sequence is completely free from interference from external factors such as probe coupling state and coupling agent differences. It solves the technical problem of incomparable features and poor repeatability caused by coupling state fluctuations in traditional ultrasonic testing, significantly improving the stability and reliability of acoustic features under complex field testing conditions.

[0027] 3. This invention breaks through the traditional static information extraction mode of spectral analysis, which only focuses on elemental content. By using time-series segmented processing, it captures the dynamic convergence behavior of spectral line intensity ratios, the attenuation slope of trace elements in the later stages, and the broadening of the full width at half maximum (FWHM) and peak shift of spectral profiles. This achieves a leap from static component detection to dynamic trend prediction, enabling the assessment of corrosion status to include not only the current thickness but also forward-looking information on whether it is spreading, providing a higher-dimensional reference for equipment maintenance decisions.

[0028] 4. This invention upgrades the fusion of acoustics and spectroscopy from traditional feature stitching to physical causal relationship modeling by calculating cross-domain response gradients and modulation coefficients on a unified spatial grid. The resulting synergistic feature set not only includes the two types of data themselves, but also contains quantitative indicators of the causal relationship between surface state and internal structure, providing physical logic support for subsequent classification decisions by deep belief networks. This solves the pain points of high false alarm rate and uninterpretable decision results in traditional black-box fusion models, significantly improving the accuracy and engineering reliability of defect type determination.

[0029] 5. This invention forms a complete end-to-end automated detection process, from synchronous data acquisition, automatic extraction of acoustic and optical features, spatial registration and collaborative fusion, to the automatic output of defect types and quantitative evaluation parameters by deep belief networks. The entire process requires no manual intervention in feature selection or threshold setting, ensuring high reproducibility of detection results and providing a reliable technical path for the standardization and digitalization of quality control in military equipment. Attached Figure Description

[0030] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention. Figure 1 This is a flowchart illustrating an equipment quality inspection method based on multi-source data. Detailed Implementation

[0031] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0032] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0033] Example 1 refer to Figure 1 This embodiment provides an equipment quality inspection method based on multi-source data, the method comprising: S1. For the same detection area on the surface of the equipment under test, a start command is issued simultaneously based on the synchronous control module (such as a digital delay generator) to drive the acoustic detection unit (such as an ultrasonic pulse transmitter / receiver) and the spectral analysis unit (such as a laser-induced breakdown spectrometer (LIBS) system) to perform data acquisition actions synchronously. The acoustic detection unit adopts an active excitation method to transmit the original excitation signal to the detection area and receive the time-domain vibration response signal returned by the detection area. This signal carries physical information on the modulation effect of the equipment's subsurface and internal structure (such as grain boundaries, crack tips, loose pores, etc.) on the sound wave propagation characteristics. Simultaneously, the spectral analysis unit excites the detection area and collects the spectral characteristic signals emitted by the surface material of that area. These signals contain information about the elemental composition, chemical state, and surface deposits or oxide layers of the surface material. The aforementioned acoustic time-domain signal and spectral characteristic signal together constitute the original multi-source data set for the detection area. This data set is then bound to its corresponding spatial coordinates and stored in a temporary data buffer for subsequent processing.

[0034] S2. Extract the acoustic stiffness features of the time-domain vibration response signal to obtain the first feature vector set: S201. Perform deconvolution processing on the time-domain vibration response signal to obtain an intrinsic transfer function sequence; A1. Taking the start time of the original excitation signal as zero, using the principle of cross-correlation, scan the first arriving interface reflection wavefront in the time-domain vibration response signal, calculate the propagation time difference between them, and use this time difference as the system delay compensation. After compensation, the excitation waveform and response signal have a strictly corresponding initial phase relationship on a unified time axis, generating a time-calibrated excitation-response data pair.

[0035] A2. Based on the excitation-response data pair, obtain the first wave interval (i.e., the direct reflection echo from near the equipment surface) that appears earliest on the time axis in the time-domain vibration response signal, and perform morphological analysis on the waveform profile of the first wave interval: The amplitude ratio of the first wave waveform in the first wave interval and the excitation waveform of the original excitation signal are calculated and the waveform similarity is compared to quantify the energy transmission loss coefficient and waveform broadening distortion factor caused by the difference in acoustic impedance of the coupling layer between the detection probe and the equipment surface; for example: Qualitative frequency domain component calculations were performed on the first waveform segment and the original excitation waveform, respectively, and the differences in energy proportions in the high-frequency, mid-frequency, and low-frequency bands were compared. If the energy proportion in the mid-to-high frequency bands of the first waveform decreased significantly compared to the original excitation, it was determined that the coupling layer selectively absorbed the high-frequency components, and the magnitude of this decrease was quantified as the energy transmission loss coefficient. By comparing the steepness of the first rising edge (i.e., the number of sampling points required to reach the peak from zero amplitude) with the steepness of the original excitation rising edge, the smoothing attenuation effect of the coupling layer on high-frequency transient components is quantified. Simultaneously, by comparing whether abnormal trailing oscillations (i.e., ringing) occur after the first falling edge, the superposition effect of multiple micro-reflections caused by microscopic gaps between the coupling layer and the equipment surface is quantified, such as the specific number of micro-reflections. These two differences are then combined (e.g., weighted fusion) to generate a waveform broadening distortion factor.

[0036] The energy transmission loss coefficient and the waveform broadening distortion factor are combined spatiotemporally to construct a dynamic system interference characteristic template. This template characterizes the inherent parasitic response mode of the entire detection link (including probe, coupling agent, and cable) under the current contact state, excluding the equipment itself. After construction, the template is temporarily stored in memory, and the corresponding spatial detection points are marked. Specific steps include: Based on the transmission loss coefficient and the waveform broadening distortion factor, the original excitation signal is subjected to reverse morphological modulation: its rising edge slope is reduced and its high-frequency components are attenuated, so that the modulated waveform is morphologically close to the standard first wave theoretical shape of healthy and defect-free equipment (this standard shape is known and represents the ideal reflection waveform of the equipment surface stiffness under perfect coupling conditions), thus obtaining the inherent reflection reference wave of the equipment surface, which represents the acoustic reflection morphology that the surface material at the test point should have under ideal coupling conditions, eliminating the influence of the random coupling state in this test.

[0037] The difference between the initial waveform and the inherent reflection reference wave on the surface of the equipment is calculated point by point to obtain the difference waveform sequence, which is the pure parasitic interference residual waveform. This residual waveform does not contain any structural information of the equipment body and is purely a distortion contribution formed by the superposition of factors such as probe clamping force fluctuations, uneven coupling agent thickness, and surface micro-roughness in the time domain during this test. After amplitude normalization processing of this residual waveform, a system interference feature template is generated.

[0038] A3. Based on the energy transmission loss coefficient, the waveform broadening distortion factor, and the system delay compensation, the time-domain vibration response signal is corrected to obtain the initial residual signal. For example, the energy transmission loss coefficient and the waveform broadening distortion factor are appended as data tags to the header metadata area of ​​the waveform file. According to the energy transmission loss coefficient, the overall amplitude of the time-domain vibration response signal is reverse-normalized and compensated: the entire time-domain waveform sequence is multiplied by a compensation gain value so that the amplitude of the first wave at each detection point is mapped to the same standard reference amplitude level. This operation strictly maintains the relative amplitude ratio between each reflected wave within the waveform and only reduces the system-level energy fluctuation introduced by the coupling state. After completing the amplitude compensation, the entire response waveform sequence is translated on the time axis with the starting time of the original excitation waveform as the absolute zero point so that the leading edge of the first wave is accurately aligned with the theoretically calculated arrival time, reducing the small time drift caused by the probe placement tilt angle or trigger delay.

[0039] Start a multi-level iterative stripping loop, with the following logical operation executed in each loop: a1. On the time axis of the initial residual signal, search for the position of the extreme point with the largest absolute value of amplitude, and take the extreme point as the position center to extract a local signal segment with the same duration as the original excitation waveform of the original excitation signal. a2. Perform morphological similarity matching between the overall envelope shape of the local signal segment and the original excitation waveform to obtain a morphological similarity value; Calculate the amplitude compression ratio and arrival time delay increment of the local signal segment relative to the original excitation waveform; If the morphological similarity value is higher than a preset first threshold, the local signal segment is determined to be an effective structural response wavelet rather than random noise, and the three effective parameters, namely the amplitude compression ratio, the delay increment, and the phase polarity (whether a flip occurs), are recorded.

[0040] a3. Obtain the parasitic oscillation component in the local signal segment that matches the shape of the system interference feature template, delete the parasitic oscillation component in the initial residual signal, and obtain a new residual signal; the specific steps include: a301. Based on the aforementioned delay increment, the system interference feature template is shifted entirely along the time axis, ensuring that the peak phase of the template completely overlaps with the peak phase of the parasitic oscillation component in the local signal segment. A preset protection window (e.g., the window length equal to one-quarter of the excitation signal width) is extended on each side of the shifted system interference feature template. Edge shaving processing is applied to the template waveform within the protection window (i.e., the amplitude at both ends of the template is smoothly reduced to zero) to prevent artificial impact spikes from being generated at the truncation boundary during subsequent stripping operations, thus generating a phase-aligned and edge-smoothed interference stripping sub-wave.

[0041] a302. Compare the amplitude compression ratio with the maximum absolute amplitude of the local signal segment to calculate a projection weight coefficient between 0 and 1. Multiply the projection weight coefficient by the interference stripping wavelet to generate an amplitude-normalized interference projection component. The amplitude level of this component is strictly anchored to the estimated intensity of the actual parasitic interference in the current local signal, rather than the original intensity of the template.

[0042] a303. Perform morphological coherence determination between the sampled points of the local signal segment in the time domain and the interference projection component: For each sampling point, determine whether the amplitude direction (positive / negative) of the local signal segment at that point is consistent with the amplitude direction of the interference projection component. If the two directions are consistent, it is determined that the energy part of the sampling point comes from the contribution of parasitic interference. From the amplitude of the sampling point in the local signal segment, only the absolute value of the amplitude of the interference projection component at that point is subtracted, but the small residual after subtraction is retained (it is not allowed to be reduced to below zero, that is, to ensure that the result after subtraction is in the same direction as the interference projection component but not in opposite phase). If the two directions are opposite, it is determined that the signal shape of the sampling point is unrelated to interference and belongs to the unique contribution of the actual structural response of the equipment body. In this case, no amplitude change is made to the sampling point, and its original value is completely preserved.

[0043] After the above point-by-point discrimination and differentiation processing, all components in the original local signal segment that are in phase with the interference template are directionally weakened, while all true response components that are out of phase or orthogonal to the interference template are preserved without loss, forming the purified local signal segment. This processing achieves non-destructive separation of useful signals and unwanted interference in the time-domain mixed state from a physical perspective.

[0044] a304. The purified local signal segments, based on their original start and end times when they were extracted, are backfilled into the global residual signal of the current round. During backfilling, only the data sequence within the time window is replaced; the signal data outside the window remains unchanged. After backfilling, a full-time-domain smooth transition processing is performed on the global residual signal: specifically, a linearly gradual splicing method is used at the edge of the backfill window to ensure no amplitude jumps at the junction of the old and new data, thereby generating an updated global residual signal.

[0045] a305. Calculate the percentage decrease in the total energy of the updated global residual signal (i.e., the sum of the squares of the amplitudes of all sampling points) relative to the total energy before the start of this iteration. If the percentage decrease is greater than the preset significant change threshold (e.g., greater than 1%), the stripping operation is deemed effective, and the current residual signal still contains considerable interference components. Record the successful removal flag and use the updated global residual signal as the new residual signal. If the percentage decrease is less than or equal to the significant change threshold, it is determined that there are no obvious parasitic interference components to be stripped from the global residual signal. Mark the current residual signal as basically pure, output the convergence command, and end the iteration loop.

[0046] a4. Return to a1, update the initial residual signal to the new residual signal, and iteratively execute a1-a4 until the maximum absolute amplitude of the current residual signal drops to below a preset percentage (e.g., 5%) of the amplitude of the first wave of the time-domain vibration response signal, and determine that the iteration has converged. A list of effective response wavelet parameters is obtained based on the effective parameters.

[0047] A4. Based on the standard wall thickness of the equipment to be tested and the known range of sound speeds, calculate the theoretical time window for a sound wave to travel through the wall thickness once.

[0048] Components whose delay increments fall outside an integer multiple (e.g., 1, 2, 3 times) of the theoretical time window in the effective response wavelet parameter list are identified as false path signals generated by side reflections or edge diffraction and are therefore removed. Components falling within integer multiples of the window are identified as multiple reflected echoes from the actual internal structure of the equipment and are retained.

[0049] All retained effective parameters are arranged in ascending order of the delay increment, and the relative acoustic impedance change trend sequence at each depth interface is reconstructed based on the amplitude compression ratio and phase polarity (essentially an impedance-depth waveform trajectory diagram along the thickness direction inside the equipment). If the initial acoustic impedance of the equipment surface is known, the following recursive steps are taken: starting from the first arranged wavelet (nearest surface), the impedance change amplitude of the current interface relative to the previous interface is calculated based on the amplitude compression ratio of that wavelet; the direction of change is calculated based on the phase polarity of that wavelet—if the polarity is reversed, it represents a sudden change in acoustic impedance from large to small (indicating defects or porosity); if the polarity is not reversed, it represents a sudden change in acoustic impedance from small to large (indicating high-density inclusions). This process is then repeated: based on the impedance value of the previous interface, the calculated change amplitude and direction are superimposed to obtain the absolute acoustic impedance value of the current interface.

[0050] The sequence is normalized so that its values ​​reflect only the difference in relative structural stiffness and are independent of the magnitude of the detection energy, thus generating an intrinsic transfer function sequence that characterizes the pure structural physical properties of the detection point.

[0051] S202. Obtain the theoretical propagation speed range of the original excitation signal in the material of the equipment under test, wherein the theoretical propagation speed range includes the longitudinal wave propagation speed range, the transverse wave propagation speed range and the surface wave propagation speed range. Based on the theoretical propagation speed range and the geometric thickness parameters of the detection area, the time threshold for each type of wave to arrive at the receiver is obtained. Based on the time threshold, the intrinsic transfer function sequence is segmented into three continuous and non-overlapping physical information windows on the time axis: the first time window is the near-surface reflection wave region, carrying structural stiffness information of the equipment surface and subsurface (approximately 0-2 mm depth); the second time window is the internal bulk wave resonance region, carrying volumetric information of the deep structure and grain boundary state of the equipment; and the third time window is the bottom surface attenuation wave region, carrying residual energy information after the sound wave penetrates the entire wall thickness.

[0052] After the segmentation is completed, waveform data within three time windows are acquired.

[0053] S203. Compare the frequency domain energy distribution within each of the physical information windows with the original excitation signal, and calculate the dynamic percentage of the additional energy offset outside the excitation frequency band (the frequency width range covered by the original excitation signal) relative to the total response energy to obtain the frequency band energy leakage factor. For dense, defect-free regions, the energy leakage factor approaches zero; when the sound wave encounters a crack interface or loose pores, a strong nonlinear modulation effect is generated, causing the factor to increase sharply; calculate the leakage factor corresponding to the three time windows respectively to form an energy leakage gradient vector characterizing the degree of nonlinearity in the depth direction of the detection area.

[0054] S204. Extract the envelope attenuation curve of the vibration waveform within the second time window, perform linear trend analysis on the envelope attenuation curve in the logarithmic domain, calculate the attenuation slope, and obtain the dynamic damping ratio, which directly reflects the rate of mechanical energy loss caused by internal friction and micro-defects of the material when the sound wave propagates inside the equipment.

[0055] The peak frequency offset of the inherent transfer function sequence is compared with the peak frequency of the standard transfer function of the same type of health equipment. The relative deviation between the two is calculated to obtain the dynamic stiffness deviation value. A positive deviation value indicates material hardening and embrittlement, while a negative value indicates material softening or microseparation.

[0056] S205. Extract the bottom attenuation coefficient of the third time window (i.e., the ratio of the reflected wave amplitude to the incident wave amplitude). The energy leakage gradient vector, the dynamic damping ratio, the dynamic stiffness deviation value, and the bottom surface attenuation coefficient are stacked vertically according to the same spatial coordinates to form a four-dimensional structural response original parameter matrix.

[0057] The matrix is ​​subjected to extremum normalization to reduce the order-of-magnitude differences between different physical dimensions and compress them into a uniform numerical range. Finally, the normalized matrix is ​​expanded into a one-dimensional feature sequence according to a preset grid division rule (such as a preset grid step size) and bound to the corresponding spatial coordinate labels to form the first feature vector set.

[0058] S3. Extract the surface stress features from the spectral feature signal to obtain the second feature vector set: The spectral characteristic signal is a sequence of multiple time-series spectral images acquired by the spectrometer within a continuous time window after pulsed laser excitation (i.e., multiple spectral curves at the same detection point at different delay times).

[0059] S301. Remove the strong continuous background (i.e., primary radiation background or bremsstrahlung background) generated by primary plasma radiation from the first frame of the spectral feature signal to obtain the effective time frame. Divide the effective time frame into three spectral line data (physical evolution stages) according to the delay time after excitation: Early transient data (early transient stage, microsecond-level early stage): In this stage, the spectral lines are dominated by ion lines, reflecting the instantaneous ionization state of surface elements under high-energy excitation; Mid-term steady-state data (mid-term steady-state stage, microsecond mid-range): In this stage, the spectral lines are mainly atomic lines, reflecting the balanced evaporation and excitation state of the main components of the surface material; Late decay data (late decay stage, microsecond-level later stage): In this stage, the signal intensity decreases but the spectral profile tends to stabilize, reflecting the slow relaxation response of the material matrix to the excitation energy.

[0060] The spectral data from the three stages are packaged independently to generate time-series segmented spectral data sets.

[0061] S302. Obtain high-energy spectral lines (sensitive to temperature) and ground-state spectral lines (insensitive to temperature) at preset fixed wavelengths. In the early transient data, the intermediate steady-state data, and the late decay data, the intensity ratio of the high-energy spectral line and the ground-state spectral line in each frame is obtained. With time as the horizontal axis and the intensity ratio as the vertical axis, the convergence trajectory curve is obtained, and the convergence time, convergence speed, and stability value of the convergence trajectory curve are obtained. If the convergence time is within a preset convergence time range, the stable value is within a preset stable range, and the convergence speed is within a preset variation range, that is, the curve smoothly decreases and quickly stabilizes into a straight line in a very short time (such as the first few frames), then the time-series segmented spectral data group is determined to be valid and marked with high confidence. If the convergence time is not within the preset convergence time range, but the convergence speed is within the preset variation range, i.e. the curve continues to slowly decrease and is still not stable after exceeding the preset time window, then it is determined that there are abnormal attachments in the detection area (such as oil stains, coatings or loose oxide layers, the laser first penetrates the covering layer, the substrate excitation is delayed or the energy is absorbed), and low confidence is marked. If the convergence time is not within the preset convergence time range and the convergence speed is not within the preset variation range, i.e. the surface is extremely uneven or there are loose attachments, causing plasma instability, then the time-series segmented spectral data set is determined to be invalid and the data set is directly discarded, prompting that the surface needs to be cleaned again and retested.

[0062] If the convergence time is within the preset convergence time range, the convergence speed is within the preset variation range, but the stable value is not within the preset stable range, that is, although there is no foreign interference, the matrix material may not conform to the standard grade (there is component segregation or mixing), then the material of the equipment under test is questionable.

[0063] In pulsed laser excitation spectroscopy (such as LIBS), when a laser strikes the surface of an equipment, it forms a plasma cloud that evolves rapidly over time. Microsecond-level front end: The plasma temperature is extremely high (tens of thousands of degrees), and the excited spectral lines are mainly high-energy ion lines, which have high energy but short lifetime; Mid-microsecond stage: The plasma expands and cools rapidly, and the temperature decreases. At this time, the excited spectral lines are mainly low-energy atomic lines, which have low energy, long lifetime and are stable.

[0064] The decay rates of high-energy ion lines and low-energy atomic lines are completely different. Therefore, the ratio of their intensities (intensity ratio) is not a fixed constant, but a dynamic variable that decreases over time and eventually approaches a stable value. This trajectory from a high value to a stable value is the convergence trajectory curve.

[0065] S303. Obtain the matrix main element (e.g., Fe in iron-based alloys or Ti in titanium alloys) and key alloying elements (e.g., Cr, Ni, Mo, V, etc.) of the equipment to be tested, obtain the internal standard reference line of the matrix main element and the target characteristic spectral line of the key alloying element, that is, set the most stable and strongest characteristic spectral line of the main element as the internal standard reference line. In the mid-term steady-state data of an effective time-series segmented spectral data set, the integral intensity value of the target characteristic spectral line is obtained, and the ratio of the integral intensity value to the internal standard intensity value of the internal standard reference line is obtained to obtain the relative intensity ratio. Obtain the theoretical standard composition table of the equipment to be tested (including the standard mass percentage range of each element), and obtain the theoretical standard relative ratio based on the theoretical standard composition table (equivalent to converting grams / kilograms into multiples relative to Fe). For the key alloying elements, calculate the percentage fluctuation of the relative strength ratio relative to the theoretical standard relative ratio; for example, if the theoretical relative ratio of Cr is 0.18 and the measured value is 0.189, then the deviation is +5%.

[0066] The floating percentages are arranged in a fixed element order to obtain the stoichiometric deviation vector.

[0067] Integrated intensity value: This refers to the total area enclosed by the peak of a characteristic spectral line on a spectrum, from the peak's start boundary to its end boundary. This area includes the sum of the true signal and the continuous background radiation. It's equivalent to measuring the entire projected area of ​​a mountain from its base to its right, including the ground (background) at the foot of the mountain.

[0068] S304, Late Decay Stage: The matrix elements decay rapidly due to plasma cooling, while the characteristic spectral lines of non-metallic elements such as oxygen, sulfur, and chlorine, although equally weak, decay at a much slower rate than those of the matrix elements. Therefore, during the late low-noise window, trace signals that were previously submerged resurface.

[0069] The presence of elements such as oxygen and sulfur directly reflects the thickness of the surface oxide layer or the degree of accumulation of corrosion products. The slope of these element signals in the later stages (whether it rises or falls, fast or slow) implies the physical structure information of the corrosion layer (whether it is a dense passivation film or a loose and porous rust layer).

[0070] In the late decay data of an effective time-series segmented spectral data set, the spectral peaks of corrosion-sensitive elements (such as oxygen, sulfur, chlorine, carbon, etc.) are obtained, and the net integral intensity value of the spectral peaks is obtained. The net integral intensity value is compared with a preset corrosion warning intensity threshold to obtain the absolute intensity level of the corrosion products (for example, divided into four levels: normal / slight / moderate / severe).

[0071] Net integrated intensity value: This refers to the area of ​​the pure peak contributed solely by the characteristic emission of the element after accurately subtracting the continuous background baseline below the peak in the spectrum. It is the total integrated area (including background) minus the integrated area of ​​the background baseline within the same wavelength range. It's analogous to measuring a mountain by first removing the base (background) and calculating only the volume (area) of the pure mountain body rising from the ground.

[0072] The net oxygen intensity value is extracted and arranged in chronological order to form an oxygen intensity-time decay sequence. The oxygen intensity-time decay sequence is then fitted with a linear trend to obtain an oxygen decay line. The slope direction and slope magnitude of the oxygen decay line are calculated (quantifying the steepness of the line's descent). Net intensity value: The amplitude (i.e., net peak height) at a single point on a given wavelength of a spectrum (the peak position of the characteristic spectral line of oxygen) after subtracting the local background from the spectrum. It is not an area (non-integral), but a single intensity value. It's analogous to measuring the height of a mountain by only reading the vertical elevation of the highest point relative to the ground.

[0073] Under normal circumstances, the oxygen signal gradually decreases as the plasma expands and dissipates (slope is negative). If the oxygen signal increases instead of decreasing (slope is positive) or remains basically flat (slope is close to zero), it indicates that oxygen elements in this region are continuously excited.

[0074] Based on the absolute strength grade of the corrosion products, the tilt direction, and the tilt amplitude, a corrosion activity index is obtained, as shown in Table 1. Table 1. Quantification of Corrosion Activity Index Low (close to normal) Negative and steep The surface is clean, free of corrosion, or has only a very thin natural passivation film. Corrosion activity index = 0 (static safe state) Medium or above Negative and steep Corrosion products are present on the surface, but the corrosion layer is dense and passivated (old corrosion, no longer expanding). Corrosion activity index = 30~50 (historical damage state) Medium or above The slope is gentle and almost horizontal. The corroded layer is loose and porous, allowing easy laser penetration, and the interior is continuously stimulated (and is deteriorating). Corrosion activity index = 60~80 (active diffusion state) High (far exceeding warning level) The slope is positive (rising) or flat. The severely loose rust layer is rapidly penetrating into the substrate. Corrosion activity index = 90~100 (high-risk accelerated state) Based on the absolute strength level of the corrosion products and the corrosion activity index, a surface corrosion state characteristic tuple is obtained.

[0075] S305. In the mid-term steady-state data of the effective time-series segmented spectral data set, select the specific dislocation sensitive spectral line (usually the diffraction corresponding spectral line with high angle or high Miller index) that is most sensitive to lattice strain as the target spectral line. Obtain a reference spectral profile template corresponding to a healthy standard part of the same material and heat treatment state as the equipment under test. The template stores the standard half-width and standard peak center position of the target spectral line under stress-free and distortion-free conditions.

[0076] The target spectral profile of the target spectral line is completely superimposed on the reference spectral profile template at the same scale (that is, after normalizing the peak amplitudes of the two to the same height, they are aligned to the same wavenumber coordinate axis). Obtain the half-width at half-maximum (WHM) value of the target spectral line (i.e., the wavenumber width corresponding to half the peak height) and the wavenumber position of the peak center. Calculate the percentage of the widening increment of the WHM value relative to the standard WHM and calculate the offset of the wavenumber position of the peak center relative to the standard peak center position. The percentage broadening increment can be mapped to a dimensionless lattice distortion tensor factor. The mapping logic can be as follows: The percentage of broadening increment is 0% (completely consistent with the standard), and the tensor factor is 0 (completely distortion-free, lattice perfect). The percentage of broadening increment is between 0% and 5%, and the tensor factor is between 0 and 30 (slight work hardening or fine precipitation, which is within the normal process range). The percentage of broadening increment is between 5% and 15%, and the tensor factor is 30-70 (significant lattice distortion, possibly with a large number of dislocation tangles or plastic deformation). The tensor factor is 70-100 (severe distortion, the material is close to fatigue damage or severe work hardening).

[0077] The larger the factor, the more disordered the grains at that point are, and the lower the material's toughness and fatigue resistance.

[0078] The peak center wavenumber location is transformed into a residual stress pointing vector, which contains two components: Stress direction component: directly determined by the offset direction. If it offsets towards the high wavenumber side, it is marked as a compressive stress state; if it offsets towards the low wavenumber side, it is marked as a tensile stress state; if there is no offset, it is marked as a stress-free equilibrium state.

[0079] Stress amplitude component: The absolute value of the offset is mapped to an equivalent residual stress estimate in megapascals according to a pre-calibrated stress-offset ratio conversion relationship (which can be pre-calibrated using standard tensile / compression specimens of the same material).

[0080] S306 The time-series segmented spectral data set marked as low confidence is multiplied by the confidence attenuation weighting coefficient to obtain the attenuated segmented spectral data set. Based on the attenuated segmented spectral data set and the time-series segmented spectral data set marked as high confidence, the cleaned data is obtained. The stoichiometric deviation vector, surface corrosion state feature tuple, lattice distortion tensor factor, and residual stress pointing vector are stacked vertically in a preset dimensional order (e.g., stoichiometric deviation vectors of the 5 principal elements of the material + surface corrosion state feature tuple + lattice distortion tensor factor + residual stress pointing vector) to obtain a fixed-length original feature row vector of the material state. The original feature vectors of the material state are divided according to the grid division rules described in S2, then normalized, expanded into a one-dimensional sequence, and bound to the corresponding spatial coordinate labels to form the second feature vector set.

[0081] S4. Merge the first feature vector set and the second feature vector set to obtain a collaborative feature set: Since the spot diameter of an acoustic probe is usually larger than that of a spectral laser, the two naturally have different sampling densities in space (the acoustic grid is coarser, and the spectral grid is finer).

[0082] The data of each fine grid node in the second feature vector set is assigned to the coarse grid node of the nearest first feature vector set according to the spatial coordinate label. All spectral data belonging to the same coarse grid are then weighted and averaged (the weight is determined by the distance between the spectral point and the center of the acoustic grid; the closer the distance, the greater the weight). The two types of data are forced to be unified into the same spatial grid coordinate system.

[0083] The feature vectors belonging to the same coarse grid are concatenated vertically to form a joint feature vector containing all dimensions of the two types of data.

[0084] For each grid node, calculate the response sensitivity of acoustic features to spectral features, and the modulation response of spectral features to acoustic features: Cross-domain response gradient: Taking the current node as the center, extract the joint feature vector of all nodes in the surrounding neighborhood (e.g., 3×3) of the center, and determine whether the internal acoustic stiffness feature (first feature vector set) changes spatially when the surface material state (second feature vector set, such as corrosion activity index) changes spatially in the surrounding neighborhood. That is, whether the spatial change directions of the two are consistent and whether the change magnitudes are proportional. If the two show a highly consistent synchronous change trend, it indicates that there is a strong physical correlation between surface corrosion and internal structural weakening in this area, and the response gradient value is high. Conversely, if the surface state changes drastically but the internal acoustics remain unchanged, it indicates that the corrosion is limited to the surface layer and does not extend to the depth, and the response gradient value is low.

[0085] Modulation coefficient: When the internal structural features (first feature vector set) change abruptly in the surrounding neighborhood, does the surface stress feature (second feature vector set) also change abruptly at the same location? If the internal acoustic anomaly happens to correspond to the local extreme point of the surface stress, it is determined that the internal defect has a modulation effect on the surface mechanical state. A high modulation coefficient indicates that the internal defect has affected the surface state.

[0086] For example, positive causality: surface stress - internal response: Independent variables: residual stress pointing vector and lattice distortion tensor factor of surface stress characteristics; Dependent variables: dynamic damping ratio and dynamic stiffness deviation of internal structural features; Related objective: To verify whether severe surface stress concentration leads to internal material softening or microcrack initiation.

[0087] Reverse causality: Internal defects - surface stress: Independent variables: energy leakage gradient vector and bottom attenuation coefficient of internal structural features; Dependent variables: surface stress characteristics and lattice distortion tensor factor; Related objective: To verify whether deep internal cracks have caused stress field distortion on the surface (i.e., the cracks have penetrated to the near-surface layer).

[0088] The joint feature vector, cross-domain response gradient, and modulation coefficients are stacked in three dimensions according to spatial coordinate labels to obtain a collaborative feature set. S5. Input the collaborative feature set into a pre-built and trained deep belief network model. This model can be an existing neural network structure based on stacked multi-layer restricted Boltzmann machines. The number of nodes in its input layer corresponds to the dimension of the collaborative feature set, and the number of nodes in its output layer corresponds to a preset defect type label. The preset defect types include, but are not limited to: surface corrosion, shallow cracks, deep porosity, through cracks, and material mixing.

[0089] Example 2

[0090] Based on Example 1, this example also provides an equipment quality inspection system based on multi-source data, the system comprising: Data Unit: Used to simultaneously issue start commands based on the synchronous control module for the same detection area on the surface of the equipment under test, driving the acoustic detection unit and the spectral analysis unit to synchronously perform data acquisition actions. The acoustic detection unit emits the original excitation signal to the detection area and receives the time-domain vibration response signal returned by the detection area. The spectral analysis unit excites the detection area and obtains the spectral characteristic signal of the surface material of the detection area under excited emission. First feature unit: used to extract the acoustic stiffness features of the time-domain vibration response signal to obtain a first feature vector set; Second feature unit: used to extract surface stress features from the spectral feature signal to obtain a second feature vector set; Fusion unit: used to fuse the first feature vector set and the second feature vector set to obtain a collaborative feature set; Result unit: used to input the collaborative feature set into a pre-built and trained deep belief network model to obtain the defect type.

[0091] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An equipment quality inspection method based on multi-source data, characterized in that, The method includes: S1. For the same detection area on the surface of the equipment under test, the synchronous control module simultaneously issues a start command to drive the acoustic detection unit and the spectral analysis unit to synchronously perform data acquisition actions. The acoustic detection unit transmits the original excitation signal to the detection area and receives the time-domain vibration response signal returned by the detection area. The spectral analysis unit excites the detection area and obtains the spectral characteristic signal of the surface material of the detection area under excited emission. S2. Extract the acoustic stiffness features of the time-domain vibration response signal to obtain the first feature vector set; S3. Extract the surface stress features of the spectral feature signal to obtain a second feature vector set; S4. Merge the first feature vector set and the second feature vector set to obtain a collaborative feature set; S5. Input the collaborative feature set into a pre-built and trained deep belief network model to obtain the defect type.

2. The equipment quality inspection method based on multi-source data according to claim 1, characterized in that, The specific steps of S2 include: S201. Perform deconvolution processing on the time-domain vibration response signal to obtain an intrinsic transfer function sequence; S202. Obtain the theoretical propagation speed range of the original excitation signal in the material of the equipment under test, and obtain the time threshold of each type of wave arriving at the receiver based on the theoretical propagation speed range and the geometric thickness parameters of the detection area. The intrinsic transfer function sequence is segmented based on the time threshold and divided into three continuous and non-overlapping physical information windows on the time axis: the first time window, the second time window, and the third time window. S203. Compare the frequency domain energy distribution within each physical information window with the original excitation signal, calculate the dynamic percentage of the extra energy offset outside the excitation frequency band to the total response energy, obtain the frequency band energy leakage factor, and obtain the energy leakage gradient vector based on the frequency band energy leakage factor. S204. Extract the envelope decay curve of the vibration waveform within the second time window, perform linear trend analysis in the logarithmic domain on the envelope decay curve to obtain the dynamic damping ratio; compare the peak frequency offset of the inherent transfer function sequence with the peak frequency of the standard transfer function of the same type of health equipment to obtain the dynamic stiffness deviation value. S205. Extract the bottom attenuation coefficient of the third time window, and stack the energy leakage gradient vector, the dynamic damping ratio, the dynamic stiffness deviation value and the bottom attenuation coefficient vertically according to the same spatial coordinate to form a four-dimensional structural response original parameter matrix; perform extreme value normalization processing on the structural response original parameter matrix, expand the normalized matrix into a one-dimensional feature sequence according to the preset grid division rules, and bind it with the corresponding spatial coordinate label to form a first feature vector set.

3. The equipment quality inspection method based on multi-source data according to claim 2, characterized in that, The specific steps of S201 include: A1. Taking the start time of the original excitation signal as zero point, scan the first arriving interface reflection wavefront in the time-domain vibration response signal, calculate the propagation time difference between the interface reflection wavefront and the start time, obtain the system delay compensation amount, and obtain the time-calibrated excitation-response data pair based on the system delay compensation amount. A2. Based on the excitation-response data pair, obtain the first wave interval that appears first on the time axis in the time-domain vibration response signal, perform morphological analysis on the waveform profile of the first wave interval, obtain the energy transmission loss coefficient and the waveform broadening distortion factor, and combine the energy transmission loss coefficient and the waveform broadening distortion factor in a time and space to construct a system interference feature template. A3. Based on the energy transmission loss coefficient, the waveform broadening distortion factor and the system delay compensation, the time-domain vibration response signal is corrected to obtain the initial residual signal, and the effective response wavelet parameter list is obtained based on the initial residual signal. A4. Based on the standard wall thickness of the equipment under test and the known sound velocity range, calculate the theoretical time window for the sound wave to travel through the wall thickness once round trip. Based on the theoretical time window, process the effective response wavelet parameter list to obtain the acoustic impedance relative change trend sequence. Normalize the amplitude of the acoustic impedance relative change trend sequence to generate the intrinsic transfer function sequence.

4. The equipment quality inspection method based on multi-source data according to claim 3, characterized in that, In step A3, the specific steps for obtaining the effective response wavelet parameter list based on the initial residual signal include: a1. On the time axis of the initial residual signal, search for the position of the extreme point with the largest absolute value of amplitude, and take the extreme point as the position center to extract a local signal segment with the same duration as the original excitation waveform of the original excitation signal. a2. Perform shape similarity matching between the overall envelope shape of the local signal segment and the original excitation waveform to obtain a shape similarity value; calculate the amplitude compression ratio and arrival time delay increment of the local signal segment relative to the original excitation waveform; if the shape similarity value is higher than a preset first threshold, determine that the local signal segment is an effective structural response wavelet, and obtain effective parameters based on the amplitude compression ratio, the delay increment, and the phase polarity; a3. Obtain the parasitic oscillation component in the local signal segment that matches the shape of the system interference feature template, delete the parasitic oscillation component in the initial residual signal, and obtain a new residual signal; a4. Return to a1, update the initial residual signal to the new residual signal, and iteratively execute a1-a4 until the maximum absolute amplitude of the current residual signal drops below a preset percentage of the amplitude of the first wave of the time-domain vibration response signal, and determine that the iteration has converged; obtain the list of effective response wavelet parameters based on the effective parameters.

5. The equipment quality inspection method based on multi-source data according to claim 4, characterized in that, The specific steps of a3 include: a301. Based on the delay increment, the system interference feature template is translated as a whole on the time axis so that the peak phase of the system interference feature template completely overlaps with the peak phase of the parasitic oscillation component in the local signal segment. A preset protection window is extended on each side of the translated system interference feature template. The template waveform within the protection window is subjected to edge shaving processing to generate interference stripping sub-wave. a302. Compare the amplitude compression ratio with the maximum absolute amplitude of the local signal segment to generate a projection weight coefficient. Multiply the projection weight coefficient by the interference stripping wavelet to generate an interference projection component. a303. The local signal segment is sampled point by point in the time domain and compared with the interference projection component to determine the morphological coherence, thereby generating a purified local signal segment; a304. The purified local signal segments are backfilled into the global residual signal of the current round based on the original start and end positions of the local signal segments when they were extracted. The global residual signal after backfilling is then subjected to a full-time domain smooth transition process to generate an updated global residual signal. a305. Calculate the percentage decrease in the total energy of the updated global residual signal relative to the total energy before the start of this iteration. If the percentage decrease is greater than the preset significant change threshold, the stripping operation is deemed effective, and the updated global residual signal is taken as the new residual signal. If the percentage decrease is less than or equal to the significant change threshold, it is determined that there are no obvious parasitic interference components that can be stripped from the global residual signal, the current residual signal is marked as basically pure, a convergence command is output, and the iteration loop ends.

6. The equipment quality inspection method based on multi-source data according to claim 5, characterized in that, The specific steps of a303 include: For each sampling point, determine whether the amplitude direction of the local signal segment at that point is consistent with the amplitude direction of the interference projection component; if the two directions are consistent, it is determined that the energy part of the sampling point comes from the contribution of parasitic interference, and the absolute value of the amplitude of the interference projection component at that point is subtracted from the amplitude of the sampling point in the local signal segment. If the two directions are opposite, the signal shape at that sampling point is determined to be unrelated to interference, and the original value is maintained.

7. The equipment quality inspection method based on multi-source data according to claim 6, characterized in that, The specific steps of S3 include: S301. Remove the strong continuous background of the spectral feature signal to obtain an effective time frame. Divide the effective time frame according to the delay time after excitation to obtain a time segmented spectral data group. The time segmented spectral data group includes early transient data, mid-term steady-state data and late decay data. S302. Obtain high-energy state spectral lines and ground-state spectral lines with preset fixed wavelengths, and obtain effective time-series segmented spectral data based on the time-series segmented spectral data set, the high-energy state spectral lines and the ground-state spectral lines. S303. Obtain the matrix main element and key alloying element of the equipment under test, obtain the internal standard reference line of the matrix main element and the target characteristic spectral line of the key alloying element, obtain the integral intensity value of the target characteristic spectral line in the mid-term steady-state data of the effective time-series segmented spectral data group, obtain the ratio of the integral intensity value to the internal standard intensity value of the internal standard reference line, and obtain the relative intensity ratio. Obtain the theoretical standard composition table of the equipment to be tested, and obtain the theoretical standard relative ratio based on the theoretical standard composition table. For the key alloying elements, calculate the percentage fluctuation of the relative strength ratio relative to the theoretical standard relative ratio. Arrange the floating percentages in the order of fixed elements to obtain the stoichiometric deviation vector. S304. In the late decay data of the effective time-series segmented spectral data group, obtain the spectral peak of the corrosion-sensitive element, obtain the net integral intensity value of the spectral peak, compare the net integral intensity value with the preset corrosion warning intensity threshold, and obtain the absolute intensity level of the corrosion product. The net oxygen intensity value is extracted and arranged in chronological order to form an oxygen intensity-time decay sequence. The oxygen intensity-time decay sequence is then fitted with a linear trend to obtain an oxygen decay line. The slope direction and slope magnitude of the oxygen decay line are then calculated. The corrosion activity index is obtained based on the absolute strength level of the corrosion products, the tilt direction, and the tilt amplitude. Based on the absolute strength level of the corrosion products and the corrosion activity index, a surface corrosion state characteristic tuple is obtained; S305. In the mid-term steady-state data of the effective time-series segmented spectral data set, select the specific dislocation-sensitive spectral line that is most sensitive to lattice strain as the target spectral line. Obtain a reference spectral profile template corresponding to a healthy standard part of the same material and heat treatment state as the equipment under test. The reference spectral profile template includes the standard half-width and standard peak center position of the target spectral line under stress-free and distortion-free state. The target spectral line profile of the target spectral line is completely superimposed on the reference spectral line profile template at the same scale. Obtain the half-width at half-maximum (WHM) value and the peak center wavenumber position of the target spectral line, calculate the percentage increase in WHM value relative to the standard WHM value, and calculate the offset of the peak center wavenumber position relative to the standard peak center position. The percentage of broadening increment is mapped to a dimensionless lattice distortion tensor factor, and the peak center wavenumber position is converted into a residual stress pointing vector, which includes stress direction components and stress amplitude components. S306. Multiply the time-series segmented spectral data group marked as low confidence by the confidence attenuation weighting coefficient to obtain the attenuated segmented spectral data group. Based on the attenuated segmented spectral data group and the time-series segmented spectral data group marked as high confidence, obtain the cleaned data. The stoichiometric deviation vector, surface corrosion state feature tuple, lattice distortion tensor factor, and residual stress pointing vector are vertically stacked according to a preset dimensional order to obtain the original feature row vector of the material state. The original feature vectors of the material state are divided according to the grid division rules, normalized, expanded into a one-dimensional sequence, and bound to the corresponding spatial coordinate labels to form the second feature vector set.

8. The equipment quality inspection method based on multi-source data according to claim 7, characterized in that, In step S302, the specific steps for obtaining a valid time-series segmented spectral data set include: In the early transient data, the intermediate steady-state data, and the late decay data, the intensity ratio of the high-energy spectral line and the ground-state spectral line in each frame is obtained. With time as the horizontal axis and the intensity ratio as the vertical axis, the convergence trajectory curve is obtained, and the convergence time, convergence speed, and stability value of the convergence trajectory curve are obtained. If the convergence time is within a preset convergence time range, the stable value is within a preset stable range, and the convergence speed is within a preset variation range, then the time-series segmented spectral data set is determined to be valid and marked with high confidence. If the convergence time is not within the preset convergence time range, but the convergence speed is within the preset variation range, then it is determined that there are abnormal attachments in the detection area, and a low confidence level is marked. If the convergence time is not within the preset convergence time range and the convergence speed is not within the preset variation range, then the time-series segmented spectral data group is determined to be invalid and the data group is discarded. If the convergence time is within a preset convergence time range, the convergence speed is within a preset variation range, but the stable value is not within the preset stable range, then the material of the equipment under test is deemed questionable.

9. The equipment quality inspection method based on multi-source data according to claim 8, characterized in that, The specific steps of S4 include: The data of each fine grid node in the second feature vector set is assigned to the nearest coarse grid node in the first feature vector set according to the spatial coordinate label, and a weighted average is performed on all spectral data belonging to the same coarse grid. Feature vectors belonging to the same coarse grid are concatenated vertically to form a joint feature vector; For each grid node, the response sensitivity of acoustic features to spectral features and the modulation response of spectral features to acoustic features are calculated, and the cross-domain response gradient and modulation coefficients are obtained respectively. By stacking the joint feature vector, cross-domain response gradient, and modulation coefficient according to spatial coordinate labels in three dimensions, a collaborative feature set is obtained.

10. An equipment quality inspection system based on multi-source data, characterized in that, The system includes: Data Unit: Used to simultaneously issue start commands based on the synchronous control module for the same detection area on the surface of the equipment under test, driving the acoustic detection unit and the spectral analysis unit to synchronously perform data acquisition actions. The acoustic detection unit emits the original excitation signal to the detection area and receives the time-domain vibration response signal returned by the detection area. The spectral analysis unit excites the detection area and obtains the spectral characteristic signal of the surface material of the detection area under excited emission. First feature unit: used to extract the acoustic stiffness features of the time-domain vibration response signal to obtain a first feature vector set; Second feature unit: used to extract surface stress features from the spectral feature signal to obtain a second feature vector set; Fusion unit: used to fuse the first feature vector set and the second feature vector set to obtain a collaborative feature set; Result unit: used to input the collaborative feature set into a pre-built and trained deep belief network model to obtain the defect type.

Images