A method for detecting defects on the inner and outer surfaces of a titanium alloy pipe

Abstract

The present application relates to the technical field of nondestructive testing of metal pipes, and discloses a method for detecting internal and external surface defects of titanium alloy pipes. The method comprises preprocessing and fusing signals of multiple detection probes to generate a multi-modal data set. Dynamic spatial registration is performed based on the real-time motion state of the pipe, and the data is mapped to the surface of a unified three-dimensional entity model. Local analysis regions that overlap with each other are divided, and feature vectors are extracted. Circumferential features at the same axial position are aggregated to construct a ring-shaped feature sequence. The sequence is cross-domain associated and matched with a corresponding manufacturing process parameter sequence to identify abnormal patterns related to specific process events. The method can overcome signal misplacement caused by pipe movement, achieve accurate spatial positioning and morphology restoration of defects, and realize defect root tracing through the association of features and processes.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology for metal pipes, specifically a method for detecting defects on the inner and outer surfaces of titanium alloy pipes. Background Technology

[0002] Titanium alloy tubing is widely used in aerospace, energy, and chemical industries, and its surface quality directly affects the safety and lifespan of the components. Current technologies primarily employ non-destructive testing methods such as eddy current and ultrasonic testing for online or offline inspection of tubing. These methods typically rely on a single or a few probes scanning along the circumferential or axial direction of the tubing to acquire signals. For tubing in motion, especially with rotation and axial feeding, the signals acquired by the probes are difficult to accurately correspond spatially and temporally. Environmental vibrations and tubing jitter introduce noise, causing signals from different probes on the inner and outer surfaces to not accurately map to the actual physical location of the tubing. This misalignment in spatial registration makes precise geometric location and morphological reconstruction of defects difficult, hindering the effective differentiation between real defects and motion artifacts.

[0003] Existing detection systems mostly focus on defect identification and alarms, with relatively isolated defect analysis functions. Detection signals and pipe manufacturing process parameters belong to different systems, resulting in weak correlation. When a defect is detected, process engineers find it difficult to quickly correlate defect characteristics with specific abnormal events in the manufacturing process. The lack of a mechanism for deep integration and analysis of online detection data and offline process data leads to low efficiency in tracing the root causes of defects and hinders feedback and optimization of the production process. The distribution pattern of circumferential defects in pipes is an important indicator of manufacturing process stability, but existing methods lack effective technical means to systematically extract and analyze these circumferential distribution characteristics and establish a quantitative correlation between them and process parameters. Summary of the Invention

[0004] The purpose of this invention is to provide a method for detecting defects on the inner and outer surfaces of titanium alloy pipes, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for detecting defects on the inner and outer surfaces of titanium alloy pipes, the method comprising:

[0006] The raw detection signals from multiple detection probes arranged on the inner and outer surfaces of titanium alloy pipes are preprocessed to remove noise interference caused by environmental vibration and generate a multimodal fusion dataset.

[0007] Based on the real-time motion state of the titanium alloy pipe, dynamic spatial registration is performed on the multimodal fusion dataset to map the detection signals under different spatial coordinates onto the surface of a unified three-dimensional solid model of the pipe.

[0008] Overlapping local analysis regions are defined on the surface of the three-dimensional solid model of the pipe. Time-frequency joint analysis is performed on the multimodal fusion dataset within each local analysis region to extract feature vectors characterizing the changes in surface micromorphology.

[0009] The feature vectors of local analysis regions belonging to the same axial position but from different circumferential angles are aggregated to construct a ring feature sequence that can reflect the distribution characteristics of circumferential defects in titanium alloy pipes.

[0010] The circular feature sequence is cross-domain correlated and matched with the manufacturing process parameter sequence of the corresponding pipe segment extracted from the production process database to identify abnormal feature patterns related to specific process events.

[0011] Preferably, the dynamic spatial registration of the multimodal fusion dataset based on the real-time motion state of the titanium alloy pipe includes:

[0012] The linear displacement and rotation angle of the titanium alloy tube measured by the encoder are obtained, and the real-time position and attitude of any point on the titanium alloy tube in the global coordinate system are calculated.

[0013] Based on the inherent installation position and angle of each detection probe, a real-time transformation relationship is established from the probe coordinate system to the global coordinate system of the titanium alloy pipe.

[0014] Based on the real-time transformation relationship, the surface feature points corresponding to the original detection signals collected by each detection probe are dynamically mapped to the precise positions on the three-dimensional solid model of the titanium alloy pipe.

[0015] Preferably, the step of dividing the surface of the three-dimensional solid model of the pipe into overlapping local analysis regions includes:

[0016] Taking any mapping point on the three-dimensional solid model of the titanium alloy tube as the center, a rectangular region of a preset size is extended along the axial and circumferential directions as the initial local analysis region.

[0017] The initial local analysis region is translated along the axial and circumferential directions, with the translation distance being less than the region size, to form a new local analysis region, ensuring that there is an overlap band of a preset width between adjacent local analysis regions;

[0018] All local analysis regions are numbered, and the center point coordinates and boundary range of each local analysis region on the three-dimensional solid model of the titanium alloy pipe are recorded.

[0019] Preferably, the step of performing time-frequency joint analysis on the multimodal fusion dataset within each local analysis region to extract feature vectors characterizing changes in surface micromorphology includes:

[0020] Empirical mode decomposition is performed on the spatially registered detection signal within the local analysis region to obtain a set of intrinsic mode function components;

[0021] Calculate the energy entropy and sample entropy of each intrinsic mode function component, and combine them to form the first feature subset describing the complexity of the signal.

[0022] The intrinsic mode function components are subjected to Hilbert transform to solve for instantaneous frequency and instantaneous amplitude, and their mean and variance are statistically analyzed to form a second feature subset describing the time-varying characteristics of the signal.

[0023] The first feature subset and the second feature subset are concatenated to form the feature vector representing the changes in surface micromorphology.

[0024] Preferably, the aggregation of feature vectors from local analysis regions belonging to the same axial position but originating from different circumferential angles to construct a ring-shaped feature sequence that reflects the circumferential defect distribution characteristics of titanium alloy pipes includes:

[0025] At the same axial section position of the three-dimensional solid model of the titanium alloy tube, extract the feature vectors of the local analysis regions at all circumferential angles;

[0026] The extracted feature vectors are arranged in ascending order of circumferential angle to form a one-dimensional data sequence with a ring structure.

[0027] The one-dimensional data sequence is smoothed to eliminate abnormal fluctuations caused by individual differences in the detection probe or instantaneous interference, generating the final ring feature sequence.

[0028] Preferably, the step of performing cross-domain association matching between the annular feature sequence and the manufacturing process parameter sequence of the corresponding pipe segment extracted from the production process database includes:

[0029] Based on the conveying speed and timestamp information of the titanium alloy pipe, the process stages experienced by the pipe segment corresponding to the annular feature sequence during the manufacturing process are determined.

[0030] The key process parameters recorded in the same process stage of the pipe section are retrieved from the production process database, including the temperature of each zone of the heating furnace, the rolling force of the rolling mill, and the cooling water pressure, and are arranged in chronological order to form a manufacturing process parameter sequence.

[0031] Calculate the cross-correlation function of the annular feature sequence and the manufacturing process parameter sequence in the time domain, find the peak point of the cross-correlation function, and thus establish the correspondence between the changes in the annular feature and the fluctuations in the process parameters.

[0032] Preferably, calculating the cross-correlation function of the annular feature sequence and the manufacturing process parameter sequence in the time domain, and finding the peak point of the cross-correlation function, includes:

[0033] The circular feature sequence and the manufacturing process parameter sequence are time-normalized to give the two sequences a unified time axis scale.

[0034] A sliding window algorithm is used to calculate the similarity metric between the circular feature sequence and the manufacturing process parameter sequence at multiple time lag positions, and to generate a cross-correlation function curve.

[0035] Traverse all data points on the cross-correlation function curve, detect local maximum points, and determine the local maximum point with the highest similarity metric as the peak point of the cross-correlation function.

[0036] Preferably, the identification of anomalous feature patterns related to a specific process event includes:

[0037] Set thresholds for amplitude change and rate of change of the circular feature sequence. When the value of the circular feature sequence or its derivative exceeds the corresponding threshold, mark that moment as a potential outlier.

[0038] Backtrack the sequence of manufacturing process parameters associated with the potential anomaly point to check for any parameter exceeding limits or drastic change events;

[0039] If it exists, the potential anomaly is identified as a defect feature pattern caused by a specific process event, and the ring feature morphology corresponding to the defect feature pattern and its associated process parameter anomaly information are recorded.

[0040] Preferably, the backtracking of the manufacturing process parameter sequence associated with the potential anomaly point time to check for parameter exceeding limits or drastic change events includes:

[0041] Centered on the potential anomaly point, extract the manufacturing process parameter sequence fragments within the preset time window before and after it;

[0042] Each parameter value in the manufacturing process parameter sequence segment is compared with a preset parameter safety range. If the parameter value exceeds the upper or lower limit of the safety range, it is marked as a parameter exceeding the limit event.

[0043] Calculate the rate of change of parameters between consecutive sampling points in the manufacturing process parameter sequence segment. If the absolute value of the rate of change exceeds a preset mutation threshold, it is marked as a violent jump event.

[0044] Preferably, the method further includes:

[0045] The confirmed defect feature patterns and their associated process parameter anomaly information are stored in the defect knowledge base.

[0046] When a new circular feature sequence is detected online again, it is matched with the historical defect feature patterns stored in the defect knowledge base.

[0047] If the matching degree exceeds the preset confidence level, the corresponding process parameter anomaly information in the defect knowledge base is directly called to guide early intervention in the production process.

[0048] Periodically perform cluster analysis on the defect feature patterns stored in the defect knowledge base, and merge duplicate or similar defect feature patterns;

[0049] Analyze the correlation rules between the characteristic patterns of each type of defect and abnormal process parameters, and extract the key process parameter combinations and critical conditions that lead to this type of defect.

[0050] Based on the extracted association rules, the amplitude change threshold and change rate threshold of the circular feature sequence are dynamically updated.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] Dynamic spatial registration based on the real-time motion of the pipe stabilizes the moving detection object onto a unified 3D model surface. Overlapping local analysis regions are then defined on this model for feature extraction. This approach eliminates signal spatial drift caused by fluctuations in pipe speed and vibration, enabling precise alignment and fusion of signals from different probes at different times based on their actual physical locations. This improves the accuracy of locating minute defects, particularly clearly reconstructing the true morphology and extent of defects on the pipe surface, avoiding missed detections or misjudgments due to signal misalignment.

[0053] This method aggregates circumferential features from the same axial cross section into a ring-shaped feature sequence and performs cross-domain correlation matching with the manufacturing process parameter sequence of the corresponding pipe section. This elevates the analytical dimension of detection information from discrete points to the entire circumferential ring, enabling the capture of periodic defect distribution patterns caused by systemic process problems. By directly correlating temporal process events with detection feature patterns, it achieves a leap from simple defect identification to defect cause diagnosis. Process engineers can intuitively understand the intrinsic connection between specific defect patterns and anomalies in upstream production processes, thus providing a direct and accurate basis for quickly adjusting process parameters and locating equipment faults. Attached Figure Description

[0054] Figure 1 This is a schematic diagram illustrating the working principle of the method for detecting internal and external surface defects in titanium alloy pipes according to the present invention.

[0055] Figure 2 Flowchart for dynamic space registration;

[0056] Figure 3 This is a flowchart of time-frequency joint analysis and feature extraction;

[0057] Figure 4 Real-time monitoring and amplitude anomaly marking curves for the annular characteristic sequence of titanium alloy pipes;

[0058] Figure 5 This is a histogram showing the distribution of the number of cluster categories for defects in titanium alloy pipes. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Please see Figure 1 This invention provides a method for detecting defects on the inner and outer surfaces of titanium alloy pipes. The method includes: preprocessing the raw detection signals collected by multiple detection probes arranged on the inner and outer surfaces of the titanium alloy pipe. These signals may contain noise interference introduced by environmental vibration. The preprocessing stage uses filtering and signal enhancement techniques to remove noise, generating a clean multimodal fusion dataset. This dataset integrates detection information from different physical modes, such as ultrasonic, eddy current, or optical signals. Based on the real-time motion state of the titanium alloy pipe, i.e., by real-time monitoring of the linear displacement and rotation angle of the pipe through an encoder, the multimodal fusion dataset is dynamically spatially registered. The signals collected by each detection probe in different spatial coordinate systems are dynamically mapped to a unified three-dimensional solid model surface of the titanium alloy pipe, ensuring that all data are spatially aligned.

[0061] Overlapping local analysis regions are delineated on the surface of the 3D solid model of the pipe, each covering a certain axial and circumferential range. Time-frequency joint analysis is performed on the multimodal fusion dataset within each region. Empirical mode decomposition and Hilbert transform are used to extract feature vectors characterizing changes in surface microstructure; these vectors describe the complexity and time-varying properties of the signal. Feature vectors from local analysis regions belonging to the same axial position but originating from different circumferential angles are aggregated and arranged in circumferential angle order to form a ring-shaped feature sequence. This sequence reflects the distribution characteristics of circumferential defects in the pipe. The ring-shaped feature sequence is then cross-domain correlated with the manufacturing process parameter sequence of the corresponding pipe segment extracted from the production process database. By calculating the cross-correlation function, abnormal feature patterns related to specific process events such as temperature fluctuations or abnormal rolling force are identified, thereby achieving causal tracing of defects.

[0062] Example 1: See Figure 2 In practical implementation, the dynamic spatial registration process relies on the collaborative work of a high-precision motion measurement device and a coordinate transformation algorithm. Rotary and linear encoders installed on the production line drive system collect the motion parameters of the titanium alloy tube in real time. The rotary encoder measures the rotation angle of the titanium alloy tube around its axis, while the linear encoder measures the linear displacement of the titanium alloy tube along its axial direction. The encoder data is transmitted to the central processing unit via a fieldbus. Based on the linear displacement and rotation angle of the titanium alloy tube measured by the encoders, the central processing unit calculates the real-time position and orientation of any point on the titanium alloy tube in the global coordinate system using a kinematic model. The kinematic model treats the titanium alloy tube as a rigid body, with its position determined by the axial displacement and its orientation by the rotation angle. The calculation process involves homogeneous coordinate transformation. Each detection probe undergoes precise calibration after installation. The calibration process records the probe's inherent installation position and angle. The inherent installation position refers to the three-dimensional coordinates of the detection probe relative to the production line reference point, and the inherent installation angle refers to the angular relationship between the probe's detection direction and the axes of the global coordinate system. Based on the inherent installation parameters of the detection probe and the real-time position and orientation of the titanium alloy pipe, a real-time transformation relationship is established from the probe coordinate system to the global coordinate system of the titanium alloy pipe. This real-time transformation relationship is represented by a 4x4 homogeneous transformation matrix, which integrates rotation and translation components. According to this real-time transformation relationship, the surface feature points corresponding to the original detection signals acquired by each detection probe are dynamically mapped to their precise positions on the 3D solid model of the titanium alloy pipe. The mapping algorithm employs the coordinate transformation principle, multiplying the signal coordinates in the probe coordinate system by the homogeneous transformation matrix to obtain the 3D coordinates in the global coordinate system, and then mapping them onto the surface of the 3D solid model of the titanium alloy pipe using a projection algorithm.

[0063] In practical implementation, the 3D solid model of the titanium alloy pipe is a pre-constructed digital representation of the pipe. The model surface is composed of a large number of triangular meshes, and the mapping process requires mapping points in the global coordinate system to the nearest triangular mesh. The dynamic spatial registration module runs continuously, and its update frequency is consistent with the encoder data acquisition frequency, usually at the kilohertz level, to ensure the real-time performance and accuracy of the mapping during the high-speed movement of the titanium alloy pipe. After completing the dynamic spatial registration, the system divides overlapping local analysis regions on the surface of the 3D solid model of the titanium alloy pipe. The division operation is centered on any mapping point on the 3D solid model of the titanium alloy pipe, which corresponds to the mapping position of a certain detection signal on the model surface. Rectangular regions of preset sizes are extended along the axial and circumferential directions as the initial local analysis regions. The preset sizes are set according to the detection accuracy requirements, with a typical axial size of 10 millimeters and a typical circumferential size of the width corresponding to a five-degree arc length. The initial local analysis region is translated a certain distance along both the axial and circumferential directions. The translation distance is less than the region size, thereby forming a new local analysis region. The axial translation distance is set to eight millimeters, and the circumferential translation distance is set to four arc lengths. This ensures that there is an overlap band of a preset width between adjacent local analysis regions. The overlap band has a width of two millimeters in the axial direction and a width of one arc length in the circumferential direction.

[0064] It is understood that the division of the local analysis region covers the entire surface to be inspected of the 3D solid model of the titanium alloy pipe. The division process is completed by iterating through all mapping points. For each initial center point, a corresponding local analysis region is generated. All local analysis regions are numbered using a hierarchical numbering system, including axial partition numbers and circumferential partition numbers. The coordinates of the center point and the boundary range of each local analysis region on the 3D solid model of the titanium alloy pipe are recorded. The center point coordinates are stored in 3D Cartesian coordinates, and the boundary range is defined by the axial start position, axial end position, circumferential start angle, and circumferential end angle. The information of the local analysis regions is stored in an index database, which supports fast querying and retrieval, providing spatial positioning basis for subsequent feature extraction. In some embodiments, the shape of the local analysis region can be adjusted to other geometric shapes according to the detection requirements, such as circles or ellipses, but rectangular regions are easy to calculate and adapt to the cylindrical surface structure of the pipe. The design of the overlapping zone aims to eliminate feature loss caused by the division boundary and improve the robustness of subsequent feature extraction. Through the data redundancy of the overlapping region, the analysis results between different regions can be smoothly transitioned.

[0065] In practical implementation, the algorithm for dividing the local analysis region needs to handle the geometric characteristics of the 3D solid model of the titanium alloy pipe. Since the 3D solid model of the titanium alloy pipe is cylindrical, the curvature effect needs to be considered when projecting a rectangular area onto the surface of the 3D model. The algorithm accurately determines the circumferential boundary of the region by calculating the geodesic of the cylindrical surface. It can be understood that the number of local analysis regions depends on the length and diameter of the titanium alloy pipe and the granularity of the division. For typical industrial-sized titanium alloy pipes, the number of local analysis regions can reach tens of thousands, requiring efficient memory management and computational resource scheduling. In some embodiments, the system allows operators to adjust the size and overlap width of the local analysis region through a human-machine interface to adapt to different specifications of titanium alloy pipes or different defect detection sensitivity requirements. The adjusted parameters are updated in real time to update the division algorithm. After the division process is completed, each local analysis region is associated with a set of spatially registered detection signal data. This data serves as input for subsequent time-frequency joint analysis, and the association is established by matching the region number with the signal coordinates.

[0066] In practical implementation, the accuracy of dynamic spatial registration directly affects the accuracy of defect location. Therefore, it is necessary to regularly calibrate the encoder and detection probe. The calibration process uses standard test blocks, and the coordinate transformation parameters are corrected by comparing the measured values ​​with the actual values. The local analysis region division strategy aims to balance computational efficiency and detection resolution. Smaller region sizes can improve spatial resolution but increase computational load, while larger region sizes will reduce resolution but improve processing speed. The determination of the preset size needs to comprehensively consider hardware performance and detection standards. Optionally, for special areas such as the ends or joints of titanium alloy pipes, a non-rectangular local analysis region division method can be adopted, such as an adaptively sized region, to cope with changes in geometry. Optionally, the system can record a local analysis region division log, which includes division parameters, timestamps, and operator information for quality traceability and process auditing. Dynamic spatial registration and overlapping local analysis region division are used.

[0067] Example 2: See Figure 3In practice, the time-frequency joint analysis process is carried out on the spatially registered detection signal within each local analysis region. The detection signal is preprocessed time-domain waveform data. Empirical Mode Decomposition (EMD) is the first step in the analysis. The EMD algorithm decomposes the signal into a set of intrinsic mode function (EMF) components through an iterative screening process. Each EMF component must satisfy the condition that the number of extrema is equal to or differs from the number of zero-crossings by at most one. The EMD process first identifies all local extrema of the detection signal, including local maxima and local minima. Then, cubic spline interpolation is used to fit the local maxima to form the upper envelope and the local minima to form the lower envelope. The mean envelope of the upper and lower envelopes is calculated, and the mean envelope is subtracted from the original detection signal to obtain candidate components. The above screening process is repeated until the candidate components meet the criteria for EMF components, thus obtaining the first EMF component. The first intrinsic mode function component is separated from the original detection signal to obtain the residual signal. The above screening process is repeated on the residual signal to extract the subsequent intrinsic mode function components in turn until the residual signal becomes a monotonic function or a constant. Finally, a set of intrinsic mode function components is obtained. The number of components depends on the complexity of the signal.

[0068] The energy entropy and sample entropy of each intrinsic mode function (EMF) component are calculated. Energy entropy is calculated based on the energy probability distribution of the EMF components. First, the total energy of each EMF component is calculated. Then, the proportion of each EMF component's energy to the total energy is calculated. Finally, the energy entropy value is calculated using the information entropy formula. Sample entropy is used to measure the complexity of the EMF component time series. Sample entropy calculation requires setting embedding dimension and tolerance threshold parameters, and estimates the regularity of the sequence by statistically matching the template vector probability. The energy entropy and sample entropy values ​​are combined to form a first feature subset describing the signal complexity. This first feature subset contains entropy features of multiple dimensions, each dimension corresponding to the entropy value of an EMF component.

[0069] In practice, a Hilbert transform is performed on each intrinsic mode function (EMF) component. The Hilbert transform is achieved through convolution, where the EMF component is convolved with a Hilbert kernel function to obtain the imaginary part of the analytic signal. Based on the analytic signal, the instantaneous frequency and instantaneous amplitude are calculated. The instantaneous amplitude is the magnitude of the analytic signal, and the instantaneous frequency is the derivative of the phase of the analytic signal. The mean and variance of the instantaneous frequency and amplitude are statistically analyzed. The mean reflects the average level of the instantaneous parameters, and the variance reflects the degree of fluctuation of the instantaneous parameters, forming a second feature subset describing the time-varying characteristics of the signal. The first and second feature subsets are concatenated to form a feature vector characterizing changes in the surface microstructure. The dimension of the feature vector is equal to twice the number of EMF components multiplied by the number of features for each component.

[0070] It is understandable that the quality of feature vector extraction directly affects the accuracy of defect identification; therefore, it is necessary to optimize the parameter settings of empirical mode decomposition and feature calculation. In some embodiments, the screening stopping criterion for empirical mode decomposition can be the standard deviation criterion, i.e., screening stops when the standard deviation of two consecutive screening results is less than a set threshold. After feature vector extraction is completed, the process of constructing a circular feature sequence begins. At the same axial section position of the three-dimensional solid model of the titanium alloy pipe, feature vectors of all local analysis regions at all circumferential angles are extracted through spatial index query. The extracted feature vectors are arranged in ascending order of circumferential angles to form a one-dimensional data sequence with a circular structure, where each element corresponds to a feature vector at a circumferential position.

[0071] In practical implementation, the one-dimensional data sequence is smoothed using digital filtering algorithms, such as moving average filtering or Savitzky-Golay filtering. Moving average filtering replaces each data point with the arithmetic mean of its neighborhood data points; the neighborhood size is called the window width. Savitzky-Golay filtering smooths the data through local polynomial fitting, better preserving the signal's peak characteristics. The smoothed data sequence generates the final circular feature sequence, which is stored in memory as an array, simultaneously recording the corresponding axial position coordinates and timestamp information. The mathematical representation of the circular feature sequence is:

[0072]

[0073] in: Indicates time Circumferential angle The circular feature sequence value at the location, This represents the number of dimensions of the feature vector. Indicates the first The weighting coefficients of each feature, Indicates time Circumferential angle The eigenvector at position 1 Individual component values. Weighting coefficients. Features are assigned higher weights based on their importance, which is pre-defined.

[0074] It is understandable that constructing a circular feature sequence allows circumferentially distributed defect features to be presented as a continuous sequence, facilitating the observation of defect distribution patterns. In some embodiments, for multimodal fusion datasets, feature vectors can be extracted from the data of each modality separately, and then the feature vectors of different modalities can be fused to construct the circular feature sequence. Optionally, the circular feature sequence can be further normalized to eliminate the influence of dimensional differences on subsequent analysis. The data structure of the circular feature sequence supports fast querying and visualization, allowing operators to observe the waveform changes of the circular feature sequence through the interface. Optionally, the system can set a storage strategy for the circular feature sequence, such as storing only the sequence data of abnormal segments to save storage space.

[0075] Example 3: In specific implementation, the cross-domain association matching process begins by determining the process stages experienced by the pipe segment corresponding to the annular feature sequence during manufacturing based on the conveying speed and timestamp information of the titanium alloy pipe. The conveying speed is measured and recorded in real time by an encoder installed on the production line, and the timestamp information is strictly synchronized with the generation time of the annular feature sequence. By aligning the timestamp of the annular feature sequence with the time axis of the manufacturing process, it can be accurately mapped to a specific process stage, such as the heating stage or the rolling stage. Key process parameters recorded at the same process stage for the pipe segment are retrieved from the production process database. The production process database uses a time-series database architecture to store manufacturing data. Key process parameters include the temperature of each zone of the heating furnace, the rolling force of the rolling mill, and the cooling water pressure. These parameters are stored in time-series form, and the sampling interval is consistent with the sampling interval of the detection signal or aligned through resampling technology. The retrieval process is based on a time range query, extracting the process parameter data corresponding to the start and end times of the annular feature sequence, and arranging them in chronological order to form a manufacturing process parameter sequence. The length of the manufacturing process parameter sequence may differ from the length of the annular feature sequence, therefore time normalization processing is required.

[0076] The cross-correlation function of the circular feature sequence and the manufacturing process parameter sequence in the time domain is calculated. The cross-correlation function quantifies the similarity between the two sequences at different time lags. Before calculation, both the circular feature sequence and the manufacturing process parameter sequence are time-normalized to ensure a unified time axis scale. The time normalization process uses a linear interpolation algorithm to resample the two sequences to the same time point, eliminating time scale differences caused by different sampling rates and ensuring the accuracy of point-to-point matching. A sliding window algorithm is used to calculate the cross-correlation function. This algorithm slides the manufacturing process parameter sequence along the time axis, calculating the similarity metric between the circular feature sequence and the manufacturing process parameter sequence at multiple time lags. The similarity metric typically uses the Pearson correlation coefficient or standardized cross-correlation value. For each lag position δ, the similarity metric is calculated and recorded, generating a cross-correlation function curve. The cross-correlation function curve is a continuous function with lag time as the x-axis and similarity as the y-axis, reflecting the correlation strength between the two sequences at different time offsets. The mathematical expression of the cross-correlation function is:

[0077]

[0078] in: Indicates the time lag The cross-correlation function value at the location, Indicates the circular feature sequence at time point The value, This indicates the sequence of manufacturing process parameters at time points. The value, This represents the mean of the circular characteristic sequence. This represents the mean of the manufacturing process parameter sequence. Indicates the length of the sequence. This represents the lag time parameter, and its value range is set according to the process response time.

[0079] The cross-correlation function curve is iterated through all data points to detect local maxima. Local maxima are identified by comparing the values ​​of each data point with its neighbors; if a data point's value is greater than the values ​​of its left and right neighbors, it is marked as a local maxima. All local maxima are sorted by similarity metric, and the local maxima with the highest similarity metric is determined as the peak point of the cross-correlation function. The peak point corresponds to the time offset of the best match between the circular feature sequence and the manufacturing process parameter sequence. A correspondence between the changes in the circular feature and the fluctuations in the process parameters is established through the peak point position. For example, the peak point lag time represents the delay time in the process parameters affecting defect features; this relationship is used for subsequent anomaly pattern recognition. In some embodiments, the cross-correlation function calculation can be accelerated using a Fast Fourier Transform algorithm, especially for long sequences, by reducing computational complexity through frequency domain transformation.

[0080] In practice, the time normalization process involves reading the timestamp arrays of the circular feature sequence and the manufacturing process parameter sequence, calculating the minimum and maximum values ​​of the two timestamp sequences, and determining a unified time axis range. The time axis is discretized into equally spaced points, with the interval size set according to the dynamic characteristics of the process, typically on the order of milliseconds. Then, a linear interpolation method is used to calculate the values ​​of the circular feature sequence and the manufacturing process parameter sequence at discrete time points. The linear interpolation formula is that for any time point t, its value is obtained by fitting a straight line to adjacent original data points, ensuring that the interpolated sequence maintains the original trend. The window size of the sliding window algorithm is set to an integer multiple of the typical fluctuation period of the process parameters. For example, if the temperature fluctuation period of the heating furnace is several minutes, the window size corresponds to a time period of several minutes. The window sliding step size is set to one-tenth to one-fifth of the window size to balance computational accuracy and efficiency. When calculating the similarity metric, for each lag position, the subsequences of the circular feature sequence and the manufacturing process parameter sequence within the window are extracted. The statistical characteristics of the subsequences, such as the mean and standard deviation, are calculated, and then the Pearson correlation coefficient formula is applied to calculate the correlation. It is understandable that the Pearson correlation coefficient can eliminate the influence of dimensions and is sensitive to linear correlations between sequences, but it may not be accurate enough for nonlinear relationships. Therefore, in some embodiments, it can be combined with other similarity measures such as mutual information or dynamic time warping. Optionally, for multi-parameter cases, the manufacturing process parameter sequence can be a multi-dimensional vector sequence. In this case, the cross-correlation function calculation needs to be extended to a multivariate version. For example, the cross-correlation function between the circular feature sequence and each process parameter sequence can be calculated, and then the weighted average or the maximum value can be taken as the comprehensive similarity.

[0081] In practical implementation, the local maximum point detection algorithm adopts a sliding window comparison method, setting a local window size, typically 5 to 11 data points. For each data point, it is checked whether it is the maximum value within the window. After the peak point of the cross-correlation function is determined, the lag time value and similarity value of the peak point are recorded and stored in the matching result database. The matching results include the ring feature sequence identifier, manufacturing process parameter sequence identifier, peak lag time, peak similarity, and timestamp information for subsequent querying and analysis. In some embodiments, the system can display the cross-correlation function curve and peak point position in real time, facilitating operators to monitor the matching process. Optionally, for periodic processes, the cross-correlation function may present multiple peak points. In this case, it is necessary to combine process knowledge to select the main peak point, for example, selecting the peak point with the shortest lag time as the main matching point.

[0082] Example 4: In specific implementation, identifying abnormal feature patterns related to specific process events begins with setting amplitude change thresholds and rate of change thresholds for the ring feature sequence. The amplitude change threshold is determined based on the historical statistical distribution of the ring feature sequence data. By calculating the mean and standard deviation of the long-term data of the ring feature sequence, the amplitude change threshold is set to the mean plus three times the standard deviation. The rate of change threshold is set by analyzing the differential sequence of the ring feature sequence. The difference between adjacent points of the ring feature sequence is calculated and divided by the time interval to obtain the instantaneous rate of change sequence. Then, the absolute value of the rate of change sequence is taken and its statistical distribution is calculated. The rate of change threshold is set to twice the mean of the absolute values ​​of the rate of change. When the value of the ring feature sequence monitored in real time exceeds the amplitude change threshold or the differential value of the ring feature sequence exceeds the rate of change threshold, the system automatically marks that moment as a potential anomaly. The marking process includes recording the timestamp of the potential anomaly, the value of the ring feature sequence exceeding the limit, and the type of exceeding the limit. The manufacturing process parameter sequence associated with the potential anomaly point is traced back. The tracing operation takes the potential anomaly point as the center point and expands forward and backward by a preset time window. The size of the time window is set according to the process response characteristics. For example, the response time of the heating process is relatively long, so the window is set to a few minutes. The response time of the rolling process is fast, so the window is set to tens of seconds. All process parameter data within the time window are extracted from the manufacturing process database to form a manufacturing process parameter sequence fragment.

[0083] In practice, the process parameter sequence is checked for parameter limit violations or drastic jumps. Limit violation checks are performed by comparing each parameter value in the sequence with a preset safety range, which is read from the process specification file. For example, the safety range for furnace temperature is ±10 degrees Celsius, and for rolling force it is ±5 percent. If a parameter value exceeds the upper or lower limit of the safety range, it is marked as a limit violation event, and the name of the violation parameter, the time of violation, the magnitude of the violation, and the duration are recorded. Dramatic jump events are checked by calculating the rate of change of parameters between consecutive sampling points in the sequence. The rate of change is equal to the current sample value minus the previous sample value divided by the sampling time interval. If the absolute value of the rate of change exceeds a preset abrupt change threshold, it is marked as a drastic jump event. The threshold is set according to the normal fluctuation range of the parameter; for example, the temperature change rate threshold is 5 degrees Celsius per second, and the rolling force change rate threshold is 10 kilonewtons per second. The results are recorded in the event log, including the event type, parameter name, event time, and change amount. The formula for calculating the rate of change of a circular characteristic sequence is:

[0084]

[0085] in: Indicates the circular feature sequence at time point rate of change, Indicates the circular feature sequence at time point The value, This indicates that the circular feature sequence at the previous time point The value, This indicates the time interval between adjacent time points.

[0086] In specific implementation, referring to Table 1, if there are parameter limit exceedance events or drastic jump events, the system will identify potential anomalies as defect feature patterns caused by specific process events. This identification process is associated with the abnormal morphology of the ring feature sequence and the abnormal process parameter event. For example, if a peak pulse appears in the ring feature sequence and the furnace temperature exceeds the limit simultaneously, then the defect pattern caused by the temperature anomaly is identified. The ring feature pattern corresponding to this defect feature pattern and its associated abnormal process parameter information are recorded. The ring feature pattern includes sequence waveform features such as peak value, pulse width, and rise slope. The abnormal process parameter information includes the abnormal parameter value, event type, and duration. The storage format uses structured data records, including timestamps, ring feature sequence data segments, process parameter data segments, and event classification labels. It is understood that the identification of defect feature patterns needs to meet the principle of time correlation, that is, the abnormal process parameter event must be highly synchronized with the abnormal ring feature sequence in time. Typically, the time difference between the events is required to be within the process response time window. In some embodiments, for complex anomalies, it may be necessary to combine multiple abnormal process parameter events for comprehensive judgment. For example, if a temperature exceedance and a rolling force jump occur simultaneously, the defect pattern is marked as a composite anomaly.

[0087] Table 1: Threshold Setting and Event Checking Parameters

[0088]

[0089] The parameter values ​​in the table are set according to the specific production line configuration. The safety range is based on the process design specifications, and the mutation threshold is based on historical normal fluctuation data statistics. The monitoring frequency is synchronized with the data acquisition system. Optionally, the system allows operators to adjust these parameters according to product specifications, such as setting different temperature safety ranges for different grades of titanium alloy pipes. It is understood that threshold parameters need to be updated periodically to adapt to changes in equipment status, typically once per production batch or monthly. In some embodiments, the system can provide an adaptive threshold adjustment function, dynamically calculating thresholds based on real-time data, such as using a sliding window to statistically analyze the latest annular characteristic sequence data to recalculate the amplitude change threshold and rate of change threshold.

[0090] In practical implementation, the potential anomaly marking algorithm adopts a real-time stream processing approach, continuously monitoring the input circular feature sequence. When an event exceeding a threshold is detected, a backtracking query is immediately triggered. Extraction of manufacturing process parameter sequence segments involves database time range queries, with the query condition being the time window before and after the potential anomaly point. Parameter limit violation event checks traverse each data point in the sequence segment, comparing it with a preset safety range and recording all violation points. Abrupt change event checks calculate the rate of change for each parameter point, compare it with a mutation threshold, and mark the violation points. After event confirmation, the system generates a defect feature pattern report, including text descriptions and data charts, for quality analysis. Optionally, for transient anomalies, the system can be configured with a de-jittering mechanism, such as requiring the anomaly to persist for multiple sampling points before confirmation, to avoid false alarms. Through automated anomaly identification and tracing, early defect detection and process improvement are achieved.

[0091] See Figure 4 This figure visually presents the dynamic changes and anomaly identification results of the annular feature sequence during the inspection of titanium alloy pipes. The blue area represents the annular feature sequence after real-time acquisition and processing; its fluctuations reflect the dynamic characteristics of the circumferential surface microstructure of the pipe. The red dashed line represents the amplitude change threshold (14.51), which is set based on historical data statistics of the annular feature sequence (mean + three standard deviations). The red scatter dots mark amplitude anomalies, i.e., the moments when the annular feature sequence value exceeds the amplitude change threshold. From a time perspective (horizontal axis, unit: seconds), the annular feature sequence shows significant fluctuations exceeding the threshold around 50 seconds and after 150 seconds, corresponding to the marked amplitude anomalies. These anomalies are marked through continuous monitoring of the annular feature sequence using a real-time stream processing algorithm. Subsequent analysis, combined with backtracking of manufacturing process parameter sequences (such as furnace temperature, rolling mill force, etc.), can confirm whether they are related to process events.

[0092] Example 5: In specific implementation, the construction and maintenance of the defect knowledge base is a crucial step in the continuous optimization of the system. Confirmed defect feature patterns and their associated process parameter anomaly information are stored in the defect knowledge base in a structured format. The defect knowledge base is implemented using a relational database management system and contains multiple data tables storing defect feature pattern descriptions, circular feature sequence data, process parameter anomaly information, and timestamp indexes, respectively. Each defect record includes a unique defect pattern identifier, a numerical array of the circular feature sequence, the name of the associated process parameter, the type of the anomaly event, the parameter limit exceedance or jump rate, the event occurrence time, the duration, and a processing suggestion comment. When the system detects a new circular feature sequence online again, the real-time pattern matching process is triggered. After preprocessing, the new circular feature sequence is compared with the historical defect feature patterns stored in the defect knowledge base for similarity calculation. The matching process uses algorithms based on distance or similarity metrics, such as calculating the cosine similarity or Euclidean distance between the new circular feature sequence and each historical circular feature sequence fragment in the knowledge base. The similarity calculation formula for pattern matching is:

[0093]

[0094] in: This represents the similarity score between the new circular feature sequence and the historical defect feature pattern. Indicates the new circular feature sequence at the th The value of each sampling point The circular feature sequence representing the historical defect feature pattern is in the first place. The value of each sampling point This represents the mean of the new circular feature sequence. The mean of the circular feature sequence representing the historical defect feature pattern. Indicates the length of the circular feature sequence. Similarity score. The value range is [-1, 1], and the closer the value is to 1, the higher the similarity.

[0095] If the calculated similarity score exceeds a preset confidence level (typically set to 0.85), the system determines that the newly detected circular feature sequence highly matches a historical defect feature pattern in the defect knowledge base. It then directly retrieves the corresponding process parameter anomaly information from the defect knowledge base, including the name of the process parameter that previously caused this type of defect, the anomaly type, and recommended handling measures. This information guides early intervention in the production process, such as automatically adjusting the furnace temperature setpoint or triggering an alarm to prompt operators to check the rolling mill status. It is understandable that the timeliness of early intervention plays a crucial role in preventing the mass generation of defects. The system can complete pattern matching and knowledge retrieval within seconds, achieving near real-time defect warnings and process adjustment suggestions.

[0096] Regularly performing clustering analysis on the defect feature patterns stored in the defect knowledge base is a core step in knowledge base optimization. Clustering analysis employs unsupervised learning algorithms, such as K-means clustering or DBSCAN density clustering, using the circular feature sequence data of all historical defect feature patterns in the defect knowledge base as the input dataset. The clustering process first requires standardizing the circular feature sequence data to eliminate the influence of dimensions. Then, based on the geometric shape, statistical characteristics, and changing trends of the sequence waveform features, the similarity distance between patterns is calculated. Repeated or similar defect feature patterns are automatically merged into the same category. Each category is represented by a representative pattern prototype; for example, the centroid of all pattern feature vectors within a category is taken as the typical defect feature pattern of that category. Data mining techniques are used to analyze the association rules between each type of defect feature pattern and abnormal process parameters. From the clustered defect categories and their associated abnormal process parameter records, frequently occurring combinations of abnormal process parameters are extracted. The conditional probability of defect occurrence is calculated, and the key process parameter combinations and critical conditions leading to this type of defect are extracted. For example, it was found that when the furnace temperature exceeds 1230 degrees Celsius and the rolling force is below 870 kN, there is a 95% probability that a specific type of fluctuation pattern will appear in the ring feature sequence. Based on the extracted association rules, the amplitude change threshold and change rate threshold of the ring feature sequence are dynamically updated. The update algorithm recalculates the thresholds based on the statistical distribution of historical data. For example, the amplitude change threshold is adjusted to the mean of normal ring feature sequence data in the most recent month plus n times the standard deviation, where the value of n is set according to the defect false alarm rate requirement, achieving adaptive optimization of the detection system.

[0097] In implementation, the defect knowledge base employs a layered data storage architecture. Original circular feature sequence data is stored in a cache layer for real-time matching, while historical aggregated data and pattern prototypes are stored in a disk database for batch analysis. The pattern matching service runs as an independent microservice, receiving new circular feature sequence data from the real-time detection system, concurrently querying the defect knowledge base, and returning the matching results to the control system via a message queue. Periodic clustering analysis tasks are triggered by a scheduler, typically weekly or monthly, depending on the amount of accumulated production data. The clustering algorithm runs on a distributed computing framework to handle large-scale historical data. Association rule mining results are stored as rule sets, supporting visualization and manual review. The threshold update module runs automatically after each clustering analysis, and the new threshold takes effect after confirmation by a quality control engineer.

[0098] See Figure 5The figure presents the distribution of defect numbers in each cluster category after clustering analysis of defect feature patterns. Specifically, the clustering analysis takes the circular feature sequence of historical defect feature patterns in the defect knowledge base as input. After standardization, the category division is completed through an unsupervised learning algorithm. The number of defects in the six cluster categories (corresponding to defect types 1 to 6) in the figure are 128, 95, 76, 58, 42, and 31, respectively. This distribution result can support subsequent association rule mining: by analyzing the abnormal records of process parameters corresponding to each cluster category, the association relationship between different defect types and process parameters can be extracted. At the same time, this distribution also provides a quantitative basis for pattern merging and dynamic threshold updating in the defect knowledge base, which is a key data support for realizing adaptive optimization of the detection system.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting defects on the inner and outer surfaces of a titanium alloy pipe, characterized by, The method comprises: Pretreatment of the original detection signals from a plurality of detection probes arranged on the inner and outer surfaces of the titanium alloy pipe, eliminating noise interference caused by environmental vibration, and generating a multimodal fusion data set; Based on the real-time motion state of the titanium alloy pipe, the multimodal fusion data set is dynamically spatially registered, and the detection signals in different spatial coordinates are mapped to the surface of a unified pipe three-dimensional entity model; Divide the pipe three-dimensional entity model surface into mutually overlapping local analysis regions, and perform time-frequency joint analysis on the multimodal fusion data set in each local analysis region to extract feature vectors representing surface micro-topography changes; Aggregate the feature vectors of local analysis regions belonging to the same axial position but from different circumferential angles to construct a ring-shaped feature sequence that can reflect the circumferential defect distribution characteristics of the titanium alloy pipe; Cross-domain correlation matching of the ring-shaped feature sequence and the manufacturing process parameter sequence of the corresponding pipe segment extracted from the production process database identifies abnormal feature patterns related to specific process events, including: According to the conveying speed and timestamp information of the titanium alloy pipe, determine the process stage experienced by the pipe segment corresponding to the ring-shaped feature sequence during the manufacturing process; Retrieve the key process parameters recorded in the same process stage for this pipe segment from the production process database, including the temperature of each zone of the heating furnace, the rolling force of the rolling mill, and the cooling water pressure, and arrange them in time sequence to form a manufacturing process parameter sequence; Calculate the cross-correlation function of the ring-shaped feature sequence and the manufacturing process parameter sequence in the time domain to find the peak point of the cross-correlation function, including: Time normalization processing of the ring-shaped feature sequence and the manufacturing process parameter sequence to make the two sequences have a uniform time axis scale; Using a sliding window algorithm, calculate the similarity measure value of the ring-shaped feature sequence and the manufacturing process parameter sequence at multiple time lag positions to generate a cross-correlation function curve; Traverse all data points on the cross-correlation function curve, detect local maximum points, and determine the local maximum point with the highest similarity measure value as the peak point of the cross-correlation function; Thus, the corresponding relationship between the ring-shaped feature changes and the process parameter fluctuations is established.

2. The method of claim 1, wherein the method is characterized by: The dynamic spatial registration of the multimodal fusion data set based on the real-time motion state of the titanium alloy pipe comprises: Obtain the linear displacement and rotation angle of the titanium alloy pipe measured by the encoder, and calculate the real-time position and attitude of any point on the titanium alloy pipe in the global coordinate system; According to the inherent installation position and angle of each detection probe, establish the real-time transformation relationship from the probe coordinate system to the global coordinate system of the titanium alloy pipe; According to the real-time transformation relationship, dynamically map the surface feature points corresponding to the original detection signals collected by each detection probe to the accurate position on the titanium alloy pipe three-dimensional entity model.

3. The method of claim 2, wherein the method is characterized by: The division of the pipe three-dimensional entity model surface into mutually overlapping local analysis regions comprises: Taking any mapping point on the titanium alloy pipe three-dimensional entity model as the center, expand a rectangular region of a predetermined size in the axial and circumferential directions as an initial local analysis region; Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents.

4. The method of claim 3, wherein the method is characterized by: Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents.

5. The method of claim 1, wherein the method is characterized by: Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. Translate the given patent specification text into English, ensuring that no details or formatting are omitted. Use formal, professional terminology and adhere to the style and norms of American patents. The method further comprises: storing the confirmed defect feature mode and its associated process parameter abnormal information into a defect knowledge base; 6. The method of claim 1, wherein the method is characterized by: ​ ​ ​ ​ 7. The method of claim 6, wherein the method is characterized by: ​ ​ ​ ​ 8. The method of claim 7, wherein the method is characterized by: ​ ​ When a new ring feature sequence is detected again online, it is pattern-matched with historical defect feature patterns stored in the defect knowledge base; If the matching degree exceeds a preset confidence level, corresponding process parameter abnormality information in the defect knowledge base is directly called to guide early intervention in the production process; Periodically, clustering analysis is performed on defect feature patterns stored in the defect knowledge base, and repeated or similar defect feature patterns are merged; The association rules between each class of defect feature patterns and process parameter abnormalities are analyzed, and key process parameter combinations and their critical conditions causing the defects of this class are refined; Based on the refined association rules, the amplitude variation threshold and the rate threshold of the ring feature sequence are dynamically updated.

Images