A method and system for detecting surface micro-defects of a semiconductor quartz tube

Abstract

This invention relates to the field of semiconductor manufacturing technology, specifically to a method and system for detecting micro-defects on the surface of semiconductor quartz tubes. The method includes the following steps: extracting the depth time axis and probe spatial coordinates to generate three-dimensional labels and splicing sequences to construct a three-dimensional matrix; constructing a feature baseline based on defect-free sequences; calculating abnormal parameters and baseline deviations to screen suspected sequences; separating surface and internal signals from suspected sequences and a reference waveform; and determining the results based on energy ratios and label generation. In this invention, by performing local reference matching and stripping on aliased suspected composite signal sequences, highly overlapping component signals are precisely decomposed, accurately distinguishing between real microcracks with minimal longitudinal physical depth differences and near-surface impurities. This effectively eliminates aliasing interference components from internal impurities on the surface response, avoiding misjudging internal impurities as surface defects or real surface flaws as internal features, thus ensuring the accuracy of micro-defect determination results under complex service conditions.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and in particular to a method and system for detecting micro-defects on the surface of a semiconductor quartz tube. Background Technology

[0002] The semiconductor manufacturing technology field mainly involves the entire process of processing silicon wafers or other semiconductor materials into integrated circuit devices. This field includes core processes such as wafer fabrication, photolithography, etching, ion implantation, thin film deposition, chemical mechanical polishing, and cleaning. It covers the entire manufacturing chain from the purification of basic materials to the patterning and transfer of complex micro-circuit structures to the final chip packaging and testing, and is highly dependent on the intervention of ultra-clean environment control and high-purity consumables. Among them, the traditional semiconductor quartz tube surface micro-defect detection method refers to the inspection process of identifying and locating micro-defects such as scratches, cracks, micropores, and particulate deposits on the surface of the quartz tube, which serves as a carrier component in semiconductor process equipment.

[0003] Traditional methods utilize probes to scan along the surface of quartz tubes and receive returned signals, relying on fixed depth thresholds to define boundaries. When the difference in longitudinal physical depth between near-surface impurities and microcracks is small, the detection signal penetrates the surface and interacts with both cracks and impurities to produce highly overlapping composite signal peaks. This makes it difficult to effectively remove the superposition interference of internal impurities on surface signals at this tiny scale. Consequently, internal impurities near the surface are incorrectly classified into surface channels, or real microcracks are misjudged as internal features. This causes non-surface features to be mixed into the defect judgment results, resulting in inaccurate classification. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a method and system for detecting micro-defects on the surface of semiconductor quartz tubes.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting micro-defects on the surface of a semiconductor quartz tube, comprising the following steps:

[0006] S1: During the semiconductor quartz tube scanning stage, continuous analog signals are acquired and depth time axis coordinates are extracted to generate an original time series. Simultaneously, the axial coordinates and circumferential angle coordinates of the scanning probe are acquired. The depth time axis coordinates are combined with the axial coordinates and circumferential angle coordinates of the scanning probe to generate three-dimensional spatial coordinate labels. Based on the three-dimensional spatial coordinate labels, the original time series is subjected to association and splicing processing to generate a three-dimensional data matrix.

[0007] S2: Delineate a sliding analysis window based on the three-dimensional data matrix, filter out local defect-free background regions, perform fusion extraction on the original time series within the local defect-free background regions to generate a reference surface response waveform, and calculate and construct a local feature baseline based on the feature parameters of the reference surface response waveform.

[0008] S3: Based on the waveform energy integral value of the original time series within the sliding analysis window, filter the set of abnormal waveform feature nodes, extract the set of abnormal morphological feature parameters, calculate the absolute deviation value between the set of abnormal morphological feature parameters and the local feature baseline, and filter the suspected composite abnormal time series based on the absolute deviation value.

[0009] S4: Call the local reference matching waveform stripping algorithm to match and strip the suspected composite anomaly time series with the reference surface response waveform to generate surface candidate component signals, and extract internal interference candidate component signals based on the difference between the suspected composite anomaly time series and the surface candidate component signals.

[0010] S5: Obtain the surface component energy ratio based on the surface candidate component signal and the internal interference candidate component signal, and determine the surface microcrack properties in combination with the three-dimensional spatial coordinate label to generate the surface microdefect detection result of the semiconductor quartz tube.

[0011] The present invention is improved in that the specific steps of S1 are as follows:

[0012] S111: Monitor the operating status of the multi-axis scanning mechanism during the semiconductor quartz tube scanning stage, collect the continuous analog signal corresponding to the reflected echo from the inner wall of the semiconductor quartz tube, extract the signal distribution points and depth time axis coordinates inside the continuous analog signal, and generate the original time series.

[0013] S112: Collect the axial coordinates and circumferential angle coordinates of the scanning probe, and combine the depth time axis coordinates, the axial coordinates and the circumferential angle coordinates of the scanning probe attached to the original time series to generate a three-dimensional spatial coordinate label;

[0014] S113: Map the three-dimensional spatial coordinate labels to the corresponding original time series, extract the three-dimensional spatial distribution coordinate features inside the three-dimensional spatial coordinate labels, and perform association splicing processing on the original time series with matching label states based on the three-dimensional spatial distribution coordinate features to generate a three-dimensional data matrix.

[0015] The present invention is improved in that the specific steps of S2 are as follows:

[0016] S211: Define a sliding analysis window for the three-dimensional data matrix, extract the peak intensity distribution values ​​associated with each of the original time series within the sliding analysis window, calculate the discrete fluctuation amplitude values ​​based on the peak intensity distribution values ​​corresponding to adjacent spatial positions within the three-dimensional data matrix, filter the analysis range of the semiconductor quartz tube based on the discrete fluctuation amplitude values, and establish a local defect-free background region.

[0017] S212: Perform a temporal alignment algorithm and an averaging algorithm on the original time series within the local defect-free background region to obtain a smoothed and denoised fused average sequence. Determine the fused average sequence as the reference surface response waveform that reflects the true surface reflection characteristics. Calculate the reference leading edge rise slope, reference half-width at half-maximum (WHM), reference main peak symmetry index, and reference tail decay time constant based on the reference surface response waveform. Combine the reference leading edge rise slope, the reference WHM, the reference main peak symmetry index, and the reference tail decay time constant into a local feature baseline.

[0018] The present invention is improved in that the process of calculating the discrete fluctuation amplitude value, selecting the analysis range of the semiconductor quartz tube based on the discrete fluctuation amplitude value, and establishing a local defect-free background region is specifically as follows:

[0019] Calculate the local mathematical mean of the wave crest intensity distribution values ​​corresponding to the adjacent spatial locations;

[0020] Calculate the numerical sequence of absolute deviations between the wave crest intensity distribution values ​​and the local mathematical mean;

[0021] Extract the maximum absolute deviation value within the absolute deviation value sequence;

[0022] The maximum absolute deviation value is determined as the discrete fluctuation amplitude value;

[0023] Extract the global distribution values ​​of peak intensity from the three-dimensional data matrix in the global scope;

[0024] Calculate the global fluctuation variance of the global distribution value of the peak intensity;

[0025] Extract the inherent background noise parameters corresponding to the hardware of the detection equipment;

[0026] Calculate the product between the global fluctuation variance and the inherent background noise parameter;

[0027] The product of the global fluctuation variance and the inherent background noise parameter is set as the fluctuation amplitude limit threshold.

[0028] The discrete fluctuation amplitude value is compared with the fluctuation amplitude limit threshold value.

[0029] If the discrete fluctuation amplitude value does not exceed the fluctuation amplitude limit threshold, extract the adjacent spatial positions that match the discrete fluctuation amplitude value.

[0030] The adjacent spatial locations that match the discrete fluctuation amplitude values ​​are defined as the analysis range of the semiconductor quartz tube.

[0031] By merging adjacent spatial locations within the analysis range, the local defect-free background region is established.

[0032] The present invention is improved in that the specific steps of S3 are as follows:

[0033] S311: Calculate the waveform energy integral values ​​associated with each of the original time series within the sliding analysis window. Based on the waveform energy integral values, filter the original time series within the sliding analysis window that exceed the background intensity corresponding to the local defect-free background region. Extract the leading edge start point, first peak point, trough point, second peak point, and tail zeroing point within the original time series that exceed the background intensity corresponding to the local defect-free background region. Extract the first peak amplitude value associated with the first peak point and the initial amplitude value associated with the leading edge start point to establish an abnormal waveform feature node set.

[0034] S312: Based on the first peak amplitude value, the initial amplitude value, the leading edge start point, the first peak point, the second peak point, and the tail zeroing point within the abnormal waveform feature node set, calculate the amplitude deviation value between the first peak amplitude value and the initial amplitude value, calculate the time span difference between the first peak point and the leading edge start point, calculate the ratio between the amplitude deviation value and the time span difference to obtain the abnormal leading edge rising slope, calculate the half-wave height duration corresponding to the first peak point and the second peak point to obtain the abnormal half-width value, extract the second local half-width value associated with the second peak point and the first local half-width value associated with the first peak point, calculate the ratio between the second local half-width value and the first local half-width value to obtain the abnormal main peak symmetry index, calculate the time span value between the tail zeroing point and the second peak point to obtain the abnormal tail decay time constant, and combine the abnormal leading edge rising slope, the abnormal half-width value, the abnormal main peak symmetry index, and the abnormal tail decay time constant into an abnormal morphological feature parameter set;

[0035] S313: Extract the baseline rising slope, baseline half-width at half-maximum (WHM), baseline main peak symmetry index, and baseline tail decay time constant from the local feature baseline, and the abnormal rising slope, abnormal WHM, abnormal main peak symmetry index, and abnormal tail decay time constant from the abnormal morphology feature parameter set. Calculate the first absolute deviation between the abnormal rising slope and the baseline rising slope, the second absolute deviation between the abnormal WHM and the baseline WHM, the third absolute deviation between the abnormal main peak symmetry index and the baseline main peak symmetry index, and the fourth absolute deviation between the abnormal tail decay time constant and the baseline tail decay time constant. Based on the first, second, third, and fourth absolute deviations, identify and filter the original time series exhibiting asymmetric broadening patterns to generate a suspected composite abnormal time series.

[0036] The present invention is improved in that the process of determining and filtering the original time series corresponding to the asymmetric broadening pattern based on the first absolute deviation value, the second absolute deviation value, the third absolute deviation value, and the fourth absolute deviation value, and generating a suspected composite anomalous time series, specifically includes:

[0037] Extract the historical distribution discrete variance of multiple features within a pre-set anomaly feature library;

[0038] Calculate the normalized reciprocal proportion of the discrete variance of the historical distribution;

[0039] The normalized reciprocal ratio is set sequentially as the first deviation weight coefficient, the second deviation weight coefficient, the third deviation weight coefficient, and the fourth deviation weight coefficient according to the corresponding feature terms;

[0040] Extract the global statistical distortion mean corresponding to historical composite anomaly samples within the pre-set anomaly feature library;

[0041] Set the global statistical distortion mean as the broadening distortion determination threshold;

[0042] The first independent weighted value is obtained by multiplying the first absolute deviation value by the first deviation weighting coefficient.

[0043] The second independent weighted value is obtained by multiplying the second absolute deviation value by the second deviation weighting coefficient.

[0044] The third independent weighted value is obtained by multiplying the third absolute deviation value by the third deviation weighting coefficient.

[0045] The fourth independent weighted value is obtained by multiplying the fourth absolute deviation value by the fourth deviation weighting coefficient.

[0046] The comprehensive distortion feature index is obtained by summing the first independent weighted value, the second independent weighted value, the third independent weighted value, and the fourth independent weighted value.

[0047] When the comprehensive distortion feature index is greater than the broadening distortion judgment threshold, the corresponding original time series is extracted to generate a suspected composite abnormal time series.

[0048] The present invention is improved in that the specific steps of S4 are as follows:

[0049] S411: Invoke the local benchmark matching waveform stripping algorithm to map the suspected composite anomaly time series to the composite object to be decomposed, map the benchmark surface response waveform to the reference matching template, shift the time axis associated with the reference matching template, calculate the maximum cross-correlation value of the leading edge between the reference matching template and the composite object to be decomposed, establish the time domain alignment starting point based on the maximum cross-correlation value of the leading edge, extract the amplitude scaling factor inside the local benchmark matching waveform stripping algorithm, adjust the response amplitude associated with the reference matching template after alignment based on the time domain alignment starting point based on the amplitude scaling factor, and generate a matching fitting waveform;

[0050] S412: Extract the test signal amplitude within the suspected composite anomaly time series and the template signal amplitude within the matched fitting waveform, calculate the mean square error between the test signal amplitude and the template signal amplitude, filter the target amplitude scaling factor and target time offset based on the condition constraint corresponding to minimizing the mean square error, adjust the time starting point associated with the reference surface response waveform based on the target time offset, adjust the amplitude corresponding to the reference surface response waveform based on the target amplitude scaling factor, and establish surface candidate component signals;

[0051] S413: Extract the current signal amplitude associated with the suspected composite anomaly time series and the base signal amplitude associated with the surface candidate component signal at the same time. Calculate the difference between the current signal amplitude and the base signal amplitude to obtain the amplitude deviation value. Based on the depth time axis coordinate, splice the amplitude deviation values ​​matched at multiple times to establish a residual signal sequence. Perform calibration processing on the residual signal sequence corresponding to the associated internal impurity response characteristics to generate internal interference candidate component signals.

[0052] The present invention is improved in that the process of selecting the target amplitude scaling factor and the target time offset based on the condition constraint corresponding to minimizing the mean square error value is specifically as follows:

[0053] Extract the extreme range of amplitude changes and the duration of waveform pulses from the historical statistical records of the detection equipment;

[0054] Set the extreme value range of the amplitude change as the amplitude scaling search boundary;

[0055] Set the duration of the waveform pulse as the time shift search boundary;

[0056] Extract the inherent time sampling resolution within the hardware parameters of the multi-axis scanning mechanism;

[0057] Set the inherent time sampling resolution to the time offset search step size;

[0058] Extract the historical standard deviation of the energy fluctuation of the probe signal;

[0059] The standard deviation of the historical fluctuation of the detected signal energy is set as the amplitude scaling search step size;

[0060] A candidate amplitude scaling coefficient sequence is generated based on the amplitude scaling search boundary and the amplitude scaling search step size;

[0061] A candidate time offset sequence is generated based on the time translation search boundary and the time offset search step size;

[0062] A two-dimensional parameter search grid is constructed by orthogonally combining the candidate amplitude scaling factor sequence and the candidate time offset sequence;

[0063] Within the two-dimensional parameter search grid, test parameter combinations, including individual candidate amplitude scaling factors and individual candidate time offsets, are extracted one by one.

[0064] For each combination of test parameters, the mean square error value under the parameter application state is extracted, and an error mapping matrix is ​​constructed according to the mean square error value associated with the grid coordinate arrangement.

[0065] Numerical extremum retrieval is performed within the error mapping matrix to extract the minimum mean square error value located at the global lower bound.

[0066] Extract the test parameter combination that matches the minimum mean square error value;

[0067] The single candidate amplitude scaling factor contained within the test parameter combination that matches the minimum mean square error value is established as the target amplitude scaling factor;

[0068] The single candidate time offset contained within the test parameter combination that matches the minimum mean square error value is established as the target time offset.

[0069] The present invention is improved in that the specific steps of S5 are as follows:

[0070] S511: Perform time-domain integration on the signal amplitude of the surface candidate component signal to obtain the surface component energy integral value, and perform time-domain integration on the signal amplitude of the internal interference candidate component signal to obtain the interference component energy integral value. Calculate the sum of the surface component energy integral value and the interference component energy integral value to obtain the total signal energy value. Calculate the ratio between the surface component energy integral value and the total signal energy value to generate the surface component energy proportion.

[0071] S512: The surface component energy ratio is compared with a preset weak signal limit. When the surface component energy ratio exceeds the preset weak signal limit, the surface candidate component signal is extracted from the adjacent scanning grid associated with the semiconductor quartz tube according to the three-dimensional spatial coordinate label. The scanning probe axial coordinate and scanning probe circumferential angle coordinate are extracted from the three-dimensional spatial coordinate label. The geometric topological structure features between the surface candidate component signal in the current test grid and the surface candidate component signal in the adjacent scanning grid are compared according to the scanning probe axial coordinate and scanning probe circumferential angle coordinate. When the geometric topological structure features are linear or planar, the attribute state of the current test grid is established, and the surface microcrack attribute is established.

[0072] S513: Extract the peak intensity value inside the candidate component signal of the surface associated with the surface microcrack attribute, obtain the ultrasonic signal transmission speed matched by the semiconductor quartz tube, extract the depth time axis coordinate inside the three-dimensional spatial coordinate label, calculate the product between the ultrasonic signal transmission speed and the depth time axis coordinate to obtain the equivalent depth value, bind the peak intensity value, the equivalent depth value and the surface microcrack attribute to construct a three-dimensional defect feature description set including defect intensity, defect depth and topological attributes, and output the three-dimensional defect feature description set as the surface micro-defect detection result of the semiconductor quartz tube.

[0073] A system for detecting micro-defects on the surface of a semiconductor quartz tube, the system being used to implement the aforementioned method for detecting micro-defects on the surface of a semiconductor quartz tube, the system comprising:

[0074] The scanning coordination module acquires continuous analog signals and extracts depth time axis coordinates during the semiconductor quartz tube scanning stage, generates an original time series, and simultaneously acquires the axial coordinates and circumferential angle coordinates of the scanning probe. It then combines the depth time axis coordinates with the axial and circumferential angle coordinates of the scanning probe to generate three-dimensional spatial coordinate labels. Based on the three-dimensional spatial coordinate labels, it performs correlation and splicing processing on the original time series to generate a three-dimensional data matrix.

[0075] The background calibration module defines a sliding analysis window based on the three-dimensional data matrix, filters out local defect-free background regions, performs fusion extraction on the original time series within the local defect-free background regions to generate a reference surface response waveform, and calculates and constructs a local feature baseline based on the feature parameters of the reference surface response waveform.

[0076] The anomaly identification module filters the set of abnormal waveform feature nodes based on the waveform energy integral value of the original time series within the sliding analysis window, extracts the set of abnormal morphological feature parameters, calculates the absolute deviation value between the set of abnormal morphological feature parameters and the local feature baseline, and filters suspected composite abnormal time series based on the absolute deviation value.

[0077] The component separation module calls the local reference matching waveform stripping algorithm to match and strip the suspected composite anomaly time series with the reference surface response waveform to generate surface candidate component signals, and extracts internal interference candidate component signals based on the difference between the suspected composite anomaly time series and the surface candidate component signals.

[0078] The crack assessment module obtains the surface component energy ratio based on the surface candidate component signal and the internal interference candidate component signal, and determines the surface microcrack attributes by combining the three-dimensional spatial coordinate labels, thereby generating the surface microdefect detection results of the semiconductor quartz tube.

[0079] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0080] In this invention, a three-dimensional coordinate label is generated by combining the depth time axis with the probe spatial coordinates. The time series is stitched together to construct a three-dimensional data matrix. A window is defined based on the matrix, and a local feature baseline is constructed by fusing and extracting a reference waveform based on a defect-free background area. Abnormal morphological parameters are extracted by combining waveform energy integration and the deviation is compared with the feature baseline to screen suspected composite abnormal sequences. A waveform stripping algorithm is called to match the composite abnormal sequence with the reference waveform and strip out surface candidate and internal interference component signals. The microcrack attributes are determined based on the energy ratio and coordinate label, accurately distinguishing microcracks with minimal depth differences from near-surface impurities. Internal impurity superposition interference is stripped to avoid misjudging the true features and ensure the accuracy of the judgment. Attached Figure Description

[0081] Figure 1 This is the main flowchart for the detection of micro-defects on the surface of semiconductor quartz tubes according to the present invention;

[0082] Figure 2 This is a flowchart of the three-dimensional data matrix construction process of the present invention;

[0083] Figure 3 This is a flowchart of the abnormal waveform identification process of the present invention;

[0084] Figure 4 This is a flowchart of the component stripping and crack assessment process of the present invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0086] Please see Figures 1 to 4 This embodiment provides a method for detecting micro-defects on the surface of a semiconductor quartz tube. In practical applications, for example, during the actual operation of a semiconductor quartz tube, which is fed axially and continuously probed around the circumference of the tube by a scanning probe, the probe receives the echo reflected from the tube wall to form a continuous analog signal. The scanning position simultaneously records the axial position, circumferential angle position, and echo depth time position. The detection process includes the following steps:

[0087] S1: During the semiconductor quartz tube scanning stage, continuous analog signals are acquired and depth time axis coordinates are extracted to generate the original time series. Simultaneously, the axial coordinates and circumferential angle coordinates of the scanning probe are acquired. The depth time axis coordinates are combined with these coordinates to generate three-dimensional spatial coordinate labels. Based on these labels, the original time series is then subjected to correlation and stitching processing to generate a three-dimensional data matrix. The continuous analog signal is the electrical signal output by the scanning probe at the tube wall echo receiver, carrying the continuous distribution of echo amplitude changes with the arrival sequence of the echoes. The depth time axis coordinates represent the temporal position of each sampling point relative to the trigger reference within the same echo sequence, used to distinguish the sequential relationship between surface reflection, near-surface response, and internal interference response within the same sequence. The original time series is an ordered set of echo amplitudes after sampling, DC drift removal, and trigger alignment, which, along with the axial and circumferential angle coordinates, enters the spatial stitching process.

[0088] S111: Monitor the operating status of the multi-axis scanning mechanism during the semiconductor quartz tube scanning stage, acquire the continuous analog signal corresponding to the reflected echo from the inner wall of the semiconductor quartz tube, extract the signal distribution points and depth time axis coordinates within the continuous analog signal, and generate the original time series. The operating status of the multi-axis scanning mechanism includes axial feed status, circumferential rotation status, and probe contact stability status. Only when the operating status is in the acquireable category will the continuous analog signal enter the sampling buffer. The signal distribution point refers to the position of the echo amplitude that can be located within the same depth time axis. The sampling buffer saves the amplitude, sampling order identifier, trigger status, and signal integrity identifier according to the trigger order. If the continuous analog signal has trigger loss, amplitude saturation, sequence breakage, or abnormal contact status, an abnormal acquisition identifier is written to the corresponding original time series, and it will not be entered into subsequent spatial stitching. The corresponding axial coordinates and circumferential angle coordinates are retained for verification.

[0089] S112: Acquire the axial coordinates and circumferential angle coordinates of the scanning probe. Combine the depth time axis coordinates, scanning probe axial coordinates, and scanning probe circumferential angle coordinates attached to the original time series to generate a 3D spatial coordinate label. The scanning probe axial coordinates are generated by the axial position feedback of the multi-axis scanning mechanism, and the scanning probe circumferential angle coordinates are generated by the circumferential rotation position feedback. Both are checked for consistency with the trigger identifier of the original time series. The 3D spatial coordinate label is a positioning data object attached to the original time series, containing axial position, circumferential angle position, depth time position, trigger status, and coordinate validity fields. If the axial position or circumferential angle position cannot match the trigger sequence, the original time series enters the alignment buffer. A valid label is generated only after the adjacent positions in the same scanning cycle provide feedback to complete the alignment. Records that still cannot be completed are entered into the abnormal trajectory log and are not used as the source of defect-free background areas.

[0090] S113: Map the 3D spatial coordinate labels to the corresponding original time series, extract the 3D spatial distribution coordinate features within the 3D spatial coordinate labels, and perform association stitching processing on the original time series with matching label states based on the 3D spatial distribution coordinate features to generate a 3D data matrix. The 3D spatial distribution coordinate features refer to the spatial arrangement relationship formed by the continuity of axial position, circumferential angle, and depth time axis. The association stitching processing determines the affiliation of each original time series in the tube wall scanning grid according to the axial position and circumferential angle position, and then arranges the echo sampling points in the same grid according to the depth time axis coordinate to form a 3D data matrix that can be read by a local window. Each matrix element in the 3D data matrix stores the echo amplitude, depth time axis coordinate, axial coordinate, circumferential angle coordinate, acquisition validity identifier, and original sequence index. Subsequent steps use this index to trace back the original acquisition records.

[0091] S2: Based on the 3D data matrix, a sliding analysis window is defined to filter out locally defect-free background regions. The original time series within these regions is then fused and extracted to generate a reference surface response waveform. A local feature baseline is constructed based on the characteristic parameters of the reference surface response waveform. The sliding analysis window is the local data range extracted along the axial, circumferential, and depth time axes of the 3D data matrix. Its boundaries are determined by the continuity of the scanning grid and valid acquisition markers. The locally defect-free background region is a set of adjacent spatial locations within the window whose peak intensity distribution remains within the stable background category and who do not exhibit abnormal acquisition markers. This set is used to construct a reference response for the current local material state of the pipe wall.

[0092] S211: A sliding analysis window is defined for the 3D data matrix. Peak intensity distribution values ​​associated with each original time series within the sliding analysis window are extracted. Based on the peak intensity distribution values ​​corresponding to adjacent spatial positions within the 3D data matrix, discrete fluctuation amplitude values ​​are calculated. The analysis range for the semiconductor quartz tube is then selected based on these discrete fluctuation amplitude values, establishing a local defect-free background region. The peak intensity distribution value is the amplitude field corresponding to the main echo peak in the same original time series, derived from valid acquisition records. The discrete fluctuation amplitude value is the amplitude field representing the deviation of peak intensity from the local background state between adjacent spatial positions. Its engineering meaning is whether the surface echo within adjacent scanning grids maintains spatial continuity. When generating this field, peak intensity distribution values ​​are first extracted from adjacent spatial positions and a local equilibrium background is established. Then, the deviation degree of the peak intensity distribution values ​​relative to this local equilibrium background is read item by item. Subsequently, the field value representing the fluctuation boundary of adjacent positions is selected from the deviation degree sequence as the discrete fluctuation amplitude value.

[0093] The fluctuation amplitude threshold is a configurable judgment field stored in the equipment calibration parameter area. Its sources include the peak intensity distribution records of the three-dimensional data matrix globally and the inherent background noise parameter of the detection equipment. The global fluctuation variance field characterizes the overall discrete state of the peak intensity distribution in the current batch of scan data. The inherent background noise parameter originates from the factory calibration and no-load calibration records of the detection equipment and is used to distinguish between hardware noise and the true response of the tube wall. The fluctuation amplitude threshold is generated after the above fields are entered into the calibration rule table and serves as the local stability judgment boundary within the sliding analysis window. When the discrete fluctuation amplitude value does not exceed this judgment boundary, the corresponding adjacent spatial position enters the semiconductor quartz tube analysis range; adjacent analysis ranges are spatially connected and have consistent acquisition validity identifiers, merging into a local defect-free background area. If there are insufficient valid background records within the window, the most recently calibrated local background field of the current scan segment is used, and the window is marked as pending update. The background area will be updated after subsequent adjacent valid records are supplemented.

[0094] S212: For the original time series within the local defect-free background region, a temporal alignment algorithm and an averaging algorithm are applied to obtain a smoothed and denoised fused average sequence. This fused average sequence is determined as the reference surface response waveform reflecting the true surface reflection characteristics. Based on the reference surface response waveform, the reference leading edge slope, reference half-width at half-maximum (HWHM), reference main peak symmetry index, and reference tail decay time constant are calculated. These parameters are then combined to form a local feature baseline. The temporal alignment algorithm in this method refers to the processing rule that corrects the time starting point of each original time series within the local defect-free background region based on the position of the surface main echo leading edge. The input is the valid original time series and its depth-time axis coordinates; the output is an aligned sequence with the consistent leading edge position. The averaging algorithm refers to smoothing the amplitude distribution at the same depth time position in the aligned sequence, preserving the common components of local surface reflection and reducing random noise components.

[0095] The reference surface response waveform is a reference waveform data object formed by the joint response of surface echoes within a locally defect-free background area. It includes a leading edge segment, a main peak segment, a tail attenuation segment, and a validity indicator. The rising slope of the reference leading edge is the rate of change of the amplitude of the leading edge segment from the background state to the main peak state. The reference half-width at half-maximum (HWHM) value is the continuous width of the main peak near the HWHM in the half-peak state. The reference main peak symmetry index is a dimensionless field transformed from the local width relationship between the left and right sides of the main peak using a unified standard. The reference tail attenuation time constant is the attenuation field of the tail returning to the background state from the rear of the main peak. The local feature baseline is the background feature record after combining the above fields, stored in the baseline cache corresponding to the current sliding analysis window, and called by the subsequent anomaly identification process. When the validity of the reference surface response waveform does not meet the requirements, the corresponding window only retains the original time series and spatial coordinate labels, does not output the suspected anomaly screening results, and triggers re-acquisition or compensation reading from adjacent windows.

[0096] S3: Based on the waveform energy integral values ​​of the original time series within the sliding analysis window, a set of anomalous waveform feature nodes is selected, an anomalous morphological feature parameter set is extracted, and the absolute deviation between the anomalous morphological feature parameter set and the local feature baseline is calculated. Suspected composite anomalous time series are then selected based on the absolute deviation values. The waveform energy integral value is the energy field formed by the cumulative amplitude response of the same original time series within the effective depth time axis range. The field source is the effective original time series in the three-dimensional data matrix. The set of anomalous waveform feature nodes consists of node data objects extracted from original time series whose energy exceeds the background intensity category. Internally, it records the leading edge start point, first peak point, trough point, second peak point, tail zeroing point, and their corresponding amplitude states.

[0097] S311: Calculate the waveform energy integral values ​​associated with each original time series within the sliding analysis window. Based on the waveform energy integral values, filter the original time series within the sliding analysis window that exceed the background intensity corresponding to the local defect-free background region. Extract the leading edge start point, first peak point, trough point, second peak point, and tail zeroing point within the original time series that exceed the background intensity corresponding to the local defect-free background region. Extract the first peak amplitude value associated with the first peak point and the initial amplitude value associated with the leading edge start point to establish an abnormal waveform feature node set. The background intensity is a background intensity field formed by the fusion of the average sequence and the peak intensity distribution in the local defect-free background region, stored in the local baseline cache. During filtering, first verify the acquisition validity identifier of the original time series, then compare its waveform energy state with the judgment category corresponding to the background intensity field. Only sequences that enter the abnormal candidate category are extracted.

[0098] The leading edge start point is the record of the position where the echo amplitude leaves the background stationary segment and enters the rising segment. The first peak point is the position of the main peak amplitude formed after the rising segment. The trough point is the local low position where the amplitude falls back after the first peak point. The second peak point is the position where the peak response re-enters after the trough point. The tail zeroing point is the position where the amplitude of the later segment returns to the background stationary category. The amplitude values ​​of the first peak and the initial amplitude values ​​are derived from the amplitude fields corresponding to the first peak point and the leading edge start point, respectively. If the main peak and the subsequent peak response cannot be distinguished during node extraction, the sequence enters the verification state and retains the energy anomaly mark; if the depth time axis is broken, the node set is written with an incomplete mark, and subsequent steps will not treat it as a suspected composite anomalous time series.

[0099] S312: Based on the amplitude values ​​of the first peak, initial amplitude, leading edge start point, first peak point, second peak point, and tail zeroing point within the abnormal waveform feature node set, calculate the amplitude deviation between the amplitude values ​​of the first peak and the initial amplitude, calculate the time span difference between the first peak point and the leading edge start point, calculate the ratio between the amplitude deviation and the time span difference to obtain the rising slope of the abnormal leading edge, calculate the half-wave height duration corresponding to the first and second peak points to obtain the abnormal half-width value, extract the second local half-width value associated with the second peak point and the first local half-width value associated with the first peak point, calculate the ratio between the second local half-width value and the first local half-width value to obtain the abnormal main peak symmetry index, calculate the time span between the tail zeroing point and the second peak point to obtain the abnormal tail decay time constant, and combine the abnormal leading edge rising slope, abnormal half-width value, abnormal main peak symmetry index, and abnormal tail decay time constant into an abnormal morphological feature parameter set. The rising slope of the abnormal front is used to record the change characteristics of the abnormal sequence's leading segment from the background state to the peak state. The abnormal half-width value is used to record the degree of broadening of the composite peak shape near the half-peak state. The abnormal main peak symmetry index is used to record the proportional relationship between the local widths of the first and second peaks. The abnormal tail decay time constant is used to record the delay characteristics of the latter segment returning to the background state.

[0100] The anomaly morphological feature parameter set is a combination of morphological fields established for the same anomaly candidate sequence. Each field contains a source node index and a coordinate label index. When duplicate identifications exist for the first peak point, the second peak point, or the tail zeroing point, deduplication is performed according to the depth time axis order and amplitude state continuity. If node conflicts still exist after deduplication, a conflict flag is written into the corresponding anomaly morphological feature parameter set and passed to the subsequent composite anomaly screening stage as an exclusion criterion. After each morphological field is generated, it undergoes a unified caliber conversion, allowing fields with different physical meanings to participate in subsequent deviation judgments based on their level status in a pre-set rule table, avoiding direct comparisons across calibers.

[0101] S313: Extract the baseline rising slope, baseline half-width at half-maximum (HWHM), baseline main peak symmetry index, and baseline tail decay time constant from the local feature baseline, and the abnormal rising slope, abnormal HWHM, abnormal main peak symmetry index, and abnormal tail decay time constant from the abnormal morphology feature parameter set. Calculate the first absolute deviation between the abnormal rising slope and the baseline rising slope; the second absolute deviation between the abnormal HWHM and the baseline HWHM; the third absolute deviation between the abnormal main peak symmetry index and the baseline main peak symmetry index; and the fourth absolute deviation between the abnormal tail decay time constant and the baseline tail decay time constant. Based on these four absolute deviations, identify the original time series exhibiting asymmetric broadening patterns and generate suspected composite abnormal time series. The absolute deviation value represents the difference between the abnormal morphology field and the corresponding baseline field under a unified standard; subsequently, only its level status and source index are read.

[0102] The pre-set anomaly feature library is a set of rule-based data formed by the detection equipment during calibration, re-inspection, and historical detection records. It stores the discrete state of features of historical composite anomaly samples, the source field of feature weights, and the global statistical distortion mean field. The historical distribution discrete variance is used to represent the stability of each morphological feature in historical composite anomaly samples. The normalized reciprocal proportion is used to convert the stability of each feature into a deviation weight coefficient. Features with higher stability receive higher weight in the comprehensive distortion judgment. The first deviation weight coefficient, the second deviation weight coefficient, the third deviation weight coefficient, and the fourth deviation weight coefficient are stored as lookup table fields in the pre-set anomaly feature library. The update condition is the completion of equipment calibration or the completion of the historical sample library review. The broadened distortion judgment threshold is derived from the global statistical distortion mean field corresponding to the historical composite anomaly samples, serving as the judgment boundary for the comprehensive distortion feature index. Each absolute deviation value and its corresponding deviation weight coefficient are entered into the distortion rule table to form an independent weighted state. These independent weighted states are then summarized into a comprehensive distortion feature index. When the comprehensive distortion feature index exceeds the broadened distortion judgment threshold corresponding to a certain category, the matched original time series is marked as a suspected composite anomalous time series and enters the stripping process carrying 3D spatial coordinate labels, anomaly node indexes, and baseline source indexes. If the preset anomaly feature library has not yet completed the current batch update during the startup phase, the previous valid calibration record is read. If the previous valid calibration record is missing, only the energy anomaly state is output, and no suspected composite anomalous time series is generated.

[0103] S4: The local benchmark matching waveform stripping algorithm is invoked to match and strip the suspected composite anomaly time series with the benchmark surface response waveform, generating surface candidate component signals. Internal interference candidate component signals are then extracted based on the difference between the suspected composite anomaly time series and the surface candidate component signals. The local benchmark matching waveform stripping algorithm is a rule-based process that decomposes the suspected composite anomaly time series into surface response components and residual response components within the same sliding analysis window. Its inputs are the suspected composite anomaly time series, the benchmark surface response waveform, and the shared three-dimensional spatial coordinate labels. The outputs are surface candidate component signals, internal interference candidate component signals, and stripping status indicators.

[0104] S411: The local benchmark matching waveform stripping algorithm is invoked to map the suspected composite anomaly time series to the composite object to be decomposed, and the benchmark surface response waveform to the reference matching template. The time axis associated with the reference matching template is shifted, and the maximum cross-correlation value of the leading edge between the reference matching template and the composite object to be decomposed is calculated. The time domain alignment starting point is established based on the maximum cross-correlation value of the leading edge. The amplitude scaling factor inside the local benchmark matching waveform stripping algorithm is extracted, and the response amplitude associated with the reference matching template after alignment based on the time domain alignment starting point is adjusted according to the amplitude scaling factor to generate a matching fitted waveform. The composite object to be decomposed is an anomaly echo sequence with an asymmetric broadened shape and spatial coordinate labels. The reference matching template is the benchmark surface response waveform generated in the local defect-free background area of ​​the current window. The maximum cross-correlation value of the leading edge is the leading edge segment shape matching status field, which comes from the similarity arrangement relationship between the leading edge segment of the reference matching template after translation and the leading edge segment of the composite object to be decomposed, and is used to determine the alignment position of their time starting points.

[0105] The amplitude scaling factor is an adjustment field for the template response amplitude. Its boundary is derived from the extreme range of amplitude changes in the historical statistical records of the detection equipment, and the field is stored in the stripping parameter cache. After the time-domain alignment start point is determined, the starting position of the time axis of the reference matching template is adjusted to match the leading edge segment of the composite object to be disassembled. Then, the template amplitude is adjusted according to the amplitude scaling factor to form a matching fitting waveform. If there is trigger drift or abnormal node conflict in the leading edge segment during the matching process, the stripping status is recorded as a category to be reviewed, and the corresponding spatial coordinate label is passed to the result output stage to save the uncertain state.

[0106] S412: Extract the test signal amplitude from the suspected composite anomaly time series and the template signal amplitude from the matched waveform. Calculate the mean square error (MSE) between the test signal amplitude and the template signal amplitude. Based on the constraint of minimizing the MSE, select the target amplitude scaling factor and the target time offset. Adjust the time starting point associated with the reference surface response waveform based on the target time offset. Adjust the amplitude of the reference surface response waveform based on the target amplitude scaling factor to establish the surface candidate component signal. The test signal amplitude comes from the suspected composite anomaly time series, and the template signal amplitude comes from the matched waveform of the reference matching template after time shifting and amplitude adjustment. The MSE value is a field representing the closeness of the residuals at the same depth and time position, used to evaluate whether the template can represent the surface candidate components.

[0107] The target amplitude scaling factor and target time offset are filtered using a two-dimensional parameter search grid. The amplitude scaling search boundary is derived from the extreme range of amplitude changes within the historical statistical records of the detection equipment; the time shift search boundary is derived from the waveform pulse duration field; the time offset search step size is derived from the inherent time sampling resolution within the hardware parameters of the multi-axis scanning mechanism; and the amplitude scaling search step size is derived from the historical fluctuation standard deviation field of the probe signal energy. All of these fields are stored in the stripping parameter cache and updated after equipment calibration or probe replacement. The candidate amplitude scaling factor sequence and candidate time offset sequence are orthogonally combined to form a two-dimensional parameter search grid. Each test parameter combination generates a corresponding matching waveform and is written into the error mapping matrix. The test parameter combination corresponding to the error state at the global lower bound in the error mapping matrix is ​​determined as the target parameter combination, and the candidate amplitude scaling factor contained therein is determined as the target amplitude scaling factor, and the candidate time offset is determined as the target time offset. The surface candidate component signal is a surface component sequence formed by adjusting the reference surface response waveform with the target time offset and target amplitude scaling factor, carrying a stripping state identifier and spatial coordinate label for subsequent energy proportion determination.

[0108] S413: Extract the current signal amplitude associated with the suspected composite anomaly time series and the base signal amplitude associated with the surface candidate component signal at the same time. Calculate the difference between the current signal amplitude and the base signal amplitude to obtain the amplitude deviation value. Based on the depth-time axis coordinates, concatenate the amplitude deviation values ​​matched at each time point to establish a residual signal sequence. Perform calibration processing on the residual signal sequence corresponding to the associated internal impurity response characteristics to generate internal interference candidate component signals. The current signal amplitude is the original response field of the suspected composite anomaly time series on the depth-time axis, the base signal amplitude is the reference response field of the surface candidate component signal at the same depth-time position, and the amplitude deviation value records the remaining responses in the composite anomaly that were not explained by the surface candidate components.

[0109] The residual signal sequence is spliced ​​together in its original order along the depth-time axis and calibrated against an internal impurity response characteristic rule table. This table is derived from the material's internal scattering response, historical re-inspection records, and equipment calibration records, internally recording residual peak shape, tail state, depth-time position continuity, and surface template matching exclusion status. When the residual signal sequence satisfies an internal interference category, candidate internal interference components are output; when the residual signal sequence contains only noise, noise residual status is output; when the residual signal sequence cannot match an internal interference category, an unclassified residual status is output, and the original sequence index is retained for re-inspection.

[0110] S5: Based on the surface candidate component signals and internal interference candidate component signals, the surface component energy ratio is obtained, and the surface microcrack attributes are determined by combining the three-dimensional spatial coordinate labels, generating the surface micro-defect detection results of the semiconductor quartz tube. The surface component energy ratio is the state field occupied by the surface candidate component in the joint response of the surface candidate component and the internal interference candidate component, and its source is the time-domain integration result of the surface candidate component signal and the internal interference candidate component signal. The surface microcrack attributes are defect attribute fields formed by combining the energy ratio state and spatial topology, used to distinguish between surface linear cracks, planar cracks and non-surface interference states.

[0111] S511: Perform time-domain integration on the signal amplitude of the surface candidate component signal to obtain the surface component energy integral value, and perform time-domain integration on the signal amplitude of the internal interference candidate component signal to obtain the interference component energy integral value. Calculate the sum of the surface component energy integral value and the interference component energy integral value to obtain the total signal energy value. Calculate the ratio between the surface component energy integral value and the total signal energy value to generate the surface component energy proportion. The surface component energy integral value is the energy field formed by the surface candidate component signal within the effective depth time axis range, and the interference component energy integral value is the energy field formed by the internal interference candidate component signal under the same depth time axis caliber. The total signal energy value is used to unify the comparison caliber between the surface component and the internal interference component. The surface component energy proportion is stored as a dimensionless state field in the result cache, along with a stripping state identifier, residual classification state, and spatial coordinate label index.

[0112] When the internal interference candidate component signal is in a noise residual state, the interference component energy field is entered into the total signal energy field according to the noise calibration rules, and the result cache is synchronously written to the noise residual source. If the surface candidate component signal or the internal interference candidate component signal has a verification mark, the surface component energy ratio is written to an uncertain state. Subsequent topology judgment still reads the spatial coordinate label and the peak position of the candidate component, but the verification state is synchronously retained in the detection result.

[0113] S512: The surface component energy ratio is numerically compared with a preset weak signal limit. When the surface component energy ratio exceeds the preset weak signal limit, the surface candidate component signals within the adjacent scanning grid associated with the semiconductor quartz tube are extracted based on the three-dimensional spatial coordinate labels. The axial coordinates and circumferential angle coordinates of the scanning probe within the three-dimensional spatial coordinate labels are extracted. The geometric topological structure features between the surface candidate component signals within the current test grid and those within the adjacent scanning grids are compared according to the scanning probe axial coordinates and circumferential angle coordinates. If the geometric topological structure features are linear or planar, the attribute state of the current test grid is established, and the surface microcrack attributes are established. The preset weak signal limit is a configurable rule field, derived from equipment calibration records, probe sensitivity records, and semiconductor quartz tube material background response records. It is stored in the weak signal judgment rule table and refreshed after equipment maintenance, probe replacement, or material batch rule updates.

[0114] The geometric topology feature is a field representing the connectivity of surface candidate component signals between the current test grid and adjacent scan grids in the axial and circumferential directions. When generating this field, the axial and circumferential angular coordinates of the current test grid are first read based on the 3D spatial coordinate labels. Then, the peak positions, continuity states, and stripping states of the surface candidate component signals within adjacent scan grids are read. Subsequently, topology classification is performed according to axial continuity, circumferential continuity, and the homogeneity of candidate components. When a continuous distribution maintains connectivity along a single extension direction, the topology is classified as a linear topology; when a continuous distribution covers adjacent axial and circumferential angular positions to form sheet-like connectivity, the topology is classified as a planar topology. When the surface component energy ratio does not exceed a preset weak signal limit, the current test grid outputs a weak surface response state; when the topology cannot establish connectivity with adjacent scan grids, an isolated candidate state is output, and the 3D spatial coordinate labels are retained for subsequent re-examination.

[0115] S513: Extract the peak intensity values ​​from the candidate surface component signals associated with surface microcrack attributes, obtain the ultrasonic signal transmission velocity matched to the semiconductor quartz tube, extract the depth-time axis coordinates from the 3D spatial coordinate labels, calculate the product between the ultrasonic signal transmission velocity and the depth-time axis coordinates to obtain the equivalent depth value, bind the peak intensity values, equivalent depth values, and surface microcrack attributes to construct a 3D defect feature description set containing defect intensity, defect depth, and topological attributes, and output the 3D defect feature description set as the surface micro-defect detection result of the semiconductor quartz tube. The peak intensity value is the main peak amplitude field in the candidate surface component signals, used to form the defect intensity field; the ultrasonic signal transmission velocity is a lookupable field matching the semiconductor quartz tube material, probe coupling state, and equipment calibration records, stored in the material acoustic parameter table; the depth-time axis coordinates are derived from the 3D spatial coordinate labels. The equivalent depth value is the defect depth field formed after matching the depth-time axis coordinates with the material acoustic parameter table, and the field is recalibrated as the material acoustic parameter table is updated.

[0116] The 3D defect feature description set is the detection result data object, containing fields for defect intensity, defect depth, topology attribute, axial coordinate, circumferential angle coordinate, peeling status, residual classification, and verification status. Before outputting the results, it is first verified whether each field has the same 3D spatial coordinate label source; if the field sources are inconsistent, the result cache is written with a coordinate conflict status and defect attribute confirmation is stopped. When the field sources are consistent and the topology attribute is valid, the 3D defect feature description set is entered into the detection result storage area and a traceability relationship is established with the original time series index, the reference surface response waveform index, and the peeling parameter cache index. The detection results are used to mark the location, intensity status, depth status, and topology status of micro-defects on the surface of semiconductor quartz tubes, and provide the same data structure output for subsequent manual re-inspection or automatic sorting interfaces.

[0117] The semiconductor quartz tube surface micro-defect detection system in the same embodiment is used to implement the aforementioned method. The system includes a scanning coordination module, a background calibration module, an anomaly identification module, a composition separation module, and a crack assessment module. The scanning coordination module receives continuous analog signals, axial position feedback, and circumferential angle position feedback, and outputs the original time series, three-dimensional spatial coordinate labels, and a three-dimensional data matrix, corresponding to the data acquisition, label combination, and matrix stitching in S1. The background calibration module reads the three-dimensional data matrix and valid acquisition identifiers, and outputs the local defect-free background area, the reference surface response waveform, and the local feature baseline, corresponding to the window delineation, background filtering, alignment fusion, and baseline construction in S2. The anomaly identification module reads the local feature baseline and the original time series within the sliding analysis window, and outputs the anomaly waveform feature node set, the anomaly morphology feature parameter set, and the suspected composite anomaly time series, corresponding to the energy filtering, node extraction, deviation judgment, and composite anomaly confirmation in S3. The composition separation module reads the suspected composite anomaly time series and the reference surface response waveform, and outputs the surface candidate component signal, the internal interference candidate component signal, and the peeling state identifier, corresponding to the template matching, parameter search, surface composition establishment, and residual calibration in S4. The crack assessment module reads surface candidate component signals, internal interference candidate component signals, three-dimensional spatial coordinate labels, and material acoustic parameter tables. It outputs the surface component energy ratio, surface microcrack attributes, and a three-dimensional defect feature description set, corresponding to the energy ratio judgment, topology attribute confirmation, and detection result construction in S5. Data transmitted between modules carries acquisition validity identifiers, coordinate validity identifiers, and traceability indexes. When an interface returns an exception, cache status expires, or field validation fails, the current data object enters a review state and returns to the corresponding processing stage along the original call chain.

[0118] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting micro-defects on the surface of a semiconductor quartz tube, characterized in that, Includes the following steps: S1: During the semiconductor quartz tube scanning stage, continuous analog signals are acquired and depth time axis coordinates are extracted to generate an original time series. Simultaneously, the axial coordinates and circumferential angle coordinates of the scanning probe are acquired. The depth time axis coordinates are combined with the axial coordinates and circumferential angle coordinates of the scanning probe to generate three-dimensional spatial coordinate labels. Based on the three-dimensional spatial coordinate labels, the original time series is subjected to association and splicing processing to generate a three-dimensional data matrix. S2: Delineate a sliding analysis window based on the three-dimensional data matrix, filter out local defect-free background regions, perform fusion extraction on the original time series within the local defect-free background regions to generate a reference surface response waveform, and calculate and construct a local feature baseline based on the feature parameters of the reference surface response waveform. S3: Based on the waveform energy integral value of the original time series within the sliding analysis window, filter the set of abnormal waveform feature nodes, extract the set of abnormal morphological feature parameters, calculate the absolute deviation value between the set of abnormal morphological feature parameters and the local feature baseline, and filter the suspected composite abnormal time series based on the absolute deviation value. S4: Call the local reference matching waveform stripping algorithm to match and strip the suspected composite anomaly time series with the reference surface response waveform to generate surface candidate component signals, and extract internal interference candidate component signals based on the difference between the suspected composite anomaly time series and the surface candidate component signals. S5: Obtain the surface component energy ratio based on the surface candidate component signal and the internal interference candidate component signal, and determine the surface microcrack properties in combination with the three-dimensional spatial coordinate label to generate the surface microdefect detection result of the semiconductor quartz tube.

2. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 1, characterized in that, The specific steps of S1 are as follows: S111: Monitor the operating status of the multi-axis scanning mechanism during the semiconductor quartz tube scanning stage, collect the continuous analog signal corresponding to the reflected echo from the inner wall of the semiconductor quartz tube, extract the signal distribution points and depth time axis coordinates inside the continuous analog signal, and generate the original time series. S112: Collect the axial coordinates and circumferential angle coordinates of the scanning probe, and combine the depth time axis coordinates, the axial coordinates and the circumferential angle coordinates of the scanning probe attached to the original time series to generate a three-dimensional spatial coordinate label; S113: Map the three-dimensional spatial coordinate labels to the corresponding original time series, extract the three-dimensional spatial distribution coordinate features inside the three-dimensional spatial coordinate labels, and perform association splicing processing on the original time series with matching label states based on the three-dimensional spatial distribution coordinate features to generate a three-dimensional data matrix.

3. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 2, characterized in that, The specific steps of S2 are as follows: S211: Define a sliding analysis window for the three-dimensional data matrix, extract the peak intensity distribution values ​​associated with each of the original time series within the sliding analysis window, calculate the discrete fluctuation amplitude values ​​based on the peak intensity distribution values ​​corresponding to adjacent spatial positions within the three-dimensional data matrix, filter the analysis range of the semiconductor quartz tube based on the discrete fluctuation amplitude values, and establish a local defect-free background region. S212: Perform a temporal alignment algorithm and an averaging algorithm on the original time series within the local defect-free background region to obtain a smoothed and denoised fused average sequence. Determine the fused average sequence as the reference surface response waveform that reflects the true surface reflection characteristics. Calculate the reference leading edge rise slope, reference half-width at half-maximum (WHM), reference main peak symmetry index, and reference tail decay time constant based on the reference surface response waveform. Combine the reference leading edge rise slope, the reference WHM, the reference main peak symmetry index, and the reference tail decay time constant into a local feature baseline.

4. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 3, characterized in that, The process of calculating discrete fluctuation amplitude values, selecting the analysis range of the semiconductor quartz tube based on these discrete fluctuation amplitude values, and establishing a local defect-free background region is as follows: Calculate the local mathematical mean of the wave crest intensity distribution values ​​corresponding to the adjacent spatial locations; Calculate the numerical sequence of absolute deviations between the wave crest intensity distribution values ​​and the local mathematical mean; Extract the maximum absolute deviation value within the absolute deviation value sequence; The maximum absolute deviation value is determined as the discrete fluctuation amplitude value; Extract the global distribution values ​​of peak intensity from the three-dimensional data matrix in the global scope; Calculate the global fluctuation variance of the global distribution value of the peak intensity; Extract the inherent background noise parameters corresponding to the hardware of the detection equipment; Calculate the product between the global fluctuation variance and the inherent background noise parameter; The product of the global fluctuation variance and the inherent background noise parameter is set as the fluctuation amplitude limit threshold. The discrete fluctuation amplitude value is compared with the fluctuation amplitude limit threshold value. If the discrete fluctuation amplitude value does not exceed the fluctuation amplitude limit threshold, extract the adjacent spatial positions that match the discrete fluctuation amplitude value. The adjacent spatial locations that match the discrete fluctuation amplitude values ​​are defined as the analysis range of the semiconductor quartz tube. By merging adjacent spatial locations within the analysis range, the local defect-free background region is established.

5. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 3, characterized in that, The specific steps of S3 are as follows: S311: Calculate the waveform energy integral values ​​associated with each of the original time series within the sliding analysis window. Based on the waveform energy integral values, filter the original time series within the sliding analysis window that exceed the background intensity corresponding to the local defect-free background region. Extract the leading edge start point, first peak point, trough point, second peak point, and tail zeroing point within the original time series that exceed the background intensity corresponding to the local defect-free background region. Extract the first peak amplitude value associated with the first peak point and the initial amplitude value associated with the leading edge start point to establish an abnormal waveform feature node set. S312: Based on the first peak amplitude value, the initial amplitude value, the leading edge start point, the first peak point, the second peak point, and the tail zeroing point within the abnormal waveform feature node set, calculate the amplitude deviation value between the first peak amplitude value and the initial amplitude value, calculate the time span difference between the first peak point and the leading edge start point, calculate the ratio between the amplitude deviation value and the time span difference to obtain the abnormal leading edge rising slope, calculate the half-wave height duration corresponding to the first peak point and the second peak point to obtain the abnormal half-width value, extract the second local half-width value associated with the second peak point and the first local half-width value associated with the first peak point, calculate the ratio between the second local half-width value and the first local half-width value to obtain the abnormal main peak symmetry index, calculate the time span value between the tail zeroing point and the second peak point to obtain the abnormal tail decay time constant, and combine the abnormal leading edge rising slope, the abnormal half-width value, the abnormal main peak symmetry index, and the abnormal tail decay time constant into an abnormal morphological feature parameter set; S313: Extract the baseline rising slope, baseline half-width, baseline main peak symmetry index, and baseline tail decay time constant from the local feature baseline, and the abnormal rising slope, abnormal half-width, abnormal main peak symmetry index, and abnormal tail decay time constant from the abnormal morphology feature parameter set. Calculate the first absolute deviation between the abnormal rising slope and the baseline rising slope, the second absolute deviation between the abnormal half-width and the baseline half-width, the third absolute deviation between the abnormal main peak symmetry index and the baseline main peak symmetry index, and the fourth absolute deviation between the abnormal tail decay time constant and the baseline tail decay time constant. Based on the first, second, third, and fourth absolute deviations, identify and filter the original time series exhibiting asymmetric broadening patterns to generate a suspected composite abnormal time series.

6. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 5, characterized in that, The process of determining and filtering the original time series that exhibits an asymmetric broadening pattern based on the first absolute deviation value, the second absolute deviation value, the third absolute deviation value, and the fourth absolute deviation value, and generating a suspected composite abnormal time series, specifically involves: Extract the historical distribution discrete variance of multiple features within a pre-set anomaly feature library; Calculate the normalized reciprocal proportion of the discrete variance of the historical distribution; The normalized reciprocal ratio is set sequentially as the first deviation weight coefficient, the second deviation weight coefficient, the third deviation weight coefficient, and the fourth deviation weight coefficient according to the corresponding feature terms; Extract the global statistical distortion mean corresponding to historical composite anomaly samples within the pre-set anomaly feature library; Set the global statistical distortion mean as the broadening distortion determination threshold; The first independent weighted value is obtained by multiplying the first absolute deviation value by the first deviation weighting coefficient. The second independent weighted value is obtained by multiplying the second absolute deviation value by the second deviation weighting coefficient. The third independent weighted value is obtained by multiplying the third absolute deviation value by the third deviation weighting coefficient. The fourth independent weighted value is obtained by multiplying the fourth absolute deviation value by the fourth deviation weighting coefficient. The comprehensive distortion feature index is obtained by summing the first independent weighted value, the second independent weighted value, the third independent weighted value, and the fourth independent weighted value. When the comprehensive distortion feature index is greater than the broadening distortion judgment threshold, the corresponding original time series is extracted to generate a suspected composite abnormal time series.

7. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 5, characterized in that, The specific steps of S4 are as follows: S411: Invoke the local benchmark matching waveform stripping algorithm to map the suspected composite anomaly time series to the composite object to be decomposed, map the benchmark surface response waveform to the reference matching template, shift the time axis associated with the reference matching template, calculate the maximum cross-correlation value of the leading edge between the reference matching template and the composite object to be decomposed, establish the time domain alignment starting point based on the maximum cross-correlation value of the leading edge, extract the amplitude scaling factor inside the local benchmark matching waveform stripping algorithm, adjust the response amplitude associated with the reference matching template after alignment based on the time domain alignment starting point based on the amplitude scaling factor, and generate a matching fitting waveform; S412: Extract the test signal amplitude within the suspected composite anomaly time series and the template signal amplitude within the matched fitting waveform, calculate the mean square error between the test signal amplitude and the template signal amplitude, filter the target amplitude scaling factor and target time offset based on the condition constraint corresponding to minimizing the mean square error, adjust the time starting point associated with the reference surface response waveform based on the target time offset, adjust the amplitude corresponding to the reference surface response waveform based on the target amplitude scaling factor, and establish surface candidate component signals; S413: Extract the current signal amplitude associated with the suspected composite anomaly time series and the base signal amplitude associated with the surface candidate component signal at the same time. Calculate the difference between the current signal amplitude and the base signal amplitude to obtain the amplitude deviation value. Based on the depth time axis coordinate, splice the amplitude deviation values ​​matched at multiple times to establish a residual signal sequence. Perform calibration processing on the residual signal sequence corresponding to the associated internal impurity response characteristics to generate internal interference candidate component signals.

8. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 7, characterized in that, The process of selecting the target amplitude scaling factor and the target time offset based on the condition constraint corresponding to minimizing the mean square error value is as follows: Extract the extreme range of amplitude changes and the duration of waveform pulses from the historical statistical records of the detection equipment; Set the extreme value range of the amplitude change as the amplitude scaling search boundary; Set the duration of the waveform pulse as the time shift search boundary; Extract the inherent time sampling resolution within the hardware parameters of the multi-axis scanning mechanism; Set the inherent time sampling resolution to the time offset search step size; Extract the historical standard deviation of the energy fluctuation of the probe signal; The standard deviation of the historical fluctuation of the detected signal energy is set as the amplitude scaling search step size; A candidate amplitude scaling coefficient sequence is generated based on the amplitude scaling search boundary and the amplitude scaling search step size; A candidate time offset sequence is generated based on the time translation search boundary and the time offset search step size; A two-dimensional parameter search grid is constructed by orthogonally combining the candidate amplitude scaling factor sequence and the candidate time offset sequence; Within the two-dimensional parameter search grid, test parameter combinations, including individual candidate amplitude scaling factors and individual candidate time offsets, are extracted one by one. For each combination of test parameters, the mean square error value under the parameter application state is extracted, and an error mapping matrix is ​​constructed according to the mean square error value associated with the grid coordinate arrangement. Numerical extremum retrieval is performed within the error mapping matrix to extract the minimum mean square error value located at the global lower bound. Extract the test parameter combination that matches the minimum mean square error value; The single candidate amplitude scaling factor contained within the test parameter combination that matches the minimum mean square error value is established as the target amplitude scaling factor; The single candidate time offset contained within the test parameter combination that matches the minimum mean square error value is established as the target time offset.

9. The method for detecting micro-defects on the surface of a semiconductor quartz tube according to claim 7, characterized in that, The specific steps of S5 are as follows: S511: Perform time-domain integration on the signal amplitude of the surface candidate component signal to obtain the surface component energy integral value, and perform time-domain integration on the signal amplitude of the internal interference candidate component signal to obtain the interference component energy integral value. Calculate the sum of the surface component energy integral value and the interference component energy integral value to obtain the total signal energy value. Calculate the ratio between the surface component energy integral value and the total signal energy value to generate the surface component energy proportion. S512: The surface component energy ratio is compared with a preset weak signal limit. When the surface component energy ratio exceeds the preset weak signal limit, the surface candidate component signal is extracted from the adjacent scanning grid associated with the semiconductor quartz tube according to the three-dimensional spatial coordinate label. The scanning probe axial coordinate and scanning probe circumferential angle coordinate are extracted from the three-dimensional spatial coordinate label. The geometric topological structure features between the surface candidate component signal in the current test grid and the surface candidate component signal in the adjacent scanning grid are compared according to the scanning probe axial coordinate and scanning probe circumferential angle coordinate. When the geometric topological structure features are linear or planar, the attribute state of the current test grid is established, and the surface microcrack attribute is established. S513: Extract the peak intensity value inside the candidate component signal of the surface associated with the surface microcrack attribute, obtain the ultrasonic signal transmission speed matched by the semiconductor quartz tube, extract the depth time axis coordinate inside the three-dimensional spatial coordinate label, calculate the product between the ultrasonic signal transmission speed and the depth time axis coordinate to obtain the equivalent depth value, bind the peak intensity value, the equivalent depth value and the surface microcrack attribute to construct a three-dimensional defect feature description set including defect intensity, defect depth and topological attributes, and output the three-dimensional defect feature description set as the surface micro-defect detection result of the semiconductor quartz tube.

10. A system for detecting micro-defects on the surface of a semiconductor quartz tube, characterized in that, The system is used to implement the method for detecting micro-defects on the surface of a semiconductor quartz tube as described in any one of claims 1-9, the system comprising: The scanning coordination module acquires continuous analog signals and extracts depth time axis coordinates during the semiconductor quartz tube scanning stage, generates an original time series, and simultaneously acquires the axial coordinates and circumferential angle coordinates of the scanning probe. It then combines the depth time axis coordinates with the axial and circumferential angle coordinates of the scanning probe to generate three-dimensional spatial coordinate labels. Based on the three-dimensional spatial coordinate labels, it performs correlation and splicing processing on the original time series to generate a three-dimensional data matrix. The background calibration module defines a sliding analysis window based on the three-dimensional data matrix, filters out local defect-free background regions, performs fusion extraction on the original time series within the local defect-free background regions to generate a reference surface response waveform, and calculates and constructs a local feature baseline based on the feature parameters of the reference surface response waveform. The anomaly identification module filters the set of abnormal waveform feature nodes based on the waveform energy integral value of the original time series within the sliding analysis window, extracts the set of abnormal morphological feature parameters, calculates the absolute deviation value between the set of abnormal morphological feature parameters and the local feature baseline, and filters suspected composite abnormal time series based on the absolute deviation value. The component separation module calls the local reference matching waveform stripping algorithm to match and strip the suspected composite anomaly time series with the reference surface response waveform to generate surface candidate component signals, and extracts internal interference candidate component signals based on the difference between the suspected composite anomaly time series and the surface candidate component signals. The crack assessment module obtains the surface component energy ratio based on the surface candidate component signal and the internal interference candidate component signal, and determines the surface microcrack attributes by combining the three-dimensional spatial coordinate labels, thereby generating the surface microdefect detection results of the semiconductor quartz tube.

Images