An ultrasonic detection-based defect type intelligent judgment system and method

Abstract

The present application belongs to the technical field of nondestructive testing, in particular to a defect type intelligent judgment system and method based on ultrasonic detection. The system comprises: a data acquisition module for acquiring ultrasonic A-scan waveform and probe position; a feature extraction module for extracting five types of multi-dimensional quantitative features of waveform, amplitude, space, dynamics and frequency domain from the waveform; a feature database for storing samples constructed according to industry standards; an intelligent classification module for loading a classification model to judge the defect type; and a result display module for outputting the type, confidence and suggestion. The method comprises: collecting and preprocessing echo signals; extracting multi-dimensional features to construct a feature vector; inputting a pre-trained model for classification; and outputting the classification result. The present application quantizes standard experience and uses artificial intelligence for judgment, solving the problems of strong subjectivity, low efficiency and poor consistency of manual judgment, and greatly improving the automation, standardization and intelligence level of ultrasonic detection defect qualification.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to an intelligent system and method for determining defect types based on ultrasonic testing. Background Technology

[0002] Currently, industrial ultrasonic testing mainly relies on inspectors to manually determine the type of defect based on the waveform, amplitude, and location of the reflected echo, combined with personal experience and industry standards (such as TSG 21-2016 and NBT 47013.3-2023). This method of judgment has the following problems: Highly subjective: Differences in experience among different testing personnel lead to inconsistent judgment results; Inefficient: Manual analysis is time-consuming, especially in batch testing; Inconsistent standard implementation: Although the standards specify the characteristics of defects, they lack quantitative and automated implementation methods; It is difficult to distinguish certain defect types: for example, planar defects (such as cracks and lack of fusion) and volumetric defects (such as pores and inclusions) overlap in echo characteristics, making manual differentiation difficult.

[0003] Existing ultrasonic testing equipment can only provide basic data such as the depth, location, and amplitude of defects, and does not have automatic classification functions. A few studies have attempted to introduce image processing or simple threshold judgment, but they generally suffer from problems such as single feature dimensions, poor model adaptability, and lack of deep integration with industry standards, which cannot meet the high reliability and high consistency requirements of industrial sites. Summary of the Invention

[0004] Based on the technical problems existing in the prior art, this invention proposes an intelligent defect type judgment system and method based on ultrasonic detection.

[0005] This invention proposes an intelligent defect type judgment system based on ultrasonic detection, comprising: The data acquisition module is used to acquire the A-scan waveform signal and probe position information output by the ultrasonic testing equipment; The feature extraction module is used to extract multi-dimensional feature vectors from the A-scan waveform signal. The multi-dimensional features are specially designed based on the physical principles of ultrasound signals and the time-domain, frequency-domain, spatial and dynamic scanning characteristics, and include at least waveform features, amplitude features, spatial features, dynamic features and frequency domain features. A feature database is used to store sample feature vectors of known defect types; The intelligent classification module is loaded with a pre-trained classification model, which is used to determine the defect type based on the feature vector. The results display module is used to output the defect type, confidence level, and handling suggestions.

[0006] Preferably, the waveform features include peak count, -6dB width, and waveform symmetry index; the amplitude features include maximum relative amplitude and attenuation slope obtained by logarithmic linear fitting of the echo envelope attenuation segment, which directly reflects the acoustic scattering intensity inside the defect; the spatial features include defect depth, indication length, and self-height calculated based on sound velocity, probe refraction angle, and geometric trigonometric relationships, where self-height is a key geometric parameter for distinguishing planar and volumetric defects; the dynamic features include echo envelope undulation during probe movement, used to quantify the continuity of the defect in the scanning direction; and the frequency domain features include echo signal main frequency shift and spectral energy distribution entropy extracted by short-time Fourier transform, used to capture the influence of the defect on the frequency components of the acoustic wave.

[0007] Preferably, the classification model is a machine learning model based on support vector machines, random forests, or convolutional neural networks.

[0008] Preferably, it also includes a preprocessing module for bandpass filtering and gain normalization of the A-scan waveform signal.

[0009] A method for intelligently determining defect types based on ultrasonic testing includes the following steps: Step S201: Acquire ultrasonic detection echo signals and probe position information; Step S202: Preprocess the echo signal; Step S203: Extract multi-dimensional features based on the physical principles of ultrasound, and construct a feature vector that integrates time domain, frequency domain, spatial and dynamic information; Step S204: Input the feature vector into the pre-trained classification model to obtain the defect type and confidence level; Step S205: Output the classification results.

[0010] Preferably, the extraction of multi-dimensional features includes: The envelope of the preprocessed A-scan signal is obtained by performing Hilbert transform, and the time-domain waveform characteristics are calculated. The attenuation slope is calculated by linear fitting of the envelope attenuation segment, and the amplitude feature is formed by combining it with the maximum relative amplitude. Based on the probe position, sound velocity, refraction angle, and trigonometric relationships, the depth, indication length, and height of the defect are calculated. During continuous scanning, multiple frames of echo amplitude sequences are extracted, and their statistical fluctuations are calculated as dynamic features. A short-time Fourier transform is performed on the echo signal to extract the main frequency offset and spectral entropy as frequency domain features.

[0011] Preferably, the classification model is trained through the following steps: Collect samples of known defect types, extract multi-dimensional feature vectors, and label the true types; The model is trained using support vector machines or random forest algorithms, and the model parameters are optimized through cross-validation. Save the trained model for subsequent classification.

[0012] Preferably, if the confidence level output by the classification model is lower than a set threshold, it is judged as "uncertain" and manual review is recommended.

[0013] Preferably, the preprocessing includes bandpass filtering and signal normalization, with the normalization formula being: s_norm[i]=(s_denoised[i]-μ) / σ, where μ is the mean and σ is the standard deviation.

[0014] Preferably, the defect types include planar defects, volumetric defects, and non-defects; wherein planar defects include cracks and lack of fusion, and volumetric defects include porosity and inclusions.

[0015] Compared with existing technologies, the present invention provides an intelligent defect type judgment system and method based on ultrasonic detection, which has the following beneficial effects: By using automated and standardized multi-dimensional feature extraction and intelligent model analysis, the time for judging defect types has been reduced from several minutes in traditional manual analysis to seconds, achieving near real-time judgment. At the same time, the system makes decisions entirely based on quantitative features and training data, completely eliminating the subjectivity and arbitrariness caused by differences in human experience, and ensuring the objectivity and fairness of the detection conclusions.

[0016] This invention transforms the qualitatively described defect features in industry standards into calculable multi-dimensional quantitative features such as waveform, amplitude, space, and dynamics. It also uses machine learning models to learn their complex and nonlinear intrinsic relationships. Through training with a large number of standard samples and field verification samples, the system can achieve a classification accuracy of over 90% for typical defects such as cracks, porosity, inclusions, and lack of fusion. In particular, its ability to distinguish between easily confused planar and volumetric defects far exceeds the average level of human intervention.

[0017] For the same defect, regardless of which inspector operates the system or when the inspection is performed, the system will output a judgment result based on the same features and model, achieving standardization and high consistency in inspection quality. In addition, the system automatically saves complete raw data, extracted feature values, model judgment results and confidence levels, forming a complete and tamper-proof electronic record chain, which greatly enhances the traceability of the inspection process and provides a solid data foundation for quality review, responsibility definition and process improvement.

[0018] This invention constructs a quantitative feature system that closely integrates the physical principles of ultrasonic acoustics and industry standards (such as NB / T 47013.3). For example, "self-height" directly corresponds to the defect height measurement in the standard, "attenuation slope" reflects the scattering characteristics of the defect, and "undulation" describes the continuity of the defect. These features have clear physical meaning and engineering interpretability, laying a solid foundation for high-precision classification.

[0019] The system's feature database and classification model are not fixed. Through online learning or offline incremental learning mechanisms, new defect samples and their features that have been subsequently verified (such as X-rays and dissections) can be continuously incorporated into the training set, enabling the model to continuously adapt to new defects that may be generated by new materials and new processes, thereby achieving self-iteration and evolution of system performance and maintaining the advanced nature and applicability of the technology.

[0020] This system can be used as a standalone host computer software to interface with existing mainstream ultrasound testing equipment, or it can be integrated into a new generation of intelligent ultrasound equipment as a functional module. Its implementation does not require changes to the existing standard testing procedures, but it can provide users with unprecedented intelligent analysis capabilities, thereby bringing significant comprehensive economic and social benefits in terms of improving testing quality and efficiency, reducing the risk of misjudgment and missed judgment, and realizing digital archiving and management. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention.

[0022] In the diagram: 101, Ultrasonic testing equipment; 102, Data acquisition module; 103, Preprocessing module; 104, Feature extraction module; 105, Feature database; 106, Intelligent classification module; 107, Result display module. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0024] like Figure 1 As shown, the intelligent defect type judgment system based on ultrasonic detection of the present invention can be physically implemented as a stand-alone industrial control computer, an embedded device, or integrated into an ultrasonic testing instrument. Its logical architecture and data flow are as follows: Ultrasonic testing equipment 101: A conventional A-type pulse-echo digital ultrasonic flaw detector, such as the Olympus EPOCH series or similar domestic equipment, is used. The probe of the equipment is a longitudinal wave angle probe (such as K2, frequency 5MHz). Its output interface (usually USB or Ethernet) should be able to transmit complete A-scan waveform data packets in real time. The data packet structure should include: time series array s_raw[i], corresponding voltage amplitude array V_raw[i], sampling rate Fs (typical value not less than 100MHz), current pulse repetition frequency PRF, and calibration parameters such as probe zero point and sound velocity v.

[0025] Data acquisition module 102: This module consists of a high-speed data acquisition card and a position encoder interface.

[0026] Data acquisition card: Receives the V_raw[i] signal from the ultrasound machine. To ensure waveform detail, the sampling rate must be at least 10 times the probe's center frequency (≥50MHz in this example). The acquisition card triggers synchronously with the ultrasound machine's transmitted pulse.

[0027] Position synchronization: The probe moves via a mechanical scanning device or manual scanning with an encoder. The encoder (such as an incremental photoelectric encoder) outputs the probe's three-dimensional coordinates (x, y, z) in the weld seam coordinate system in real time, where x is along the weld seam direction, y is perpendicular to the weld seam direction, and z is the depth direction (calculated from the probe angle and sound path). A dedicated channel of the acquisition card synchronously acquires the encoder pulses, converts them into spatial coordinates through software, and aligns them with the timestamp of each frame of A-scan data s_raw[i] to form a data pair (s_raw[i], x, y, z).

[0028] Preprocessing module 103: This module performs purification and standardization on the original signal s_raw[i], as follows: Bandpass filtering: Design a 5th-order Butterworth bandpass filter with a passband frequency range of [f_center*0.5, f_center*1.5], i.e. [2.5MHz, 7.5MHz], to retain the main frequency information and suppress high-frequency noise and low-frequency interference. The filtered signal is denoted as s_filt[i].

[0029] Noise suppression: Wavelet thresholding denoising is employed, using the db8 wavelet basis and a 5-level decomposition. A soft thresholding function η(d) = sign(d) * max(|d| - T, 0) is applied to the detail coefficients of each level. The threshold T is calculated according to the Donoho-Johnstone criterion σ * sqrt(2 * log(N)), where σ is the noise standard deviation of the median absolute deviation estimate of the first-level detail coefficients, and N is the signal length. The denoised signal is denoted as s_denoised[i].

[0030] Gain Normalization: To eliminate amplitude differences caused by different detection sensitivity (dB gain) settings, amplitude normalization is performed on the entire A-scan signal. First, the maximum absolute value of the amplitude, V_max = max(|s_denoised[i]|), is found across the entire screen (excluding the initial pulse blind zone). Then, all data points are scaled proportionally to the reference range [-1, 1]: s_norm[i] = s_denoised[i] / V_max. This step ensures the comparability of amplitude features extracted under different detection conditions.

[0031] Feature extraction module 104: This module takes the preprocessed signal s_norm[i] and its corresponding spatial coordinates (x, y, z) as input. After the operator locks onto a defect echo through a gate on the instrument, the module automatically analyzes the data segment within the gate (time window t1 to t2, corresponding to sample indices i1 to i2). It calculates 17 features across four categories, forming the feature vector F: 1. Waveform characteristics (time domain analysis): Peak count (Np): Calculate the signal envelope. First, perform a Hilbert transform on s_norm[i] to obtain the analytic signal, and take its magnitude to obtain the envelope E[i]. Set the dynamic threshold V_th=0.3*max(E[i]), and count the number of independent peaks in E[i] that exceed V_th. For volumetric defects (such as porosity groups), Np is usually greater than 2; while for simple planar defects (such as cracks), Np is usually less than or equal to 2.

[0032] -6dB width (W6dB): Find the maximum value point i_max of the envelope E[i] and its amplitude E_max. Search to the left and right, find the points i_left and i_right where the amplitude first drops to E_max*10^(-6 / 20)≈0.5*E_max, then W6dB=(i_right-i_left) / Fs (unit: µs). This feature is related to the acoustic projection width of the defect.

[0033] Waveform symmetry (Sym): Using the aforementioned i_max as the dividing point, calculate the ratio of the envelope energy of the first half to the second half.

[0034] Sym = sum(E[i1:i_max]^2) / sum(E[i_max:i2]^2). A Sym value that deviates significantly from 1 (e.g., >2.5 or <0.4) usually indicates that the echo waveform is asymmetrical, which is a characteristic of some tilted cracks or lack of fusion.

[0035] 2. Amplitude characteristics (energy level analysis): Relative amplitude (A_rel): A_rel=20*log10(E_max) (unit: dB), directly reflects the reflection intensity of the defect.

[0036] Attenuation slope (α): A linear fit is performed on the envelope attenuation segment to the right of the peak point i_max (e.g., from i_max to i2), and the fitting equation is log(E[i])≈k*i+b, then the attenuation slope α=-k*Fs (unit: MHz). This feature is one of the key points of this invention. Internal scattering of volumetric defects (such as pores) leads to rapid attenuation of acoustic energy (large α value), while planar defects (such as cracks) attenuate slowly (small α value).

[0037] 3. Spatial characteristics (geometric positioning): Defect depth (D): D=v*(i_max / Fs) / 2 (direct method) or calculated based on the geometric relationship of the angled probe. This is the basis for defect location.

[0038] Indication length (L): During continuous scanning, the probe movement distance when the echo amplitude of the defect exceeds the evaluation line (e.g., -12dB) is determined by analyzing the changes in A_rel within the defect gate in a series of continuous A-scan frames. The system finds the start and end coordinates x_start and x_end of A_rel>(A_rel_max-12), and then L=|x_end-x_start|.

[0039] Height (H): The endpoint diffraction method or the 6dB method is used. This example uses the 6dB method: At the probe position where the defect echo is strongest, a gate is placed over the defect echo. The sound paths S1 and S2 corresponding to the two endpoints where the amplitude drops by 6dB are found. Based on the probe refraction angle β and the workpiece thickness, H = |S1*cosβ - S2*cosβ| is calculated using geometric trigonometric relationships. H is key to distinguishing between planar (small height) and volumetric (relatively large height) defects.

[0040] 4. Dynamic characteristics (scanning process analysis): Echo envelope undulation (F): During the calculation of the indicated length L, the defect relative amplitude sequence A_seq[j] of all valid A-scan frames within that length is obtained, j=1...M. Its undulation is calculated as: F=std(A_seq) / mean(A_seq). Stable and continuous crack echoes have lower F values; while discontinuous and irregularly shaped inclusion echoes have higher F values.

[0041] 5. Frequency domain characteristics (spectral analysis): Dominant frequency shift (Δf): Perform a short-time Fourier transform on the defect echo band s_norm[i1:i2], calculate its spectrum, find the frequency f_defect corresponding to the maximum amplitude of the spectrum, and compare it with the dominant frequency f_ref of the reference signal in the defect-free area. Δf = f_defect - f_ref. Scattering bodies such as pores may cause the dominant frequency to shift to lower frequencies.

[0042] Spectral entropy (SE): Calculate the normalized energy distribution P(k) of the above defect echo spectrum. Then the spectral entropy SE = -ΣP(k)log2(P(k)). A high entropy value indicates that the spectral energy is dispersed, which may correspond to defects with complex shapes (such as inclusions).

[0043] Finally, the feature vector F = [Np, W6dB, Sym, A_rel, α, D, L, H, F, Δf, SE, ... (other derived features)]^T.

[0044] Feature Database 105: This is a structured electronic database containing two main parts: Standard Feature Library: Based on the typical defect maps and descriptions in Appendix NBT 47013.3-2023, feature vectors are artificially synthesized or extracted from a large number of measured typical defects, and are accompanied by defect type labels marked by experts (such as “Crack”, “Pore”, “Slag”).

[0045] User history database: Stores the features of defect samples that have been finally verified (e.g., by X-ray, dissection) during each on-site inspection and their final determination type, for continuous optimization and incremental learning of subsequent models.

[0046] Intelligent classification module 106: Loads a pre-trained classification model to determine the defect type based on feature vectors. The classification model is a machine learning model based on support vector machine, random forest, or convolutional neural network. Its training data comes from typical defect samples defined in TSG21-2016 and NBT 47013.3-2023 standards. This embodiment adopts a hybrid cascaded classification model to fully utilize the discriminative power of different features, balancing accuracy and real-time performance.

[0047] Level 1: Binary Classifier (Shear Defects vs. Volumetric Defects). A linear kernel Support Vector Machine (SVM) is used, with the input being a feature subset F1=[Sym,H,F]. Training data is selected from a standard feature library, focusing on clearly defined shear defects (cracks, lack of fusion) and volumetric defects (porosity, inclusions). This level of the model is computationally fast and can quickly perform coarse classification of defects.

[0048] Second level: Fine classifier. Based on the results of the first level, different models are selected for fine classification.

[0049] If the defect is classified as a planar defect, the input feature F2=[Np,Sym,W6dB,α,L / H] is fed into a RandomForest classifier, which further distinguishes between "cracks", "lack of fusion", and "incomplete penetration". The RandomForest contains 100 decision trees and uses the Gini index as the splitting criterion.

[0050] If the defect is classified as a volumetric defect, the input feature F3=[Np,A_rel,α,F, SE] is fed into another random forest classifier to distinguish between "pores", "clinker", and "dense pores / clinker". Here, the frequency domain feature SE and the dynamic feature F are introduced, which can effectively distinguish volumetric defects with different scattering characteristics.

[0051] Model training process: Randomly allocate 80% of the data from the feature database as the training set and 20% as the test set. Use grid search and 5-fold cross-validation to optimize the SVM penalty factor C and hyperparameters such as the maximum depth and minimum number of leaf node samples in the random forest. After training, save the model file (e.g., .pkl or .onnx format) for the classification module to load.

[0052] Confidence calculation: For SVM, the PlattScaling method is used to convert the decision distance into probability; for random forest, the voting ratio is used as the confidence score; the final output of the system is {defect type, confidence score P}, where P∈[0,1].

[0053] Result Display Module 107: This module provides a graphical user interface (GUI), the main areas of which include: Real-time A / B / C scan display area: Displays the currently acquired A-scan waveform and highlights the analyzed area with a color gate.

[0054] Feature value and dynamic curve display area: Real-time display of calculated key feature values, and displays the changes of features such as A_rel and Sym in the form of curves during probe movement.

[0055] Defect Classification Results Panel: The classification results are displayed in a prominent font, such as: [Judgment Result: Crack (Surface Defect)], and the confidence bar and value [Confidence: 93%] are displayed below.

[0056] Handling suggestion box: Based on the defect type, size, and standard, it automatically provides suggestions, such as: "According to NB / T 47013.3, the crack indicator length is 8mm and its height is 1.2mm. It is recommended to assess it as Class IV and conduct a re-inspection using radiographic testing." Data management and report generation: All test data, feature values, model judgment results and original waveforms are automatically saved, and test reports that meet the TSG21-2016 format requirements can be generated with one click.

[0057] Reference Figure 2 This invention discloses an intelligent defect type judgment method based on ultrasonic testing. This embodiment, combined with a specific inspection case (a longitudinal weld seam of a pressure vessel cylinder, 30mm thick, scanned using a K1 angle probe), details the entire process from inspection to judgment: S201: Acquire ultrasonic echo signals and probe position information: This step is performed by the data acquisition module 102. During the scan, at the probe position x=152mm, the A-scan waveform shows a significant echo at a sound path of 25µs, exceeding the set alarm threshold; the data acquisition module 102 captures the waveform s_raw (length 1024 points, Fs=100MHz) and the current coordinates (152,5,0); the system automatically surrounds the echo with the gate (gate width 10µs, gate height 50%), and records the data indices i1=2400, i2=2600 inside the gate.

[0058] S202: Preprocess the echo signal: This step is performed by the preprocessing module 103, which includes bandpass filtering and signal normalization.

[0059] Bandpass filtering: Apply a [2.5,7.5]MHz Butterworth bandpass filter to s_raw[2400:2600] to obtain s_filt.

[0060] Signal normalization: Wavelet denoising is performed on s_filt to obtain s_denoised. The maximum absolute value of s_denoised in this segment, V_max = 1.85V, is calculated. According to the normalization formula s_norm[i] = (s_denoised[i] - μ) / σ, where μ is the mean and σ is the standard deviation, or the simplified normalization s_norm = s_denoised / V_max in this example is used to obtain the normalized signal s_norm, which has a uniform dynamic range.

[0061] S203: Extract multi-dimensional features and construct feature vectors: This step is performed by the feature extraction module 104.

[0062] Calculate waveform characteristics: Calculate the envelope E[i] for s_norm and find a single peak, so the number of peaks Np=1. Calculate the -6dB width and get W6dB=0.4µs. Calculate the symmetry Sym=sum(E[2400:2500]^2) / sum(E[2500:2600]^2)=0.31.

[0063] Calculation of amplitude characteristics: Maximum relative amplitude A_rel=20*log10(0.92)≈-0.7dB, and the attenuation slope α=0.8MHz is obtained by fitting the attenuation segment.

[0064] Calculation of spatial characteristics: Based on the sound velocity and sound path, the defect depth D≈18.7mm was calculated, the indicator length L=6mm was measured along the scanning direction, and the height H=1.5mm was measured using the 6dB method.

[0065] Calculate dynamic characteristics: Analyze the amplitude sequence during the scanning process, with echo envelope undulation F=0.15.

[0066] Frequency domain characteristics were calculated: FFT analysis was performed on s_norm[2400:2600], and the main frequency offset Δf = -0.15MHz and the spectral entropy SE = 2.3 were obtained.

[0067] Finally, a feature vector F containing all five types of information for the defect is constructed.

[0068] S204: Input the feature vector into the pre-trained classification model to obtain the defect type and confidence level: This step is performed by the intelligent classification module 106.

[0069] Model application: Input the feature vector F into the pre-trained hybrid cascade classification model. The first-level SVM judges it as "surface defect" based on [Sym,H,F]; the second-level random forest judges it as "crack" based on [Np,Sym,W6dB,α,L / H].

[0070] Confidence output: The model also outputs the confidence score of the judgment, P_final=0.93. The system internally sets the confidence threshold to 0.85. Since 0.93>0.85, the system directly outputs the judgment result. If the confidence score is lower than this threshold, the system will mark "uncertain" in the result display module 107 and flash a prompt suggesting manual review.

[0071] S205: Output classification results: This step is performed by the results display module 107. The system interface displays the final judgment result [Crack (Planar Defect)] and a confidence level of 93%, and generates processing suggestions.

[0072] The training of the classification model of this invention is an independent offline or online preparation process, specifically implemented as follows: Sample preparation: A large number of samples of known defect types (covering planar defects, volumetric defects and non-defects, with planar defects including cracks and lack of fusion, and volumetric defects including porosity and inclusions) were collected from feature database 105. Each sample includes its 17-dimensional feature vector F_i and an expert-confirmed defect type label y_i.

[0073] Model training and optimization: Using machine learning libraries such as Scikit-learn, the dataset was divided into training and test sets in an 8:2 ratio. Support Vector Machine (SVM) and Random Forest (RF) algorithms were used for training respectively. Grid Search was used to optimize the hyperparameters of SVM, such as the penalty factor C and kernel function parameter γ, as well as the number of trees and maximum depth of RF. Five-fold cross-validation was used to evaluate the performance and prevent overfitting.

[0074] Model saving and loading: The trained optimal model parameters are serialized and saved as files. When the intelligent classification module 106 is initialized, these model files are loaded and can be used for subsequent real-time classification and judgment.

[0075] System Deployment and Operation: Install the system software on an industrial touchscreen computer and connect the ultrasonic instrument and encoder via USB. After startup, the system is first calibrated (sound velocity, probe zero point, encoder scale). Operators only need to perform scanning as with a conventional ultrasonic instrument. When a defect echo is detected, the system automatically completes signal acquisition, preprocessing, feature extraction, model judgment, and displays the results in real time, greatly simplifying the operation process.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A defect type intelligent judgment system based on ultrasonic detection, characterized in that, include: The data acquisition module (102) is used to acquire the A-scan waveform signal and probe position information output by the ultrasonic testing device (101); The preprocessing module (103) is used to perform bandpass filtering and gain normalization on the A-scan waveform signal; The feature extraction module (104) is used to extract multi-dimensional feature vectors from the A-scan waveform signal. The multi-dimensional features are specially designed based on the physical principles of ultrasound signals and the time-domain, frequency-domain, spatial and dynamic scanning characteristics, and include at least waveform features, amplitude features, spatial features, dynamic features and frequency-domain features. The waveform features include peak count, -6dB width, and waveform symmetry index; the amplitude features include maximum relative amplitude and attenuation slope obtained by logarithmic linear fitting of the echo envelope attenuation segment, which directly reflects the acoustic scattering intensity inside the defect; the spatial features include defect depth, indication length, and self-height calculated based on sound velocity, probe refraction angle, and geometric trigonometric relationships, where self-height is a key geometric parameter for distinguishing planar and volumetric defects; the dynamic features include echo envelope undulation during probe movement, used to quantify the continuity of the defect in the scanning direction; the frequency domain features include echo signal main frequency shift and spectral energy distribution entropy extracted by short-time Fourier transform, used to capture the influence of the defect on the frequency components of the acoustic wave; The feature database (105) is used to store sample feature vectors of known defect types; The intelligent classification module (106) is loaded with a pre-trained classification model, which is used to determine the defect type based on the feature vector. The classification model is a hybrid cascaded classification model. The first level is a binary classifier based on support vector machine, which is used to distinguish between planar defects and volumetric defects. The second level is a fine classifier based on random forest, which is used to further distinguish specific defect types. The results display module (107) is used to output the defect type, confidence level, and handling suggestions.

2. A method for intelligently determining defect types based on ultrasonic testing, characterized in that, Includes the following steps: Step S201: Acquire ultrasonic detection echo signals and probe position information; Step S202: Preprocess the echo signal; Step S203: Extract multi-dimensional features based on the physical principles of ultrasound, and construct a feature vector that integrates time domain, frequency domain, spatial and dynamic information; The extraction of multi-dimensional features based on ultrasonic physics principles includes: The envelope of the preprocessed A-scan signal is obtained by performing Hilbert transform, and the time-domain waveform characteristics are calculated. The attenuation slope is calculated by linear fitting of the envelope attenuation segment, and the amplitude feature is formed by combining it with the maximum relative amplitude. Based on the probe position, sound velocity, refraction angle, and trigonometric relationships, the depth, indication length, and height of the defect are calculated. During continuous scanning, multiple frames of echo amplitude sequences are extracted, and their echo envelope undulation is calculated as a dynamic feature. Short-time Fourier transform is performed on the echo signal to extract the main frequency offset and spectral entropy as frequency domain features; Step S204: Input the feature vector into the pre-trained classification model to obtain the defect type and confidence level; The classification model is trained through the following steps: Collect samples of known defect types, extract multi-dimensional feature vectors, and label the true types; The model is trained using support vector machines or random forest algorithms, and the model parameters are optimized through cross-validation. Save the trained model for subsequent classification; Step S205: Output the classification results.

3. The intelligent defect type judgment method based on ultrasonic detection according to claim 2, characterized in that, If the confidence level output by the classification model is lower than the set threshold, it is judged as "uncertain" and manual review is recommended.

4. The intelligent defect type judgment method based on ultrasonic detection according to claim 2, characterized in that, The preprocessing includes bandpass filtering, noise suppression, and signal normalization. The normalization formula is: s_norm[i]=(s_denoised[i]-μ) / σ, where s_denoised[i] is the A-scan signal after noise reduction, s_norm[i] is the normalized A-scan signal, μ is the mean, and σ is the standard deviation.

5. The intelligent defect type judgment method based on ultrasonic detection according to claim 2, characterized in that, The defect types include planar defects, volumetric defects, and non-defects; among which planar defects include cracks and lack of fusion, and volumetric defects include porosity and inclusions.

Images