A numerical control machining tool wear online monitoring and life prediction method

Abstract

The application discloses a kind of numerical control processing tool wear online monitoring and life prediction method, belong to intelligent manufacturing equipment technical field.The application is realized by FPGA multi-source sensor signal hardware level synchronous acquisition, and adaptive time-frequency feature extraction is completed by Bayesian optimization VMD and multi-head self-attention, and hybrid neural network embedding wear physical mechanism is constructed;Concept drift detection is realized by combining KS test, model incremental learning is completed using EWC algorithm, and remaining life uncertainty quantitative prediction is realized by Monte Carlo and particle filtering;Finally, through EtherCAT bus and numerical control system closed-loop linkage, process adaptive compensation and automatic tool changing are completed.The application solves the defects of traditional technology signal synchronization difference, weak model generalization, no closed-loop control, high monitoring accuracy, strong real-time, can effectively reduce tool cost, reduce unplanned downtime, suitable for multi-axis numerical control processing tool intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the fields of CNC machining condition monitoring, industrial Internet of Things (IIoT) and predictive maintenance of equipment, specifically to a method for online monitoring and life prediction of tool wear in multi-axis CNC milling, turning, drilling, and composite machining centers. This method is applicable to continuous machining scenarios of high-precision, high-value-added parts such as aerospace structural components, automotive precision parts, mold cavities, and key components of energy equipment. It enables real-time perception, dynamic evaluation, accurate prediction of remaining tool life, and closed-loop adaptive control of the machining process, belonging to the interdisciplinary technical field of intelligent manufacturing equipment and industrial intelligent operation and maintenance.

[0002] Explanation of core terms and abbreviations:

[0003] RUL: Remaining Useful Life, refers to the remaining machining time of the tool from the current monitoring moment until it reaches the maximum allowable wear level;

[0004] FPGA: FieldProgrammableGateArray, used to implement hardware-level synchronous triggering and nanosecond-level timestamp marking of multi-source sensor signals;

[0005] AE: Acoustic Emission, refers to the elastic wave signal released by material strain during tool cutting / wearing, used for wear and tool breakage monitoring;

[0006] DTW: Dynamic Time Warping, used for millisecond-level time alignment of multi-channel sensor signals;

[0007] VMD: Variational Mode Decomposition, used for adaptive time-frequency decomposition of non-stationary cutting signals;

[0008] IMF: Intrinsic Mode Function, a single-component stationary signal obtained after VMD decomposition;

[0009] BiLSTM: Bidirectional Long Short-Term Memory, a bidirectional long short-term memory network used to capture the forward and backward correlations of tool wear temporal features;

[0010] EtherCAT: Ethernet control automation technology, an industrial real-time bus used for closed-loop command interaction between monitoring systems and CNC systems;

[0011] KS test: Kolmogorov-Smirnov statistical test, used for concept drift detection in online processing conditions;

[0012] EWC: ElasticWeightConsolidation, used for incremental model learning to avoid catastrophic forgetting;

[0013] PMC: Programmable Machine Controller, used to issue process compensation and tool change macro commands;

[0014] VB: Back face wear width, a core evaluation indicator of tool wear, and a commonly used parameter in the industry;

[0015] Concept drift: Changes in processing conditions, materials, and parameters cause shifts in signal data distribution, leading to a decrease in model accuracy;

[0016] Uncertainty quantification: Upgrading lifetime point prediction to probability interval prediction and providing the confidence level of the results;

[0017] Hardware-level synchronization: Multi-sensor synchronous data acquisition is triggered by FPGA hardware pulses, which is different from traditional software timestamp alignment;

[0018] PHM2010: The publicly available benchmark dataset for tool wear released at the 2010 IEEE International Conference on Prediction and Health Management contains multi-source sensor signals throughout the milling lifecycle and the true value of tool flank wear (VB). It serves as a general validation dataset for the field of tool life prediction. Background Technology

[0019] In modern discrete manufacturing systems, CNC machining equipment is the core carrier for achieving high-precision material removal. The cutting tool, as the actuator that directly contacts the workpiece and completes the cutting action, has its wear state as a key factor determining machining accuracy, surface quality, production efficiency, and operational safety. During the cutting process, the cutting tool undergoes three stages: initial break-in, stable wear, and rapid wear. If the tool is not replaced in time during the rapid wear stage, it can easily lead to problems such as workpiece dimensional deviations, excessive surface roughness, increased machine tool vibration, and abnormal spindle load. In severe cases, tool chipping or breakage can occur, resulting in equipment damage, production interruptions, and even safety accidents. Conversely, premature tool replacement leads to insufficient tool utilization, increased consumable costs, and more unplanned downtime, directly increasing the company's overall production costs.

[0020] Industry data shows that unplanned downtime of CNC machining equipment caused by tool failure accounts for 22% to 31% of total downtime, and tool procurement, maintenance, and replacement costs account for 16% to 42% of total machining costs. Inadequate tool management has become a core bottleneck restricting the efficiency and product consistency of high-end CNC machining. Therefore, achieving online real-time monitoring of tool wear and accurate prediction of remaining tool life (RUL) is a key technological breakthrough for improving the intelligence level of CNC machining, realizing predictive maintenance of equipment, and reducing manufacturing costs.

[0021] Current tool condition monitoring technologies are mainly divided into two categories: offline detection and online indirect monitoring. Offline detection often uses equipment such as tool microscopes, laser profilometers, and 3D vision measurement devices to collect wear morphology data and calculate flank wear on disassembled tools after shutdown. This method offers high measurement accuracy but suffers from significant lag, making it unsuitable for continuous and flexible production needs. Furthermore, it requires manual intervention for measurement and data entry, hindering its integration into automated production systems. Online indirect monitoring is the mainstream approach in current industrial applications. It involves deploying sensors for vibration, acoustic emission (AE), spindle current, and cutting force on the machine tool spindle, tool holder, worktable, or power supply circuit to collect physical signals during machining. After signal preprocessing and feature extraction, these signals are combined with threshold judgment or machine learning models to achieve wear state inversion and life prediction.

[0022] Although existing online monitoring technologies have achieved initial engineering applications, several technical shortcomings still exist in real-world industrial scenarios, which are as follows:

[0023] First, the acquisition of signals from multiple sensors lacks a hardware-level time synchronization mechanism. Different sensors have inherent differences in sampling frequency, triggering logic, and transmission delay. Relying solely on software timestamp alignment will lead to phase shifts and timing errors in the signals, directly causing distortion in subsequent multimodal feature fusion and significantly reducing the reliability of state recognition and lifetime prediction. Existing publicly available technologies such as CN112846887A and CN113536560A all use software synchronization schemes, which cannot solve the timing error problem at its root.

[0024] Secondly, signal feature extraction relies excessively on manually designed time-domain and frequency-domain statistical features, such as root mean square, kurtosis, peak factor, and FFT spectral energy. These features are extremely unsuitable for complex variable working conditions such as changes in cutting parameters, fluctuations in material hardness, and intermittent cutting impacts. Furthermore, they have high feature dimensions and a lot of redundant information, which can easily lead to model overfitting and the curse of dimensionality, and cannot accurately characterize the evolution of tool wear.

[0025] Third, the prediction models mostly adopt traditional machine learning or shallow deep learning architectures driven by pure data, without incorporating the physical evolution mechanism of tool wear, such as the intrinsic mechanism of abrasive wear, adhesive wear, and thermo-coupled diffusion wear. This leads to a sharp drop in the generalization ability of the model under small sample size, extreme working conditions, and cross-material / cross-tool scenarios, and the prediction results lack interpretability and cannot provide theoretical support for operation and maintenance decisions. The algorithms accompanying the PHM2010 public dataset are all pure data-driven models and do not embed physical prior knowledge.

[0026] Fourth, the model lacks online adaptive update capability. In actual processing, there is conceptual drift between the distribution of working conditions and the distribution of training data. After long-term operation, the model accuracy continues to decline, requiring manual recalibration and training, resulting in extremely high operation and maintenance costs. At the same time, existing technologies mostly output deterministic lifetime point estimates without providing a quantitative range of prediction uncertainty. Operation and maintenance personnel cannot judge the confidence level of the prediction results, and decision-making lacks risk boundaries.

[0027] Fifth, the monitoring system and the CNC system control logic are isolated from each other, and can only realize status warning and data display. It cannot transform the monitoring results into closed-loop control actions such as process parameter adjustment and tool change command triggering. The system integration is low and it is difficult to meet the engineering requirements of unmanned and intelligent machining.

[0028] In summary, existing technologies have not yet formed a complete technical solution that integrates hardware-level signal synchronization, adaptive feature extraction, physical mechanism and data-driven fusion modeling, online adaptive updating, uncertainty quantification prediction, and closed-loop linkage of CNC systems. This makes it unsuitable for applications requiring complex and variable working conditions, high real-time performance, and high reliability in industrial settings. Therefore, developing an online tool wear monitoring and life prediction method that can overcome these technical bottlenecks and has strong engineering applicability has become an urgent need in the high-end CNC machining field. Summary of the Invention

[0029] To address the shortcomings of existing technologies, such as large synchronization errors of multi-source signals, poor feature robustness, insufficient model generalization, lack of online adaptive capability, lack of confidence intervals for prediction results, and lack of closed-loop linkage with CNC systems, this invention aims to provide an online monitoring and life prediction method for CNC machining tool wear. This method utilizes an FPGA (Field-Programmable Gate Array) hardware-level multi-source signal synchronous acquisition architecture, an adaptive multi-scale time-frequency feature extraction strategy, a hybrid neural network model embedded with physical mechanisms, an online concept drift detection and incremental learning mechanism, uncertainty quantification prediction, and an edge-cloud collaborative inference system. This enables accurate identification of tool wear status, high-confidence dynamic prediction of remaining life (RUL), and closed-loop adaptive compensation of the machining process, comprehensively improving the intelligence level and production safety of the CNC machining process.

[0030] To achieve the above objectives, the present invention adopts the following technical solution: a method for online monitoring and life prediction of CNC machining tool wear, comprising the following steps:

[0031] S100 Multi-Source Heterogeneous Signal Hardware Synchronous Acquisition and Time Alignment

[0032] A sensor array is deployed in a non-invasive installation location on the CNC machining equipment. This array includes a triaxial piezoelectric vibration sensor on the outer wall of the spindle box, a broadband acoustic emission (AE) sensor on the tool holder cone surface, and a Hall current sensor for the spindle drive circuit. This eliminates the need for traditional invasive cutting force sensors, obtaining cutting force characteristics through decoupling of current and vibration signals, thus reducing the difficulty of machine tool modification and hardware deployment costs. An FPGA hardware trigger synchronization system is built, with the hardware trigger pulse generator linked to the CNC system's M-code. A rising edge trigger signal is output 50ms before the tool enters the workpiece. The FPGA receives the signal and generates a 10μs-wide global synchronization pulse. Each acquisition channel synchronously initiates analog-to-digital conversion. Each sampling data packet is appended with a 64-bit nanosecond-level timestamp with an accuracy of no less than 10μs. The raw data is stored in the FPGA's built-in circular buffer to prevent data loss and timing errors. Signal segments are extracted using a preset sliding window with a length of 2048 sampling points and a window overlap rate of 50%, corresponding to approximately 8ms physical time at a 256kHz sampling rate. This parameter setting balances real-time performance with feature integrity. A Dynamic Time Warping (DTW) algorithm is used to perform millisecond-level time alignment on the multi-channel signals, controlling the signal alignment error within ±2 sampling points. The final output is a synchronous multimodal signal matrix with dimensions [3, 2048], providing a high-quality data foundation for subsequent feature extraction. The hardware synchronization system in this step is compatible with mainstream CNC systems such as Siemens, FANUC, Huazhong CNC, and GSK. The FPGA synchronization control board supports 16-channel synchronous acquisition and is adaptable to various types of CNC machining equipment.

[0033] S200 Adaptive Multi-Scale Time-Frequency Feature Extraction and Dimensionality Reduction Fusion

[0034] Variational mode decomposition (VMD) is performed on the synchronized multimodal signal matrix. A Bayesian optimization algorithm is used to automatically find the optimal number of modes K and the penalty factor α within a preset search space. The search space is set to K∈[3,8] and α∈[1000,10000], a range determined based on the frequency domain distribution characteristics of the tool wear signal, covering the core frequency band of wear features. The optimization objectives are minimizing the modal envelope entropy and minimizing the overlap between adjacent modal frequency bands. The number of Bayesian optimization iterations is set to 50 to ensure that the optimal VMD decomposition parameters are found in a short time, improving online processing efficiency. After decomposition, the energy spectrum, frequency band centroid, and instantaneous amplitude features of each intrinsic mode function (IMF) are extracted, and core modes highly correlated with wear are selected for subsequent processing. A convolutional-multi-head self-attention time-frequency feature encoder was constructed, consisting of four layers of residual convolutional blocks and two layers of multi-head self-attention modules concatenated together. The convolutional blocks are used to extract local time-frequency texture features of the signal, while the multi-head self-attention modules are used to capture global long-range dependent features across time steps and frequencies. The encoder output features are reduced in dimensionality by global average pooling to obtain a single-modal high-dimensional feature vector. A two-factor dynamic weighted fusion strategy is adopted, and a constrained optimization problem is constructed based on the signal-to-noise ratio and wear sensitivity. The fusion weights of each modality signal are calculated in real time, and the weights are updated every 10 sliding windows to ensure that the signal channels with high signal-to-noise ratio and high wear sensitivity dominate the feature fusion process. Adaptive weight fusion is used to remove redundant features, and the multimodal features are integrated into a 256-dimensional low-dimensional high-discriminative tool state feature vector, which significantly improves the adaptability of the features to complex and variable working conditions.

[0035] Construction of a Hybrid Neural Network Embedded with S300 Physical Mechanisms

[0036] A hybrid neural network architecture integrating data-driven and physical priors was constructed. This architecture includes a data-driven branch, a physical mechanism branch, and a gating fusion module, achieving deep coupling between wear mechanisms and data patterns. The data-driven branch employs a cascaded structure of a bidirectional long short-term memory network (BiLSTM) and a Transformer encoder. The BiLSTM layer captures the forward and backward temporal evolution of feature vectors, while the Transformer encoder extracts global contextual features, adapting to the nonlinear fitting requirements of long-term signals. The physical mechanism branch directly embeds the Arcard wear equation and the thermo-coupled wear dynamics model into the forward propagation layer of the neural network, constructing a physical regularization constraint term to replace the traditional soft-constraint loss function. This directly injects the physical laws of wear into the model training process, rather than merely using it as a loss term to assist optimization, fundamentally improving the model's interpretability and generalization ability. The initial value of the physical regularization loss weight is set to 0.1, and it is automatically adjusted during training as the model converges. The dual-branch output features are nonlinearly integrated by a gated fusion module to construct a dual-task model for wear state classification and remaining lifetime (RUL) regression. The total loss function adopts a weighted combination of cross-entropy loss, mean squared error loss, physical regularization loss and L2 regularization loss. Each loss weight is automatically optimized through backpropagation to balance physical mechanism constraints and data fitting accuracy.

[0037] S400 Model Offline Pre-training and Lightweight Edge Deployment

[0038] A training sample set was constructed by collecting historical machining data from enterprises and the publicly available PHM2010 tool wear dataset. This set included complete lifecycle data for different cutting parameters, workpiece materials, and tool models. The AdamW optimizer (a well-known optimization technique) was used for offline pre-training of the hybrid neural network, with an initial learning rate of 5e-4 and a batch size of 64. An early stopping strategy was employed to avoid model overfitting. During the inference phase, Monte Carlo Dropout and deep ensemble strategies were used to quantify prediction uncertainty. The Monte Carlo Dropout rate was kept at 0.2, and 50 forward propagations were performed to generate the prediction probability distribution. Deep ensemble used three models with different initialization seeds. Finally, the prediction mean and 95% confidence interval were calculated, upgrading traditional point prediction to probabilistic prediction. A joint strategy of 8-bit integer quantization and knowledge distillation was adopted to achieve model lightweighting. The original floating-point precision hybrid network was used as the teacher model and the lightweight MobileNetV3-LSTM variant was used as the student model. Under the premise of less than 3% accuracy loss, the model size was compressed to less than 65MB and the edge inference latency was reduced to less than 8ms. The lightweight model was deployed to industrial edge AI computing nodes to complete the initialization of the online monitoring system and meet the millisecond-level real-time inference requirements of industrial sites.

[0039] S500 Online Wear Condition Monitoring and Real-time Inference

[0040] During CNC machining, edge computing nodes collect and process multi-source signals from the current machining cycle in real time. After synchronization and feature extraction, the signals are input into a lightweight hybrid neural network model, which outputs tool wear status categories (normal, light wear, moderate wear, severe wear) and predicted RUL values. Combined with the safe wear thresholds set in the process specifications, an adaptive risk index calculation model is constructed to replace the traditional fixed threshold early warning mechanism. The risk index is jointly calculated from the predicted wear amount, the lower limit of the life confidence interval, and the cutting load. When the risk index exceeds the dynamic warning line, the system sends an early warning signal to the CNC system PLC module via EtherCAT (Ethernet Automation Technology) industrial Ethernet. A new transient tool breakage early warning function is added. The tool breakage early warning threshold is defined as the acoustic emission energy peak exceeding the benchmark value by 3 times and the vibration kurtosis mutation rate exceeding 200%, with dual judgment reducing the false alarm rate. Tool breakage faults are determined by jointly analyzing the sudden energy peak of the acoustic emission signal and the vibration kurtosis mutation rate. The fault identification response time is ≤10ms, triggering an emergency stop command to avoid equipment and workpiece damage, thus overcoming the shortcomings of traditional wear monitoring, which can only monitor progressive wear and cannot identify sudden faults.

[0041] S600 Online Concept Drift Detection and Incremental Model Update

[0042] The system continuously caches feature vectors and prediction results during online inference, performing concept drift detection every 2 hours. It employs the Kolmogorov-Smirnov (KS) statistical test to compare the difference between the current online feature distribution and the initial reference distribution, setting a dynamic judgment threshold based on the online feature variance. When the test statistic exceeds the threshold, concept drift is determined to have occurred in the processing condition. An incremental learning mechanism is triggered, selecting high-confidence samples with a confidence level greater than 0.9 and a prediction error less than 5% from the cached data. This high-confidence sample is then combined with representative difficult examples from the top 10% of prediction variance to construct a fine-tuning dataset. The Elastic Weight Consolidation (EWC) algorithm is used to apply regularization constraints to key model parameters, preventing catastrophic forgetting during incremental learning. The fine-tuning batch size is set to 32, the learning rate uses a cosine annealing strategy with an initial value of 1e-4, and the regularization coefficient λ is set to 500 to ensure online update stability. The updated model parameters are synchronized to the cloud knowledge base after hash consistency verification. After the cloud completes the model version management, the new version model is sent back to the edge computing node to replace the old version. The entire update process takes ≤3 seconds and does not affect the normal machining monitoring process. A new multi-tool collaborative update mechanism has been added, in which tools of the same model share incremental learning data, which greatly improves the model adaptation speed under small samples and new working conditions. Incremental learning adopts an experience replay pool mechanism to store recent high-confidence samples and difficult samples to avoid forgetting historical knowledge.

[0043] S700 Remaining Life Dynamic Prediction and Uncertainty Quantification

[0044] The updated hybrid neural network model outputs a multi-step RUL prediction sequence with a prediction step size of 20 steps, corresponding to approximately 160 seconds of processing time. This prediction sequence is then input into a particle filter module with 1000 particles. The state transition equation uses an exponential decay model with a random walk term, and the resampling threshold is set to 0.5% of the effective particle count. System resampling is performed when the effective particle count falls below the threshold. The filtered state sequence is then used to generate the RUL probability density function through kernel density estimation. The kernel density estimation bandwidth is adaptively calculated using the Silverman criterion, outputting the upper and lower limits of lifetime prediction at a 95% confidence level, and generating a lifetime prediction curve that evolves over time. Combining the workshop processing task scheduling plan and tool and spare parts inventory status, the optimal tool change time is automatically calculated, with a minimum safe processing margin set at 5 minutes, adapting to the tool change preparation time requirements of most continuous processing scenarios. The wear curve, lifetime prediction range, warning status, and tool change suggestions are visualized through a human-machine interface, providing maintenance personnel with intuitive decision-making support.

[0045] S800 Closed-Loop Control Feedback and Process Adaptive Compensation

[0046] Tool wear monitoring, life prediction, and early warning results are fed into the CNC machining adaptive control system as state feedback. A process parameter compensation model is constructed based on a fuzzy PID controller (a common knowledge). The input variables are the RUL decay rate and the current wear gradient, while the output variables are the feed rate compensation coefficient and the spindle speed fine-tuning. The compensation range is limited to ±5% of the process card's allowable fluctuation range, maintaining cutting force and cutting temperature within a safe range and slowing down the rate of tool wear deterioration. When the predicted lower limit of RUL falls below the minimum safe machining margin, the system automatically generates an M06 tool change macro program command, which triggers an automatic tool change after operator confirmation. After the tool change is completed, the system automatically records the tool's cumulative cutting time, peak load, wear evolution curve, and tool change reason, archiving this information to the equipment's full lifecycle management platform. This forms a closed-loop control process of "monitoring-prediction-decision-execution-archiving," achieving intelligent management of the tool's entire lifecycle. The process compensation command is sent to the CNC system via PMC (Programmable Controller) macro variables, requiring no modification to the machine tool hardware and is ready to use immediately.

[0047] The beneficial effects achieved by adopting the above structure in this invention are as follows:

[0048] 1. Achieve hardware-level nanosecond-level synchronous acquisition of multi-source signals, completely solving the problems of signal phase offset and feature distortion caused by traditional software alignment. Non-invasive sensor deployment does not require modification of machine tool structure, reducing hardware costs and installation difficulty, and providing high-quality data support for multimodal feature fusion.

[0049] 2. An adaptive feature extraction strategy combining Bayesian optimized VMD and multi-head self-attention encoder is adopted to overcome the limitations of manual feature design. Dynamic weight fusion removes redundant information, significantly improving feature robustness and discriminative power. Feature stability is improved by more than 40% under complex and variable working conditions, and the redundant feature removal rate reaches 92%.

[0050] 3. It pioneered a hybrid neural network architecture that directly embeds physical mechanisms into the network layer, integrating the Archard wear equation and thermo-dynamic coupling dynamics into model training. This fundamentally improves the interpretability and generalization ability of the model, solves the industry problem of pure data-driven models failing under extreme working conditions, and significantly enhances the generalization ability across materials and tooling scenarios.

[0051] 4. Construct an online concept drift detection and incremental learning closed loop, and combine EWC regularization and multi-tool collaborative update mechanism to achieve continuous adaptive evolution of the model without forgetting historical knowledge. After running for 100 hours, the accuracy decay rate is only 3.8%, which is far better than the 18.5% of the traditional model.

[0052] 5. By introducing Monte Carlo uncertainty quantification and particle filter lifetime prediction, deterministic point prediction is upgraded to probabilistic prediction with a 95% confidence interval, providing a risk boundary for tool change decision-making, reducing the false alarm rate to 2.1% and the false negative rate to less than 1%.

[0053] 6. A new transient warning function for tool breakage has been added, enabling full-dimensional monitoring of progressive wear and sudden fracture failures, triggering emergency protection within 10ms, and greatly improving machining safety.

[0054] 7. Adopting an edge-cloud collaborative architecture and lightweight model design, the edge inference latency is ≤8ms, meeting the real-time requirements of industry. At the same time, it realizes closed-loop linkage with the CNC system to complete process compensation and automatic tool changing. The system has a high degree of integration and engineering implementation, and can be directly applied to high-end manufacturing fields such as aerospace, automobile manufacturing, and mold processing, effectively reducing tool costs by 15%~40% and reducing unplanned downtime by more than 20%. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the overall system architecture and data flow of the method of the present invention;

[0056] Figure 2 This is a flowchart of the multi-source signal synchronous acquisition and adaptive feature extraction process of the present invention;

[0057] Figure 3 This is a flowchart illustrating the inference process of embedding the physical mechanism of the present invention into a hybrid neural network.

[0058] Figure 4 This is a flowchart of the online concept drift detection and incremental model update process of this invention;

[0059] Figure 5 This is a diagram showing the industrial field hardware deployment and CNC system communication topology of the present invention. Detailed Implementation

[0060] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Example 1: Three-axis vertical CNC milling for machining 45# steel quenched and tempered parts

[0062] This embodiment uses a VMC-850 three-axis vertical CNC milling center to machine 45# steel tempered parts as the application scenario. A Φ16mm uncoated carbide four-flute end mill is used, with a maximum allowable flank wear width (VB) of 0.3mm. The cutting parameters are set as follows: cutting speed vc = 150m / min, feed rate fz = 0.12mm / tooth, axial depth of cut ap = 2.0mm, radial depth of cut ae = 4.0mm, emulsion cooling flow rate 10L / min, and continuous milling until tool failure.

[0063] 1. Hardware System Deployment and Signal Acquisition: The hardware system consists of three parts: a sensor unit, a synchronous acquisition unit, and an edge computing unit. A PCB356A15 triaxial piezoelectric vibration sensor with a frequency response range of 0.5-15kHz and a sensitivity of 100mV / g is installed on the outer wall of the spindle box 50mm from the tool clamping surface. A PACR15a broadband acoustic emission (AE) sensor with a resonant frequency of 100-500kHz and a preamplifier gain of 40dB is installed in the transition area of ​​the tool holder's tapered surface. A Hall-type current sensor with a range of 0-50A and a bandwidth of DC-5kHz is added to the three-phase power line of the spindle drive motor. All sensor output signals are connected to a 16-bit high-speed data acquisition card via shielded cables, with a single-channel sampling rate set to 256kHz. The FPGA synchronous control board is linked with the Siemens 840Dsl CNC system M-code. A rising edge trigger signal is generated 50ms before the tool enters the workpiece. The FPGA outputs a 10μs global synchronization pulse, and each channel synchronously starts analog-to-digital conversion. Each data packet is appended with a 64-bit nanosecond-level timestamp, and the data is stored in the FPGA's circular buffer. The edge computing node uses an Intel Core i7-12700E industrial computer equipped with an NVIDIA Jetson AGX Orin AI acceleration module. It extracts data in a 2048-point sliding window, and outputs a [3,2048] synchronization signal matrix after alignment by the DTW algorithm. The alignment error is controlled within ±2 sampling points.

[0064] 2. Adaptive Multi-Scale Feature Extraction: After synchronization, the vibration, acoustic emission, and current signals are subjected to VMD decomposition. The Bayesian optimization algorithm is iterated 50 times to determine the optimal number of modes K=5 and the penalty factor α=5000. The energy spectrum and frequency band centroid features of each IMF are extracted, and three core modes are selected for subsequent processing. The time-frequency map of the core modes is input into a convolutional multi-head self-attention encoder. A 4-layer residual convolutional block extracts local time-frequency features, and a 2-layer 8-head self-attention module captures long-range dependent features. The single-mode feature vector is obtained by global average pooling. Dynamic fusion weights are calculated based on signal-to-noise ratio and wear sensitivity: vibration signal weight 0.42, acoustic emission signal weight 0.35, and current signal weight 0.23. The weights are updated every 10 windows. Finally, a 256-dimensional tool state feature vector is output, with a redundant feature removal rate of 92%.

[0065] 3. Hybrid Neural Network Construction and Offline Training: The data-driven branch of the hybrid neural network uses a 2-layer BiLSTM (128 hidden units) and a 3-layer Transformer encoder (8 heads, 256 feedforward dimensions) in series. The physical mechanism branch embeds the Archard wear equation and a thermo-coupled model to construct physical regularization constraints. The gated fusion module integrates the features of the two branches and outputs wear classification and RUL regression results. The training sample set is constructed by mixing 12,800 sets of historical processing data from enterprises with the PHM2010 public dataset. The AdamW optimizer is used with an initial learning rate of 5e-4, a batch size of 64, and an early stopping patience value of 15 rounds. After training, Monte Carlo Dropout and deep ensemble are used to generate 95% confidence intervals. After 8-bit quantization and knowledge distillation for lightweighting, the model size is 62MB, and the edge inference latency is 7.6ms, meeting the real-time requirements of industry.

[0066] 4. After the online monitoring and model adaptive update system went online, it provided real-time inference output of wear status and RUL value, achieving a wear classification accuracy of 98.7%. At 150 processing cycles, the KS test detected concept drift caused by operating condition fluctuations, automatically selecting 120 high-confidence samples to construct a fine-tuning dataset. Incremental learning was performed using the EWC algorithm, completing model updates within 3 seconds with no reduction in prediction accuracy after the update. During processing, tool breakage faults were monitored in real-time, with dual judgment based on acoustic emission and vibration kurtosis; no false alarms or missed alarms occurred throughout the process. When the risk index exceeded the threshold, an early warning signal was sent to the PLC via the EtherCAT bus, allowing maintenance personnel to view the wear curve and life confidence interval in real time.

[0067] 5. The Remaining Life Prediction and Closed-Loop Control particle filter module generates the RUL probability density function, outputting a 95% confidence interval, predicting RMSE = 4.2 minutes. Combined with the machining schedule, the system automatically calculates the optimal tool change time. When the lower limit of RUL falls below the 5-minute safety margin, the M06 tool change command is triggered. The fuzzy PID controller adjusts the feed rate to 95% and fine-tunes the spindle speed by -30 rpm based on the wear gradient and RUL decay rate, with compensation within the ±5% allowable range, effectively delaying wear deterioration. After tool change, the system automatically archives the tool's cumulative cutting time, wear data, and tool change reason to the cloud platform, completing closed-loop management of the entire process.

[0068] Example 2: CNC turning of aerospace aluminum alloy 2A12 parts

[0069] This embodiment uses a CK6150 CNC lathe and a GSK 980TDb CNC system to machine aerospace aluminum alloy 2A12 parts. The cutting tool is a Φ20mm coated carbide turning tool with a maximum permissible flank wear width (VB) of 0.25mm, a cutting speed of vc of 200m / min, and a feed rate of f of 0.15mm / r. Sensors are deployed at the lathe spindle box and tool holder. The FPGA hardware synchronization system is compatible with the lathe CNC system. Feature extraction parameters are adaptively matched to the turning conditions, and the model uses a multi-tool collaborative update mechanism. The measured results show a wear classification accuracy of 98.2%, a RUL prediction RMSE of 3.8 minutes, and a long-term accuracy decay rate of 3.5%, verifying the versatility of this invention in turning, aerospace materials, and coated tool scenarios.

[0070] Comparative experiments to verify

[0071] Under the same operating conditions, the present invention was compared with the traditional threshold method, SVM + manual feature method, and pure LSTM model in 50 repeated comparative experiments. The results are as follows: Wear classification accuracy: the present invention 98.7%, the threshold method 68%, the SVM method 85.5%, and the pure LSTM method 93.5%; RUL prediction RMSE: the present invention 4.2 minutes, the pure LSTM method 8.1 minutes, and the SVM method 12.3 minutes; False alarm rate: the present invention 2.1%, the threshold method 35%, the SVM method 12%, and the pure LSTM method 5%; Accuracy decay rate after 100 hours of long-term operation: the present invention 3.8%, and the pure LSTM method 18.5%; Edge inference latency: the present invention 7.6ms, meeting the millisecond-level real-time requirements of industry.

[0072] Experimental results show that the present invention is significantly superior to the prior art in terms of monitoring accuracy, prediction stability, anti-drift capability, and engineering applicability. It can stably adapt to complex processing scenarios in industrial sites and has outstanding creativity, novelty and practicality.

[0073] The technical solution of this invention is logically sound, has a clear implementation path, reasonable parameter settings, and sufficient experimental data. It can be directly applied to tool condition monitoring and predictive maintenance of various CNC machining equipment, and has broad industrial application prospects and market promotion value.

[0074] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for online monitoring and life prediction of CNC machining tool wear, characterized in that: Includes the following steps: The S100 uses an FPGA hardware-triggered synchronization system to acquire nanosecond-level timestamps of multi-source heterogeneous signals such as spindle vibration, acoustic emission, and spindle current, and align them with DTW at the millisecond level, outputting a synchronous multimodal signal matrix. After the synchronization signal is decomposed by Bayesian optimization VMD, the S200 extracts features through a convolutional-multi-head self-attention encoder and dynamically weights and fuses them to obtain the tool state feature vector. S300 constructs a hybrid neural network that embeds the Archard wear equation and a thermo-coupled model, integrating data-driven and physical mechanism branch features; The S400 model is pre-trained offline and then deployed to edge computing nodes in a lightweight manner. The S500 online inference outputs wear status and RUL prediction values, triggering an early warning. S600 uses the KS test to detect concept drift and employs the EWC algorithm to complete incremental model updates. S700 combined with particle filtering enables quantitative prediction of RUL uncertainty; The S800 is linked with the CNC system via EtherCAT bus to complete process compensation and automatic tool changing closed-loop control.

2. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: The sensors mentioned in step S100 are non-invasive triaxial piezoelectric vibration sensors, broadband acoustic emission sensors, and Hall current sensors, which are respectively installed on the machine tool spindle box, tool holder tapered surface, and spindle drive circuit.

3. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S100, the FPGA hardware trigger pulse is linked with the CNC system M code. The trigger signal is output 50ms before the tool enters the workpiece. The sampled data packet is appended with a 64-bit nanosecond-level timestamp, and the timestamp accuracy is not less than 10μs.

4. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S200, the optimal number of modes K∈[3,8] and the penalty factor α∈[1000,10000] of the VMD decomposition are optimized by Bayesian optimization with the goal of minimizing the mode envelope entropy.

5. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S300, the hybrid neural network includes a BiLSTM-Transformer data-driven branch, a physical mechanism branch, and a gated fusion module, and the total loss function incorporates the physical regularization loss.

6. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S400, the model is lightweighted by 8-bit integer quantization and knowledge distillation, with a size ≤65MB and an edge inference latency ≤8ms.

7. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: Step S500 adds a transient warning for broken blades, which is determined by both the peak value of acoustic emission energy and the abrupt change rate of vibration kurtosis, and the fault response time is ≤10ms.

8. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S600, incremental learning adopts an experience replay pool and a multi-tool collaborative update mechanism, and the entire model update takes ≤3 seconds.

9. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S700, the RUL prediction outputs a 95% confidence interval, which is combined with the machining schedule to automatically calculate the optimal tool change time, with a minimum safe machining margin of 5 minutes.

10. The method for online monitoring and life prediction of CNC machining tool wear according to claim 1, characterized in that: In step S800, the process compensation range is limited to ±5%, and compensation instructions are sent to the CNC system through PMC macro variables.

Images