Real-time identification method for PCB drilling defects based on acoustic emission signals

Abstract

The present application relates to a PCB drilling defect real-time identification method based on acoustic emission signals, aiming to realize early automatic warning of defect states such as micro-crack initiation and hole wall tearing. The scheme continuously collects and slides to intercept acoustic emission sensor signal fragments at high frequency, adopts multi-scale envelope extraction, topological dynamic similarity calculation and graph convolution aggregation to obtain local timing subgraphs describing the micro-damage evolution path of the drilling hole, and uses the generated feature vector to distinguish the evolution stages of the drilling process. Combined with defect state sequence and specific transition mode identification, the evolution trend from normal drilling to tearing acceleration can be accurately captured, and warning signals and feature heat maps can be output in time. This method improves the drilling defect monitoring sensitivity and the accuracy of evolution stage identification, and helps to ensure process quality and operation safety.

Description

Technical Field

[0001] This invention relates to the field of acoustic emission signal analysis and intelligent monitoring technology of electronic manufacturing processes, and in particular to a real-time identification method for PCB drilling defects based on acoustic emission signals. Background Technology

[0002] Current mainstream technologies for PCB drilling defect identification generally employ static segment analysis of voltage signals acquired by acoustic emission sensors. These analyses typically utilize methods such as traditional time-series window truncation, frequency domain envelope extraction, feature statistics, or deep learning-based time-series modeling to determine abnormal states during the drilling process. Mainstream approaches rely heavily on long time series or static signal segments. Typical applications include threshold determination based on parameters such as acoustic emission energy thresholds, envelope curves, and event frequencies, as well as dynamic monitoring of defect states using time-series models (such as LSTM / RNN structures). Most publicly available industry technologies simplify the PCB drilling process signal into a one-dimensional time-series sequence or static feature vector, identifying defect stages through empirical models or data-driven learning, focusing on improving classification accuracy and anomaly detection sensitivity. Some advanced technologies attempt to integrate multimodal sensor data, change point detection, fractal dimension, attention strategies, or knowledge graph reasoning to enrich the scenarios and data associations for defect identification, but time-series modeling remains the core.

[0003] Existing technologies have significant shortcomings. First, static signal segment analysis or long-time-window stacking methods cannot effectively capture the dynamic transformation and evolution trends of defects during the drilling process, making it difficult to identify early defect initiation stages with high sensitivity. Second, long-time-window statistics and depth-time-series models require a large amount of historical data input, causing online response delays and failing to meet the real-time monitoring requirements of high-speed drilling processes. Third, existing classification algorithms mostly rely on global signal features and cannot locate local evolution paths for single signal segments, thus lacking foresight and having weak defect early warning capabilities. Furthermore, current technologies generally use single discriminant models, lacking structured historical memory and dynamic trend modeling of defect state evolution stages, resulting in classification results that can only remain at the static state discrimination level and cannot assist operators in making process decisions. With the increasing demand for real-time intelligent monitoring in the electronics manufacturing industry, existing acoustic emission signal defect identification schemes lack the ability to understand the dynamic initiation, expansion, and critical transformation processes of defects. Online systems can only rely on static signal segments or global model predictions, leading to delayed early warnings and untimely responses to critical defects, thus restricting the level of intelligent quality control and equipment maintenance. Summary of the Invention

[0004] This application provides a real-time identification method for PCB drilling defects based on acoustic emission signals, aiming to solve one of the problems or issues of the prior art mentioned in the background section above.

[0005] The real-time PCB drilling defect identification method based on acoustic emission signals provided in this application specifically includes: S1: During the PCB drilling process, the raw voltage signal output by the acoustic emission sensor is continuously acquired at a fixed sampling rate, and the raw voltage signal is processed by sliding slicing according to a preset ultra-short time window to generate a high-density acoustic emission signal segment stream.

[0006] S2: Perform adaptive amplitude normalization and multi-scale envelope extraction processing on each signal segment in the high-density acoustic emission signal segment stream to obtain a local response curve reflecting the transient energy distribution.

[0007] S3: Based on the number of adjacent local response curves, calculate the topological connection weight between each pair of curves according to the slope of amplitude change, the difference in zero crossing rate and the peak delay relationship, and construct a dynamic similarity matrix as the initial adjacency relationship data.

[0008] S4: Using the local response curve corresponding to the current signal segment as the root node, backtrack a preset number of historical local response curves, and construct a local time series subgraph focusing on short-range evolutionary correlation using the initial adjacency relationship data with introduced time decay factor.

[0009] S5: Perform graph convolution operation on the local temporal subgraph to aggregate neighbor node information and implicitly encode the relative position of the current segment in the defect development path as a topological evolution representation vector.

[0010] S6: Input the topological evolution representation vector into a lightweight classifier for stage discrimination processing, and output the probability distribution of defect states representing four types of states: normal drilling, microcrack initiation, accelerated hole wall tearing, or abnormal drill bit wear.

[0011] S7: Update the stage transition memory of the rolling stage based on the probability distribution of the defect states generated multiple times in succession, and record the real-time state sequence composed of different defect states.

[0012] S8: Determine whether the real-time state sequence contains a specific state transition pattern. If the specific state transition pattern is satisfied, trigger an early warning signal and simultaneously push the characteristic heat map of the corresponding stage.

[0013] The real-time PCB drilling defect identification method based on acoustic emission signals provided in this application has the following advantages: (1) By constructing a dynamic modeling mechanism based on ultra-short time window sliding slices and local time series subgraphs, the problem of high response delay and poor real-time performance caused by relying on long sequence cumulative analysis in traditional PCB drilling status monitoring is effectively overcome. Existing technologies generally use time series networks such as LSTM and RNN for global modeling, which requires waiting for the input of complete signal segments before inference can be performed, making it difficult to meet the millisecond-level industrial control requirements. However, this solution abandons the large-scale stacking of historical data and instead uses the current segment as the root node to backtrack a limited number of historical segments to construct a lightweight local time series subgraph. Combined with the time decay factor, it dynamically suppresses the weight of far-end connections and focuses on short-range evolution correlation, which significantly improves the response sensitivity of status recognition. On this basis, adaptive amplitude normalization and multi-scale envelope extraction are introduced to enhance the identifiability of weak defect signals in complex background noise, so that early abnormal features such as microcrack initiation and accelerated hole wall tearing can be stably captured within 0.3~0.5 seconds, greatly advancing the early warning time and effectively ensuring processing quality and equipment safety.

[0014] (2) A state discrimination architecture based entirely on the intrinsic structural evolution law of pure acoustic emission signals is proposed, which does not require the introduction of multimodal fusion, image assistance, or external physical parameter modeling, thus avoiding the defects of high system complexity, high deployment cost, and poor cross-platform compatibility in existing solutions. Traditional methods often rely on acoustic-visual joint perception, fractal dimension calculation, or knowledge graph reasoning to enhance judgment capabilities, which not only increases hardware configuration requirements but also makes the stability susceptible to environmental interference. This invention starts from the original voltage signal, constructs topological connection relationships through a dynamic similarity matrix, quantifies the evolution trend between segments using low-order statistical features such as amplitude change slope, zero-crossing rate difference, and peak delay, generates adjacency weights with clear physical meaning, and achieves implicit encoding of defect development paths through lightweight graph convolution aggregation. This approach significantly reduces algorithm complexity without sacrificing discrimination accuracy, and a single inference can be completed within hundreds of milliseconds. It supports embedded edge deployment and has good engineering practicality and scalability.

[0015] (3) A rolling update stage transition memory and patterned transition detection mechanism are designed to achieve a leap from isolated state recognition to dynamic behavior understanding, enhancing the system's ability to capture progressive fault evolution and improve decision reliability. Unlike static classification models that only output instantaneous state probability distributions, this scheme maintains the state sequence composed of continuous discrimination results online. When a stage transition pattern with typical precursor significance, such as "normal drilling → microcrack initiation → accelerated hole wall tearing," is detected, a high-level warning is triggered and the corresponding heat map is pushed simultaneously for manual review, effectively suppressing the false alarm rate and improving the reliability of the alarm. At the same time, the entire process does not use attention mechanism weighted aggregation, change point detection, or complex parameter tuning strategies. The model training and inference process is highly concise and has strong robustness and plug-and-play characteristics. Thus, a closed-loop, adaptive, and interpretable drilling state monitoring system is constructed, which is not only suitable for high-precision processing monitoring in PCB manufacturing scenarios, but can also be extended to other micro-destruction detection fields based on acoustic emission signals, with broad application prospects and industrial value.

[0016] In summary, this solution innovatively integrates local temporal subgraph modeling, dynamic topology aggregation, and stage transition identification mechanisms. Without relying on multimodal information fusion and complex deep network structures, it achieves millisecond-level accurate identification and early warning of key defect states during PCB drilling. This significantly improves the real-time performance, accuracy, and engineering applicability of the monitoring system, forming a unique technical advantage that distinguishes it from existing technical approaches. It possesses outstanding creativity and practical application value. Attached Figure Description

[0017] Figure 1 This is the main flowchart of a real-time PCB drilling defect identification method based on acoustic emission signals.

[0018] Figure 2 This is a sub-flowchart of a method for real-time identification of PCB drilling defects based on acoustic emission signals.

[0019] Figure 3 This is another sub-flowchart of the real-time identification method for PCB drilling defects based on acoustic emission signals. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0022] like Figure 1 As shown, this application provides a real-time identification method for PCB drilling defects based on acoustic emission signals, specifically including: S1: During the PCB drilling process, the raw voltage signal output by the acoustic emission sensor is continuously acquired at a fixed sampling rate, and the raw voltage signal is processed by sliding slicing according to a preset ultra-short time window to generate a high-density acoustic emission signal segment stream.

[0023] S2: Perform adaptive amplitude normalization and multi-scale envelope extraction processing on each signal segment in the high-density acoustic emission signal segment stream to obtain a local response curve reflecting the transient energy distribution.

[0024] S3: Based on the number of adjacent local response curves, calculate the topological connection weight between each pair of curves according to the amplitude change slope, zero-crossing rate difference and peak delay relationship, and construct a dynamic similarity matrix as the initial adjacency relationship data.

[0025] S4: Using the local response curve corresponding to the current signal segment as the root node, backtrack a preset number of historical local response curves, and construct a local time series subgraph focusing on short-range evolutionary correlation using the initial adjacency relationship data with introduced time decay factor.

[0026] S5: Perform graph convolution operation on the local temporal subgraph to aggregate neighbor node information and implicitly encode the relative position of the current segment in the defect development path as a topological evolution representation vector.

[0027] S6: Input the topological evolution representation vector into a lightweight classifier for stage discrimination processing, and output the probability distribution of defect states representing four types of states: normal drilling, microcrack initiation, accelerated hole wall tearing, or abnormal drill bit wear.

[0028] S7: Update the stage transition memory of the rolling stage based on the probability distribution of the defect states generated multiple times in succession, and record the real-time state sequence composed of different defect states.

[0029] S8: Determine whether the real-time state sequence contains a specific state transition pattern. If the specific state transition pattern is satisfied, trigger an early warning signal and simultaneously push the characteristic heat map of the corresponding stage.

[0030] Step S1: During the PCB drilling process, the raw voltage signal output by the acoustic emission sensor is continuously acquired at a fixed sampling rate, and the raw voltage signal is sliced ​​according to a preset ultra-short time window to generate a high-density acoustic emission signal segment stream. Specifically, this includes: S1.1: Perform high-frequency synchronous sampling processing on the continuous analog voltage signal output by the acoustic emission sensor during the PCB drilling process to obtain a discrete original voltage signal sequence with a unified time base.

[0031] When performing high-frequency synchronous sampling processing on the continuous analog voltage signal output by the acoustic emission sensor during PCB drilling, a sampling system with a unified time reference is established, using the sensor signal as input. The clock source of the sampling system undergoes temperature compensation and phase locking to ensure stable time intervals between sampling points. The signal is pre-filtered by an anti-aliasing filter to remove frequency components higher than the sampling Nyquist frequency, avoiding aliasing distortion. For the filtered analog signal, a high-precision analog-to-digital converter is used to quantize it according to a preset sampling frequency, converting the voltage value into discrete data points. A trigger synchronization mechanism is introduced during sampling to align the sampling trigger time with the drilling machine's operating status signal, eliminating timing offsets caused by mechanical phase differences. The sampling results are arranged in chronological order to form an original voltage signal sequence, with timestamps added to ensure that subsequent analysis accurately corresponds to the acquisition time of each data point. Through high-frequency synchronous sampling and time alignment processing, unstable analog signals are transformed into discrete original voltage signal sequences with a unified time reference, achieving the goals of real-time signal stream digitization and timing consistency.

[0032] For example, in high-speed PCB drilling operations, the analog voltage signal output by the acoustic emission sensor has a bandwidth of approximately 350kHz. The sampling system is configured with a sampling frequency of 1MHz, employing a 24-bit analog-to-digital converter, and using a temperature-compensated crystal oscillator as the clock source, with a frequency stability better than ±0.5ppm. The anti-aliasing filter cutoff frequency is set to 450kHz, and it is a 7th-order elliptic filter. In the trigger synchronization mechanism, the drill spindle speed sensor signal is synchronized with the sampling clock via a hardware trigger unit, ensuring that the sampling points correspond to the drilling cycle. Under this configuration, 1 million raw voltage data points per second are sampled, with each point accompanied by a timestamp recorded in microseconds. The sampled data has been verified to have a timing interval of 1μs, a maximum deviation of no more than 0.1μs, and a quantization error within 1.5μV, meeting the high precision and high consistency requirements for subsequent ultra-short time window slicing and transient feature extraction.

[0033] S1.2: Based on the discretized original voltage signal sequence, set an ultra-short time window length parameter of less than or equal to two milliseconds to generate sliding window configuration parameters for limiting the granularity of signal analysis.

[0034] The input condition is a discrete original voltage signal sequence with a unified time reference obtained after high-frequency synchronous sampling processing in S1.1. This sequence serves as the direct basis for setting the time window length. The number of sampling points within a single sampling interval of the discrete original voltage signal sequence is counted, and the minimum time length required for analysis granularity is determined by combining this with the energy distribution characteristics of the acoustic emission signal spectrum. Based on the determined time length and signal sampling rate, the corresponding number of time window points is calculated to ensure that the window length does not exceed two milliseconds, while simultaneously meeting the requirement of capturing high-frequency transient features. Using this number of time window points as the core parameter, a sliding window configuration data structure including window length, step size, and overlap rate is constructed to ensure stable time alignment characteristics in subsequent slicing processing. Parameter consistency checks are performed on the sliding window configuration data structure to verify the matching relationship between the window length / number of points and the sampling rate, as well as the logical constraints of the step size and overlap rate, to prevent time drift or data omissions during overlapping sliding truncation. Through the above processing method, the discrete signal sequence from the previous step is transformed into ultra-short time window sliding window configuration parameters that limit the signal analysis granularity, achieving precise control over the subsequent segmentation boundaries.

[0035] For example, in a real-time PCB drilling monitoring scenario, the sampling rate of the acoustic emission sensor is set to 500 kHz. The discretized original voltage signal sequence obtained after processing by S1.1 contains sampling points with uniform intervals. Based on the maximum window length constraint of two milliseconds, the number of window length points is calculated as follows: ,get Window length configuration for each point. Considering that high-frequency pulse components during drilling are mainly concentrated in the 80–200 kHz range, 0.5 milliseconds was chosen as the analysis granularity, and the number of points calculated for the window length was [number missing]. ,get The window length is defined as follows: the number of points for the window length, the number of points for the step size (set to 125 points), and the overlap rate of 0.5 are encapsulated as sliding window configuration parameters. When performing consistency checks on these parameters, it is verified that the matching relationship between the 125-point step size and the 0.5 overlap rate conforms to the sampling rate constraint. Applying these configuration parameters enables the stable generation of a 250-point long, temporally continuous, and overlapping half-window signal segment stream in subsequent slices, effectively avoiding the loss of transient features due to improper window length settings, and significantly improving the accuracy and real-time performance of defect stage modeling.

[0036] S1.3: Perform an overlapping sliding truncation operation on the discretized original voltage signal sequence using the sliding window configuration parameters to output a series of short-time acoustic emission signal segments arranged in chronological order.

[0037] The sliding window configuration parameters generated based on step S1.2 are correlated and analyzed with the discretized original voltage signal sequence output by step S1.1 to determine the correspondence between the starting sampling index and the ending sampling index of each truncation operation.

[0038] The starting sampling index is shifted step-by-step according to the time base sequence. The shift step size is smaller than the window length and is set by the overlap ratio in the sliding window configuration parameters. Obtain the index displacement value, where This represents the number of sampling points corresponding to the window length. This is the overlap coefficient.

[0039] The discretized original voltage signal sequence is continuously truncated according to the above-mentioned translation step size. Each truncation generates a short signal segment with a fixed number of sampling points, and the truncated data vector is stored in the segment buffer queue in chronological order.

[0040] During each segmentation process, a time series index is established using the segment start timestamp and sequence number marker to ensure that subsequent time series processing modules can be called strictly in the acquisition order.

[0041] The generated fragment buffer queue is subjected to integrity verification to verify the coverage of consecutive fragments on the timeline and the consistency of overlapping segments, ensuring that there are no frame drops or duplicate frames.

[0042] By using an overlapping sliding truncation operation, the discretized original voltage signal obtained in the previous step is transformed into a series of short acoustic emission signal segments arranged in time sequence with high sampling density, thus achieving the expected technical effect of balancing signal granularity refinement and temporal continuity.

[0043] For example, in an online monitoring scenario for a PCB drilling production line, the analog signal output by the acoustic emission sensor is sampled at high frequency to obtain a discrete data sequence with a sampling rate of 500kHz. The sliding window parameters are set to a window length of 1 millisecond (500 points) and an overlap coefficient of 0.5, according to the formula... A translation step size of 250 points was obtained. During the truncation operation, the first segment was truncated from point 0 to point 500, followed by the second segment from point 250 to point 750, and this process was iterated until the end of the signal, forming a total of 200 short segments. For signals under different drilling conditions, the peak amplitude of the segments during early microcrack initiation showed slight fluctuations in the overlapping section. Integrity verification confirmed that all segments achieved 100% coverage on the time axis without any dropped frames. This output sequence, after being indexed and marked, was provided to the multi-scale envelope extraction module. In real-time testing, this overlapping truncation method significantly improved the ability to capture defect evolution trends under millisecond-level inference delay constraints.

[0044] S1.4: Perform streaming encapsulation and indexing on the series of short acoustic emission signal segments arranged in chronological order to generate a high-density acoustic emission signal segment stream containing complete temporal context information.

[0045] Step S2: Perform adaptive amplitude normalization and multi-scale envelope extraction processing on each signal segment in the high-density acoustic emission signal segment stream to obtain a local response curve reflecting the transient energy distribution. Specifically, this includes: S2.1: Perform sliding window peak detection processing on the current acoustic emission signal segment in the high-density acoustic emission signal segment stream to obtain the local peak amplitude parameter characterizing the instantaneous maximum amplitude of the segment.

[0046] For the current acoustic emission signal segment input in the high-density acoustic emission signal segment stream, a sliding window group with a fixed length and overlap rate is established to limit the time range of local peak search.

[0047] Perform absolute value calculation on discrete sampled values ​​within each sliding window to convert all amplitude values ​​within that window into a non-negative amplitude sequence for subsequent peak determination.

[0048] The maximum value extraction operation is performed on the non-negative amplitude sequence, and the position and amplitude of the peak sampling point are determined by comparing the amplitude of each sampling point within the window.

[0049] A threshold screening is performed on the determined peak sampling point amplitude to discard peaks below the preset noise tolerance, so as to avoid false peaks from participating in subsequent amplitude parameter calculations.

[0050] A local peak amplitude parameter is constructed using the filtered peak amplitude data, and this parameter is used as the amplitude benchmark for subsequent adaptive amplitude normalization processing.

[0051] By using the above-mentioned sliding window peak detection and amplitude parameter generation processing method, the high-density acoustic emission signal segment stream of the previous step is transformed into a local peak amplitude parameter that characterizes the instantaneous maximum amplitude of the current segment, thereby realizing a quantitative description of the transient energy peak.

[0052] S2.2: Perform adaptive amplitude normalization processing on the current acoustic emission signal segment based on the local peak amplitude parameter to generate a standardized acoustic emission signal sequence that eliminates the effects of drill bit wear and rotation speed fluctuations.

[0053] S2.3: Perform multi-order Hilbert transform processing on the standardized acoustic emission signal sequence to obtain a set of complex analytic signals covering different frequency bands.

[0054] S2.4: Perform modulus operation and low-pass smoothing filtering based on the complex analytic signal set to extract multi-scale envelope component groups that characterize the transient energy envelope of each frequency band.

[0055] S2.5: Perform time-domain weighted fusion and time-series alignment processing on the multi-scale envelope component group to generate a local response curve that reflects the transient energy distribution characteristics.

[0056] The sub-step of performing time-domain weighted fusion and time-series alignment on the multi-scale envelope component group is used in the main step to integrate the transient energy envelopes of each frequency band into a single response curve data with a unified time reference, thereby providing a consistent input object for subsequent topology connection calculations. The input object is a multi-scale envelope component group obtained by multi-order Hilbert transform and modulus operation, and low-pass smoothing filtering. Each component contains transient energy change sequences of different frequency bands. For this input object, a weighted coefficient sequence is first established based on the sensitivity coefficients of each frequency band in the borehole defect detection task. By multiplying each component point by the corresponding weight in the time domain, a weighted envelope data group is generated while retaining the original time axis index. Then, cumulative integral normalization processing is performed on the weighted envelope data group to eliminate baseline offset caused by frequency band energy differences and ensure that the energy contributions of each scale envelope are within a comparable range during fusion. Subsequently, a time-delay cross-correlation detection method is employed to calculate the relative delay between each weighted envelope and the reference-scale envelope. This delay is used as a time-series offset value to perform interpolation translation or scaling operations on the envelopes, achieving time synchronization of waveform peaks and key zero-crossing positions. Further, the time-aligned weighted envelopes are then subjected to a weighted superposition operation in the time domain. The response curve generated by this fusion operation is the local response curve reflecting the multi-band integrated transient energy distribution. Through the above processing method, the multi-scale envelope component group from the previous step is transformed into a local response curve with a unified time reference and frequency band weighting characteristics, achieving the expected technical effect of providing a stable, sensitive, and high-fidelity description of energy changes for subsequent topology weight calculations.

[0057] For example, in a short-time segment of an acoustic emission signal from a PCB drilling hole, the input multi-scale envelope component group contains three frequency bands: low frequency (0~20kHz), mid frequency (20~60kHz), and high frequency (60~120kHz). Based on previous feature sensitivity analysis, the weighting coefficients for the low, mid, and high frequency bands are set to 0.5, 1.0, and 1.5, respectively. The transient energy sequence of each frequency band is multiplied point-by-point in the time domain by the corresponding coefficient to generate weighted envelope data. Normalization is performed on each weighted envelope by dividing the cumulative integral by the sum of the integrals to obtain a baseline-consistent energy curve. A cross-correlation method is used to detect a 0.25ms delay between the high-frequency and mid-frequency envelopes, and a 0.15ms delay between the mid-frequency and low-frequency envelopes. Spline interpolation is used to shift and correct each envelope, aligning the peak values ​​on the time axis. The aligned envelopes were fused according to the above formula, with a total weight of 3.0. The combined energy at the peak of the fused result was approximately 0.87 times that of the original high-frequency peak energy, and the waveform details were significantly improved due to the contributions of each frequency band. The output local response curve showed significant sensitivity to the energy abrupt change during the microcrack initiation stage in the subsequent amplitude and slope calculation, verifying the effectiveness of the fusion and alignment processing.

[0058] like Figure 2 As shown, step S3 involves: based on a predetermined number of adjacent local response curves, calculating the topological connection weights between each pair of curves according to the amplitude change slope, zero-crossing rate difference, and peak delay relationship, and constructing a dynamic similarity matrix as initial adjacency relationship data. Specifically, this includes: S3.1: Perform first-order differential processing on the local response curves of the current moment and the preset number of adjacent curves to obtain the amplitude change slope sequence that characterizes the transient energy change characteristics, and use this amplitude change slope sequence as the basic input variable for subsequent calculation of topological connectivity strength.

[0059] Perform a first-order difference operation on adjacent preset number of local response curve inputs to obtain the amplitude change sequence of each curve on the time axis.

[0060] Divide the amplitude change sequence by the corresponding time interval parameter to generate the amplitude change rate sequence to characterize the intensity of local energy abrupt changes in the curve.

[0061] A centering method is used to normalize the amplitude change rate sequence to zero mean, thereby eliminating the offset effect caused by the difference in absolute amplitude of different curves.

[0062] Moving average filtering is used to smooth the normalized amplitude change rate sequence to reduce high-frequency noise interference while preserving the energy mutation characteristics.

[0063] An amplitude change slope sequence is constructed based on the smoothed amplitude change rate sequence, and this sequence is used as the basic input variable for calculating the topological connectivity strength.

[0064] Through the above chain processing method, the original amplitude data of the local response curve is transformed into amplitude change slope sequence data that can be used for similarity measurement, thereby achieving accurate quantification of transient energy mutation characteristics.

[0065] S3.2: Based on the amplitude change slope sequence, perform zero-crossing detection processing, count the time offset of the zero-crossing position between adjacent local response curves, generate a zero-crossing rate difference parameter that reflects the degree of waveform phase mismatch, and use the zero-crossing rate difference parameter as a key indicator for measuring the consistency of signal shape.

[0066] S3.3: Use envelope peak search technology to locate the extreme points of the local response curve, calculate the time interval between the occurrence of the main peaks of adjacent curves, generate peak delay relationship data that characterizes the energy propagation lag characteristics, and use the peak delay relationship data as the core basis for evaluating the temporal causality of defect evolution.

[0067] An envelope peak search method is used to locate extreme points on adjacent local response curves of a predetermined number, identifying the time coordinate parameters of the main peak in the multi-scale envelope component group of each curve. The time difference between adjacent curves is calculated for the located main peak time coordinates to obtain initial time interval values ​​reflecting differences in energy envelope propagation speed. These initial time interval values ​​are then refined using a peak time offset correction method based on interpolation fitting to eliminate quantization errors introduced by sampling rate limitations, ensuring the time intervals are within sub-millisecond accuracy. The corrected time interval values ​​are then normalized to adjust the scale differences caused by different time window lengths and sampling configurations to a unified comparison benchmark. The normalized time interval values ​​are mapped to peak delay relationship data, and the delay values ​​between curves are structured and stored by constructing an energy propagation hysteresis characteristic index matrix for subsequent multi-dimensional feature deviation calculations. By evaluating the temporal causality of borehole defect evolution using peak delay relationship data as the core, the local response curve characteristics of the previous step are transformed into structured temporal characteristics that characterize the lag in energy fluctuation propagation, thus achieving the expected technical effect of identifying the causal path in the process of focusing defect initiation to expansion.

[0068] For example, in a real-time PCB drilling monitoring scenario, the acoustic emission signal sampling rate is configured to 250 kHz, the sliding window length is 2 ms, and a local response curve with 500 points is generated. The envelope peak search method is used to locate the main peak in the multi-scale envelope component group; for example, the main peaks are found at point 120 of the low-frequency envelope component and point 135 of the high-frequency envelope component. The time interval between the main peaks of two adjacent curves is calculated as follows: seconds, after interpolation fitting correction is Seconds. Normalize this time interval and define a normalized delay value. Multiply by the reference scale factor The obtained delay index was 0.06. This delay index is stored in the peak delay relationship data matrix and is used to quantify the energy propagation hysteresis characteristics. When there are significant differences in the peak delay range caused by different defect types, this matrix can effectively map the temporal causal evolution path from the normal state to the crack initiation state and then to the hole wall tearing state during the drilling process. The verification results show that the method can significantly improve the identification accuracy of the early defect stage and meet the millisecond-level real-time judgment requirements.

[0069] S3.4: Based on the amplitude change slope sequence, zero-crossing rate difference parameter and peak delay relationship data, the weighted Euclidean distance metric method is used to calculate the multidimensional feature deviation between each pair of local response curves, generate the original connection weight value of the similarity level between the curves, and use the original connection weight value as the direct material for constructing the edge weight of the graph structure.

[0070] The input conditions include the amplitude change slope sequence, zero-crossing rate difference parameter, and peak delay relationship data generated in steps S3.1 to S3.3, all of which have been established in pairs according to the local response curves corresponding to the time series. The amplitude change slope sequence serves as a quantitative indicator of transient energy change, the zero-crossing rate difference parameter serves as a quantitative benchmark for waveform phase consistency, and the peak delay relationship data serves as the causal basis for energy propagation timing.

[0071] The three features mentioned above are combined to construct a multidimensional feature vector for each pair of local response curves, and a feature weighting matrix is ​​formed according to the preset weight coefficients. The weight coefficients are set based on the importance ranking of the features in relation to defect evolution.

[0072] Perform a difference operation on the corresponding elements of each feature vector to obtain a set of numerical values ​​for the difference in amplitude change slope, the difference in zero-crossing rate, and the difference in peak delay, which serve as the input components for Euclidean distance calculation.

[0073] Using the weighted Euclidean distance metric, the multidimensional feature deviation between curves is obtained through the following formula: in, For the difference in the slope of the amplitude change, For the difference in zero-crossing rate, For peak latency differences, , , These are the weighting coefficients for each feature component. is the normalization factor for the sum of weights.

[0074] The calculated multidimensional feature deviation is converted into connection weight values ​​according to the preset connectivity reverse measurement rules, ensuring that the smaller the deviation value, the larger the connection weight, so as to directly reflect the curve similarity level.

[0075] The above connection weight values ​​are normalized to the range [0,1] so as to unify the edge weight scale and optimize the computational stability in the subsequent graph structure construction process.

[0076] By using a weighted Euclidean distance metric, the results of the previous step are transformed into quantitative curve similarity data, thus achieving the goal of preparing basic data for constructing graph structure edge weights.

[0077] For example, in PCB drilling acoustic emission monitoring, a feature vector is generated by combining N=5 adjacent local response curves. The differences in amplitude change slopes are 0.12, 0.08, 0.15, 0.10, and 0.09, the differences in zero-crossing rates are 0.05, 0.06, 0.04, 0.07, and 0.05, and the differences in peak delay are 0.02, 0.03, 0.02, 0.04, and 0.03. The weighting coefficients are set to... =0.5、 =0.3、 =0.2, weights sum =1.0. Substitute the above data into the formula to calculate the deviation of the first pair of curves: The calculation result is The values ​​are converted to connection weights ≈ 0.912 according to the reverse metric rule, and after normalization and mapping to the [0,1] interval, they are used for dynamic similarity matrix filling. In multi-set curve calculations, this method can significantly improve the accuracy and stability of temporal local topological associations, ensuring that the construction of subsequent local temporal subgraphs focuses on short-range evolutionary associations.

[0078] S3.5: Perform a matrix arrangement operation based on the original connection weight values ​​to map the discrete pairwise connection weights into a two-dimensional square matrix structure, generate a dynamic similarity matrix that fully describes the topological association strength between all adjacent local response curves, and output the dynamic similarity matrix as the initial adjacency relationship data for the local time series subgraph construction module to call.

[0079] Based on the original set of connection weights calculated in substep S3.4, pairs of values ​​containing the connection strengths between all pairwise local response curves are selected as the input for matrix arrangement. The original connection weights are mapped row-column according to the curve time index, mapping the weight value between the i-th and j-th local response curves in the time index sequence to the i-th row and j-th column of the two-dimensional matrix. To ensure matrix symmetry and the integrity of adjacency data, a mirror filling operation is performed on the matrix elements, simultaneously assigning the value of the element in the i-th row and j-th column to the j-th row and i-th column, forming a matrix structure of undirected topological relationships. During the construction of the two-dimensional matrix, the values ​​of the main diagonal elements are set to zero to eliminate invalid connection weights of their own nodes. To enhance the resolvability of the matrix during the construction of local temporal subgraphs, normalization is performed on the matrix elements, mapping all weight values ​​to the closed interval [0,1], facilitating weighted correction when introducing time decay factors later. The normalization calculation is performed using the following formula: in, These are the original connection weight values. It is the minimum value among all weights. The maximum value among all weights. These are the normalized weight values. Through the above matrix arrangement and normalization process, the discrete pairwise connection weights are mapped into a two-dimensional square matrix structure, generating a dynamic similarity matrix that fully describes the strength of topological associations between all adjacent local response curves. This matrix is ​​then output as the initial adjacency relationship data for use by the local time series subgraph construction module in step S4, achieving a smooth transition from time series signals to graph structure data.

[0080] For example, when constructing a dynamic similarity matrix for N=5 consecutive local response curves, the original set of connection weight values ​​includes × The weights are combined pairwise. After matrix transformation, a 5×5 weight matrix is ​​obtained, with the main diagonal elements set to 0. Assuming the original weights have a minimum of 0.12 and a maximum of 0.87, during matrix element normalization, the weight value of 0.45 in the 2nd row and 4th column is normalized as follows: ≈ The normalized matrix can intuitively reflect the similarity ratio between local response curves. In the subsequent construction of local temporal subgraphs, this matrix is ​​used as the initial adjacency relation data and combined with the time decay factor to achieve weight correction that focuses on short-range evolutionary associations. This significantly improves the temporal feature fidelity and association strength resolution capability of the defect evolution path graph model in millisecond-level responses.

[0081] like Figure 3As shown, step S4 involves: using the local response curve corresponding to the current signal segment as the root node, tracing back a preset number of historical local response curves, and constructing a local time-series subgraph focusing on short-range evolutionary correlations using the initial adjacency relationship data with an introduced time decay factor. Specifically, this includes: S4.1: Perform root node mapping processing on the local response curve corresponding to the current signal segment, and perform neighbor node mapping processing on the historical local response curves that are traced back a preset number of times, so as to generate a local time-series subgraph node set containing the root node and multiple neighbor nodes.

[0082] Using a dynamic similarity matrix as the input data source, a unique root node identifier is generated in the node mapping layer for the local response curve corresponding to the current acoustic emission signal segment. The feature embedding vector and timestamp information of this curve are then bound to ensure the temporal accuracy of the subsequent topological structure. The index positions in the initial adjacency data are traced back a predetermined number of historical local response curves in chronological order. Corresponding neighbor node identifiers are generated in the node mapping layer for each, and their respective feature embedding vectors and timestamp parameters are bound to ensure that the neighbor node set covers the temporal range required for short-range evolutionary associations. Node set initialization processing is performed on the root node and each neighbor node, uniformly mapping all node indices to the global node table of the local temporal subgraph to establish a unified node reference structure for subsequent edge weight calls. Using the edge weight values ​​in the dynamic similarity matrix, reference information is added to the topological connection relationships between the root node and neighbor nodes in the node set, and the source of the connection values ​​is marked to accurately locate the edge weights that need correction during time-decay weighting processing. By synchronously binding the node set with the index mapping relationship, the dynamic similarity matrix of the previous step is transformed into a local temporal subgraph node set containing the root node and multiple neighbor nodes, thereby realizing the structured association between the current signal segment and its short-range historical local response curve.

[0083] By using node mapping and set binding, the dynamic similarity matrix from the previous step is transformed into a set of local temporal subgraph nodes with temporal indexes and feature vector bindings, thereby realizing the preparation of foundational data for local subgraph construction and the explicit expression of short-range evolutionary associations.

[0084] For example, in a PCB drilling acoustic emission monitoring scenario, when the sampling frequency is set to 250kHz and the preset backtracking quantity is 5, the local response curve of the current segment generates a root node numbered N0 in the node mapping layer. The bound feature embedding vector length is set to 32 dimensions, and the timestamp accuracy is at the microsecond level. The five backtracked historical response curves generate neighbor nodes numbered N1 to N5 respectively, with the feature vector uniformly set to 32 dimensions and the timestamps decreasing in order of sampling period. When performing set initialization on the root node and neighbor nodes, the node index is mapped to the global node table of the local time series subgraph, with the index positions arranged consecutively from 0 to 5. The connection weights between the root node and neighbor nodes in the dynamic similarity matrix are 0.82, 0.76, 0.65, 0.58, and 0.51 respectively. These values ​​are referenced through node set binding to provide accurate positioning for the time decay factor weighting in the subsequent S4.2 sub-step. In this configuration, the local temporal subgraph node set output contains an index structure and feature binding information of 6 nodes. The verification results show that this set can preserve the transient energy distribution features of short-range historical fragments during graph convolution aggregation and significantly improve the ability of defect stage discrimination to capture early states.

[0085] S4.2: Calculate the time interval difference based on the timestamp information of each node in the local time-series subgraph node set, and use the exponential decay function to perform nonlinear transformation on the time interval difference to generate a time decay factor sequence that represents the decreasing weight of the influence of historical nodes on the current node over time.

[0086] Perform interpolation on the timestamp information of each node in the local time series subgraph node set to obtain a time interval difference sequence relative to the root node timestamp.

[0087] The time interval difference sequence is used as the input variable of the exponential decay function. The decay coefficient parameter is preset, and the shape of the decay curve is determined according to the real-time requirements of the process and the signal sampling accuracy.

[0088] For each time interval difference, a nonlinear exponential function transformation is performed. The attenuation factor for node 1 is calculated using the formula: ≈0.6065, node 2 is ≈0.3679, node 3 is ≈0.2231, node 4 is ≈0.1353. This time decay factor sequence effectively suppresses the edge weights of remote low-correlation nodes in the subsequent weighted adjacency matrix correction, significantly improving the sensitivity and discrimination accuracy of the local subgraph to the short-range evolution stage of defects.

[0089] S4.3: Perform weighted correction processing on the dynamic similarity matrix elements in the initial adjacency relationship data, and use the time decay factor sequence to perform dynamic suppression operation on the topological connection weights between each pair of nodes to generate a weighted adjacency matrix focusing on short-range evolutionary associations.

[0090] Targeted data retrieval processing is performed on the dynamic similarity matrix elements in the initial adjacency relationship data. All elements of the matrix are reordered according to the node index and timestamp matching rules to ensure the consistency of the mapping between the matrix row and column correspondence and the node set of the local time series subgraph.

[0091] For each edge weight element in the matched dynamic similarity matrix, the association binding process of the time decay factor is performed. The time decay factor sequence output in step S4.2 is matched one by one with the topological connection weight at the corresponding position in the matrix to establish a direct mapping relationship between the time decay parameter and the weight value.

[0092] Dynamic suppression operations are performed on the bound weight values. The time decay factor is multiplied by the original connection weights using the multiplication suppression mode to calculate the corrected connection weight values.

[0093] The corrected connection weight matrix is ​​normalized by scaling all weight values ​​proportionally to the maximum value of the row or column, ensuring that the weights are distributed within the [0,1] interval, thus improving the numerical stability of subsequent graph convolution operations.

[0094] The normalized connection weight matrix is ​​subjected to matrix encapsulation processing for focusing on proximal evolutionary associations. High-weight proximal node pairs are marked and the weight priority of remote node pairs is reduced to generate a weighted adjacency matrix for focusing on proximal evolutionary associations.

[0095] Through the above weighted correction and normalization process, the dynamic similarity matrix of the previous step is transformed into a weighted adjacency matrix with time decay constraints and short-range focusing characteristics, thereby realizing high-precision construction and evolutionary correlation focusing of edge weights in local temporal subgraphs.

[0096] For example, in a real-time monitoring scenario for PCB drilling, the initial dynamic similarity matrix is ​​set to a 5×5 two-dimensional square matrix, with matrix elements ranging from 0.15 to 0.92, and the time decay factor sequence ranging from 0.3 to 1.0. Multiplication suppression is performed on each edge weight element. For example, when the original connection weights... Corresponding time decay factor At that time, the corrected weight calculation result is = Perform row normalization on the entire matrix. If the maximum value of a certain row is... Then the elements in that row will be scaled proportionally, as shown above. Scaling = After normalization, the high weights of the matrix are concentrated on node pairs with a time interval of ≤2 milliseconds, while the low weights correspond to node pairs with a time interval of ≥8 milliseconds. This effectively highlights the short-range evolutionary correlation features and significantly improves the sensitivity and stability of defect stage discrimination in subsequent graph convolution aggregation.

[0097] S4.4: Based on the local temporal subgraph node set and weighted adjacency matrix, perform graph structure assembly processing, and topologically bind the node feature vectors with the corrected edge weights to generate a local temporal subgraph focusing on short-range evolutionary associations.

[0098] Using a set of nodes in a local temporal subgraph as input, the feature embedding vectors of each node in the set are read in chronological order and a node index sequence is built.

[0099] For each element in the weighted adjacency matrix, perform an index mapping process to map the matrix row and column indices to the node sequence positions, so as to ensure that the edge weights match the actual association between the nodes.

[0100] Perform topology binding operations to bidirectionally associate the weight value of each edge with the corresponding node feature vector, and introduce an index verification mechanism during the binding process to prevent erroneous connections, ensuring that the edge weight correction data is accurately applied to the target node pair.

[0101] The bound node features and edge weights are encapsulated into a graph structure data object. Node attribute fields and edge attribute fields are explicitly defined in the data object to form structured subgraph data that meets the requirements of graph convolution operations.

[0102] By encapsulating the output with structured subgraphs, the time-decayed edge weights and node feature vectors are transformed into local temporal subgraphs that can be directly called, thus achieving the effect of graph structure construction with a focus on short-range evolutionary associations.

[0103] For example, in PCB drilling acoustic emission monitoring, a local timing subgraph node set is set to include one root node and four neighboring nodes. Each node's feature vector has a dimension of 128, and the time-decayed weighted adjacency matrix is ​​5×5, with off-diagonal edge weights ranging from 0.05 to 0.87. The node index sequence is arranged in ascending order of timestamps as {0,1,2,3,4}, corresponding to the matrix row and column positions. In the binding operation, edge weight 0.82 is bound to the feature vector pair corresponding to node index (0,1), edge weight 0.24 is bound to the feature vector pair corresponding to node index (0,3), and so on. During encapsulation, the feature vector is recorded in the node attribute field of the subgraph data object. For example, the first component value of the feature vector of node index 0 is... The edge attribute field records the corresponding edge weight, such as the edge weight at index (0,1). The local temporal subgraph generated after this step can significantly improve the root node's ability to capture short-term historical evolution information in subsequent graph convolution aggregation, and achieve the expected high response speed and stable judgment accuracy in real-time defect discrimination.

[0104] Step S5: Perform a lightweight graph convolution operation on the local temporal subgraph to aggregate neighbor node information, implicitly encoding the relative position of the current segment in the defect development path as a topological evolution representation vector. Specifically, this includes: S5.1: Perform linear projection transformation on the root node feature embedding vector and neighbor node feature embedding vector in the local temporal subgraph to map the original feature space to a high-dimensional hidden feature space based on a learnable weight matrix, thereby generating an initial node hidden state representation with nonlinear expressive capabilities.

[0105] S5.2: Based on the initial node hidden state representation and the initial adjacency relationship data with introduced time decay factor, perform attention coefficient calculation processing to quantify the dynamic association strength between the root node and each historical neighbor node using the leaked rectified linear unit activation function and the normalized exponential function, and generate a standardized edge direction attention weight distribution.

[0106] S5.3: Perform a weighted aggregation operation on the standardized edge direction attention weight distribution and the neighbor node feature embedding vector to fuse the transient energy distribution features of multi-source historical fragments based on the connection strength difference, and generate an intermediate layer node aggregation feature vector containing short-range evolutionary association information.

[0107] A matching index relationship is established between the standardized edge-direction attention weight distribution and the neighbor node feature embedding vector, ensuring that each attention weight corresponds only to its associated neighbor node feature vector. Element-wise multiplication is performed on the matched attention weights and feature vectors to amplify the feature components of high-weight connections and suppress the feature components of low-weight connections in the feature space. Vector summation is performed on all weight-adjusted neighbor node feature vectors to aggregate the transient energy distribution features of multi-source historical fragments in the same dimension. The aggregation result is normalized, and the vector norm is used to control the numerical scale of the aggregated feature vector of the intermediate-layer node to avoid numerical explosion in subsequent convolution calculations. A nonlinear activation function is used to perform feature mapping on the normalized vector, enabling it to describe the nonlinear pattern of defect development while maintaining short-range evolutionary correlation information. Through weighted aggregation, the attention weight information from the previous step is fused with the neighbor node features to obtain the aggregated feature vector of the intermediate-layer node containing short-range evolutionary correlation information, achieving a local semantic characterization of the defect development path stage of the current node.

[0108] For example, in the PCB drilling acoustic emission signal processing, the standardized edge-direction attention weight distribution is set as follows: the weights corresponding to the three neighboring nodes are 0.65, 0.25, and 0.10, respectively, and the feature embedding vectors of the neighboring nodes are [0.8, 0.3, 0.4], [0.6, 0.5, 0.2], and [0.1, 0.9, 0.7], respectively. Element-wise multiplication is performed on each feature vector: the first neighbor receives [0.52, 0.195, 0.26], the second neighbor receives [0.15, 0.125, 0.05], and the third neighbor receives [0.01, 0.09, 0.07]. Vector summation is performed on the above results to obtain the aggregate vector [0.68, 0.41, 0.38]. Normalization is then performed on this vector: its L2 norm is calculated. The normalization factor is approximately 0.889. Dividing each element of the aggregated vector by this factor yields [0.765, 0.462, 0.427]. The normalized result is then subjected to a nonlinear mapping using a leaky rectified linear unit activation function, with negative values ​​set to zero and positive values ​​remaining unchanged, resulting in the final aggregated feature vector of the intermediate layer nodes [0.765, 0.462, 0.427]. This vector significantly improves the ability to capture short-range evolution features of the current borehole state in subsequent multi-level graph convolution processing, enhancing the accuracy and foresight of defect stage determination.

[0109] S5.4: Based on the aggregated feature vectors of the intermediate layer nodes, perform multi-level feature splicing and nonlinear activation processing to abstract the local topological structure of the acoustic emission signal layer by layer through multiple cascaded lightweight graph convolutional layers, and generate deep semantic defect evolution path feature embedding.

[0110] S5.5: Perform global pooling and dimensionality compression on the deep semantic defect evolution path feature embedding to extract a low-dimensional dense vector that reflects the relative position attributes of the current segment throughout the entire process from microcrack initiation to hole wall tearing acceleration, and finally generate a topological evolution representation vector for stage discrimination.

[0111] The relative position attribute refers to the stage position of the current acoustic emission signal segment in the defect evolution time path, including but not limited to the starting position of the normal drilling stage, the transition position of the microcrack initiation stage, and the key turning point of the hole wall tearing acceleration stage.

[0112] When performing global pooling and dimensionality compression on the feature embedding of deep semantic defect evolution paths, the input object is set as a high-dimensional feature tensor abstracted by multiple layers of lightweight graph convolution, which contains local topological pattern information after the fusion of the root node and its neighbor nodes. Based on this high-dimensional feature tensor, the global pooling operator is first called to aggregate the feature components of all nodes. A joint strategy of average pooling and max pooling is adopted to ensure that both the global energy distribution trend is preserved and the features of abnormal peak nodes are highlighted. The comprehensive global description vector is obtained by simultaneously calculating the mean and extreme values ​​along the feature axis and introducing weights for fusion. Secondly, the feature covariance matrix and eigenvalue decomposition are performed on the above comprehensive global description vector to construct the principal component mapping matrix for subsequent compression. The compression parameters are set according to the cumulative contribution rate threshold to ensure that all key information required for defect evolution stage discrimination is still covered after dimensionality reduction. Thirdly, the comprehensive global description vector is linearly transformed using the mapping matrix to project the high-dimensional feature vector to a preset low-dimensional space, forming a dense vector structure to reduce the input size of the classifier and reduce computational latency. Subsequently, normalization is performed on the compressed dense vector to map each component value to a unified dimension, achieving portability and robustness for cross-sample comparisons. Through a joint processing method of global pooling and dimensionality compression, the deep graph convolution output of the previous step is transformed into a low-dimensional dense topological evolution representation vector reflecting the relative positional attributes of the current segment throughout the entire process from microcrack initiation to accelerated hole wall tearing, thus achieving accurate input data for defect stage discrimination.

[0113] For example, in a real-time PCB drilling inspection scenario, global pooling is performed on the 256-dimensional deep feature embedding obtained after three layers of lightweight graph convolution. The average pooling and max pooling weight coefficients are set to 0.6 and 0.4, respectively, to obtain a fused 256-dimensional comprehensive global description vector. A feature covariance matrix is ​​constructed for this vector, and its top 10 eigenvalues ​​and corresponding eigenvectors are solved, with a cumulative contribution rate of 0.92. These 10 principal components are selected to form a mapping matrix. The comprehensive global description vector is linearly transformed using this mapping matrix to obtain a 10-dimensional dense vector. Z-score normalization is used, and the mean and standard deviation are obtained from global statistics of the training set. After this processing, the defect stage discrimination delay of the output 10-dimensional vector on the validation set is significantly reduced compared to the original 256-dimensional input, the real-time response performance of the classifier is greatly improved, and the recognition accuracy in the early microcrack initiation stage is significantly improved.

[0114] Step S6: Input the topological evolution representation vector into a lightweight classifier for stage discrimination processing, and output the defect state probability distribution representing four types of states: normal drilling, microcrack initiation, accelerated borehole wall tearing, or abnormal drill bit wear. Specifically, this includes: S6.1: Perform a fully connected layer linear transformation on the topology evolution representation vector to obtain an initial state score vector containing the original scores of four types of potential defect states.

[0115] The low-dimensional dense topological evolution representation vector from step S5.5 is subjected to input mapping binding processing, and the vector is used as the source of input sample data for a lightweight classifier.

[0116] The lightweight classifier uses a multilayer perceptron (MLP) architecture, which includes the following hierarchical structure: Input layer: Receives the topological evolution representation vector output from step S5 as input, with the input dimension matching the dimension of the topological evolution representation vector (set to d). in =12 to 64, preferably 12). The input layer does not perform non-linear transformations, but only performs data transfer and dimension alignment functions.

[0117] Hidden layer (optional): To enhance the non-linear expressive power of the model, up to one hidden layer can be inserted between the input and output layers. The number of neurons in the hidden layer is set to 0.5 to 2 times the input dimension, preferably the same as the input dimension. The hidden layer uses the ReLU activation function.

[0118] Output Layer: The output layer is a fully connected layer with the number of neurons matching the number of defect state categories (corresponding to normal drilling, microcrack initiation, accelerated borehole wall tearing, and abnormal drill bit wear, respectively). The output layer does not have an activation function; it directly outputs the raw score vectors for each category.

[0119] Output processing: The original score vector of the output layer is normalized by the Softmax function to obtain the probability distribution of the four types of defect states, and then the final result is output after temperature scaling calibration.

[0120] To achieve the goal of "lightweight" design, the total number of parameters of the classifier is controlled below 5000, and the number of floating-point operations (FLOPs) per inference is controlled below 10. 4 Magnitude.

[0121] The lightweight classifier is optimized by combining offline training with online fine-tuning. Offline training: A training sample set was constructed based on historical borehole experimental data. Each sample contained a topological evolution representation vector and its corresponding defect state label. The cross-entropy loss function and Adam optimizer were used for iterative parameter optimization. The number of training epochs was set to 50-200, the initial learning rate was set to 0.001, and a cosine annealing strategy was used to decay the learning rate.

[0122] Hyperparameter tuning: Determine the optimal combination of hyperparameters, including hidden layer dimension, learning rate, batch size, etc., through grid search or Bayesian optimization methods to minimize the classification error on the validation set.

[0123] Online fine-tuning: During model deployment and operation, new samples can be collected to incrementally learn the classifier, and the weight parameters can be fine-tuned using mini-batch gradient descent to enhance the model's adaptability to new defect patterns.

[0124] The lightweight classifier achieves a fast and accurate mapping from topological evolution representation vectors to defect state probability distributions, providing high-confidence classification input for subsequent stage transition memory updates and state transition pattern recognition.

[0125] The source of the input sample data is constructed and matrix multiplication is performed in combination with the preset weight matrix of the fully connected layer to achieve linear transformation of the feature space and generate linear combination results corresponding to the four types of defect states.

[0126] A bias vector is introduced into the linear combination result to perform element-wise summation, thereby introducing a class-related fixed translation in the feature projection to ensure the model's discriminative sensitivity under different states.

[0127] The result after the above summation is subjected to numerical precision truncation and overflow protection operations to avoid outliers in floating-point operations from interfering with subsequent probability normalization.

[0128] The numerically processed linear combination vector is output as the initial state score vector, where each element corresponds to the original score value of a potential defect state.

[0129] By using linear transformation of the fully connected layer, the topological evolution representation vector from the previous step is transformed into data containing the original scores of four types of potential defect states, thus mapping the complex topological features into low-dimensional state score indicators that can be probabilistically determined.

[0130] For example, in the PCB drilling defect detection scenario, the input to the lightweight classifier is a topological evolution representation vector of length 12, where each element is a floating-point number and has been normalized using Z-score. The size of the fully connected layer weight matrix is ​​set to... The bias vector has a length of 4, and the weights and bias parameters are iteratively updated on the training set using the Adam optimization algorithm. In actual operation, the original score calculation formula for the first type of state (normal drilling) is as follows: in The element in the first row and i-th column of the weight matrix. Let i be the i-th component of the input vector. This represents the bias value corresponding to category 1. Operational test data shows that the original scores for the four states—normal drilling, microcrack initiation, accelerated borehole wall tearing, and abnormal drill bit wear—are 0.84, 0.84, and 0.84, respectively. 0.21, 0.57, 0.05, subsequent normalization can significantly improve the sensitivity of the classification model to changes in evolutionary paths.

[0131] S6.2: Perform exponential normalization operation based on the initial state score vector to eliminate dimensional differences and generate an uncalibrated state probability distribution vector that reflects the relative confidence level.

[0132] S6.3: The uncalibrated state probability distribution vector is smoothed and corrected using a preset temperature scaling parameter to optimize the sharpness of the class boundaries and output the calibrated state probability distribution vector.

[0133] S6.4: Perform maximum likelihood estimation based on the numerical values ​​of each element in the calibrated state probability distribution vector to determine the most likely single defect state label at the current moment.

[0134] S6.5: Combine the single defect state label with the calibrated state probability distribution vector and perform structured encapsulation processing to generate the final four types of state defect state probability distribution data.

[0135] The single defect state label and the calibrated state probability distribution vector are subjected to dual data source binding processing, and a one-to-one correspondence is established between the discrete label data representing the most likely defect state at the current moment and the continuous numerical parameters of the corresponding probability distribution in a unified structured data model.

[0136] The discrete label data is transformed into four-state encoded identifiers according to the category enumeration index rules, and the encoded identifiers are embedded into the category dimension metadata field of the probability distribution vector to maintain semantic consistency.

[0137] The probability distribution vector is processed to unify the precision of each element value. A hybrid storage scheme of fixed-point and floating-point formats is adopted to ensure the accuracy of the representation of boundary probabilities and extreme confidence levels. Confidence boundary markers are added to the structured data model to support the rapid interpretation of subsequent pattern matching.

[0138] Based on the structured data model, multi-field encapsulation processing is performed, combining category codes, confidence scores, label text descriptions, and timestamp information into a multi-layered nested data structure, and applying hash verification to generate a unique identifier to prevent duplication or misassociation during the memory update process in the rolling phase.

[0139] The multi-layered nested data structure is encoded into an ordered byte stream using a predefined serialization protocol, which is then transformed into the final four-class defect state probability distribution data and output to the subsequent stage transfer memory update module. Through this structured encapsulation process, the classification and judgment results of the previous step are transformed into highly consistent data objects that can be directly used for time-series pattern analysis, realizing standardized storage and fast retrieval of defect state probability distribution.

[0140] For example, in a real-time PCB drilling monitoring scenario, the lightweight classifier outputs a calibrated state probability distribution vector of [0.72, 0.18, 0.06, 0.04] in a single inference, corresponding to the categories of normal drilling, microcrack initiation, accelerated hole wall tearing, and abnormal drill wear, respectively. Maximum likelihood estimation determines the current defect state label as normal drilling, with a category code of 0. Through dual data source binding, the code 0 is embedded into the first category metadata field of the probability vector, and confidence boundary markers are added to the probability values ​​of each element, for example, marking high confidence at 0.72. During multi-field encapsulation, the data structure is constructed as follows: {"Category Code":0,"Label":"Normal Drilling","Probability Vector":[0.72,0.18,0.06,0.04],"Timestamp":1682547894,"Confidence Boundary":{"High":0.72,"Low":0.04}}, and a hash checksum is calculated for this data structure. In the serialization stage, a protocol using fixed-length fields and timestamp compression encoding is employed to convert the above structure into an ordered byte stream of 48 bytes, which is then transmitted to the stage transition memory update module. In this embodiment, the probability distribution data of the four types of state defects generated in this way can be directly indexed and retrieved in subsequent state transition pattern matching, significantly reducing data reading latency, ensuring real-time requirements, and maintaining the numerical consistency of the probability distribution during transmission and storage.

[0141] Step S7: Update the stage transition memory of the rolling stage based on the probability distribution of the defect states generated multiple times consecutively, and record the real-time state sequence composed of different defect states. Specifically, this includes: S7.1: Perform maximum likelihood estimation on the probability distribution of the defect state output by the lightweight classifier at the current moment to determine the current defect state label with the highest confidence from the four states: normal drilling, microcrack initiation, accelerated hole wall tearing, or abnormal drill bit wear.

[0142] The output of the lightweight classifier at the current moment is parsed to obtain the calibrated probability distribution vector containing four types of defect states as the initial processing object of this step.

[0143] An element position index scan is performed on the probability distribution vector, and each probability value is compared with a preset confidence threshold one by one to obtain a set of candidate states that are higher than the threshold, providing a candidate range for subsequent maximum likelihood estimation.

[0144] The index position of the maximum probability is calculated for the set of candidate states using the maximum likelihood estimation formula. in, This is the current defect status label. For the set of candidate states, The calibrated state probability function, Input features to the classifier, and output the class index corresponding to the maximum probability. .

[0145] The category index is mapped to a predefined four-category status label encoding table to obtain the corresponding label text or code form, ensuring that it can be directly called during the data maintenance process.

[0146] Perform format validation and timestamp binding on the current defect status label obtained from the mapping to generate a status label data packet with the current time identifier attribute.

[0147] By using the maximum likelihood estimation and mapping process described above, the probability distribution result of the previous step is transformed into the current defect status label with the highest confidence, thereby achieving accuracy and real-time performance in status judgment.

[0148] For example, in a real-time monitoring scenario for PCB drilling, the lightweight classifier outputs a calibrated probability distribution vector of [0.15, 0.62, 0.18, 0.05] at t=125.6 ms, corresponding to normal drilling, microcrack initiation, accelerated hole wall tearing, and abnormal drill wear, respectively. With a confidence threshold of 0.5, the candidate state set C={1} is obtained. Applying the maximum likelihood estimation formula yields i=1, corresponding to the state label "microcrack initiation". This label is then bound to a timestamp to generate a data packet (label: "microcrack initiation", time: 125.6ms). For different drilling materials, if the classifier outputs a probability distribution of [0.48, 0.33, 0.09, 0.10] and the threshold is set to 0.4, the candidate state set is C={0}. Maximum likelihood estimation yields S=0, corresponding to the label "normal drilling". The data packet is automatically associated with this label and the current time. This embodiment achieves accurate tag extraction under different materials and conditions through formula calculation and threshold matching, and maintains the stability and effectiveness of state judgment even under diverse signal characteristics.

[0149] S7.2: Perform a time-series splicing operation based on the current defect state label and the previous defect state label recorded at the previous moment to generate a candidate real-time state sequence fragment containing the latest evolution information.

[0150] S7.3: Perform sliding window truncation processing on the candidate real-time state sequence fragments to remove outdated state data that exceeds the preset memory depth and generate a standardized real-time state sequence that meets the fixed length requirement.

[0151] The stage transition memory is a cyclic data structure used to store and continuously update the defect state sequence at consecutive moments during the drilling process, supporting real-time matching and early warning triggering of subsequent state transition patterns. This memory uses a fixed-length queue (FIFO, First-In-First-Out) as its underlying storage model, ensuring that only the most recently recorded state within a preset memory depth is retained in each update cycle, avoiding the interference of unlimited accumulation of historical data on pattern recognition.

[0152] The data structure of the stage transfer memory includes the following fields: State Label: Integer encoding (0-3) representing four defect states: normal drilling, microcrack initiation, accelerated borehole wall tearing, and abnormal drill bit wear; Timestamp: A sampling timestamp accurate to milliseconds, used to calculate the temporal interval between states; Confidence Value: The maximum probability value in the calibrated probability distribution corresponding to the state, used for weight evaluation during subsequent pattern matching; Probability Distribution Snapshot (optional): A complete probability distribution vector for the four states, used for deep pattern analysis and offline backtracking. The storage capacity of this memory is determined by the preset memory depth parameter N. mem The value is typically between 8 and 32, with 20 being preferred, meaning that at most the most recent N values ​​are retained at any given time. mem Each status record.

[0153] The stage transfer memory uses an atomic write-up strategy, and the specific operation process is as follows: Step A: Combine the defect status label determined by maximum likelihood estimation at the current moment with the corresponding timestamp and confidence value to form a new status record unit; Step B: Append the state record unit to the tail of the memory queue; Step C: If the queue length exceeds the preset memory depth N mem If so, the earliest state record is removed from the head of the queue; Step D: Perform a time sequence continuity check on the updated queue to ensure that the time interval between adjacent records is consistent with the preset sampling interval; Step E: Use the updated queue as the post-update stage transfer memory.

[0154] The stage transfer memory can be deployed in the memory of edge computing nodes, implemented using a circular buffer to reduce memory footprint and improve read / write efficiency. The memory data can be periodically persisted to local storage or a cloud database for offline process analysis and model iteration optimization.

[0155] The storage length of the candidate real-time state sequence fragment formed by splicing the current defect state label with the previous defect state label is calculated to determine the total number of elements in the current sequence fragment.

[0156] The preset memory depth parameter is compared with the total number of current elements to generate a set of outdated state data indexes that are out of range.

[0157] The old state data index set is used to perform a deletion operation from the candidate real-time state sequence fragments, removing all historical state tags that exceed the preset memory depth.

[0158] The remaining state labels are rearranged according to their timestamps to satisfy the strict temporal continuity constraint under a fixed length.

[0159] A standardized real-time state sequence data structure is generated based on the rearranged set of state labels, ensuring that the data structure maintains the same length and time ordering characteristics in each update cycle.

[0160] By using sliding window truncation and rearrangement, the candidate real-time state sequence fragments generated in the previous step are transformed into standardized real-time state sequences that meet the fixed length requirements and have complete temporal continuity, thereby stabilizing and ensuring structural consistency of the input data of the stage transition memory.

[0161] For example, in a real-time monitoring scenario for PCB drilling, the candidate real-time state sequence segment has a length of 12 state tags, a preset memory depth parameter of 8, and sequence elements are recorded at a sampling period of 0.5 seconds. Four tags exceeding the length are determined through length calculation, with an index set of [1,2,3,4], corresponding to timestamps 4 seconds earlier than the current time. After performing a deletion operation, the 8 most recent state tags are retained, arranged sequentially from earliest to latest timestamp as [microcrack initiation, normal drilling, normal drilling, microcrack initiation, accelerated hole wall tearing, accelerated hole wall tearing, abnormal drill bit wear, abnormal drill bit wear]. This set is encapsulated into a standardized real-time state sequence of fixed length 8 and transmitted to the stage transition memory update module. In this embodiment, fixed-length truncation avoids excessive accumulation of historical states, ensuring the effectiveness and real-time performance of stage transition pattern matching, thereby significantly improving the accuracy and response speed of early defect warning.

[0162] S7.4: Utilize the standardized real-time state sequence to perform atomic write update operations on the stage transfer memory of the rolling phase, so as to replace the old historical state records in the memory and generate an updated stage transfer memory that reflects the latest drilling process evolution trajectory.

[0163] S7.5: Perform sequence integrity verification processing based on the continuous defect states stored in the updated stage transition memory to confirm that the real-time state sequence is ready for subsequent steps to perform specific state transition pattern matching and judgment.

[0164] Index position analysis is performed on the continuous defect state records in the updated stage transfer memory to extract a fixed-length real-time state sequence and confirm that its time span is consistent with the preset memory depth.

[0165] The time interval difference between adjacent status tags is calculated based on the real-time status sequence. The continuity of the time interval is verified by linear scanning and there are no missing records. When an abnormal interval occurs, the integrity flag is set to fail.

[0166] A tag set consistency check is performed on the real-time state sequence. The validity of the state tags at each position is compared using a Boolean mask to ensure that all tags belong to one of the four states: normal drilling, microcrack initiation, accelerated hole wall tearing, or abnormal drill bit wear.

[0167] Calculate the logical continuity index of the real-time state sequence to quantify the continuous state coverage.

[0168] A threshold comparison operation is performed based on the coverage index. If the coverage is greater than the preset threshold, the integrity flag is set to pass; otherwise, it is set to fail.

[0169] Through sequence integrity verification, the post-update stage transfer memory of the previous step is transformed into a standardized real-time state sequence with temporal continuity and label validity, thus providing reliable input data for subsequent specific state transition pattern matching and judgment.

[0170] For example, during PCB drilling, the sampling interval is set to 1 millisecond, and the memory depth is set to 20 records. After the update, the stage transition memory contains the 20 most recent status tags and their corresponding timestamps. The status tags are, in order, normal drilling, microcrack initiation, etc., and the timestamps are continuous values ​​accurate to milliseconds. After performing index position analysis, a sequence of length 20 is extracted. Linear scanning confirms that the interval between adjacent records is 1 millisecond, and Boolean mask checks show that all tags are one of the four states. When calculating the coverage index, an indicator function is executed for the time difference between all adjacent records. When all time differences are equal to 1 millisecond, the indicator function outputs 1, and the sum is 20, indicating the coverage. Setting the value to 1 significantly improves the reliability of the sequence, ultimately setting the integrity flag to pass, and providing this standardized real-time status sequence to the pattern matching module for use, thus ensuring sufficient conditions for subsequent early warning judgments.

[0171] Step S8: Determine whether the real-time state sequence contains a specific state transition pattern. If the specific state transition pattern is satisfied, trigger an early warning signal and simultaneously push the corresponding stage's feature heatmap. Specifically, this includes: S8.1: Perform sliding window truncation processing on the probability distribution of continuous defect states in the real-time state sequence to obtain a combination of state labels containing the current time and multiple previous historical times, forming a state transition segment to be analyzed.

[0172] S8.2: Based on the state transition fragment, perform a predefined rule engine matching operation, and use the stored specific state transition pattern library to perform logical verification on the jump path of adjacent state labels to generate a pattern matching result that represents whether the target evolution path is hit.

[0173] The state transition segments are initialized using a rule engine environment variable, mapping the state label combinations to be analyzed into an internal encoding sequence recognizable by the rule matching module. Adjacent state label relationships are parsed using this internal encoding sequence to generate a state transition path vector arranged chronologically, which serves as input data for logical verification. Based on the state transition path vector, a pre-set specific state transition pattern library is invoked, and hash indexing is used to locate the corresponding pattern entry and extract the pattern rule set. A step-by-step logical matching process is performed on the extracted pattern rule set, verifying the equality and directionality of adjacent label pairs in the path vector against the expected label pairs in the pattern rule set, recording the result state flag bit for each match. A comprehensive judgment operation is performed using the matching flag bit sequence to calculate the matching score. The matching score is compared with a preset hit threshold to generate pattern matching result data representing whether the target evolution path has been hit. Through the rule engine matching process, the state transition segments from the previous step are transformed into pattern matching results containing hit / miss information, enabling rapid identification of symbolic patterns in the borehole defect evolution path.

[0174] For example, in a PCB drilling online monitoring system, the length L of the state transition segment is set to 5, and adjacent label pairs include (normal drilling → microcrack initiation), (microcrack initiation → accelerated hole wall tearing), and (accelerated hole wall tearing → accelerated hole wall tearing). The rule engine internally enumerates and maps the labels as {0: normal drilling, 1: microcrack initiation, 2: accelerated hole wall tearing, 3: abnormal drill wear}. The evolution path of a specific target in the pattern library is defined as the encoded sequence [0,1,2], requiring continuous unidirectional jumps. After path vector parsing, the encoded sequence [0,1,2,2,2] is generated, with a matching flag of [1,1,0,0]. The calculated matching score is 0.4, higher than the preset hit threshold of 0.3, indicating a successful pattern match. This process triggers subsequent early warning signal output in the system. Verification results show that it can achieve significant early warning before defects enter the accelerated hole wall tearing stage, improving the early protection response speed.

[0175] S8.3: Perform Boolean logic judgment processing based on the pattern matching result. When it is detected that the status label has undergone a continuous unidirectional transition from the normal drilling state through the microcrack initiation state to the hole wall tearing acceleration state, an activation command to trigger the early warning signal is output.

[0176] Binarization is performed on the matching results data of specific state transition patterns to generate a Boolean variable matrix containing only hit and miss judgment values.

[0177] The current hit flag recorded in the Boolean variable matrix is ​​compared element-wise with the preset logic condition vector to form a matching compliance vector representing whether the current state sequence conforms to the warning trigger path.

[0178] The matching conformity vector is summed cumulatively. When the cumulative value reaches a predetermined threshold, a logical true condition identifier is generated, indicating that the unidirectional transition path from the normal drilling state through the microcrack initiation state to the hole wall tearing acceleration state is continuously satisfied.

[0179] A conditional decision function is used to confirm the state of the logical true condition flag, ensuring that there is no reverse transition or interruption event when the threshold is reached.

[0180] When the condition determination result is true, an activation instruction data packet containing the trigger timestamp and the current sequence index is constructed as input for subsequent feature backtracking module calls.

[0181] By using Boolean logic judgment and condition verification, the pattern matching results are transformed into executable early warning trigger commands, thereby realizing the early warning function of defects based on real-time status sequences.

[0182] For example, in a real-time monitoring scenario for PCB drilling, a specific state transition mode library defines a continuous unidirectional transition path between the normal drilling state (N), the microcrack initiation stage (M), and the hole wall tearing acceleration state (T) as the warning trigger condition. A matching compliance threshold of 2 is set, meaning that the sequence N→M→T needs to be satisfied twice consecutively to trigger the warning. In this scenario, the real-time state sequence is [N, M, T, N, M, T], which, after binarization, forms a Boolean matrix [1, 1, 1, 0, 1, 1]. This matrix is ​​compared element-wise with the logical condition vector [1, 1, 1]. The first round of matching compliance vector is [1, 1, 1]. The cumulative summation result... The threshold condition is met, and a logical true identifier is generated. The second round of matching uses a conformity vector [0, 1, 1], and the cumulative summation result is... The threshold condition is also met. The condition determination function confirms that no reverse transition or interruption occurred in either round, and finally constructs an activation instruction data packet containing the trigger timestamp (2024-06-01 10:15:23) and sequence index (0, 3), and calls the feature backtracking module at that moment. Verification shows that this embodiment can issue an early warning signal before changes in the acoustic emission signal significantly affect the processing quality during drilling, significantly improving the forward-looking judgment capability of the monitoring system.

[0183] S8.4: In response to the activation instruction, the feature backtracking module is invoked to extract the topological evolution representation vector and associated weight data corresponding to the current hit state transition mode from the node attributes of the local temporal subgraph, so as to construct the original feature mapping set.

[0184] S8.5: Perform multi-dimensional data visualization rendering processing on the original feature mapping set, map the topological connection strength and temporal evolution position attributes into color gradients and spatial coordinates, generate heat map data that intuitively displays the defect development stage and push it to the monitoring terminal.

[0185] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0186] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0187] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A real-time PCB drilling defect identification method based on acoustic emission signals, specifically including: S1: During PCB drilling, the raw voltage signal output by the acoustic emission sensor is continuously acquired at a fixed sampling rate, and the raw voltage signal is slicing according to a preset ultra-short time window to generate a high-density acoustic emission signal fragment stream. S2: Perform adaptive amplitude normalization and multi-scale envelope extraction processing on each signal segment in the high-density acoustic emission signal segment stream to obtain the local response curve reflecting the transient energy distribution; S3: Construct a dynamic similarity matrix based on a preset number of adjacent local response curves as initial adjacency relationship data; S4: Using the local response curve corresponding to the current signal segment as the root node, backtrack a preset number of historical local response curves, and construct a local time series subgraph focusing on short-range evolution correlation using the initial adjacency relationship data with the introduction of a time decay factor. S5: Perform graph convolution operation on the local temporal subgraph to aggregate neighbor node information and implicitly encode the relative position of the current segment in the defect development path as a topological evolution representation vector. S6: Input the topological evolution representation vector into a lightweight classifier for stage discrimination and output the probability distribution of defect states; S7: Update the stage transition memory of the rolling stage based on the probability distribution of defect states generated multiple times in succession, and record the real-time state sequence composed of different defect states; S8: Determine whether the real-time state sequence contains a specific state transition pattern. If it is determined that the specific state transition pattern is met, trigger an early warning signal and simultaneously push the characteristic heat map of the corresponding stage.

2. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 1, characterized in that, The method of constructing a dynamic similarity matrix based on a preset number of adjacent local response curves as initial adjacency relationship data involves calculating the topological connection weight between each pair of curves based on the amplitude change slope, zero-crossing rate difference, and peak delay relationship, and then constructing a dynamic similarity matrix as initial adjacency relationship data.

3. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 1, characterized in that, The probability distribution of defect states represents four types of states: normal drilling, microcrack initiation, accelerated hole wall tearing, or abnormal drill bit wear.

4. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 1, characterized in that, Step S3 specifically includes: The first-order differential is performed on the local response curves of the current moment and the preset number of adjacent curves to obtain the amplitude change slope sequence that characterizes the transient energy change. This amplitude change slope sequence is then used as the basic input variable for subsequent calculation of topological connectivity strength. Generate zero-crossing rate difference parameters based on the amplitude change slope sequence; The extreme points of the local response curves are located by using the envelope peak search technique, the time interval between the occurrence of the main peaks of adjacent curves is calculated, peak delay relationship data characterizing the energy propagation hysteresis characteristics is generated, and this peak delay relationship data is used as the core basis for evaluating the temporal causality of defect evolution. Based on the amplitude change slope sequence, zero-crossing rate difference parameter and peak delay relationship data, the weighted Euclidean distance metric method is used to calculate the multidimensional feature deviation between each pair of local response curves, generate the original connection weight value of the quantification curve similarity level, and use the original connection weight value as the direct material for constructing the edge weight of the graph structure. A dynamic similarity matrix is ​​generated based on the original connection weight values.

5. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 4, characterized in that, The method of generating zero-crossing rate difference parameters based on amplitude change slope sequence specifically involves performing zero-crossing detection processing based on amplitude change slope sequence, statistically analyzing the time offset of zero-crossing point positions between adjacent local response curves, generating zero-crossing rate difference parameters that reflect the degree of waveform phase mismatch, and using these zero-crossing rate difference parameters as a key indicator for measuring signal morphology consistency.

6. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 4, characterized in that, The process of generating a dynamic similarity matrix based on the original connection weight values ​​involves performing a matrix arrangement operation based on the original connection weight values, mapping the discrete pairwise connection weights into a two-dimensional square matrix structure, generating a dynamic similarity matrix that fully describes the strength of the topological association between all adjacent local response curves, and outputting this dynamic similarity matrix as the initial adjacency relationship data for use by the local time series subgraph construction module.

7. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 1, characterized in that, Step S4 specifically includes: Root node mapping is performed on the local response curve corresponding to the current signal segment, and neighbor node mapping is performed on the historical local response curves that are traced back a preset number of times to generate a local time-series subgraph node set containing the root node and multiple neighbor nodes. The time interval difference is calculated based on the timestamp information of each node in the local temporal subgraph node set, and the time interval difference is processed by nonlinear transformation using the exponential decay function to generate a time decay factor sequence that represents the decreasing weight of the influence of historical nodes on the current node over time. We perform weighted correction processing on the dynamic similarity matrix elements in the initial adjacency relationship data, and use the time decay factor sequence to perform dynamic suppression operation on the topological connection weights between each pair of nodes to generate a weighted adjacency matrix that focuses on short-range evolutionary associations. Graph structure assembly is performed based on the local temporal subgraph node set and weighted adjacency matrix. The node feature vectors are topologically bound to the corrected edge weights to generate a local temporal subgraph focusing on short-range evolutionary associations.

8. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 1, characterized in that, Step S5 specifically includes: A linear projection transformation is performed on the root node feature embedding vector and the neighbor node feature embedding vector in the local temporal subgraph. Based on the learnable weight matrix, the original feature space is mapped to a high-dimensional latent feature space to generate an initial node hidden state representation with nonlinear expressive power. A standardized edge direction attention weight distribution is generated based on the initial node hidden state representation and the initial adjacency relationship data with the introduction of a time decay factor. A weighted aggregation operation is performed on the standardized edge direction attention weight distribution and the neighbor node feature embedding vector to fuse the transient energy distribution features of multi-source historical fragments based on the connection strength difference, thereby generating an intermediate layer node aggregation feature vector containing short-range evolutionary association information. Deep semantic defect evolution path feature embedding is generated based on aggregated feature vectors from intermediate layer nodes; Global pooling and dimensionality compression are performed on the deep semantic defect evolution path feature embedding to extract a low-dimensional dense vector that reflects the relative position attributes of the current segment throughout the entire process from microcrack initiation to hole wall tearing acceleration, and finally generate a topological evolution representation vector for stage discrimination.

9. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 8, characterized in that, The process of generating a standardized edge-direction attention weight distribution based on the initial node hidden state representation and the initial adjacency relationship data with the introduction of a time decay factor involves performing attention coefficient calculation processing based on the initial node hidden state representation and the initial adjacency relationship data with the introduction of a time decay factor, so as to quantify the dynamic association strength between the root node and each historical neighbor node using the leaky rectified linear unit activation function and the normalized exponential function, thereby generating a standardized edge-direction attention weight distribution.

10. The real-time PCB drilling defect identification method based on acoustic emission signals according to claim 8, characterized in that, The method of generating deep semantic defect evolution path feature embedding based on aggregated feature vectors of intermediate layer nodes specifically involves performing multi-level feature concatenation and nonlinear activation processing based on aggregated feature vectors of intermediate layer nodes. This process abstracts the local topological structure of acoustic emission signals layer by layer through cascading multiple lightweight graph convolutional layers, thereby generating deep semantic defect evolution path feature embedding.

Images