PCB drill bit micro-chipping online detection and early warning method

Abstract

The present application relates to a kind of PCB drill bit micro-chipping online detection and early warning method.For the problem of difficult tool state identification under different materials and working conditions, a kind of high-speed vision, current, vibration and displacement signal is synchronously collected, a high-resolution multi-modal data set covering typical micro-chipping evolution stage is constructed, and based on normalized feature fusion, lightweight time coding and fingerprint database establishment, high-fidelity representation and online dynamic matching of micro-chipping state are realized.The system accurately identifies and predicts the evolution trend through dynamic time warping and Bayesian inference, and realizes adaptive early warning feedback loop by linking machining control system, while supporting incremental optimization of fingerprint database based on artificial review.The scheme effectively improves the sensitivity and discrimination accuracy of drill bit health state identification, and provides technical support for efficient and intelligent machining process monitoring.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and intelligent early warning technology for PCB manufacturing processes, and in particular to an online detection and early warning method for micro-chipping of PCB drill bits. Background Technology

[0002] With the increasing demands for precision and high integration density in the PCB manufacturing industry, real-time detection and early warning of micro-chipping anomalies in the drilling process has become a crucial direction for the intelligent transformation of production lines. Currently, most mainstream PCB drill bit health monitoring and anomaly early warning systems use fixed threshold methods or empirical criterion models as the basis for online judgment of drill bit micro-chipping conditions. Common solutions include setting static thresholds for drilling cycle force signals, current envelopes, acceleration spectra, or image features, or adjusting judgment criteria based on statistical analysis under specific operating conditions. These technical approaches are easy to deploy and can achieve basic anomaly alerts under single material, constant speed, and thickness conditions, possessing a certain degree of engineering universality. However, with the expansion of the drilling process material spectrum (such as FR-4, high-frequency ceramics, composite substrates, etc.), and the rise of batch-oscillating board thicknesses, diverse drilling speed ranges, and complex tool load modes, the fixed threshold method has revealed bottlenecks such as poor adaptability, frequent false alarms, potential for missed alarms, and inability to track the entire process of tool failure evolution in a timely manner.

[0003] Specifically, the existing technology exhibits significant shortcomings in the following aspects: First, the static threshold method for judging the health status of drill bits is highly dependent on the absolute value of a single physical quantity (such as peak current, root mean square value, edge changes in a single frame image, etc.). Once the processing material is changed, the thickness fluctuates, or the spindle speed is adjusted, the judgment standard becomes invalid, resulting in the inability to guarantee the consistency of product quality within and outside the batch. This problem is particularly prominent on PCB production lines for high-frequency microvias and complex composite materials, which not only increases the cost of manual inspection and offline verification, but also restricts the automated continuous production capability of the production line.

[0004] Second, some improved dynamic threshold technologies attempt to adaptively adjust criteria through historical distribution or use clustering algorithms for anomaly classification. However, when faced with the multimodal, multivariate, and highly time-varying evolution of drill wear and micro-chipping processes, they often fall into complex dilemmas such as model drift, inaccurate criteria, and parameter accumulation. Especially in actual production, tool failure modes exhibit strong temporal and physical evolution stage characteristics. Simple numerical limit exceedance detection cannot effectively distinguish physical state transitions such as initial disturbances, subcritical initiation, and critical expansion, making it difficult to provide factories with reliable, segmented intelligent early warning capabilities.

[0005] Third, some existing equipment integrates simple multi-sensor fusion algorithms, but lacks explicit modeling of the micro-chipping evolution stage of drill bit at the feature fusion and state discrimination levels. It cannot effectively correspond dynamic data across working conditions and multiple cycles with physical evolution laws, resulting in weak generalization ability of the early warning system when facing new materials or new processes, leading to high false alarm rates or missed detection of key anomalies. Summary of the Invention

[0006] This application provides an online detection and early warning method for micro-chipping of PCB drill bits, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0007] This application provides an online detection and early warning method for micro-chipping of PCB drill bits, specifically including: S1: For drilling processes of various materials such as copper foil, FR-4 and high-frequency ceramic substrates under different thicknesses and drilling speeds, the micro-chipping image sequence, spindle current transient waveform, Z-axis feed vibration spectrum and drilling displacement trajectory of the drill bit throughout its entire life cycle are collected simultaneously to construct a multimodal raw dataset containing four types of evolution state annotations, from sub-pixel level edge perturbation to local edge stripping precursors.

[0008] S2: Based on the working condition labels of different material and drilling speed combinations in the multimodal original dataset, the gray-scale gradient change rate of image blocks, the frequency of abrupt changes in the slope of the current envelope, and the energy decay trend of the Z-axis acceleration power spectrum in a specific frequency band are jointly normalized to generate a multidimensional time-series feature vector that eliminates the differences in working condition background.

[0009] S3: Using the multidimensional temporal feature vector as training samples, a high-fidelity multimodal stage fingerprint database is constructed by mapping the features within multiple consecutive drilling cycles into fixed-length embedded representations through a lightweight temporal encoder, which can characterize the physical evolution of micro-chipping edges from their inception to critical failure.

[0010] S4: Based on the original multimodal joint temporal pattern input in real time during online runtime, start a lightweight temporal encoder to extract a real-time temporal semantic vector with a length of 64 dimensions.

[0011] S5: Using an improved dynamic time warping algorithm, the real-time temporal semantic vector is compared with the evolutionary stage template in the high-fidelity multimodal stage fingerprint database to perform similarity retrieval, and the state matching result including matching confidence and stage transition probability is output.

[0012] S6: Based on the matching confidence and stage transition probability in the state matching result, activate the processing logic of the corresponding evolution stage. If it is identified as the single-tooth microcrack initiation stage and the matching confidence meets the preset threshold, the dual-window sliding verification mechanism is enabled. If it is identified as the multi-tooth collaborative collapse and expansion stage and the stage transition probability increases sharply in a short period of time, the secondary warning instruction is directly triggered.

[0013] S7: Based on the secondary warning instruction or the primary warning signal confirmed by the dual-window sliding verification mechanism, push the drill bit replacement suggestion to the MES work order system and execute the corresponding processing pause or speed reduction control action to complete the adaptive warning feedback closed loop based on state transition.

[0014] S8: Receive manual verification and confirmation information from the operator regarding the early warning execution results, encrypt and upload samples confirmed as genuine anomalies along with their original multimodal time-series data to the edge training node, and incrementally update the boundary criteria of the high-fidelity multimodal stage fingerprint database based on the newly added samples to optimize the fingerprint self-evolution capability under subsequent working conditions.

[0015] The online detection and early warning method for micro-chipping of PCB drill bits provided in this application has the following beneficial effects: (1) By constructing a high-fidelity multimodal stage fingerprint database and introducing an evolution stage identification mechanism based on temporal semantic matching, this scheme significantly improves the accuracy and adaptability of drill bit micro-chipping state identification. Compared with the traditional method that relies on a single threshold criterion of spindle current or vibration signal, which is easily affected by changes in material type, plate thickness and rotation speed, resulting in false alarms and missed alarms, this invention divides micro-chipping into four evolution stages with clear physical meaning, from "sub-pixel level edge disturbance" to "precursor of local edge stripping". It also combines multi-dimensional features such as image grayscale gradient, frequency of abrupt changes in current envelope slope and high-frequency vibration energy attenuation along the Z-axis to form a 64-dimensional temporal semantic vector with strong discriminative power. This vector achieves high-precision template matching in a pre-set fingerprint database through an improved dynamic time warping algorithm. It not only effectively captures the nonlinear temporal pattern of micro-chipping development, but also overcomes the matching inaccuracy problem caused by signal amplitude drift and phase shift under different processing conditions. This enables the system to maintain stable discrimination capability in scenarios with different materials (such as copper foil, FR-4, and high-frequency ceramic substrates) and varying parameters (0.2–3.2 mm thickness, 80k–180k rpm drilling speed), which greatly improves the sensitivity and robustness of early micro-defect detection.

[0016] (2) Through the hierarchical decision-making logic driven by the design phase and the dual-window sliding verification mechanism, this solution realizes the dynamic, refined and reliable early warning response, and effectively solves the common problems in the existing technology such as rigid early warning strategies, high false alarm rate and delayed response in key stages. In view of the different risk levels in different evolution stages, the system adopts differentiated decision rules: for the initial abnormal state of "single tooth microcrack initiation", continuous periodic consistency verification is further introduced on the basis of matching confidence meeting the standard to avoid false triggering caused by occasional noise or transient disturbances; when "multi-tooth collaborative collapse expansion" is identified and the stage transition probability rises rapidly in a short period of time, the secondary early warning is immediately activated and directly connected to the MES work order system, skipping the delayed confirmation process to ensure the timeliness of response. This intelligent decision-making model, based on evolutionary trends rather than static thresholds, ensures the stability of early warnings in low-risk phases and guarantees that high-risk phases can be captured in milliseconds in advance. It fully meets the stringent requirements of PCB production lines for a true positive rate of over 99.2% and an extremely low false alarm rate, significantly outperforming the performance boundaries that traditional fixed thresholds or simple statistical models struggle to achieve.

[0017] (3) By integrating a fingerprint self-evolution module and a structured knowledge graph storage architecture, this solution constructs a closed-loop diagnostic system that is continuously optimized, highly interpretable, and easily scalable, effectively overcoming the problem that existing models are difficult to adapt to new processes and substrates after deployment, leading to degradation in generalization ability. Whenever an operator completes manual review and confirms that the warning is effective, the corresponding sample and its original multimodal data will be encrypted and uploaded to the edge training node for local incremental update of the boundary criteria in the fingerprint database, especially strengthening the learning ability of micro-chipping features under new composite substrates such as PTFE+ceramic hybrid plates, thereby realizing the online evolution of the model. At the same time, all stage fingerprints are organized in the form of a knowledge graph, including five types of attribute nodes such as typical texture patterns, sensitive frequency bands, verification window length, and false alarm suppression strategies, supporting fast retrieval and strategy reloading by material, rotation speed, or historical performance, greatly improving the maintainability and engineering adaptation efficiency of the system. This design gets rid of the dependence of traditional methods on complex parameter tuning and centralized training, and has the advantages of lightweight inference and long-term operational stability, providing reliable tool support with self-learning capabilities for intelligent manufacturing of high-density interconnect boards.

[0018] In summary, this solution establishes a multimodal stage fingerprint system oriented towards the physical evolution process, integrating temporal semantic matching, stage-aware decision-making, and edge self-evolution mechanism. This achieves a paradigm shift from "signal alarm" to "state understanding," significantly improving the accuracy, real-time performance, and cross-condition adaptability of micro-chipping identification. Furthermore, it constructs an intelligent monitoring closed loop with continuous learning capabilities, high interpretability, and industrial-grade robustness, fully meeting the core requirements of modern PCB production lines for millisecond-level response, high anode rate early warning, and low maintenance costs in precision machining processes. Attached Figure Description

[0019] Figure 1 This is the main flowchart of an online detection and early warning method for micro-chipping of PCB drill bits.

[0020] Figure 2 This is a sub-flowchart of a method for online detection and early warning of micro-chipping of PCB drill bits.

[0021] Figure 3 This is another sub-flowchart of a method for online detection and early warning of micro-chipping of PCB drill bits. Detailed Implementation

[0022] 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.

[0023] 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.

[0024] like Figure 1 As shown, this application provides an online detection and early warning method for micro-chipping of PCB drill bits, specifically including: S1: For drilling processes of various materials such as copper foil, FR-4 and high-frequency ceramic substrates under different thicknesses and drilling speeds, the micro-chipping image sequence, spindle current transient waveform, Z-axis feed vibration spectrum and drilling displacement trajectory of the drill bit throughout its entire life cycle are collected simultaneously to construct a multimodal raw dataset containing four types of evolution state annotations, from sub-pixel level edge perturbation to local edge stripping precursors.

[0025] S2: Based on the working condition labels of different material and drilling speed combinations in the multimodal original dataset, the gray-scale gradient change rate of image blocks, the frequency of abrupt changes in the slope of the current envelope, and the energy decay trend of the Z-axis acceleration power spectrum in a specific frequency band are jointly normalized to generate a multidimensional time-series feature vector that eliminates the differences in working condition background.

[0026] S3: Using the multidimensional temporal feature vector as training samples, a high-fidelity multimodal stage fingerprint database is constructed by mapping the features within multiple consecutive drilling cycles into fixed-length embedded representations through a lightweight temporal encoder, which can characterize the physical evolution of micro-chipping edges from their inception to critical failure.

[0027] S4: Based on the original multimodal joint temporal pattern input in real time during online runtime, start a lightweight temporal encoder to extract a real-time temporal semantic vector with a length of 64 dimensions.

[0028] S5: Using an improved dynamic time warping algorithm, the real-time temporal semantic vector is compared with the evolutionary stage template in the high-fidelity multimodal stage fingerprint database to perform similarity retrieval, and the state matching result including matching confidence and stage transition probability is output.

[0029] S6: Based on the matching confidence and stage transition probability in the state matching result, activate the processing logic of the corresponding evolution stage. If it is identified as the single-tooth microcrack initiation stage and the matching confidence meets the preset threshold, the dual-window sliding verification mechanism is enabled. If it is identified as the multi-tooth collaborative collapse and expansion stage and the stage transition probability increases sharply in a short period of time, the secondary warning instruction is directly triggered.

[0030] S7: Based on the secondary warning instruction or the primary warning signal confirmed by the dual-window sliding verification mechanism, push the drill bit replacement suggestion to the MES work order system and execute the corresponding processing pause or speed reduction control action to complete the adaptive warning feedback closed loop based on state transition.

[0031] S8: Receive manual verification and confirmation information from the operator regarding the early warning execution results, encrypt and upload samples confirmed as genuine anomalies along with their original multimodal time-series data to the edge training node, and incrementally update the boundary criteria of the high-fidelity multimodal stage fingerprint database based on the newly added samples to optimize the fingerprint self-evolution capability under subsequent working conditions.

[0032] Step S1: For drilling processes of various materials such as copper foil, FR-4, and high-frequency ceramic substrates under different thicknesses and drilling speeds, simultaneously collect micro-chipping image sequences, spindle current transient waveforms, Z-axis feed vibration spectra, and drilling displacement trajectories throughout the entire lifecycle of the drill bit. Construct a multimodal raw dataset containing annotations for four evolutionary states, from sub-pixel-level edge perturbation to precursors of local edge stripping. Specifically, this includes: S1.1: For processing scenarios involving various substrates such as copper foil, FR-4, and high-frequency ceramic substrates with different thicknesses and drilling speed combinations, a vision acquisition unit consisting of a high-speed linear array camera and a stroboscopic light source is configured to capture continuous frame images of the drill bit edge area during the drilling process, generating a raw data stream of micro-chipping image sequences containing sub-pixel level details.

[0033] For processing scenarios involving various substrates such as copper foil, FR-4, and high-frequency ceramic substrates under different thicknesses and drilling speed combinations, a vision acquisition unit consisting of a high-speed linear array camera and a stroboscopic light source is configured. The optimal focal length adaptive adjustment method (parameter: the focal plane position is automatically adjusted based on the maximum gradient value of the cutting edge area) is adopted to achieve precise focusing and capture of the cutting edge area.

[0034] By using a high-frequency synchronous triggering technique (parameter: the pulse period of the stroboscopic light source is matched with the camera's line scan period to ±0.1µs), the transient motion freeze of the cutting edge under high-speed rotation is achieved, resulting in continuous frame images without motion blur.

[0035] A subpixel optical calibration method is adopted (parameters: calibration board resolution is not less than 2µm, distortion compensation coefficient is updated in real time according to radial distortion model) to achieve accurate mapping and correction of pixel position and generate local area image data of the cutting edge that meets geometric accuracy requirements.

[0036] By using a multi-scale edge enhancement method (parameters: Laplace enhancement combined with Gaussian difference, scale parameter is 0.5–2.0 mm), the micro-chipping morphology of the cutting edge is significantly visualized, and a high-contrast feature texture image is generated.

[0037] A continuous frame time series capture method (parameters: sampling frequency ≥ 40kHz, frame buffer window length of 5–8 drilling cycles) is used to achieve dynamic tracking of the evolution of the cutting edge state and form a raw data stream of micro-chipping image sequence containing sub-pixel level detail change patterns.

[0038] Through multi-stage collaborative processing of the high-speed vision acquisition unit, the optical capture results of the previous step are transformed into micro-crack image data with complete time series and sub-pixel spatial accuracy, thus achieving the basic input effect for the subsequent construction of multimodal temporal data blocks.

[0039] For example, in drilling FR-4 sheet metal with a thickness of 1.6 mm and a spindle speed of 120 rpm, a high-speed linear scan camera was configured with an exposure time of 5 µs, a focal length adjusted to 28.3 mm to match the sharp gradient peaks in the cutting edge region, and a stroboscopic light source pulse period set to 25 µs, synchronized with the camera scanning cycle to ±0.05 µs. The calibration board resolution was 2 µm, and the distortion compensation coefficient was dynamically updated to 0.0025 to correct radial aberrations. Multi-scale enhancement was performed using the Laplace operator and a Gaussian difference filter with σ=1.25 mm. A sequence of cutting edge images for seven consecutive drilling cycles was captured at a sampling frequency of 40 kHz, with a resolvable chipping boundary change of approximately 0.8 µm per frame within the sequence. This image sequence was verified to possess high contrast and sub-pixel-level spatial positioning accuracy, and significantly improves the sensitivity of subsequent force-motion coupling feature extraction.

[0040] S1.2: Based on the acquisition timestamp of the original data stream of the micro-chipping image sequence, the high-precision current sensor, triaxial accelerometer and grating ruler displacement sensor are synchronously triggered to acquire the transient waveform data of the spindle current, the Z-axis feed vibration spectrum data and the drilling displacement trajectory data at the corresponding time, forming a multi-source physical signal set that is strictly time-aligned with the image data.

[0041] Based on the acquisition timestamp of the original data stream of the micro-blow-off image sequence, a time-triggered multi-source signal synchronization method (parameters: acquisition timestamp accuracy ≤ 1μs, synchronization delay tolerance ≤ 2μs) is adopted to realize the sampling start control of the high-precision current sensor.

[0042] By employing a sensor synchronization strategy that combines static calibration and dynamic correction (parameters: calibration cycle 10 min, dynamic drift compensation coefficient 0.98), the triaxial accelerometer is started synchronously, achieving millisecond-level alignment with the image acquisition time, and obtaining the raw data of the Z-axis feed vibration spectrum.

[0043] By using a phase-locked coding triggering method (parameters: grating ruler resolution 0.1μm, phase-locking accuracy ≤0.05°), the sampling synchronization of the grating ruler displacement sensor is achieved, the borehole displacement trajectory data at the corresponding moment is obtained, and strict time alignment with the image and other signals is ensured.

[0044] The event tag embedding mechanism inside the multi-source acquisition controller (parameters: event tag word length 64 bits, timestamp bit width 48 bits) is adopted to achieve deep binding between the acquisition timestamp and various physical signal metadata, and generate a structured multi-source physical signal raw dataset index.

[0045] By using a multi-channel time alignment verification method (parameters: maximum allowable deviation ≤ 500ns, cross-correlation coefficient threshold ≥ 0.99), the transient current waveform, vibration spectrum and displacement trajectory from the previous step are strictly time aligned with the micro-blow-off image data to form a complete multi-source physical signal set, thereby achieving the expected technical effect of synchronization of multi-modal sensing input.

[0046] For example, under processing conditions of a copper foil substrate thickness of 0.8 mm and a drilling speed set to 120,000 rpm, the high-speed camera acquisition timestamp accuracy for the micro-chipping image sequence is configured to 0.5 μs, the sampling frequency of the high-precision current sensor is 500 kHz, the sampling frequency of the triaxial accelerometer is 200 kHz, and the sampling frequency of the grating ruler displacement sensor is 50 kHz. During execution, current sampling is initiated at each image acquisition using a time-triggered synchronization method, and the vibration data time axis is corrected using a dynamic drift compensation coefficient of 0.98. Grating ruler displacement acquisition is precisely initiated using a phase-locked coding trigger method to obtain the trajectory curve of displacement change within each drilling cycle. To verify time alignment, a cross-correlation coefficient is used. Calculation formula: in, For the i-th sample point of the current transient waveform sequence, For the i-th sample point of the vibration spectrum energy sequence, The average value of the transient current waveform sequence. The average value of the vibration spectrum energy sequence. The number of sample points is given. The calculated cross-correlation coefficient reached 0.995, far exceeding the threshold of 0.99, verifying the strict temporal consistency between the multi-source physical signal set and the image data. The final output multi-source physical signal set contains complete current transient waveforms, Z-axis vibration spectrum, and displacement trajectory, which significantly improves the accuracy of the working condition feature representation in the subsequent S1.3 multidimensional data fusion.

[0047] S1.3: Using a time synchronization mechanism, the original data stream of the micro-chipping image sequence and the multi-source physical signal set are subjected to multi-dimensional data fusion processing to construct a multi-modal time-series data block of the drill bit's entire life cycle containing visual texture information and force-motion coupling information, ensuring that each data block fully represents the comprehensive working condition characteristics within a single drilling cycle.

[0048] The input conditions are the original data stream of the micro-collision image sequence obtained by step S1.1 and the set of multi-source physical signals that are strictly aligned with its timestamp obtained by step S1.2.

[0049] A time synchronization matching technique (parameters: acquisition timestamp, sensor sampling interval) is used to achieve precise alignment of image data and physical signals on a unified time axis.

[0050] By using a multi-dimensional feature joint encoding technique (parameters: image grayscale matrix, current waveform amplitude sequence, vibration spectrum power density, displacement trajectory coordinate sequence), structured encapsulation of cross-modal data is achieved, and aligned data units of each modal feature within each drilling cycle are obtained.

[0051] A cross-modal feature fusion technique (parameters: feature fusion weights are based on material type and processing speed) is adopted to realize the combined mapping of visual texture features and force-motion coupling features in a multi-dimensional tensor space, and generate a fusion feature matrix containing the interrelationships of each modality.

[0052] By using a time-series block-based technique (parameters: drilling cycle segmentation boundary, number of cycles), the fused feature matrix is ​​divided and stored according to a single drilling cycle, forming multiple independent drilling cycle multimodal time-series data blocks.

[0053] Through consistency verification (parameters: synchronization deviation tolerance, matching success rate threshold), the results of the previous step are transformed into data blocks that have been verified by synchronization accuracy, so as to achieve the expected technical effect that each data block fully represents the comprehensive working condition characteristics within a single drilling cycle.

[0054] For example, in a high-density FR-4 substrate drilling scenario, a high-speed linear scan camera captures images of the cutting edge area at a sampling frequency of 4000 frames per second, a current sensor sampling frequency of 50kHz, a triaxial accelerometer sampling frequency of 10kHz, and a grating ruler displacement sensor sampling frequency of 1kHz. The image sampling timestamp interval is... Seconds, using time synchronization matching technology, the maximum phase deviation tolerance is set to . Milliseconds are used to ensure that the sampling time error of all modal signals within the same drilling cycle does not exceed this value. Multi-dimensional feature joint encoding is used to encapsulate the image grayscale matrix (512×512 pixel sub-pixel level cutting edge texture), current waveform amplitude sequence (2500 sampling points within a single cycle), vibration spectrum power density (power values ​​in the 12–28kHz frequency band with a resolution of 100Hz), and displacement trajectory coordinate sequence (Z-axis displacement accuracy of 0.5μm) into a three-dimensional tensor. This is based on the drilling cycle length of... The data is divided into blocks with a second-level threshold. Each block contains synchronized data from three channels: image, force signal, and motion parameters. Consistency verification confirms that the synchronization deviation tolerance meets the requirements. The final output multimodal time-series data block fully reflects the changes in tool edge morphology and machining physical state within the cycle, providing reliable basic data for subsequent evolution state annotation.

[0055] S1.4: Based on the micro-chipping evolution criteria defined by the expert knowledge base, the multimodal time-series data blocks of the drill bit's entire life cycle are jointly annotated by manual and algorithm-assisted methods to identify and mark four typical evolutionary state labels: sub-pixel level edge disturbance, single-tooth micro-crack initiation, multi-tooth collaborative chipping propagation, and local edge stripping precursors, generating an annotated dataset with state semantic attributes.

[0056] For the drill bit full life cycle multimodal time series data block obtained by time synchronization processing, a pattern recognition technology method based on expert knowledge base (parameters: evolution criterion rule set, stage feature template) is adopted to achieve preliminary classification of multimodal signals in the physical evolution stage.

[0057] By using an image texture feature matching method (parameters: grayscale gradient distribution, edge perturbation amplitude), sub-pixel level perturbation patterns in the cutting edge region are identified, and sub-pixel level cutting edge perturbation label data is obtained.

[0058] By using a force signal spectral domain anomaly detection method (parameters: transient slope threshold of spindle current, energy decay rate of Z-axis acceleration power spectrum), the initiation stage of single-tooth microcracks is determined, and single-tooth microcrack initiation label results are generated.

[0059] By using a cross-period correlation analysis method (parameters: frequency of multi-tooth positional coordinated disturbance, phase consistency of vibration spectrum), the identification of the multi-tooth coordinated collapse expansion stage is achieved, and multi-tooth coordinated collapse expansion label data is obtained.

[0060] By using an edge contour integrity detection method (parameters: blade length truncation ratio, image boundary gap area), the early stage of local blade stripping is determined, and a label result of the early stage of local blade stripping is generated.

[0061] By using a combination of manual and algorithm-assisted review and processing, the labeling results from the previous step are transformed into a labeled dataset with state semantic attributes, enabling accurate labeling of multimodal time series data on four typical evolutionary states.

[0062] For example, in high-density FR In the 4-substrate processing scenario, the drill speed is set to... rpm, plate thickness is mm, the high-speed linear array camera acquires image blocks of the cutting edge region with a resolution of [missing information]. × Pixels. Based on the evolutionary criteria of the expert database, the grayscale gradient perturbation recognition threshold is set to [value missing]. The threshold for the transient slope change of the spindle current is A / ms, Z-axis acceleration power spectrum energy attenuation threshold is After collecting continuous data In the multimodal data of a drilling cycle, the image feature matching module detected that the amplitude of the grayscale gradient perturbation at the same position of the cutting edge exceeded the set threshold, and the force signal spectral domain analysis module detected that the frequency of sudden changes in the single-tooth current slope stabilized within a certain range. After repeated testing and manual review, the signal was confirmed to possess characteristics of single-tooth microcrack initiation. Ultimately, this data block was labeled as exhibiting single-tooth microcrack initiation and included in a labeled dataset with state semantic attributes. This significantly improved the model's ability to identify early crack states during subsequent classification training.

[0063] S1.5: Based on the working condition metadata index in the labeled dataset with state semantic attributes, the labeled data blocks under different material types, thickness specifications and drilling speed ranges are classified, stored and encapsulated in a structured manner. Finally, a multimodal original dataset containing four types of evolution state labels is constructed as a standardized input source for subsequent model training and fingerprint database establishment.

[0064] Step S2: Based on the working condition labels of different material and drilling speed combinations in the multimodal original dataset, the gray-level gradient change rate of image blocks, the frequency of abrupt changes in the slope of the current envelope, and the energy attenuation trend of the Z-axis acceleration power spectrum in a specific frequency band are jointly normalized to generate a multidimensional time-series feature vector that eliminates differences in working condition background. Specifically, this includes: S2.1: Based on the working condition metadata index in the multimodal original dataset, extract the substrate type identifier, plate thickness value and spindle speed setting value corresponding to the current data block as working condition background parameters. Use statistical distribution analysis to calculate the benchmark mean and standard deviation of the image block grayscale gradient change rate, current envelope slope abrupt change frequency and Z-axis acceleration power spectrum energy decay trend under each working condition category, and generate a working condition background benchmark vector containing the statistical characteristics of multidimensional physical quantities.

[0065] Based on the working condition metadata index in the multimodal raw dataset, a working condition parameter extraction method (parameters: material type identifier, plate thickness value, spindle speed setting value) is used to parse the working condition background parameters of the current data block.

[0066] By using statistical distribution analysis (parameter: working condition category classification strategy), we can model the sample distribution of the gray-level gradient change rate sequence of image blocks under each working condition category and obtain the corresponding benchmark mean and standard deviation.

[0067] By using statistical distribution analysis (parameter: operating condition classification strategy), we can model the sample distribution of the frequency sequence of abrupt changes in the slope of the current envelope and obtain the corresponding benchmark mean and standard deviation.

[0068] By using statistical distribution analysis (parameter: working condition classification strategy), we can model the sample distribution of the Z-axis acceleration power spectrum energy decay trend sequence and obtain the corresponding benchmark mean and standard deviation.

[0069] A multi-dimensional physical quantity vector construction method (parameters: image gradient, slope change frequency, vibration energy decay trend, and three types of benchmark mean and standard deviation) is adopted to generate the working condition background benchmark vector and obtain a standardized reference vector containing the statistical characteristics of the three types of physical quantities.

[0070] By constructing a benchmark vector for the operating condition background, the results of the previous step are transformed into benchmark parameter data that are comparable across materials and rotational speeds, thereby achieving the expected technical effect of eliminating differences in operating condition background.

[0071] S2.2: Based on the baseline mean and standard deviation in the background baseline vector, perform zero-mean normalization transformation on the original acquired image block grayscale gradient change rate sequence to eliminate the absolute influence of surface texture of materials with different reflectivity on the gradient amplitude, and generate a normalized image gradient feature sequence with a unified dimension scale.

[0072] S2.3: Based on the dynamic load reference range in the working condition background reference vector, perform piecewise linear mapping processing on the original collected current envelope slope change frequency data, compress the current transient response under different drilling speed ranges to a standard unit range, and generate a normalized current slope feature sequence that eliminates the rotational speed coupling effect.

[0073] S2.4: Based on the vibration transfer function characteristics in the background reference vector of the working condition, logarithmic domain differential normalization is performed on the energy attenuation trend data of the original Z-axis acceleration power spectrum in the 12 to 28 kHz frequency band to offset the acoustic attenuation baseline drift caused by the change in plate thickness, and generate a normalized vibration energy characteristic sequence with consistent frequency response.

[0074] In the input conditions, the original Z-axis acceleration power spectrum energy decay trend data are in the 12 to 28 kHz frequency band, with the plate thickness and vibration transfer function characteristics as the parameter sources for the working condition background reference vector.

[0075] The frequency band energy integration technique (parameter: frequency band range 12–28kHz) is used to extract the local energy curve of the original Z-axis acceleration power spectrum curve within the specified frequency band and obtain the frequency domain discrete energy sequence.

[0076] By using the logarithmic domain mapping technique (parameter: natural logarithm base e), the frequency domain discrete energy sequence is converted into a logarithmic energy value sequence, and vibration energy logarithmic spectrum data with compressed dynamic range is obtained.

[0077] The differential calculation technique (parameter: first-order central difference window length of 3 points) is used to calculate the time series change rate of vibration energy logarithmic spectrum data and obtain a differential value sequence reflecting the energy decay trend.

[0078] The difference value sequence is normalized by using a reference drift correction method (parameters: thickness reference attenuation value μ and standard deviation σ), and the normalized difference value is calculated.

[0079] A frequency response consistency modeling method (parameter: transfer function gain compensation coefficient γ) is adopted to perform gain compensation and phase consistency adjustment on the normalized difference value sequence, and output the final normalized vibration energy characteristic sequence.

[0080] By using gain compensation and normalization, the results of the previous step are transformed into multidimensional vibration energy characteristics with consistent frequency response, thereby achieving the expected technical effect of offsetting the acoustic attenuation baseline drift caused by changes in plate thickness.

[0081] For example, in the case of drilling into an FR-4 substrate with a thickness of 1.6 mm, the original Z-axis acceleration power spectrum has 256 discrete energy points in the frequency band of 12–28 kHz, with the energy value of each point in m / s². After extracting the frequency domain energy sequence, a natural logarithmic transformation is used to generate logarithmic energy values ​​ranging from... 4.2 to Logarithmic spectral data between 1.8 and 1.8. The rate of change of the logarithmic spectral data was calculated using a 3-point central difference, with the difference value D ranging from 1.8 to 1.8. 0.35 to 0.12. The thickness reference attenuation value μ is set to... 0.20, standard deviation σ is set to 0.05, and the difference value is entered, for example, when D = At 0.35, the normalized difference value is 3.0. A gain compensation coefficient γ = 1.15 is used to adjust the frequency response of the normalized difference value, and the output normalized vibration energy feature sequence maintains a consistent energy attenuation trend across substrates of different thicknesses. In this scenario, during subsequent multi-dimensional temporal feature vector fusion, this normalized vibration energy feature significantly improves the identification stability of substrates of different thicknesses, ensuring the discrimination accuracy of the state matching engine under cross-thickness conditions.

[0082] S2.5: Based on the normalized image gradient feature sequence, normalized current slope feature sequence, and normalized vibration energy feature sequence generated in the previous steps, the feature data with the same timestamp from the three channels are dimensionally fused using the multi-channel tensor splicing method to generate a multi-dimensional time-series feature vector that eliminates differences in working conditions and has cross-scene comparability.

[0083] like Figure 2As shown, step S3 involves using the multi-dimensional temporal feature vector as training samples, and mapping the features within multiple consecutive drilling cycles into fixed-length embedded representations using a lightweight temporal encoder. This constructs a high-fidelity multimodal stage fingerprint database capable of characterizing the physical evolution of micro-chipping edges from their inception to critical failure. Specifically, this includes: S3.1: Based on the multi-dimensional temporal feature vector generated in the previous steps to eliminate the differences in working conditions and background, the vector is structured and grouped according to material type, drilling speed range and micro-chipping evolution stage labels. A standardized training sample set containing four typical states is constructed, including sub-pixel level cutting edge perturbation, single-tooth micro-crack initiation, multi-tooth collaborative chipping propagation and local cutting edge stripping precursor, in order to establish the supervised learning input basis of the lightweight temporal encoder.

[0084] Based on the multi-dimensional time-series feature vector generated in the previous steps to eliminate differences in working conditions and backgrounds, a material category screening technique (parameter: substrate type identifier) ​​is used to initially group the feature vectors according to substrate type.

[0085] By using the drilling speed interval segmentation method (parameter: spindle speed setting value), the same material category is divided into subcategories according to the drilling speed interval, and a multi-dimensional time series feature set corresponding to each subcategory is obtained.

[0086] An evolution stage label separation technique (parameter: micro-crack evolution stage label) is adopted to further subdivide the feature set within each drilling speed range into four categories of state labels: subpixel-level cutting edge disturbance, single-tooth micro-crack initiation, multi-tooth collaborative chipping and propagation, and local cutting edge stripping precursors, and generate a structured grouped feature subset.

[0087] By using a multi-dimensional feature standardization encapsulation method (parameter: unified dimension description rule), the feature vectors, working condition labels, and status labels of samples in the same group are integrated into training sample entries with a unified field format, and a standardized training sample matrix is ​​obtained.

[0088] A data index mapping method (parameters: material type, drilling speed range, status label index) is adopted to achieve bidirectional association storage between structured grouping results and training sample matrix, forming a complete standardized training sample set.

[0089] By using structured grouping and standardized matrix encapsulation, the results of the previous step are transformed into training data that can be directly used by a lightweight temporal encoder to perform supervised learning, thus building a highly reliable input foundation.

[0090] For example, in a drilling cycle of FR-4 sheet metal with a thickness of 0.8 mm and a spindle speed of 120,000 rpm, the multidimensional temporal feature vector obtained based on the preceding steps includes a grayscale gradient change rate of 0.015–0.021, a current envelope slope abrupt change frequency of 2–4 times / cycle, and a Z-axis acceleration power spectrum energy decay trend value of 0.32–0.37. The material category screening method incorporates this feature vector into the FR-4 category; the drilling speed range segmentation method categorizes it according to the 110,000–130,000 rpm range; and the evolution stage label separation method classifies it as the single-tooth microcrack initiation stage based on expert annotations. The multidimensional feature standardization encapsulation method forms training sample entries from the above feature vector, working condition labels, and status labels according to a unified dimensional description rule, such as... Where 0.018 is the mean normalized grayscale gradient change rate, 3 is the mean normalized current slope abrupt change frequency, and 0.35 is the mean normalized vibration energy attenuation trend. Finally, by establishing an index relationship with material type, drilling speed range, and state label through a data index mapping method, a standardized training sample set containing the sample is formed, enabling the lightweight time encoder to significantly improve the stability of state discrimination under multiple working conditions.

[0091] S3.2: For the multi-dimensional temporal feature vector sequence of multiple consecutive drilling cycles in the standardized training sample set, a lightweight temporal encoder architecture is designed and initialized. A multi-head self-attention mechanism is used to model the temporal dependency relationship of the gray-level gradient change rate of image blocks, the frequency of abrupt changes in the slope of the current envelope, and the energy decay trend of the Z-axis acceleration power spectrum, so as to generate a hidden layer state sequence containing long-range evolution information.

[0092] S3.2: For the multi-dimensional temporal feature vector sequence of multiple consecutive drilling cycles in the standardized training sample set, a lightweight temporal encoder architecture is designed and initialized. A multi-head self-attention mechanism is used to model the temporal dependency relationship of the gray-level gradient change rate of image blocks, the frequency of abrupt changes in the slope of the current envelope, and the energy decay trend of the Z-axis acceleration power spectrum, so as to generate a hidden layer state sequence containing long-range evolution information.

[0093] For the multidimensional temporal feature vector sequence of multiple consecutive drilling cycles in the standardized training sample set, the input object is a training sample matrix with a unified field format generated by structured grouping in step S3.1. Its feature components include the normalized image patch grayscale gradient change rate sequence, the normalized current envelope slope abrupt change frequency sequence, and the normalized Z-axis acceleration power spectrum energy decay trend sequence. A lightweight temporal encoder architecture design method is adopted, with 3 input channels, variable sequence time step length, and 128 embedding dimensions, to achieve functional matching of the model structure and support for variable-length input. The network architecture initialization method is called, with 16 filters in the convolutional pre-layer, a kernel size of 3×1, and a stride of 1, to achieve primary extraction of local temporal features at the input end, so as to reduce the redundant pressure of the original input on the subsequent self-attention calculation. A multi-head self-attention mechanism is adopted to construct the encoder, with 8 attention heads and 16 query / key / value dimensions per head, to establish cross-modal correlation quantization channels for the three types of features—image gradient, current slope, and vibration energy—changing with time step within the encoder main layer. The self-attention weight calculation process utilizes the scaled dot product formula: Where S is the attention weight score. For querying the matrix, The key matrix has an embedding dimension of 128. Square root scaling is applied to the dot product result to stabilize the gradient distribution. Based on the weight matrix, a softmax function is applied to normalize the global dependency weights across time steps, and then applied to the value matrix. A weighted fusion operation is applied to generate an attention output tensor containing long-range evolutionary information. The attention output tensor is then stably fused with the primary features extracted from the pre-convolutional layer using residual connections and layer normalization (parameter: normalization axis set to the feature dimension), resulting in a fused feature tensor with temporal consistency and a unified cross-modal expression. A hidden state generation technique (parameters: fully connected transformation layer dimension 256, activation function ReLU) is invoked to perform a nonlinear mapping on the fused feature tensor and compress redundant information, generating a hidden state sequence as the direct input to step S3.3. Through a multi-head self-attention mechanism and the collaborative processing of convolutional and fully connected modules, the multi-dimensional temporal feature vector sequence from the previous step is transformed into hidden state data capable of representing the long-range physical evolution of micro-collapsing blades, achieving the encoder's expected technical effect of mapping local features to global semantics.

[0094] For example, under drilling cycle conditions of FR-4 substrate, 1.6mm thickness, and spindle speed of 120,000 rpm, the standardized training sample matrix contains normalized image gradient features with a mean range of 0.012–0.018, a current slope abrupt change frequency range of 3–5 times / cycle, and a vibration energy attenuation trend range of 0.29–0.34 for 8 consecutive cycles. The lightweight timing encoder has 3 input channels, an embedding dimension of 128, 16 filters in the convolutional pre-layer, and a kernel size of 3×1. The multi-head self-attention mechanism is configured with 8 heads, each with a query / key / value dimension of 16, and the dot product is calculated using the formula: S = After scaling, the maximum value of the weight matrix is ​​obtained. The minimum value is After softmax normalization, the distribution stabilizes at the mean. Standard deviation Within the specified range, the fused features are compressed to dimension 256 through a fully connected transform layer. After ReLU activation, the length of the output hidden state sequence is equal to the length of the input time step (8). This hidden state successfully preserves the cross-cycle correlation information between image gradient, current slope, and vibration energy in the subsequent pooling aggregation stage. This reduces the encoder's recognition delay for the multi-tooth cooperative collapse expansion stage on the validation set to the cycle level, significantly improving the ability to capture long-range evolutionary features and the stability of state discrimination.

[0095] S3.3: The hidden layer state sequence is compressed using the pooling aggregation method to map the variable-length multimodal temporal dynamic features into a fixed-length 64-dimensional real-time temporal semantic vector, so as to eliminate the influence of the difference in drilling cycle duration on feature expression and obtain a unified-dimensional state embedding representation.

[0096] S3.4: Based on the 64-dimensional real-time temporal semantic vector and the corresponding micro-collapsing blade evolution stage labels, perform a contrastive loss function to optimize the training process, adjust the network parameters of the lightweight temporal encoder to bring the samples of the same evolution stage closer in the embedding space and push away the samples of different classes, and output a finely tuned lightweight temporal encoder model with strong discriminative ability.

[0097] Based on the 64-dimensional real-time temporal semantic vector output from the preceding steps and the corresponding micro-collapsing blade evolution stage labels, a contrastive loss function optimization training method (parameters: positive sample aggregation coefficient α, negative sample separation coefficient β, embedding space radius γ) is adopted to achieve the aggregation of samples of the same stage and the separation of samples of different stages within the embedding space.

[0098] The gradient backpropagation technique (parameters: learning rate η, gradient clipping threshold δ) is used to calculate the partial derivative of the loss function with respect to the weights of the lightweight temporal encoder, and obtain the gradient matrix used for updating the network parameters.

[0099] The algorithm of stochastic gradient descent (parameters: batch size B, momentum factor μ) is used to iteratively adjust the weights of the encoder's multi-head self-attention structure layer and fully connected projection layer, and generate the embedding space distribution after round-by-round optimization.

[0100] The L2 regularization method (parameter: regularization coefficient λ) is adopted to add a weight penalty term to the loss function to suppress the overfitting tendency of the encoder model and obtain an updated weight vector that balances discriminative ability and generalization performance.

[0101] Through convergence detection techniques (parameter: convergence criterion) The validation set matching accuracy threshold (τ) transforms the performance changes of the validation set in each round of training into a convergence signal, thereby achieving dynamic termination control of the training process.

[0102] Through the above processing method, the real-time temporal semantic vector and evolution stage label of the previous step are transformed into lightweight temporal encoder network parameters with stable discrimination ability, so as to realize the separability of cross-stage states and the high cohesion of states in the same stage in the embedding space.

[0103] For example, under the conditions of FR-4 material, plate thickness of 1.6 mm, and spindle speed of 120 rpm, the input sequence length is 8 drilling cycles, and the corresponding 64-dimensional real-time temporal semantic vector is used as training samples. The positive sample aggregation coefficient α is set to 0.8, the negative sample separation coefficient β is set to 1.3, and the embedding space radius γ is set to 2.5. The contrastive loss function is used. in, Let be the label of the i-th sample pair. For Euclidean distance, Let i be the embedding vector of the i-th sample. To and The embedding vectors of paired positive / negative samples. The learning rate η is set to 0.002, the batch size B to 64, the momentum factor μ to 0.9, and the regularization coefficient λ to 0.0005. During training, the gradient clipping threshold δ is set to 5.0 to prevent gradient explosion. The validation set matching accuracy stabilizes at τ = 0.95 and converges for 5 consecutive rounds. When the value is less than 1e-4, the iteration stops. The final output of the fine-tuned lightweight temporal encoder model achieves high cohesion and cross-stage separability of samples at each evolution stage within the embedding space under this condition, effectively enhancing the stage discrimination capability of micro-collapsing anomalies.

[0104] S3.5: Call the fine-tuned lightweight temporal encoder model to perform batch inference calculations on the full standardized training sample set, extract the 64-dimensional real-time temporal semantic vector corresponding to each sample and associate it with the original evolution stage annotation for storage, and build a multimodal stage fingerprint database that covers the entire life cycle and has high fidelity.

[0105] The finely tuned lightweight temporal encoder model is invoked on the full standardized training sample set, and a batch inference computation technique (parameters: batch size 256, inference accuracy FP16) is used to achieve efficient embedding and mapping processing of multi-dimensional temporal feature vectors.

[0106] By using a batch inference computation method (parameters: sequence length of 64, model weights are the configuration optimized in step S3.4), temporal-cross-modal fusion of multi-channel feature inputs for each training sample is achieved, resulting in a real-time temporal semantic vector of length 64.

[0107] The feature association storage technique (parameters: index key is the unique ID of the sample, association value is the evolution stage label) is adopted to realize the one-to-one correspondence binding of the 64-dimensional real-time temporal semantic vector with the original micro-collapse blade evolution stage label, and generate fingerprint entries with semantic association capabilities.

[0108] By using a multimodal indexing construction method (parameters: the index structure is an inverted table, and the fields include material type, drilling speed range, and status label), the structured storage of the binding results is achieved, forming a stage fingerprint data structure that can be retrieved according to conditions.

[0109] The lifecycle coverage verification technique (parameter: traversing the entire training sample set and verifying that there are corresponding fingerprint entries for four typical states) is adopted to confirm the coverage of the fingerprint database and obtain a multimodal stage fingerprint database with complete lifecycle representation capabilities.

[0110] Through the structured storage and coverage verification processing method, the embedding vector and state label results of the previous step are transformed into searchable, scalable, high-fidelity stage fingerprint data, realizing the expected technical effect of stable state matching and adaptive early warning judgment under cross-operating conditions.

[0111] For example, in the high-density FR-4 plate drilling scenario, the batch inference calculation parameters are set to a batch size of 128, a sequence length of 64, FP32 precision mode, and the model weights are the lightweight encoder optimized by step S3.4. The original normalized feature vector contains 24 dimensions of image gradient features, 20 dimensions of current slope features, and 20 dimensions of vibration energy features, for a total of 64 input dimensions. During batch inference, the encoder performs long-range dependency modeling through a multi-head self-attention mechanism, outputting the embedding vector corresponding to each sample. An inverted index structure is adopted, using the material type "FR-4", drilling speed range "120k–150k rpm", and state label "single-tooth microcrack initiation" as index keys to associate and store the embedding vectors. Traversing the entire training set confirms that all four states have complete fingerprint entries. When the fingerprint database internally calls the improved DTW algorithm during the online matching stage, it can quickly retrieve the corresponding state template, significantly improving the cross-working-condition state recognition speed and reducing the risk of false alarms.

[0112] like Figure 3 As shown, step S4 involves: based on the original multimodal joint temporal pattern input during online runtime, starting a lightweight temporal encoder to extract a 64-dimensional real-time temporal semantic vector. Specifically, this includes: S4.1: Obtain the image block grayscale gradient change rate sequence, current envelope slope abrupt change frequency sequence, and Z-axis acceleration power spectrum energy decay trend sequence within the continuous drilling cycle under the current working condition as the original multimodal joint time series pattern. Perform sliding window truncation processing on the original multimodal joint time series pattern to generate a standardized time series data window containing data from the most recent five drilling cycles.

[0113] S4.2: Based on the standardized time-series data window, the image feature channel, force signal feature channel, and motion parameter feature channel are synchronized and calibrated on the time axis using a multi-channel feature alignment method to eliminate phase deviations caused by differences in sensor sampling frequencies and generate a multi-channel synchronized time-series tensor with strict time axis alignment.

[0114] Based on the image feature channel, force signal feature channel, and motion parameter feature channel in the standardized time-series data window, a multi-channel feature alignment method (parameters: time synchronization marker, sampling frequency ratio matrix) is adopted to realize the synchronous calibration function of different channel data on a unified time axis.

[0115] By using the interpolation resampling technique (parameter: target sampling frequency setting), the data from the low sampling frequency channel is compensated to the time scale of the high sampling frequency channel, and a reconstructed dataset of continuous time series is obtained.

[0116] The phase difference between each channel signal is quantized by using a phase deviation calculation method (parameter: cross-correlation function window length), and a phase correction factor matrix is ​​generated.

[0117] By using a time delay compensation processing technique (parameter: phase correction factor matrix), the signals of each channel are offset at the microsecond level on the time axis to eliminate timing misalignment caused by differences in sampling frequency.

[0118] By using the amplitude consistency normalization method (parameter: boundary values ​​of amplitude intervals for each channel), the amplitude of the calibrated signal is normalized, and a multi-channel synchronous timing tensor with strictly aligned time axes and consistent amplitudes is generated.

[0119] By using a multi-channel feature alignment method, the standardized time-series data window result from the previous step is transformed into a multi-channel synchronous time-series tensor with strict time axis alignment, thereby achieving the effect of cross-modal precise synchronization technology at the input of the subsequent lightweight time encoder.

[0120] For example, in a machining scenario with FR-4 material thickness of 1.6 mm, drilling speed of 120 k rpm, spindle current sampling frequency of 5 kHz, image sampling frequency of 100 Hz, and Z-axis acceleration sampling frequency of 10 kHz, a standardized time-series data window of the most recent 5 drilling cycles is obtained. The image feature channels are resampled to a target frequency of 5 kHz using cubic spline interpolation, consistent with the time scale of the force signal and motion parameter channels. The phase deviation between each channel and the force signal channel is calculated using a cross-correlation function. Assuming a window length of 1024 points, a phase correction factor matrix is ​​obtained, where the image channel phase deviation is... Seconds, motion parameter channel phase deviation is Seconds. Through time delay compensation processing, the image channels are delayed by a corresponding millisecond offset, and the motion parameter channels are advanced by a corresponding millisecond offset, achieving complete alignment of the three-channel time scale. Amplitude consistency normalization processing maps the amplitude range of each channel to [-1,1]. The output multi-channel synchronous temporal tensor significantly improves the extraction accuracy of cross-modal correlation features in subsequent encoder inputs, and achieves a performance improvement of reducing the recognition delay of micro-bumping anomalies to the millisecond level in the validation set.

[0121] S4.3: Based on the encoder weight parameter configuration in the high-fidelity multimodal stage fingerprint database constructed in the previous steps, load the lightweight temporal encoder model, and input the strictly time-axis aligned multi-channel synchronous temporal tensor into the input layer of the lightweight temporal encoder model to start the feature extraction and dimensionality reduction mapping process.

[0122] The input condition is a multi-channel synchronous temporal tensor with strictly aligned time axis output from the previous step S4.2. This tensor contains a synchronous data sequence of image feature channels, force signal feature channels, and motion parameter feature channels.

[0123] A model weight loading method (parameter: encoder weight configuration in high-fidelity multimodal stage fingerprint database) is adopted to realize the initialization and network structure recovery of lightweight temporal encoder.

[0124] By verifying the weights (parameters: hash checksum, version number), the integrity of the encoder weight file is verified, and a reliable set of model parameters is obtained.

[0125] An input tensor preprocessing method (parameters: input layer size, channel number matching rules) is adopted to achieve dimension matching and data type conversion of multi-channel synchronous temporal tensors, and obtain a standardized input matrix that meets the requirements of the encoder input layer.

[0126] Through the data batch processing interface (parameters: batch size, time step length), batch encapsulation and time step allocation of multi-channel synchronous time series tensors are realized, and efficient model input batches that can be processed in parallel are generated.

[0127] By using a model inference scheduling method (parameters: GPU / CPU allocation strategy, number of inference threads), the standardized input matrix is ​​passed into the input layer of the lightweight temporal encoder model, the feature extraction and dimensionality reduction mapping process is initiated, and the encoder hidden layer state sequence is generated.

[0128] By coordinating model loading and input data preprocessing, the time axis strictly aligned result of the previous step is transformed into a batch of time-series data that can be efficiently processed by the encoder, enabling real-time extraction of online state features and preparation for subsequent dimensionality reduction.

[0129] For example, on a high-density PCB manufacturing line, a multi-channel synchronous timing tensor containing the most recent five drilling cycles is obtained. The image feature channel is a 128×128 pixel grayscale gradient matrix, the force signal feature channel is the frequency sequence of slope abrupt changes in the main axis current envelope (length = 512), and the motion parameter feature channel is the Z-axis acceleration power spectrum energy decay trend sequence (length = 512). A lightweight timing encoder weight file (version 2.3) from the high-fidelity multimodal stage fingerprint database is called, and the hash checksum match is successful. Dimension matching processing is performed on the input tensor, adjusting the image feature channel to 64×64, and normalizing the force signal and motion parameter feature channels to the [-1, 1] interval, padding with zeros to a length of 1024. The batch size is set to 1, and the time step length is 5. Using GPU inference mode, 4 TensorRT engine threads are allocated, and the input batch is sent to the encoder input layer to obtain the hidden state sequence. This processing flow is completed within the cycle, ensuring that subsequent dimensionality reduction mapping can be performed in real time, significantly improving the response speed and stability of online state representation.

[0130] S4.4: Inside the lightweight temporal encoder model, the long-range dependencies of cross-modal features in the strictly time-aligned multi-channel synchronous temporal tensor are captured through a multi-head self-attention mechanism. The long-range dependencies are then processed by a fully connected projection layer to compress redundant information and fuse cross-domain features, generating a compact temporal embedding representation with a length of 64 dimensions.

[0131] The multi-channel synchronous temporal tensor with strict time axis alignment is input into the multi-head self-attention computation module of the lightweight temporal encoder model (parameters: 8 attention heads, 128 feature embedding dimensions) to achieve global dependency capture of cross-modal information.

[0132] By calculating the self-attention weight matrix (parameter: scaling factor is the inverse of the square root of the embedding dimension), the correlation between image features, force signal features and motion parameter features in long-range time series is quantified, and an attention output tensor containing cross-modal interaction information is obtained.

[0133] By using residual connections and layer normalization techniques (parameter: normalization axis is the feature dimension), we achieve stable fusion of attention output and original encoded input, and obtain a normalized fused feature tensor.

[0134] By performing nonlinear transformation processing through a fully connected projection layer (parameters: hidden layer dimension 256, activation function ReLU), the dimensionality of attention fusion features is compressed and redundant information is removed, and an intermediate representation vector with reduced dimensionality is generated.

[0135] By combining a multi-head self-attention mechanism with a fully connected projection layer, cross-modal long-range dependent features are compressed and fused into a compact temporal embedding representation of length 64, thereby achieving a unified encoding representation of cross-domain features.

[0136] For example, for a multi-channel synchronous timing tensor under the conditions of a copper foil substrate thickness of 0.8 mm and a drilling speed of 120 k rpm, the multi-head self-attention module is configured with 8 attention heads and an embedding dimension of 128. The scaling factor calculation formula is as follows: Where 128 is the embedding dimension, and the scaling factor is used to stabilize the attention distribution. In the attention weights, the maximum correlation value between image features and force signal features is... The maximum correlation value between motion parameter features and image features is After residual connection and feature dimension normalization, the mean of the feature tensor stabilizes at... The standard deviation is stable at In the fully connected projection layer, the hidden layer dimension is set to 256, and the intermediate representation vector is output using the ReLU activation function. After dimensionality reduction and compression to a 64-dimensional embedding vector, this embedding representation significantly improves the similarity score of the template for the "single-tooth microcrack initiation" stage, verifying that this processing method improves the stability and response speed of anomaly recognition under this working condition.

[0137] S4.5: Perform vector normalization processing on the compact temporal embedding representation to eliminate the influence of dimensions and enhance the stability of feature distribution. Finally, output a standardized real-time temporal semantic vector and transmit the standardized real-time temporal semantic vector to the retrieval interface to complete the construction and delivery of online state representation.

[0138] Step S5: Using an improved dynamic time warping algorithm, the real-time temporal semantic vector is compared with the evolutionary stage templates in the high-fidelity multimodal stage fingerprint database to perform a similarity search, outputting a state matching result including matching confidence and stage transition probability. Specifically, this includes: S5.1: Based on the standardized real-time temporal semantic vector output from the preceding steps, extract the 64-dimensional feature components as the query sequence to be retrieved. At the same time, retrieve the reference evolution stage template set covering four types of micro-collapsing blade evolution states from the high-fidelity multimodal stage fingerprint database, and construct the comparison data pair containing the query sequence and multiple reference templates to establish the input basis for dynamic time warping calculation.

[0139] S5.2: For the query sequence and each reference evolution stage template in the comparison data pair, perform the local distance matrix construction operation in the improved dynamic time warping algorithm, use the Euclidean distance metric function to calculate the point-to-point difference value between each feature component in the query sequence and the corresponding feature component in the reference template, and generate a local cost matrix that represents the instantaneous state deviation, so as to quantify the basic discreteness of real-time conditions and historical fingerprints in the feature space.

[0140] Evolutionary stage templates refer to reference time-series feature sequences that are pre-stored in a high-fidelity multimodal stage fingerprint database and represent typical patterns of different evolutionary stages of tool micro-chipping.

[0141] Each evolution stage template is a fixed-length (e.g., 70 time points) temporal vector sequence formed by extracting multimodal features from the evolution process of a specific micro-chipping state (e.g., initial stage, development stage, severe stage). Each time point corresponds to a 64-dimensional feature vector, which integrates information from multiple sensors such as vibration, acoustic emission, and force, and can comprehensively describe the cutting state at that moment.

[0142] A feature alignment matrix generation method (parameters: query sequence length 64, template sequence length m) is adopted to realize the distance calculation of each feature component of the query sequence to be retrieved and the reference evolution stage template.

[0143] The instantaneous difference between each query feature component and the template feature component is calculated using the Euclidean distance metric function (parameter: dimension index i∈[1,64]), and a list of local difference values ​​is obtained to form the basic cost unit.

[0144] A matrix construction method (parameter: matrix size 64×m) is adopted to arrange all basic cost units in a dual manner according to dimensional index and time series index, thereby generating a local cost matrix skeleton that represents the feature differences at each time point.

[0145] Processing via structural grouping (parameter: subexpression containing feature component operations) This involves explicitly combining the numerator, denominator, or exponent operations in the Euclidean distance calculation formula to ensure stable parsing of matrix element values.

[0146] The Euclidean distance calculation formula is used: in, To query the feature component value of the i-th dimension of the sequence, Let be the feature component value of the template sequence in the i-th dimension.

[0147] The calculation results from the above formula are filled into the corresponding positions in the matrix skeleton to achieve the complete construction of the local cost matrix.

[0148] The dynamic time warping preprocessing module transforms the results of the previous step into a two-dimensional numerical matrix that characterizes the spatial dispersion of real-time operating conditions and historical fingerprint features, thereby enabling precise input conditions for similarity retrieval.

[0149] For example, under the conditions of copper foil substrate, thickness 0.8mm, and drilling speed 120k rpm, the query sequence length is set to 64 dimensions, and the reference template sequence length is 70 dimensions. Euclidean distance is calculated for the i-th feature (i=12), with a query value of 0.235 and a template value of 0.198. The difference in the formula is... The square operation yields The sum of all dimensions is obtained. The local distance value after taking the square root is This value is filled into the template index column corresponding to row 12 of the matrix skeleton. After repeating the full-dimensional operation, a 64×70 local cost matrix is ​​generated. This matrix exhibits a low average difference distribution in subsequent global cumulative path search, significantly improving matching accuracy and effectively suppressing false alarms under dynamic conditions.

[0150] S5.3: Based on the local cost matrix, perform global cumulative path search processing, introduce slope constraint factor and step size penalty term to modify the traditional dynamic programming recursive formula, and find the minimum cumulative distance path connecting the starting point to the ending point of the matrix under the premise of allowing nonlinear scaling of the time axis, generate the optimal alignment path and the corresponding minimum cumulative distance value that characterize the overall similarity of the time sequence pattern, so as to eliminate the phase misalignment interference caused by the fluctuation of the drilling cycle rate.

[0151] S5.4: Using the minimum cumulative distance value, perform normalized similarity mapping transformation, convert the distance value into a matching confidence score between zero and one through an exponential decay function, and compare and analyze the score with a preset category boundary threshold to generate a preliminary matching label that identifies the current real-time temporal semantic vector that is closest to the evolution stage type, so as to realize the logical conversion from geometric distance to semantic confidence.

[0152] Based on the minimum cumulative distance value obtained from the previous steps, a normalized similarity mapping technique (parameter: exponential decay coefficient λ) is used to convert the distance value into a standardized similarity score.

[0153] By using the exponential decay function mapping method (parameter: λ depends on the global mean and standard deviation of the template distance distribution at each stage in the fingerprint database), a nonlinear mapping from distance value to similarity score is achieved, and a matching confidence score between zero and one is obtained.

[0154] The matching confidence score is calculated using a formula, where the normalized mapping formula is: in, To match the confidence score, The minimum cumulative distance value. This is the exponential decay coefficient.

[0155] By using a category boundary threshold comparison method (parameter: minimum acceptable confidence level setting for each stage type), the system compares and analyzes the matching confidence score with the preset threshold, and generates stage candidate label data that meets the criteria.

[0156] A category matching identifier generation method (parameters: stage type index, current score value, threshold criterion result) is adopted to convert stage candidate labels into preliminary matching labels and output the data object of the label.

[0157] By using a geometric distance to semantic confidence conversion process, the minimum cumulative distance result in the previous step is transformed into a preliminary matching label that can be directly used for decision rule invocation, thus achieving the technical effect of moving from distance measurement to semantic judgment.

[0158] For example, under the conditions of FR-4 material, plate thickness of 1.6mm, and drilling speed of 120k rpm, the minimum cumulative distance value calculated by dynamic time warping is: The global mean distance for the same stage within the fingerprint database is The standard deviation is The exponential decay coefficient λ is calculated using the empirical formula λ = 1 / mean. Substitute this distance value into the formula to calculate the match confidence score: The rating result is After comparison with the preset stage category threshold of 0.35, it is determined to pass, and a preliminary matching label "single-tooth microcrack initiation" is generated. This label will be used in the next step of the Bayesian inference module to perform transition probability estimation. The application effect is to significantly improve the stability of matching judgment under this working condition and effectively reduce misjudgment.

[0159] S5.5: Based on the preliminary matching labels and their corresponding historical state transition statistics, perform Bayesian probability inference calculation, combine the current matching confidence score and the prior state transition matrix, estimate the probability value of the micro-chipping state undergoing a hierarchical jump within a future preset time window, and generate a state matching result containing the final matching confidence and stage transition probability to complete the quantitative prediction output of the drill bit health evolution trend.

[0160] Based on the initial matching labels, the historical state transition statistics module is invoked to load the prior state transition matrix for the corresponding evolutionary stage, establishing the probabilistic basis for state transitions. A joint quantization method using matching confidence scores and prior state transition matrices (parameters: evolutionary stage category, matching confidence value, transition matrix elements) is employed to achieve real-time state probability fusion calculation, obtaining a joint distribution vector representing the current state's credibility. A Bayesian probability inference method (parameters: prior state transition matrix, observation likelihood function configuration) is used to solve for the posterior state transition probability based on the current matching confidence, generating predicted probability values ​​for state level jumps within a future preset time window. The posterior probability is calculated by combining the prior and likelihood using Bayes' theorem, as follows: in, This is an evolutionary stage state. For the current matching confidence observation data, In the state The observed data The likelihood probability, Let be the prior state probability. This is a full probability normalization factor. Using a state transition trend analysis method (parameters: posterior probability sequence, time window length), the rate of change of transition probabilities is extracted, and a dynamic rate of change index containing predicted fluctuation information is generated. Through a risk quantification mapping method (parameters: final matching confidence value, posterior transition probability value, rate of change index), the results of the previous step are transformed into a state matching result data package, achieving a quantitative and controllable output of the drill bit health evolution trend.

[0161] For example, under continuous machining conditions of FR-4 material, plate thickness of 0.6 mm, and spindle speed of 120 rpm, the initial matching label is single-tooth microcrack initiation, with a matching confidence score of 0.84. The prior transition probability from single-tooth microcrack initiation to multi-tooth cooperative collapse propagation in the historical state transition matrix is ​​0.18. A time window of 15 minutes is set, and the observation likelihood function uses a Gaussian distribution model with a mean μ = 0.85 and a standard deviation σ = 0.05. Substituting the matching confidence score of 0.84 into the Bayesian formula, the posterior transition probability is calculated to be 0.31. Through rate of change extraction, it is calculated that this posterior probability increased from 0.21 to 0.31 within the last 15 minutes, with a rate of change value of 0.66. The risk quantification mapping method integrates the matching confidence level of 0.84 with the posterior transition probability of 0.31 and the rate of change of 0.66 to generate a state matching result data packet. The output stage label is single-tooth microcrack initiation, final matching confidence level of 0.84, stage transition probability of 0.31, and the transition trend is marked as significantly increasing, so as to achieve accurate triggering of subsequent processing logic.

[0162] Step S6: Based on the matching confidence and stage transition probability in the state matching result, activate the processing logic for the corresponding evolution stage. If it is identified as the single-tooth microcrack initiation stage and the matching confidence meets the preset threshold, then the dual-window sliding verification mechanism is enabled. If it is identified as the multi-tooth collaborative collapse and expansion stage and the stage transition probability increases sharply in a short period of time, then a secondary warning instruction is directly triggered. Specifically, this includes: S6.1: The matching confidence and stage transition probability in the state matching result are analyzed to identify the specific evolution stage type of the current micro-collision blade and extract the corresponding risk criterion parameters, and generate a decision context environment containing stage identifiers and dynamic threshold factors.

[0163] The matching confidence and stage transition probability in the state matching results are subjected to structured parsing (parameters: evolution stage template index, threshold mapping configuration table, risk level mapping table) to achieve multi-dimensional data deconstruction of matching confidence and stage transition probability values.

[0164] By using a state classification parsing method (parameter: stage label mapping matrix), the matching confidence and stage transition probability are used to jointly determine the specific evolutionary stage type of the current micro-collision blade, and obtain stage identifier data.

[0165] By employing risk criterion extraction (parameters: historical risk threshold library, working condition adaptation coefficient), the corresponding risk criterion parameters can be retrieved based on the stage identifier, and a risk criterion parameter set containing risk level, triggering conditions and dynamic correction coefficients can be generated.

[0166] By using a dynamic threshold calculation method (parameters: matching confidence value, stage transition probability value, and operating condition adaptation coefficient), the risk criterion threshold is adaptively calculated, and a dynamic threshold factor that can be used for subsequent early warning triggering logic is obtained.

[0167] A context construction method (parameters: stage identifier, risk criterion parameter set, dynamic threshold factor) is adopted to generate a complete decision context environment containing stage identifier and dynamic threshold factor.

[0168] By using the above processing method, the state matching results of the previous step are transformed into structured decision context data, achieving the technical effect of evolution stage identification and risk criterion extraction.

[0169] For example, under the conditions of FR-4 substrate, 0.8mm plate thickness, and spindle speed of 120,000 rpm, the input is a state matching result with a matching confidence of 0.84 and a stage transition probability of 0.56. During structured parsing, the evolution stage template index and threshold mapping configuration table are called to determine that the current stage is the single-tooth microcrack initiation stage. Risk criterion extraction and reverse lookup of historical risk thresholds show that the basic triggering condition for this stage is a matching confidence ≥ 0.82. Combined with the working condition adaptation coefficient of 1.05, a corrected threshold is obtained. = The dynamic threshold calculation method weights the matching confidence level of 0.84 and the stage transition probability of 0.56. Dynamic threshold calculated comprehensively = The context construction method encapsulates the stage identifier "initiation of single-tooth microcrack", the risk criterion "matching confidence ≥ 0.861" and the dynamic threshold factor 0.762 into a decision context environment for subsequent dual-window sliding verification mechanism to call, achieving the technical effect of adaptive binding of criteria and threshold.

[0170] S6.2: Execute logical branch judgment based on the stage identifier in the decision context environment. If it is determined to be the single-tooth microcrack initiation stage, call the dual-window sliding verification strategy configuration instruction to generate a set of time consistency verification rules for filtering random noise interference.

[0171] Based on the stage identifier in the decision context, the conditional branch control technique (parameters: stage identifier, risk criterion parameter) is used to classify and judge the evolution stage type of micro-collision blade.

[0172] By using pattern matching techniques (parameters: stage identifier, evolution stage template set), the accuracy of determining the initiation stage of single-tooth microcracks is verified, and a stage matching validity flag is obtained.

[0173] By using a strategy loading technique (parameters: stage matching validity flag, strategy library index), the dual-window sliding verification strategy configuration command is invoked, and a set of strategy execution parameters for the current operating conditions is generated.

[0174] By using a rule set construction technique (parameters: policy execution parameter set, noise interference characteristic data), a time-series consistency verification rule set is generated, resulting in multi-cycle reproducibility verification rules with noise interference filtering capabilities.

[0175] By using rule set optimization techniques (parameters: multi-period reproducibility verification rules, historical false alarm rate data), the results of the previous step are transformed into a data structure with stable time-series reproducibility criteria, thereby achieving a highly reliable filtering effect against random noise interference.

[0176] For example, in a high-density multilayer PCB drilling task, the stage identifier resolution result is the single-tooth microcrack initiation stage, and the dynamic threshold factor in the risk criterion parameter is set to 0.82. The conditional branch control module loads the single-tooth microcrack template from the evolution stage template set and determines the matching validity flag as TRUE. Based on the TRUE flag, the strategy loading module calls the dual-window sliding verification strategy in the preset strategy library index, sets the window length to two drilling cycles (each cycle lasting 0.85 seconds), the sampling frequency to 12kHz, and generates a strategy execution parameter set. The rule set construction module uses the strategy execution parameter set and noise interference characteristic data (noise spectrum concentrated in 4~6kHz, average amplitude of 0.15g) to generate consistency verification rules, where the comparison index is the amplitude change rate of the microstructure disturbance enhancement signal. The rule set optimization module combines historical false alarm rate data (false alarm count of 12, total detection count of 240) and adjusts the amplitude change rate threshold in the reproducibility criterion to 0.12, achieving a significant filtering effect on noise disturbance. The rule set is executed to verify the disturbance signals in the current and two subsequent periods. The output results have stable reproducibility indicators under noise interference, and finally form a time consistency verification rule set with high reliability and targeted noise filtering capabilities.

[0177] S6.3: The time sequence consistency verification rule set is used to perform overlapping window comparison processing on the microstructure disturbance enhancement signals in the subsequent two consecutive drilling cycles to confirm the spatiotemporal reproducibility of abnormal features and calculate the comprehensive confidence score, and output a first-level early warning trigger signal that has been double-verified.

[0178] S6.4: Based on the stage transition probability in the decision context environment, the change rate monitoring method is executed. If it is determined to be a multi-tooth cooperative collapse expansion stage and the probability value is detected to show a sharp upward trend within a preset time window, the delay verification process is skipped and a level 2 emergency warning instruction is directly generated.

[0179] Based on the stage transition probability value input in the decision context, a rate of change monitoring method (parameters: time window length, transition probability threshold, rate of change threshold) is used to achieve quantitative analysis of the state change rate in the multi-tooth cooperative collapse expansion stage.

[0180] Using a sliding time window processing method (parameter: window length T) w This allows for continuous sampling of the transition probabilities of different stages, and the resulting probability sequence P(t) of each sampling point within the window. i ).

[0181] The discrete-time difference method (parameter: sampling interval Δt) is used to calculate the rate of change of the probability sequence and generate the rate of change sequence R(t). i ).

[0182] Using a threshold comparison method (parameter: rate of change threshold R) th This allows for the processing of rate of change sequences greater than R. th The sampling points were detected, and the event flag set E of the sudden increase in rate of change was obtained. r .

[0183] An event counting and conditional judgment method is used (parameter: threshold N for the number of bursts). th The threshold P for the absolute value of the transition probability th ), for E r The number of bursts and the absolute value of the current probability are used for parallel conditional judgment. If the number of bursts is ≥ N, the condition is determined. th And the transition probability at the current stage is ≥ P th If so, a trigger flag is generated to skip the delay verification process.

[0184] By triggering a flag-driven early warning command generation method, the secondary emergency early warning template library is called to generate a secondary emergency early warning command data package associated with the current tool's unique identifier, thus achieving rapid conversion from rate of change detection results to immediate early warning signals.

[0185] By using a rate-of-change monitoring method, the stage transition probability data from the previous step is transformed into a mutation event index, enabling real-time capture and rapid early warning of the rapid evolution trend in the multi-toothed synergistic collapse expansion stage.

[0186] For example, under the condition of processing FR-4 plate with a thickness of 1.6mm and a drilling speed of 120k rpm, the time window length T w The time interval is set to 600 seconds, the sampling interval Δt is set to 5 seconds, and the rate of change threshold R is set to 5 seconds. th Set to 0.04 s -1 Threshold N for sudden increase th Set to 3, threshold P for absolute value of transition probability th The value is set to 0.75. In the transition probability sequence during the continuous sampling phase, P(t) is obtained. i The values ​​were 0.35, 0.41, 0.48, 0.56, 0.61, and 0.79, respectively, corresponding to the rate of change R(t). i The calculation is as follows: , , , , All results are greater than R. th And at the last sampling point P(t) i When ) is 0.79, P is satisfied. th Condition: The number of events with a sudden increase in the rate of change is 5 ≥ N. th If the triggering conditions are met, the system skips the delay check and directly calls the secondary emergency warning template to generate an instruction data packet, which is then pushed to the MES system, achieving an emergency intervention effect that significantly improves the response speed to tool failure trends.

[0187] S6.5: Execute graded response action mapping according to the first-level early warning trigger signal or the second-level emergency early warning command to bind the corresponding processing control strategy and message push template, and output the final adaptive early warning feedback command containing drill bit replacement suggestions and speed reduction control parameters.

[0188] For Level 1 warning trigger signals confirmed by the dual-window sliding verification mechanism or Level 2 emergency warning instructions generated directly, a hierarchical response action mapping method (parameters: stage type, risk criterion, control priority) is adopted to map different levels of warning signals to corresponding processing control strategies and message template binding functions.

[0189] By using rule matching and priority parsing techniques (parameters: dynamic threshold for risk criteria, stage identifier), the safety boundary conditions of the processing environment where the warning signal is located are retrieved and analyzed, and the control mode type and execution sequence data matching the stage are obtained.

[0190] By binding the machining strategy (parameters: control mode type, equipment control interface protocol), the spindle drive command, feed rate adjustment command and deceleration curve fitting function under the target control mode are combined with the message push template to generate a machining strategy binding report containing machining action parameters and information push content.

[0191] Using a deceleration control parameter calculation method (parameters: material type, drilling speed range, micro-chipping evolution stage), the safe deceleration threshold is calculated using the following formula: in, To achieve a safe speed reduction threshold, This is the maximum permissible rotational speed during the current processing stage. The material's deceleration coefficient. This represents the variation range of micro-chipping characteristics. This refers to the drilling time per cycle.

[0192] By using a message push template adaptation method (parameters: message content structure, MES work order interface fields), the unique drill bit identifier, recommended replacement suggestions, and speed reduction parameters are encapsulated into a structured early warning feedback data packet that conforms to the work order system's receiving standards.

[0193] By using hierarchical response action mapping and strategy binding processing, the results of the previous step are transformed into a final adaptive early warning feedback command that includes drill bit replacement suggestions and speed reduction control parameters, thereby achieving precise intervention and information synchronization during the machining process.

[0194] For example, in a scenario involving the processing of an FR-4 substrate with a thickness of 1.6 mm, a spindle speed of 120,000 rpm, and a multi-tooth cooperative chipping propagation identification stage, the stage transition probability of the warning matching result increases from 0.42 to 0.78 within a 10-minute window. Based on the graded response mapping, a level-two warning command should be triggered, and the control mode is emergency replacement and immediate speed reduction. A speed reduction coefficient is used. , characteristic change range Single-cycle drilling time seconds, maximum permissible speed rpm, substituting into the formula to calculate the safe speed reduction threshold is: ≈119972.7rpm. This value matches the deceleration curve in the processing strategy binding report, generating a standardized early warning feedback data package containing drill ID#A125, recommended replacement priority of high, and deceleration to 119972rpm. This data package is then pushed synchronously through the MES interface, enabling timely execution of processing pause and tool replacement. The equipment operation status is switched to controlled safety mode, significantly improving production line stability.

[0195] Step S7: Based on the secondary warning instruction or the primary warning signal confirmed by the dual-window sliding verification mechanism, push the drill bit replacement suggestion to the MES work order system and execute the corresponding processing pause or speed reduction control action to complete the adaptive warning feedback closed loop based on state transition. Specifically, this includes: S7.1: Obtain the first-level early warning trigger signal confirmed by the dual-window sliding verification mechanism or the directly generated second-level emergency early warning command as the input control source, and use the hierarchical response action mapping to perform logical parsing and strategy matching processing on the input control source to generate a structured adaptive early warning feedback command containing a unique drill bit identifier, recommended replacement priority and target control mode.

[0196] S7.2: Based on the unique drill bit identifier and target control mode in the structured adaptive early warning feedback instruction, the manufacturing execution system communication interface protocol is invoked to encapsulate and convert the data of the structured adaptive early warning feedback instruction, so as to generate a standardized drill bit replacement suggestion message that conforms to the receiving standard of the manufacturing execution system work order system.

[0197] S7.3: The standardized drill bit replacement suggestion message is sent to the manufacturing execution system work order system via the industrial Ethernet bus. The work order system message queue mechanism is used to asynchronously deliver and verify the status of the standardized drill bit replacement suggestion message to generate system interaction confirmation information containing a work order creation success flag and an expected response timestamp.

[0198] S7.4: Based on the target control mode and system interaction confirmation information in the structured adaptive early warning feedback instruction, the CNC system sends a speed setpoint modification command or emergency stop holding signal to the spindle drive unit through the real-time control channel to execute the machining pause action or dynamically adjust the spindle speed to the safe deceleration threshold, thereby generating the machining equipment operating parameters in a controlled state.

[0199] S7.5: Collect the operating parameters of the processing equipment under control and the system interaction confirmation information, and record them together. Use closed-loop status monitoring logic to perform integrity verification processing on the early warning execution results, so as to generate an adaptive early warning feedback closed-loop log marked as completed, realizing full-process traceability from anomaly identification to physical intervention.

[0200] Step S8: Receive manual verification confirmation information from the operator regarding the early warning execution result; encrypt and upload samples confirmed as genuine anomalies along with their original multimodal time-series data to the edge training node; incrementally update the boundary criteria of the high-fidelity multimodal stage fingerprint database based on the newly added samples to optimize the fingerprint self-evolution capability under subsequent operating conditions. Specifically, this includes: S8.1: Obtain the manual review and confirmation information of the operator inputting through the human-computer interaction interface for the secondary warning command or the primary warning trigger signal, and use data verification to verify the authenticity and completeness of the manual review and confirmation information in order to filter out the valid review records marked as real abnormal states and generate a confirmed real abnormal sample index set containing abnormal type identifier and timestamp.

[0201] S8.2: Based on the confirmed real abnormal sample index set, retrieve the corresponding original multimodal time series data block from the local storage cache, and use an asymmetric encryption method to encrypt and encapsulate the original multimodal time series data block and the associated state matching result to ensure privacy and anti-tampering characteristics during data transmission, and generate an encrypted abnormal data packet with secure transmission attributes.

[0202] S8.3: The encrypted abnormal data packet is uploaded to the edge training node via the industrial Ethernet protocol. The encrypted abnormal data packet is decrypted and isolated by the decryption module and the security sandbox mechanism to restore the plaintext multimodal temporal feature vector sequence that can be used for model training and generate an incremental training sample set to be integrated into the fingerprint database.

[0203] S8.4: Based on the distribution characteristics of the 64-dimensional real-time temporal semantic vector in the incremental training sample set, the feature space centroid position of the corresponding evolution stage category is recalculated using the cluster center offset calculation method, so as to quantify the degree of disturbance of the original category boundary by the new sample and generate a dynamic boundary criterion adjustment amount that reflects the latest working condition characteristics.

[0204] S8.5: Based on the dynamic boundary criterion adjustment amount, perform weighted fusion update processing on the processing logic parameters of the four typical states stored in the high-fidelity multimodal stage fingerprint database, namely sub-pixel level cutting edge disturbance, single-tooth microcrack initiation, multi-tooth cooperative chipping propagation, and local cutting edge stripping precursor, in order to correct the identification deviation caused by material batch changes or tool wear mode evolution, and generate an optimized high-fidelity multimodal stage fingerprint database with self-evolution capability.

[0205] 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.

[0206] 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, "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" involved in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0207] 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 method for online detection and early warning of micro-chipping in PCB drill bits, specifically including: S1: For drilling processes with various materials of different thicknesses and drilling speeds, simultaneously collect micro-chipping image sequences, spindle current transient waveforms, Z-axis feed vibration spectrum, and drilling displacement trajectories throughout the entire life cycle of the drill bit, and construct a multimodal raw dataset with four types of evolution state annotations. S2: Based on the original multimodal dataset, the energy attenuation trend of multidimensional physical quantities in a specific frequency band is jointly normalized to generate a multidimensional time-series feature vector; S3: Using multi-dimensional temporal feature vectors as training samples, a lightweight temporal encoder is used to map the features of multiple consecutive drilling cycles into fixed-length embedded representations, thereby constructing a high-fidelity multimodal stage fingerprint database that can characterize the physical laws of the entire evolution process of micro-chipping. S4: Extract a 64-dimensional real-time temporal semantic vector from the original multimodal joint temporal pattern based on online real-time input; S5: Perform similarity retrieval between the 64-dimensional real-time temporal semantic vector and the evolutionary stage template in the fingerprint database, and output the state matching result containing the matching confidence and stage transition probability; S6: Activate the processing logic of the corresponding evolution stage based on the matching confidence and stage transition probability. If it is the single-tooth microcrack initiation stage and the matching confidence meets the preset threshold, then enable dual-window sliding verification. If it is the multi-tooth collaborative collapse and expansion stage and the stage transition probability changes beyond the preset threshold within a preset time, then trigger a level 2 warning. S7: Based on a Level 2 warning or a verified Level 1 warning signal, push out a drill bit replacement suggestion and execute a processing pause or speed reduction.

2. The method for online detection and early warning of micro-chipping of PCB drill bits according to claim 1, characterized in that, Step S7 is followed by step S8: receiving manual verification and confirmation information from the operator regarding the early warning execution results, encrypting and uploading samples confirmed as genuine anomalies along with their original multimodal time-series data to the edge training node, and incrementally updating the boundary criteria of the high-fidelity multimodal stage fingerprint database based on the newly added samples to optimize the fingerprint self-evolution capability under subsequent operating conditions.

3. The online detection and early warning method for micro-chipping of PCB drill bits according to claim 1, characterized in that, Step S2 specifically includes: Based on the working condition metadata index of the multimodal raw dataset, the substrate type, plate thickness and spindle speed are extracted as working condition background parameters. The benchmark mean and standard deviation of the energy decay trend of multidimensional physical quantities under each working condition category are calculated to generate the working condition background benchmark vector. Zero-mean normalization is performed on the original image patch gray-level gradient change rate sequence to eliminate the influence of material surface texture on gradient magnitude and generate a normalized image gradient feature sequence. A piecewise linear mapping is performed on the original current envelope slope abrupt change frequency data to compress the current transient response under different drilling speed ranges to a standard unit range, generating a normalized current slope feature sequence that eliminates the rotational speed coupling effect. Logarithmic domain difference normalization was performed on the energy decay trend data of the original Z-axis acceleration power spectrum in the 12 to 28 kHz frequency band to offset the baseline drift caused by the change in plate thickness and generate a normalized vibration energy characteristic sequence. The feature data from three sources with the same timestamp are dimensionally fused to generate a multi-dimensional time-series feature vector that eliminates differences in working conditions and backgrounds and is comparable across scenarios.

4. The online detection and early warning method for micro-chipping of PCB drill bits according to claim 3, characterized in that, The dimensional fusion of the feature data with the same timestamp from the three channels is specifically achieved by dimensional fusion through multi-channel tensor splicing.

5. The online detection and early warning method for micro-chipping of PCB drill bits according to claim 1, characterized in that, Step S4 specifically includes: The energy decay trend sequence of multidimensional physical quantities within a continuous drilling cycle under the current working condition is obtained as the original multimodal joint time series model. The standardized time series data window containing the data of the last five drilling cycles is generated by truncating the data through a sliding window. By performing time axis synchronization calibration on the feature channels of image, force signal and motion parameters through multi-channel feature alignment, the phase deviation caused by the difference in sensor sampling frequency is eliminated, and a multi-channel synchronized time tensor with strict time axis alignment is generated. Load a lightweight temporal encoder model that matches the high-fidelity multimodal stage fingerprint database, input multi-channel synchronous temporal tensors into the model input layer, and start the feature extraction and dimensionality reduction mapping process; The model internally captures long-range dependencies of cross-modal features through a multi-head self-attention mechanism, performs nonlinear transformations via a fully connected projection layer, compresses redundant information and fuses cross-domain features to generate a compact temporal embedding representation with a length of 64 dimensions. The compact temporal embedding representation is normalized to eliminate the influence of dimensions and enhance the stability of feature distribution. The normalized real-time temporal semantic vector is output and transmitted to the retrieval interface.

6. The method for online detection and early warning of micro-chipping of PCB drill bits according to claim 1, characterized in that, Step S5 specifically includes: Based on standardized real-time temporal semantic vectors, 64-dimensional feature components are extracted as query sequences. A set of reference templates covering four types of micro-collapsing blade evolution states is retrieved from a high-fidelity multimodal stage fingerprint database to construct comparison data pairs and establish the basis for dynamic time warping calculation input. For the query sequence and each reference template, a local distance matrix construction operation is performed to calculate the point-to-point difference value, generate a local cost matrix representing the instantaneous state deviation, and quantify the basic discreteness. Based on the search of the global cumulative path using the local cost matrix, the traditional dynamic programming recursive formula is corrected. Under the premise of allowing nonlinear scaling of the time axis, the path with the minimum cumulative distance is found, the optimal alignment path and the corresponding minimum cumulative distance value are generated, and phase misalignment interference is eliminated. The minimum cumulative distance value is used to perform a normalized similarity mapping transformation, which is then converted into a matching confidence score between 0 and 1 through an exponential decay function. This score is then compared with a preset threshold to generate an initial matching label. Based on the preliminary matching labels and historical state transition statistics, the probability of micro-collision blade state level jump is estimated, and state matching results including final matching confidence and stage transition probability are generated.

7. The online detection and early warning method for micro-chipping of PCB drill bits according to claim 6, characterized in that, The point-to-point difference between each feature component in the query sequence and the corresponding feature component in the reference template is calculated using the Euclidean distance metric function.

8. The online detection and early warning method for micro-chipping of PCB drill bits according to claim 6, characterized in that, The modified traditional dynamic programming recursive formula is specifically modified by introducing a slope constraint factor and a step size penalty term.

9. The online detection and early warning method for micro-chipping of PCB drill bits according to claim 6, characterized in that, The estimation of the micro-collision blade state level transition probability based on the preliminary matching labels and historical state transition statistics is specifically achieved by performing Bayesian probability inference calculations and combining the current matching confidence score with the prior state transition matrix to estimate the micro-collision blade state level transition probability.

10. The method for online detection and early warning of micro-chipping of PCB drill bits according to claim 1, characterized in that, Step S6 specifically includes: The matching confidence and stage transition probability in the state matching results are analyzed to identify the micro-collision blade evolution stage and extract risk criterion parameters, and generate a decision context environment containing stage identifiers and dynamic threshold factors. The logic branch judgment is based on the stage identifier. If it is the single-tooth microcrack initiation stage, the dual-window sliding verification strategy configuration instruction is called to generate a time sequence consistency verification rule set. The overlapping window comparison of the enhanced microstructure disturbance signals in the next two consecutive drilling cycles is performed using the temporal consistency verification rule set to confirm the spatiotemporal reproducibility of the abnormal features and calculate the comprehensive confidence score, and output a first-level early warning trigger signal with double verification. If it is in the multi-tooth collaborative collapse expansion stage and the detected probability value shows a sharp upward trend within the preset time window, then skip the delay verification process and directly generate a level 2 emergency warning command. Based on the first-level warning trigger signal or the second-level emergency warning command, the hierarchical response action mapping is executed, the corresponding machining control strategy and message push template are bound, and the final adaptive warning feedback command containing drill bit replacement suggestions and speed reduction control parameters is output.

Images