Titanium plate welding position automatic detection system and method

Abstract

The application relates to the technical field of nondestructive testing, and particularly discloses a titanium plate welding position automatic detection system and method, which acquires original physical field signals of a to-be-detected position under single energy field excitation, extracts spatial energy attenuation gradient and time phase lag distribution through wavelet packet decomposition and Hilbert transformation, constructs a dynamic propagation model describing signal propagation path evolution rules, and performs space-time registration and vector difference calculation with a preset ideal reference model to generate a difference evolution atlas, identifies and confirms a stable abnormal mode caused by defects through high-dimensional topological feature mapping, density clustering and multi-scale persistence analysis on the atlas, traces back a dynamic evolution path of the abnormal mode, extracts defect core parameters, and combines process information to complete three-dimensional positioning, type classification and safety level evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nondestructive testing, in particular to a titanium plate welding site automatic detection system and method. BACKGROUND

[0002] Titanium and titanium alloys are widely used in aerospace, shipbuilding, chemical equipment and other high-end industrial fields due to their high strength, light weight and excellent corrosion resistance. These key application scenarios have put forward extremely strict requirements on the safety and reliability of titanium plate welding structures. Therefore, the quality detection of the welding site is an important link in the manufacturing process. At present, the automatic detection system based on nondestructive testing technology, especially the scheme combining machine vision and deep learning, has become a research and application hotspot in this field. Such technology usually relies on feature extraction and pattern recognition of collected image or signal data to realize automatic judgment of welding defects. However, the core logic of the existing technical method is mostly based on the "data-driven" paradigm, and its performance is highly dependent on the size and completeness of the training data set.

[0003] The existing technology has the following disadvantages:

[0004] The data-driven detection method is too sensitive to atypical weak interference outside the training data distribution, which leads to catastrophic misjudgment or missed detection. The existing method cannot effectively distinguish the signal changes caused by real defects from the changes caused by non-structural interference such as condensate film, because the model decision lacks the essence understanding of the physical process of signal propagation in the material.

[0005] The present application provides a solution that can fundamentally eliminate interference and accurately identify internal defects, especially for the weak interference scenarios similar to the optical characteristics of micro-cracks and other defects that occur in special working conditions of titanium plate welding sites. SUMMARY

[0006] The purpose of the present application is to provide a titanium plate welding site automatic detection system and method to solve the problems in the above background.

[0007] The purpose of the present application can be achieved by the following technical solutions:

[0008] A titanium plate welding site automatic detection method, comprising the following steps:

[0009] S1: obtaining the original physical field signal of the titanium plate welding site to be detected under the excitation of a single energy field;

[0010] S2: extracting the propagation dynamics characteristics of the original physical field signal to obtain the energy attenuation gradient of the original physical field signal in the spatial dimension and the phase lag distribution in the time dimension;

[0011] S3: Construct a dynamic propagation model describing the evolution law of the original physical field signal on the propagation path in the material based on the energy attenuation gradient and the phase lag distribution;

[0012] S4: Compare the dynamic propagation model with a preset ideal reference model representing the signal propagation of the standard titanium plate welding part without defects, generate a difference evolution map by calculating the evolution difference of the two at a specific space-time node;

[0013] S5: Perform pattern mining on the difference evolution map to identify abnormal patterns caused by signal propagation path disturbance caused by defects, and determine the defect information of the welding part.

[0014] As a further scheme of the application: the extraction process of the energy attenuation gradient comprises:

[0015] Wavelet packet decomposition is performed on the original physical field signal, and the original physical field signal is decomposed into a plurality of sub-band signals of different frequency ranges, each sub-band signal corresponding to a specific scale space;

[0016] For each sub-band signal, the energy value of each sampling point in the spatial dimension is calculated, and the energy value is obtained by calculating the square sum of the signal amplitude, forming a spatial energy distribution sequence of each sub-band;

[0017] The spatial energy distribution sequence of each sub-band is subjected to spatial gradient operation, and the energy change rate between adjacent sampling points is calculated by using the central difference method, to obtain the local energy gradient of each sub-band;

[0018] The local energy gradients of all sub-bands are weighted and fused, and the weight is determined based on the energy contribution degree of each sub-band signal, to finally generate a comprehensive energy attenuation gradient.

[0019] As a further scheme of the application: the extraction process of the phase lag distribution comprises:

[0020] The original physical field signal is subjected to Hilbert transform processing, and an analytical expression of the original physical field signal is constructed to extract the instantaneous phase information of the original physical field signal;

[0021] The geometric center point of the welding part is taken as a phase reference datum, and the instantaneous phase difference value of each spatial sampling point relative to the reference point is calculated;

[0022] The calculated phase difference value is subjected to spatial interpolation processing to generate a continuous phase lag distribution.

[0023] As a further scheme of the application: the construction process of the dynamic propagation model comprises:

[0024] The energy attenuation gradient is mapped to the energy dissipation rate attribute of the nodes on the space-time mapping grid, and the phase lag distribution is mapped to the signal propagation delay attribute of the nodes on the space-time mapping grid, so that each node on the grid contains a pair of energy and phase coupled attribute parameters;

[0025] Based on the space-time mapping grid, a signal propagation correlation rule between nodes is constructed, and an optimal path search algorithm for signal propagation from a grid source point to any node is defined based on the energy dissipation rate of an upstream node and the signal propagation delay of a downstream node as constraint conditions.

[0026] Based on the signal propagation correlation rule, dynamic evolution calculation is performed on the space-time mapping grid, and the attribute parameters of each node are iteratively updated to simulate the signal propagation process in the actual material, and finally a dynamic propagation model is generated.

[0027] As a further scheme of the application, the input and output of the dynamic propagation model specifically include:

[0028] The input of the dynamic propagation model is the energy attenuation gradient and the phase lag distribution;

[0029] The output of the dynamic propagation model is a dynamic evolution sequence of the signal on the propagation path in the material, and the dynamic evolution sequence is presented in the form of a space-time distribution map, which includes an abnormal convergence area of signal energy attenuation and a wavefront distortion area of phase propagation; the spatial coupling relationship between the abnormal convergence area and the wavefront distortion area directly corresponds to the geometric features and position information of the internal defects of the welding part.

[0030] As a further scheme of the application, the process of generating the difference evolution map specifically includes:

[0031] The dynamic propagation model and the ideal reference model are time-space registered;

[0032] On the time-space grid after registration, the vector difference of the signal energy amplitude and phase value of the dynamic propagation model relative to the ideal reference model is calculated node by node, and the vector difference includes the amplitude difference and the phase shift direction;

[0033] The vector difference of each space-time node is calculated and directionally encoded to generate a space-time difference vector field containing difference intensity and difference direction information;

[0034] The space-time difference vector field is subjected to anisotropic diffusion filtering to retain the edge features of the difference area while suppressing isolated noise points, and finally a difference evolution map is output.

[0035] As a further scheme of the application, the mode mining of the difference evolution map specifically includes:

[0036] convert the difference evolution map into a data point distribution in a high-dimensional topological feature space, wherein each data point corresponds to a local region in the map, and the feature vector of the data point contains the average difference intensity, gradient direction consistency and spatial distribution tightness of the corresponding region;

[0037] identify data point clusters with abnormal characteristics in the topological feature space, and the local patterns of the original map corresponding to these clusters are abnormal patterns caused by potential defects;

[0038] perform topological persistence analysis on each identified abnormal pattern, calculate its stability index at different feature scales, and screen out stable abnormal patterns as defect feature patterns;

[0039] perform topological similarity matching between the screened stable abnormal patterns and a pre-established defect disturbance feature library to determine the defect type confidence of each abnormal pattern, and complete abnormal pattern confirmation.

[0040] As a further scheme of the application, the defect information of the welding position is determined, specifically including:

[0041] perform dynamic evolution path backtracking on the confirmed abnormal pattern, analyze its propagation trajectory and evolution law in the time-space dimension, and establish a complete development path of the abnormal pattern from generation to stability;

[0042] Based on the dynamic evolution path of the abnormal pattern, extract the core feature parameters of the defect, including the starting position of the defect, the spatial expansion direction, the influence range growth rate, and the interaction strength with the material structure;

[0043] Combine the extracted defect feature parameters with the welding process parameters and material properties to perform defect hazard level evaluation and classification determination;

[0044] Generate a complete defect information report containing defect three-dimensional positioning, type classification, size quantification and safety level evaluation, and present it stereoscopically through a visual interface.

[0045] A titanium plate welding position automatic detection system based on titanium plate welding position, comprising:

[0046] An original signal acquisition module is configured to acquire an original physical field signal of a titanium plate welding position to be detected under excitation of a single energy field;

[0047] A propagation dynamics feature extraction module is configured to extract the propagation dynamics features of the original physical field signal to obtain the energy attenuation gradient of the original physical field signal in the spatial dimension and the phase lag distribution of the original physical field signal in the time dimension;

[0048] a dynamic propagation model construction module, which constructs a dynamic propagation model describing the evolution law of the original physical field signal on the propagation path in the material based on the energy attenuation gradient and phase lag distribution;

[0049] a difference evolution atlas generation module, which compares the dynamic propagation model with a preset ideal reference model representing the signal propagation of the defect-free standard titanium plate welding site, generates a difference evolution atlas by calculating the evolution difference of the two at a specific spatiotemporal node, and

[0050] a defect mode recognition module, which performs pattern mining on the difference evolution atlas, identifies abnormal patterns caused by signal propagation path disturbance caused by defects, and determines the defect information of the welding site.

[0051] The beneficial effects of the present application are:

[0052] (1) By using a specific physical energy field excitation and matching sensor to collect the original physical field signal of the titanium plate welding site, and using advanced wavelet packet decomposition and Hilbert transform technology to extract the propagation dynamics characteristics, a dynamic propagation model is constructed, which can accurately capture the response of the internal structure characteristics of the welding site to the external excitation. Compared with the traditional detection method, this method can more carefully reflect the energy attenuation gradient and phase lag distribution inside the material, thereby improving the recognition ability of small defects and the accuracy of the detection results, and effectively avoiding misjudgment and missed detection.

[0053] (2) The present application proposes a complete automation process from original signal collection, feature extraction, model construction to defect recognition and quantitative analysis, and uses pattern mining technology and a multi-criteria decision system to realize intelligent recognition of abnormal patterns and automatic determination of defect information. This method not only greatly reduces the need for manual intervention, reduces the complexity of operation and cost, but also can quickly and accurately generate a structured report containing three-dimensional positioning of defects, type classification and safety level evaluation, improve the detection efficiency and consistency of the results, and provide strong support for subsequent maintenance decisions. BRIEF DESCRIPTION OF DRAWINGS

[0054] The present application will be further described below with reference to the accompanying drawings.

[0055] Figure 1 is a flowchart of the method of the present application;

[0056] Figure 2 is a flowchart of the system in the present application. DETAILED DESCRIPTION

[0057] Clearly, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present application.

[0058] Please refer to Figure 1 The present application is an automatic detection method for titanium plate welding parts, which comprises the following steps:

[0059] S1: Obtain the original physical field signal of the titanium plate welding part to be detected under the excitation of a single energy field;

[0060] S2: Extract the propagation dynamics characteristics of the original physical field signal to obtain the energy attenuation gradient of the original physical field signal in the spatial dimension and the phase lag distribution in the time dimension;

[0061] S3: Based on the energy attenuation gradient and the phase lag distribution, a dynamic propagation model is constructed to describe the evolution law of the original physical field signal on the propagation path in the material;

[0062] S4: Compare the dynamic propagation model with the ideal reference model representing the signal propagation of the standard titanium plate welding part without defects, and generate a difference evolution map by calculating the evolution difference of the two at a specific space-time node;

[0063] S5: Perform pattern mining on the difference evolution map to identify abnormal patterns caused by signal propagation path disturbance due to defects, and determine the defect information of the welding part.

[0064] In S1, the original physical field signal of the titanium plate welding part to be detected under the excitation of a single energy field is obtained, specifically including:

[0065] The original physical field signal specifically refers to the initial response data received by a special sensor without deep processing under controlled conditions by applying a specific type of physical energy field to the titanium plate welding part. This data directly records the physical effects after the interaction of the energy field and the material, and its essence is the objective mapping of the internal structural characteristics of the material (such as discontinuity, density change, thermal physical property difference, and electrical conductivity non-uniformity) to external excitation. This signal is the starting point and data basis for all subsequent analysis and processing, and its quality directly determines the reliability of the final detection result. The process of obtaining this signal is a systematic process that includes excitation, sensing, conversion, and preliminary standardization.

[0066] First, the appropriate physical energy field generator and its associated sensor system need to be selected and configured according to the specific inspection purpose and the geometry and material properties of the titanium plate weld joint. The choice of energy field type depends mainly on the physical nature of the target defects and their location in the structure. For example, if volumetric defects such as porosity, inclusions, or lack of fusion inside the weld need to be detected, an ultrasonic field is typically chosen because it is sensitive to changes in acoustic impedance when propagating in an elastic medium. In this case, the excitation source is a piezoelectric ultrasonic probe that can convert an electrical pulse into mechanical vibrations. If surface cracks or near-surface microstructure changes are the focus of attention, an eddy current field is more appropriate because it is sensitive to changes in electromagnetic properties at the surface and near-surface of a conductor. In this case, the excitation source is a coil with a high-frequency alternating current. In addition, if the thermal uniformity of the material needs to be evaluated or if layered defects need to be detected, a pulsed thermal excitation field is an effective means to heat the surface with a transient heat source such as a flash lamp. The sensor system must be precisely matched to the selected energy field: a piezoelectric probe is used for ultrasonic detection (both transmission and reception), a detection coil is used for eddy current detection, and an infrared thermal imager is used for thermal excitation detection. The position, angle, and coupling state of all sensors relative to the workpiece need to be precisely set and maintained stable to ensure that the response from the entire weld area and heat-affected zone can be captured completely and consistently.

[0067] Second, after the system configuration is complete, the energy field is started and the titanium plate weld area is excited, while the sensor synchronously collects the original physical response. This process needs to be carried out under the premise that environmental interference is effectively controlled. Specifically, when an ultrasonic field is used, the ultrasonic probe is in good contact with the titanium plate surface through a coupling agent (such as water or gel), and an electrical pulse of a specific frequency and energy is applied to it. The pulse is converted into an ultrasonic wave by the probe and transmitted into the workpiece. When the ultrasonic wave propagates in the material, it encounters interfaces with acoustic impedance differences (such as defects or boundaries), which produce reflected waves or cause transmitted wave energy to attenuate. These acoustic signals containing internal structural information are captured by the receiving probe (or the same probe during the transmission interval) and converted back into an electrical signal. This time-varying voltage signal is a form of the original physical field signal. When an eddy current field is used, the excitation coil is brought close to the surface of the workpiece and a high-frequency current is passed through it, generating an alternating magnetic field, which in turn induces an eddy current in the titanium plate. This eddy current field itself generates a secondary magnetic field. If the workpiece has defects, it will disturb the eddy current path and the secondary magnetic field, which is perceived by the detection coil as a change in its impedance. The recorded complex impedance (including resistance and reactance components) is the original signal. When a pulsed thermal excitation field is used, a short and uniform thermal pulse is applied to the weld area, and then an infrared thermal imager is used to continuously record the process of surface temperature decay over time. Internal defects will hinder heat flow, causing abnormal temperature distribution at the corresponding surface. This series of thermal images arranged in time sequence constitutes the original physical field signal.

[0068] Then, the analog signals output directly by the sensors are conditioned and digitized to convert them into a standardized digital format that can be processed by a computer. The raw electrical signals output by the sensors are usually small in amplitude and contain noise. Therefore, they need to be adjusted in amplitude by a preamplifier to adapt to the input range of the subsequent acquisition equipment. Subsequently, an analog filter (such as a low-pass or band-pass filter) is used to filter out noise components outside the working frequency band and prevent spectral aliasing during digitization. The conditioned analog signals are sampled and quantized by an analog-to-digital converter to convert them into discrete digital sequences. For ultrasonic signals, this is time series data; for thermal excitation, it is image sequence data containing spatial dimensions. For image data, preprocessing such as non-uniformity correction of pixel response is also required to ensure the consistency of the response of each sensor unit. The final digitized data is the original physical field signal directly used for feature extraction in step S2.

[0069] Finally, the acquired original physical field signal is evaluated for quality to confirm its effectiveness and usability. The evaluation includes but is not limited to checking whether the signal-to-noise ratio meets the minimum requirement, observing whether the signal baseline is smooth, and verifying whether the signal clearly contains the expected response features from the weld area. For example, the ultrasonic signal should be able to observe clear front surface echoes and bottom waves; the thermal image sequence should be able to show a clear thermal diffusion process. If the signal quality is not up to standard, the cause needs to be investigated, such as adjusting the excitation parameters, improving the sensor coupling conditions, or eliminating environmental interference, and re-acquiring.

[0070] In S2, the original physical field signal is subjected to propagation dynamics feature extraction to obtain the energy attenuation gradient in the spatial dimension and the phase lag distribution in the time dimension of the original physical field signal, which specifically includes:

[0071] The extraction process of the energy attenuation gradient:

[0072] The extraction of the energy attenuation gradient is first realized by wavelet packet decomposition of the original physical field signal. Wavelet packet decomposition is a time-frequency analysis method that can finely divide the full frequency band of a signal. In specific implementation, a specific wavelet basis function needs to be selected, and the present invention selects the Daubechies series wavelet and determines the number of decomposition layers, which is 3 layers. The decomposition process is completed by iteratively filtering and down-sampling the original signal with a set of high-pass and low-pass filters. Each layer of decomposition decomposes the input signal into a low-frequency approximation component and a high-frequency detail component, and both components will be further decomposed in the next layer. After decomposition to the preset number of layers, the original signal is represented as a linear combination of a series of sub-band signals, each covering a specific frequency range and corresponding to a unique scale space index. These sub-band signals collectively constitute a complete representation of the original signal in multiple resolution scales.

[0073] After the wavelet packet decomposition is completed and each subband signal is obtained, the energy value of each subband signal at each sampling point in the spatial dimension is calculated. The method of calculating the energy value is that, for each spatial sampling point, the time sequence segment corresponding to the point in the subband signal is taken, the square of the amplitude of all data points in the time sequence segment is calculated, and then the sum of these square values is calculated. The result is the energy value of the spatial point under this subband. Sequentially performing this calculation on all spatial sampling points forms a sequence representing the spatial distribution of the energy of the subband signal, which is called the spatial energy distribution sequence of the subband. This process condenses the time domain information of the signal into the energy distribution characteristics in the spatial domain.

[0074] Subsequently, the spatial gradient operation is performed on the spatial energy distribution sequence of each subband to quantify the rate of change of energy in space. The central difference method is used for calculation. For each internal point (i.e. non-boundary point) in the spatial energy distribution sequence, the gradient value is obtained by calculating the difference between the energy value of the next point and the energy value of the previous point, and then dividing the spatial distance between the two points. For the start point and the end point of the sequence, the forward difference and backward difference methods are used for processing, respectively. Through this operation, the local energy gradient sequence corresponding to each subband signal can be obtained, which clearly shows the degree of attenuation or enhancement of energy at each point in space.

[0075] Finally, the local energy gradients calculated for all subbands are weighted and fused to generate a comprehensive energy attenuation gradient map. The key to weighted fusion is to determine the weight of each subband gradient. The weight is determined based on the energy contribution of each subband signal, that is, the total energy of each subband signal (the sum of all values in the spatial energy distribution sequence) is calculated first, and then the proportion of the total energy of each subband in the sum of the total energies of all subbands is calculated. This proportion is used as the fusion weight of the subband gradient. The local energy gradient of each subband is multiplied by its corresponding weight, and then superimposed. The final comprehensive energy attenuation gradient map can fully reflect the spatial attenuation characteristics of signal energy under the joint action of different frequency components.

[0076] Extraction process of phase lag distribution:

[0077] The extraction of the phase-lag distribution starts with the Hilbert transform of the original physical field signal. The Hilbert transform is a linear integral transform whose purpose is to construct the analytic signal of the original signal. The specific calculation process is to perform a convolution operation on the original physical field signal with a specific kernel function, which is represented as a function of the time inverse in the time domain. The result of the convolution operation is called the Hilbert transform signal of the original signal. Using the original signal and its Hilbert transform signal, a complex signal, i.e. the analytic signal, can be constructed. The real part of the analytic signal is the original signal itself, and the imaginary part is its Hilbert transform signal. The amplitude of the analytic signal is the instantaneous amplitude of the signal, and the phase angle is the instantaneous phase of the signal. By calculating the time series of the phase angle of the analytic signal, the instantaneous phase information of the original physical field signal can be accurately extracted, which reflects the fine phase change of the signal in the time dimension.

[0078] After obtaining the instantaneous phase information of each spatial sampling point, a phase reference benchmark needs to be established to calculate the relative phase lag. The geometric center of the welding site is usually chosen as the phase reference point. The instantaneous phase difference of all other spatial sampling points relative to the reference point is calculated. Specifically, at the same time, the instantaneous phase value of each point is subtracted from the instantaneous phase value of the reference point, and the difference obtained is the instantaneous phase difference of the point relative to the reference point. This phase difference directly reflects the time delay effect experienced by the signal propagating from the reference point to the point, and is a direct measure of the phase lag phenomenon in the wave propagation process.

[0079] Since the spatial sampling points are discrete, the phase difference values obtained directly are also discretely distributed on the spatial sampling points. In order to obtain a continuous phase lag distribution view on the continuous space, spatial interpolation processing needs to be performed on the discrete phase difference values. Methods such as Kriging interpolation algorithm can be used. Kriging interpolation is a spatial interpolation method based on variogram theory, which can make optimal unbiased estimation of the value of the estimated point according to the spatial correlation of the known sampling points. Through interpolation processing, a continuous phase lag distribution surface or image covering the entire region of interest can be generated. The distribution clearly shows the spatial variation of the phase lag.

[0080] The final phase-lag distribution is a key characteristic quantity that describes the propagation dynamics of the original physical field signal in the titanium plate welding site, providing accurate input for subsequent dynamic propagation model construction. The entire feature extraction process converts the original, seemingly complex physical field signal into a feature parameter with clear physical meaning through rigorous mathematical transformation and calculation, ensuring the accuracy and reliability of subsequent analysis and judgment.

[0081] In S3, based on the energy attenuation gradient and the phase-lag distribution, a dynamic propagation model is constructed to describe the evolution of the original physical field signal on the propagation path in the material, specifically including:

[0082] Firstly, a spatio-temporal mapping grid covering the welding site and its surrounding area to be detected needs to be established. The grid is a discretized computational space composed of nodes distributed uniformly or non-uniformly in space, each node corresponding to a specific spatial location. The grid resolution should be determined according to the detection resolution requirements, for example, the space can be divided into 1 mm square grid cells. After the grid is established, the energy attenuation gradient data obtained in step S2 is mapped to each node of the grid as the "energy dissipation rate" attribute value of the node, which represents the degree of energy attenuation when the signal propagates to this location. At the same time, the phase lag distribution data is mapped to the grid nodes as the "signal propagation delay" attribute value of the node, which represents the time lag of signal propagation to this location relative to the reference point. Through this mapping process, each grid node has two core attribute parameters that are coupled with each other, laying a data foundation for subsequent propagation simulation.

[0083] Secondly, based on the established spatio-temporal mapping grid, the signal propagation correlation rules between nodes are constructed. The rules are used to define how the signal propagates from one node to its adjacent nodes. The core of the rules is to establish an evaluation function that considers both the energy dissipation rate attribute of the upstream node (signal source direction) and the signal propagation delay attribute of the downstream node (signal propagation direction). Specifically, starting from the signal injection point (i.e. the excitation point) of the grid, it is taken as the source point. Then, a kind of optimal path search algorithm is defined, for example, an improved algorithm based on Dijkstra principle, which calculates the path cost when searching for the propagation path from the source point to any target node in the grid. The path cost is not only related to the geometric distance between nodes, but more importantly, it is associated with the energy dissipation rate and signal propagation delay attributes of the nodes on the way. The algorithm tends to choose the path with lower energy dissipation rate (i.e. slower signal attenuation) and reasonable signal propagation delay increment as the optimal propagation path. Through this step, abstract physical attributes are transformed into specific rules that control the behavior of signal propagation.

[0084] Finally, based on the defined signal propagation rules, the whole spatio-temporal mapping grid is calculated dynamically to simulate the signal propagation process in the actual material. The evolutionary computation is performed in an iterative manner. At the initial time, only the signal injection point is activated, and its state parameters are initialized. In the first iteration, according to the propagation rules, the signal propagates from the injection point to its adjacent nodes, and the energy dissipation rate and signal propagation delay properties of these nodes are updated according to the rules to reflect the state changes after the signal arrives. In each subsequent iteration, the signal continues to propagate from the currently activated nodes to the next level of adjacent nodes, and the node attribute parameters are also updated according to the rules. This process continues until the signal propagates to the boundary of the grid, or the state of all nodes tends to be stable. Finally, by recording the state parameters of all nodes after each iteration (such as signal strength, arrival time, etc.), a dynamic propagation model is generated that can fully describe the spatio-temporal evolution rules of the signal propagation in the entire welding site. Essentially, this model is a data set containing four-dimensional information (three-dimensional space plus one-dimensional time).

[0085] The input parameters of the dynamic propagation model are clear and directly derived from the previous processing results. The main inputs include the energy attenuation gradient and phase lag distribution calculated accurately by step S2. The energy attenuation gradient, as one of the key driving parameters in the model, determines the simulation method of energy attenuation when the signal propagates between grid nodes. The phase lag distribution, as another key input, provides the basic constraints for the timing relationship of signal propagation. In addition, according to the specific algorithm used, some basic physical parameters may also need to be input, such as the initial energy value of the signal, the estimated range of propagation speed, etc. These parameters can be set according to the detection physical field used (such as ultrasonic speed). All these input data together provide the necessary initial conditions and boundary conditions for the construction and operation of the model.

[0086] The output of the dynamic propagation model is a dynamic evolution sequence of the signal propagation path inside the material. This sequence is presented in the form of a spatiotemporal distribution map. Specifically, for each time step (corresponding to the number of iterations) of the model simulation, an energy field map reflecting the spatial distribution of signal energy at that moment and a phase field map reflecting the spatial distribution of signal wavefront phase can be output. Combining all these time-sequentially arranged field maps constitutes the dynamic evolution sequence. By analyzing this sequence, one can clearly observe how the signal energy expands, attenuates, and reflects or scatters when encountering defects, etc. inside the weld site. In particular, in the output spatiotemporal distribution map, there will be some characteristic regions. For example, regions where signal energy is abnormally concentrated, called abnormal convergence zones, which usually correspond to the reflection or focusing effect of defects on signals. There will also be regions where the phase wavefront is distorted, called wavefront distortion zones, which are usually caused by changes in local propagation speed due to defects. The coupling relationship in space between abnormal convergence zones and wavefront distortion zones, i.e., whether they appear at the same or adjacent locations, as well as their geometric features such as shape and size, have a direct correspondence with the type, size, and exact location of defects (such as pores, cracks, incomplete fusion, etc.) inside the weld site. Therefore, the output of this dynamic evolution sequence provides an essential information basis for subsequent accurate identification and positioning of defects.

[0087] In S4, the dynamic propagation model is compared with a preset ideal reference model representing the signal propagation of a defect-free standard titanium plate weld site. By calculating the evolution difference between the two at a specific spatiotemporal node, a difference evolution map is generated, which includes:

[0088] The process of generating the difference evolution map is a process of fine comparison between the dynamic propagation model corresponding to the titanium plate weld site to be detected and an ideal reference model representing the signal propagation characteristics of a defect-free standard titanium plate weld site. This ideal reference model can be obtained by simulating a defect-free standard test block, or by statistical modeling of a large amount of known qualified workpiece detection data. The purpose is to highlight the abnormal signal propagation pattern caused by internal defects by quantifying the deviation between the model to be tested and the ideal state. This process mainly includes four key links: spatiotemporal registration, vector difference calculation, vector field generation, and filtering optimization. Finally, a difference evolution map that can clearly indicate the location of defects is output.

[0089] First, a high-precision spatiotemporal registration between the dynamic propagation model and the ideal reference model is required. Due to the possible minor differences in coordinate systems or positional deviations during data acquisition, direct node comparison will lead to errors. The purpose of spatiotemporal registration is to eliminate such systematic spatial position and temporal origin inconsistencies. The registration process adopts an elastic registration algorithm based on intrinsic feature points of the models. Specifically, a set of feature points with clear physical meaning and easy to identify, such as signal excitation points, weld geometric center points, and several specific base material points away from the weld area, are selected in the data sets of the dynamic propagation model and the ideal reference model, respectively. Then, by calculating, an optimal spatial transformation (usually including translation, rotation, and scaling parameters) is found that minimizes the average distance between the set of feature points in the dynamic propagation model and the corresponding feature points in the ideal reference model. This optimal transformation can be obtained through iterative optimization algorithms (such as the least squares method). Applying the obtained transformation parameters to all spatiotemporal node coordinates of the dynamic propagation model can achieve accurate alignment of the two models in the spatiotemporal coordinate system, laying the foundation for subsequent node-by-node difference calculation. The registration in the time dimension mainly ensures that the two models are consistent in the starting time point and time step of signal propagation.

[0090] After completing the accurate spatiotemporal registration, the vector difference between the dynamic propagation model and the ideal reference model can be calculated node by node on the aligned spatiotemporal grid. The difference here is not a simple scalar difference, but a vector that contains both amplitude and direction information. For each node in the spatiotemporal grid, the calculation of the vector difference is divided into two parts. The first part is to calculate the difference in signal energy amplitude, that is, to subtract the signal energy amplitude of the ideal reference model at the same node from the signal energy amplitude of the dynamic propagation model at that node. The second part is to calculate the offset direction of the phase, that is, to calculate the difference between the signal phase values of the two models at that node. This phase difference not only reflects how much the phase is delayed, but also indicates whether the phase is advanced or delayed by its sign. Finally, these two scalar information (amplitude difference and phase offset) are combined into a two-dimensional vector, which completely describes the comprehensive deviation of the test signal relative to the ideal signal in intensity and timing at a specific spatiotemporal node. Performing this operation on all nodes results in a preliminary distribution of vector differences covering the entire area of interest.

[0091] The calculated vector difference at each spatiotemporal node is then modulus calculated and direction coded to generate a spatiotemporal difference vector field that is more amenable to analysis and visualization. The modulus calculation is a scalar quantity that is computed for each node's vector difference as the square root of the sum of the squares of the amplitude difference and the phase shift. This modulus value represents the strength of the total difference at that node and is a scalar quantity. The direction coding is a discrete classification of the direction of the vector difference. For example, the direction can be simply coded as "phase lag" and "phase lead" according to the sign of the phase shift. Or the direction can be more finely classified, such as by combining the sign of the amplitude difference (enhancement or reduction) and the direction of the phase shift to classify the difference vector into one of eight main directions (such as north, northeast, east, etc.). Through modulus calculation and direction coding, the complex vector information at each node is converted into a strength value (modulus) and a category label (direction). This spatiotemporal difference vector field, which contains both the difference strength and difference direction information, clearly shows the spatial distribution and nature of the abnormal signal propagation patterns.

[0092] The generated spatiotemporal difference vector field is then anisotropically diffused to effectively suppress isolated difference points caused by random noise while preserving the significant difference regions caused by real defects that have continuous edges. Anisotropic diffusion is a nonlinear filtering technique whose core idea is that the strength of diffusion depends on the local gradient of the image. On the difference intensity map (i.e., the modulus map), for flat regions with small gradients (which can correspond to noise), strong diffusion is allowed to smooth the noise; while for edge regions with large gradients (which can correspond to defect boundaries), weak diffusion or no diffusion is performed to preserve the sharpness of the edges. The process is usually iterated 10 to 20 times or until the signal-to-noise ratio of the difference field reaches a satisfactory level. After this filtering step, a smooth and continuous difference evolution map is finally output, with noise suppressed and defect contours clear. This map directly reveals the significant difference regions between the tested weld and the noise-free standard state, which are closely related to the presence of internal defects.

[0093] In S5, the difference evolution map is pattern mined to identify abnormal patterns that are consistent with the signal propagation path perturbation caused by defects, and to determine the defect information of the weld, specifically including:

[0094] The first step of pattern mining is to map the two-dimensional difference evolution map to a higher-dimensional topological feature space, so as to reveal its internal structure more clearly. In implementation, the difference evolution map is divided into multiple local regions that overlap with each other, for example, a sliding window (e.g., 10 pixels x 10 pixels in size) is used to traverse the entire map with a step size of 5 pixels. For each local region covered by the window, a set of feature values that can represent its statistical and geometric characteristics is calculated to form a feature vector. The feature vector mainly includes three components: the first is the arithmetic mean of the difference intensity values of all pixels in the region, which reflects the overall abnormality of the region; the second is the standard deviation or consistency measure of the gradient direction of the pixels in the region, which describes whether the direction of difference change is consistent (defect edges usually have consistent gradient directions); the third is the spatial distribution tightness of the high-difference pixel points in the region, which can be quantified by calculating the ratio of the eigenvalues of the coordinate covariance matrix of these pixel points. The larger the ratio, the more dispersed the distribution; the smaller the ratio, the more concentrated and compact the distribution. Each local region is thus represented as a data point in a high-dimensional topological feature space.

[0095] In this high-dimensional topological feature space, identify the data point clusters that have abnormal characteristics. This is usually achieved by a density peak-based clustering algorithm. The algorithm first calculates two key quantities for each data point: local density and distance to higher density points. Local density refers to the number of other points contained in the sphere centered at the point with a given cutoff distance as the radius. The distance to higher density points refers to the minimum distance from the point to all points with higher local density. For those points with higher local density and farther distance to higher density points, they are identified as potential cluster centers. Then the remaining points are assigned to the class to which the nearest point with higher density belongs. The clusters found by this process, which correspond to local pattern regions with similar abnormal characteristics in the difference evolution map, are considered as abnormal patterns caused by potential defects.

[0096] Perform topological persistence analysis on each identified abnormal pattern to evaluate its robustness and filter out false patterns that may be generated by noise. Topological persistence analysis observes whether the pattern persists at different scales by changing the scale parameter during feature extraction (e.g., changing the size of the previous sliding window, using windows of 8 pixels x 8 pixels, 12 pixels x 12 pixels, and 15 pixels x 15 pixels respectively to re-extract features and repeat clustering). Calculate its stability index, for example, the degree to which the core data points included in the pattern remain clustered at multiple scales, or the offset distance of its cluster center at multiple scales. Set a stability threshold, only when the stability index of the pattern exceeds the threshold, it is considered a stable abnormal pattern with topological persistence, and is screened as a defect feature pattern.

[0097] The identified stable abnormal pattern is then compared to a pre-established library of defect signatures. The library contains templates of patterns that are typically observed in the difference evolution map for various types of defects (e.g. porosity, crack, lack of fusion, etc.). The similarity between the feature vector of the abnormal pattern to be identified and each template in the library is calculated, typically using the cosine similarity or Mahalanobis distance. Based on the calculated similarity values, the abnormal pattern is assigned a most likely defect type label and a confidence score (e.g. the highest similarity value normalized), thus completing the abnormal pattern identification.

[0098] After the abnormal pattern is identified, the quantitative defect information determination phase is entered. First, the dynamic evolution path of the identified abnormal pattern is traced back. Using the spatio-temporal data output by the dynamic propagation model, the shape and location of the abnormal pattern at each time step before it was identified in the difference evolution map is traced back. By analyzing the movement trajectory of the centroid and the expansion law of the influence range, the complete development path of the abnormal pattern from its initial generation, gradual evolution to final stability is established. This path reveals the dynamic process of the interaction between the defect and the signal.

[0099] Based on the dynamic evolution path of the abnormal pattern, core feature parameters that describe the defect itself are extracted. These parameters include: the starting position of the defect, i.e. the coordinates of the starting point of the backtracking path in three-dimensional space; the spatial expansion direction of the defect, determined by principal component analysis of the path trajectory; the growth rate of the defect influence range, obtained by calculating the rate of change of the area of the abnormal pattern region with time step; the interaction strength between the defect and the material structure, which can be quantified by the average difference intensity value in the abnormal pattern region or the degree of change in the signal propagation path caused by the pattern.

[0100] The extracted core feature parameters of the defect, combined with known welding process parameters (such as welding current, speed, heat input) and the characteristics of the base material and welding material (such as grade, thickness), are input into a multi-criteria decision system for defect hazard level evaluation and classification determination. The decision system has a built-in rule library established according to different application scenarios and standards (such as aerospace, pressure vessel standards). The rule library may contain logical rules such as "when the defect length exceeds one-third of the wall thickness and is located in a high stress area and is a crack, it is determined to be a dangerous defect". The system integrates all information to classify the defect (e.g. slight, moderate, severe) and finally determine the type.

[0101] The system generates a structured complete defect information report. The report lists the three-dimensional spatial positioning coordinates of the defect, the defect type classification based on the matching results, the defect size quantification value (such as length, width, equivalent size) calculated according to the abnormal mode range, and the final safety level evaluation conclusion in the form of data and charts. All these information can be presented stereoscopically through the three-dimensional visualization interface, which intuitively shows the location, shape and rating of the defect on the workpiece model to the detection personnel, providing accurate and reliable basis for subsequent maintenance decision.

[0102] Referring to Figure 2 The titanium plate welding part automatic detection system comprises:

[0103] An original signal acquisition module is configured to acquire an original physical field signal of a titanium plate welding part to be detected under excitation of a single energy field;

[0104] A propagation dynamics feature extraction module is configured to extract propagation dynamics features of the original physical field signal to obtain an energy attenuation gradient of the original physical field signal in a spatial dimension and a phase lag distribution of the original physical field signal in a time dimension;

[0105] A dynamic propagation model construction module is configured to construct a dynamic propagation model describing evolution rules of the original physical field signal on a propagation path in a material based on the energy attenuation gradient and the phase lag distribution;

[0106] A difference evolution atlas generation module is configured to compare the dynamic propagation model with a preset ideal reference model representing signal propagation of a standard titanium plate welding part without defects, and generate a difference evolution atlas by calculating evolution differences of the two models at specific space-time nodes;

[0107] A defect mode recognition module is configured to perform pattern mining on the difference evolution atlas, recognize abnormal patterns caused by signal propagation path disturbance of defects, and determine defect information of the welding part.

[0108] The working principle of the present application is as follows: under controlled conditions, the welding part of the titanium plate is excited by a single physical energy field such as ultrasonic wave, eddy current or pulse thermal excitation, and the original physical field signal reflecting the internal structure characteristics of the material is synchronously collected by the matched sensor, and after signal conditioning and digital conversion, a digital sequence for analysis is formed; secondly, the propagation dynamics characteristics of the original signal are extracted, the energy attenuation gradient of the spatial dimension is calculated by wavelet packet decomposition, and the phase lag distribution of the time dimension is obtained by combining Hilbert transform and spatial interpolation method; based on the above characteristics, a dynamic propagation model is constructed, the energy attenuation gradient and the phase lag distribution are mapped into the energy dissipation rate and signal propagation delay attribute of the node respectively by establishing the space-time mapping grid, and the propagation correlation rule of coupling energy and phase is defined, and the dynamic model describing the propagation law of the signal in the material is generated by iterative evolution; then, the dynamic propagation model is space-time registered with the preset ideal reference model, and the vector difference containing amplitude difference and phase offset is calculated node by node, and after modulus calculation, direction coding and anisotropic diffusion filtering, a clear difference evolution map is generated; the difference evolution map is pattern mined, and its local area is mapped to a high-dimensional topological feature space, the potential abnormal pattern is identified by density clustering, the stable pattern is screened by multi-scale topological persistence analysis, and the defect type is confirmed by similarity matching with the defect disturbance feature library, and then the defect starting position, expansion direction, growth rate and other core parameters are extracted by backtracking the dynamic evolution path of the abnormal pattern, the multi-criteria decision is made combined with the process and material information, the three-dimensional positioning, size quantification, type classification and safety level evaluation of the defect are completed, and finally a structured report is generated and three-dimensional visualization is supported.

[0109] The above describes one embodiment of the present application in detail, but the content described is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the patent coverage range of the present application.

Claims

1. An automatic detection method for welded parts of titanium plates, characterized in that, Includes the following steps: S1: Obtain the original physical field signal of the welded part of the titanium plate to be tested under a single energy field excitation; S2: Extract the propagation dynamics features of the original physical field signal to obtain the energy attenuation gradient in the spatial dimension and the phase lag distribution in the time dimension of the original physical field signal; S3: Based on the energy attenuation gradient and phase lag distribution, a dynamic propagation model is constructed to describe the evolution law of the propagation path of the original physical field signal in the material; The construction process of the dynamic propagation model is as follows: A spatiotemporal mapping grid is established, which maps the energy decay gradient to the energy dissipation rate attribute on the grid node, and maps the phase lag distribution to the signal propagation delay attribute on the grid node, so that each grid node contains a pair of energy and phase coupled attribute parameters. Based on the spatiotemporal mapping grid, a signal propagation association rule between nodes is constructed. With the energy dissipation rate of the upstream node and the signal propagation delay of the downstream node as constraints, an optimal path search algorithm for the signal propagation from the grid source point to any node is defined. Based on signal propagation association rules, dynamic evolution calculations are performed on the spatiotemporal mapping mesh. By iteratively updating the attribute parameters of each node, the propagation process of the signal in the actual material is simulated, and finally a dynamic propagation model is generated. The inputs and outputs of the dynamic propagation model specifically include: The inputs to the dynamic propagation model are the energy decay gradient and the phase lag distribution; The output of the dynamic propagation model is the dynamic evolution sequence of the signal propagation path inside the material. The dynamic evolution sequence is presented in the form of a spatiotemporal distribution map, which includes an abnormal convergence area of ​​signal energy attenuation and a wavefront distortion area of ​​phase propagation. The spatial coupling relationship between the abnormal convergence area and the wavefront distortion area directly corresponds to the geometric features and location information of the internal defects of the welded part. S4: Compare the dynamic propagation model with the ideal reference model that pre-determines the signal propagation at the welding part of the defect-free standard titanium plate, and generate a difference evolution map by calculating the evolution difference between the two at spatiotemporal nodes. The process of generating the differential evolution map specifically includes: Spatiotemporal registration of the dynamic propagation model and the ideal reference model; On the registered spatiotemporal grid, the vector difference between the signal energy amplitude and phase values ​​of the dynamic propagation model and the ideal reference model is calculated node by node. The vector difference includes both the amplitude difference and the phase offset direction. The vector difference of each spatiotemporal node is modulated and directionally encoded to generate a spatiotemporal difference vector field containing information on difference intensity and difference direction. Anisotropic diffusion filtering is applied to the spatiotemporal difference vector field to suppress isolated noise points while preserving the edge features of the difference region, and finally outputs the difference evolution map. S5: Perform pattern mining on the differential evolution map to identify abnormal patterns that match the signal propagation path disturbances caused by defects, and determine the defect information of the welding part; The pattern mining of the differential evolution map specifically includes: The differential evolution map is transformed into a data point distribution in a high-dimensional topological feature space, where each data point corresponds to a local region in the map, and its feature vector contains the mean difference intensity, gradient direction consistency and spatial distribution density of the corresponding region. In the topological feature space, identify data point clusters with anomalous characteristics. The local patterns of the original map corresponding to these regions are the anomalous patterns caused by potential defects. For each identified anomalous pattern, topological persistence analysis is performed to calculate its stability index at different feature scales, and stable anomalous patterns are selected as defect feature patterns. The selected stable anomaly patterns are matched with the pre-established defect perturbation feature library by topological similarity to determine the confidence level of the defect type corresponding to each anomaly pattern, thus completing the anomaly pattern confirmation. The defect information for determining the welded area specifically includes: The confirmed anomalous patterns are dynamically traced back to their evolutionary paths, and their propagation trajectories and evolutionary patterns in the spatiotemporal dimensions are analyzed to establish a complete development path from the generation to the stabilization of the anomalous patterns. Based on the dynamic evolution path of the anomaly pattern, the core characteristic parameters of the defect are extracted, including the defect's starting position, spatial expansion direction, influence range growth rate, and interaction strength with the material structure. The extracted defect feature parameters are combined with welding process parameters and material properties to assess and classify the defect hazard level. Generate a complete defect information report that includes 3D defect location, type classification, size quantification, and safety level assessment, and present it in a three-dimensional form through a visual interface.

2. The automatic detection method for welded parts of titanium plates according to claim 1, characterized in that, The process of extracting the energy decay gradient includes: The original physical field signal is processed by wavelet packet decomposition, which decomposes the original physical field signal into multiple sub-band signals with different frequency ranges, and each sub-band signal corresponds to a scale space. For each sub-band signal, the energy value at each sampling point in the spatial dimension is calculated. The energy value is obtained by calculating the sum of squares of the signal amplitude, forming a spatial energy distribution sequence for each sub-band. Spatial gradient calculation is performed on the spatial energy distribution sequence of each sub-band, and the energy change rate between adjacent sampling points is calculated using the central difference method to obtain the local energy gradient of each sub-band. The local energy gradients of all sub-bands are weighted and fused, with the weights determined based on the energy contribution of each sub-band signal, to finally generate a comprehensive energy attenuation gradient.

3. The automatic detection method for welded parts of titanium plates according to claim 1, characterized in that, The extraction process of the phase lag distribution includes: The original physical field signal is processed by Hilbert transform to construct an analytical expression for the original physical field signal and extract the instantaneous phase information of the original physical field signal. Using the geometric center point of the welded area as the phase reference, the instantaneous phase difference between other spatial sampling points and the reference point is calculated. Spatial interpolation is performed on the calculated phase difference values ​​to generate a continuous phase lag distribution.

4. An automatic detection system for welded parts of titanium plates, characterized in that, An automatic detection method for titanium plate welded joints as described in any one of claims 1-3, comprising: The original signal acquisition module is used to acquire the original physical field signal of the welded part of the titanium plate to be detected under a single energy field excitation. The propagation dynamics feature extraction module is used to extract propagation dynamics features from the original physical field signal to obtain the energy attenuation gradient in the spatial dimension and the phase lag distribution in the time dimension of the original physical field signal. A dynamic propagation model construction module, which constructs a dynamic propagation model based on energy attenuation gradient and phase lag distribution to describe the evolution law of the propagation path of the original physical field signal in the material; The differential evolution map generation module compares the dynamic propagation model with a preset ideal reference model representing the signal propagation of the welding part of the defect-free standard titanium plate, and generates a differential evolution map by calculating the evolution difference between the two at spatiotemporal nodes. The defect pattern recognition module performs pattern mining on the difference evolution spectrum, identifies abnormal patterns that match the signal propagation path disturbances caused by defects, and determines the defect information of the welding part.

Images