Cutter wear monitoring and predicting method

Abstract

The invention relates to the field of tool wear in the numerical control machining process, in particular to a tool wear monitoring and predicting method which comprises the steps of signal collection, multi-domain feature extraction, multi-modal feature fusion, tool wear prediction model construction, real-time monitoring and wear prediction. And a multi-signal fusion monitoring and prediction model is constructed, so that the precision of tool wear monitoring and the reliability of prediction are remarkably improved, and the defects of a traditional method are effectively overcome.

Description

Technical Field

[0001] The present invention relates to the field of tool wear during numerical control machining, and particularly to methods for tool wear monitoring and prediction. Background Art

[0002] In the machining industry, tool wear has a crucial impact on machining accuracy, production efficiency, and cost control. Traditional time-based tool wear monitoring and prediction methods, which are based on fixed time thresholds (such as replacing the tool every 2 hours of machining) or cumulative machining time for wear monitoring and prediction, do not rely on real-time physical signals and only indirectly infer the wear state through time intervals. They do not consider the dynamic changes in machining conditions (such as feed rate, cutting depth, material hardness differences) and the non-linear characteristics of actual tool wear (such as slow initial wear and accelerated wear in the later stage), resulting in a disconnection between the monitoring results and the actual wear process. This method cannot accurately match the actual wear process of the tool. In the context of the current high-precision machining requirements, the drawbacks of this traditional method are becoming increasingly prominent and it is difficult to meet the strict requirements of modern manufacturing industries for efficient and precise machining. For example, in the manufacturing of precision molds, slight tool wear may cause dimensional deviations in the mold, seriously affecting product quality and increasing production costs. Therefore, it is urgent to develop a more accurate and efficient tool wear monitoring and prediction technology.

[0003] In view of this, the present invention provides a method for tool wear monitoring and prediction. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for tool wear monitoring and prediction in view of the deficiencies of the prior art.

[0005] To solve the above technical problems, the following technical solutions are adopted: A method for tool wear monitoring and prediction, characterized by comprising the following steps: S1. Signal acquisition: Acquire the cutting displacement signals of the tool and the workpiece during machining, denoted as D(t), and measure the displacement error data by obtaining the position change between the cutting point of the tool and the workpiece; acquire the cutting signals during machining, where the cutting signals include cutting vibration signals and cutting energy consumption signals, denoted as V(t) and E(t) respectively; measure the wear value W of the flank face of the tool , , ,

[0005] , real ,

[0006] , .

[0006] S2. Multi-domain feature extraction: Perform feature extraction on the acquired cutting displacement signals, cutting vibration signals, and cutting energy consumption signals respectively, screen out displacement-sensitive features highly correlated with displacement error to form a displacement-sensitive feature set, screen out vibration-sensitive features highly correlated with abnormal vibration to form a vibration-sensitive feature set, and screen out energy consumption-sensitive features highly correlated with abnormal energy consumption to form an energy consumption-sensitive feature set.

[0007] S3. Multi-modal Feature Fusion: Based on the selected displacement-sensitive features, vibration-sensitive features, and energy consumption-sensitive features, use the SVM model to establish the non-linear mapping relationships between the displacement-sensitive features and the vibration-sensitive features, and between the displacement-sensitive features and the energy consumption-sensitive features respectively; further, through the Bayesian optimization algorithm, establish the three-dimensional non-linear mapping relationship among the displacement-sensitive features, vibration-sensitive features, and energy consumption-sensitive features.

[0008] S4. Tool Wear Prediction Model Construction: Use the processed cutting displacement signal as the input and the tool wear value measured by a hand-held microscope as the label, establish a tool wear prediction model based on the GRU network, and incorporate the attention mechanism in the dimension of the cutting displacement for model training; use the tool wear monitoring model to monitor and output the real-time wear value of the tool.

[0009] S5. Real-time Monitoring: Input the real-time collected signal into the trained tool wear prediction model to monitor the tool wear status in real time.

[0010] S6. Wear Prediction: Predict the remaining service life of the tool according to the monitoring results and model predictions.

[0011] Due to the adoption of the above technical solutions, the following beneficial effects are achieved: By providing a tool wear monitoring and prediction method, the present invention constructs a monitoring and prediction model for multi-signal fusion by integrating the cutting displacement signal, cutting vibration signal, and cutting energy consumption signal, thereby significantly improving the accuracy of tool wear monitoring and the reliability of prediction, and effectively making up for the defects of traditional methods.

[0012] Multi-signal fusion improves monitoring accuracy: Integrate the three-modal signals of displacement, vibration, and energy consumption, covering multiple physical mechanisms of wear. The monitoring accuracy is improved by more than 50% compared with the traditional single-signal method, and the dimensional deviation is reduced by 50% especially in the precision machining scenario.

[0013] Advanced model enhances prediction reliability: The GRU network combines Bayesian optimization and the attention mechanism to effectively process complex spatio-temporal data. The remaining life prediction error ≤ 10%, reducing the risk of unexpected tool wear.

[0014] Full-process real-time performance and adaptability: End-to-end low latency (≤ 100ms) from signal acquisition to prediction output, supporting multiple materials and multiple working conditions (such as variable curvature paths, high-load cutting), and the model has strong generalization ability.

[0015] Intelligent production support: Through real-time monitoring, dynamic alarm, and life prediction, realize the closed-loop of "state perception - trend prediction - proactive maintenance" of tool wear, improve the intelligent level of the manufacturing process, reduce manual intervention, and improve production efficiency.

[0016] Effectively make up for the defects of traditional methods: Solve the core problem of the traditional time-based monitoring method that "it cannot fit the actual wear process", and achieve the technological leap from "timed maintenance" to "condition-based predictive maintenance". Brief Description of the Drawings

[0017] The present invention will be further described below with reference to the accompanying drawings: FIG. 1 is an overall flowchart of the tool wear monitoring and prediction method of the present invention.

[0018] FIG. 2 is a diagram of the multi-modal feature fusion process of the present invention.

[0019] FIG. 3 is a flowchart of the dynamic weight calculation in step S3 of the present invention.

[0020] FIG. 4 is an architecture diagram of the tool wear prediction model of the present invention.

[0021] FIG. 5 is a schematic diagram of the sequence construction in step S4 of the present invention.

[0022] FIG. 6 is an architecture diagram of the spatio-temporal hybrid GRU network in step S4 of the present invention.

[0023] Figure 7 is an architecture diagram of the real-time monitoring system of the present invention.

[0024] Figure 8 is a real-time monitoring stream processing timing diagram of step 5 of the present invention.

[0025] Figure 9 is a schematic diagram of the HMI interface of the present invention. Detailed Embodiments

[0026] To make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below through the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the scope of the present invention. In addition, in the following description, the descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0027] Refer to Figures 1-9 , the tool wear monitoring and prediction method includes the following steps: S1. Signal acquisition: Collect the cutting displacement signals of the tool and the workpiece during machining of the workpiece, denoted as D(t), and measure the displacement error data of the position change between the cutting point of the tool and the workpiece; collect the cutting signals during machining of the workpiece, and the cutting signals include cutting vibration signals and cutting energy consumption signals, denoted as V(t) and E(t) respectively; measure the wear value W of the flank face of the tool real .

[0028] Specifically, in this embodiment, a displacement sensor, an acceleration sensor, a current sensor, and a power analyzer are installed on a numerically controlled machine tool.

[0029] Among them, the acquisition method of the cutting displacement signal includes installing a displacement sensor on the tool of the machine tool, or adding IO signals indicating the start and end of the monitoring process segment in the machining program of the machine tool, or the machine tool obtaining the segment signals of the start or end of the machining through communication with the NC system. Then, the absolute cutting displacement of the tool and the workpiece during the machining process can be obtained.

[0030] The acquisition method of the cutting vibration signal includes installing a piezoelectric acceleration sensor on the numerically controlled machine tool.

[0031] The acquisition method of the cutting energy consumption signal includes installing a corresponding power sensor or current sensor on the outlet side of the power driver of the machining axis of the machine tool.

[0032] Measurement principle: The energy consumption generated by the tool cutting the workpiece is within a set range. Through the displacement sensor or the IO signals at the start and end of the process segment in the machining program of the machine tool, the actual coordinate change amount in each machining process segment and each segment can be known, and thus the cutting displacement signal of the tool and the workpiece during the machining of the workpiece in this machining interval can be known. At the same time, by measuring the cutting vibration signal and the cutting energy consumption signal in this machining interval, if both the cutting vibration signal and the cutting energy consumption signal are within the set threshold range, it can be determined that the tool is within the normal service life range and can be used normally; if the cutting vibration signal and the cutting energy consumption signal are not within the set threshold range, through the constructed wear prediction model, the tool wear monitoring model can be used to monitor and output the real-time wear value of the tool, and the remaining service life of the tool can be predicted according to the monitoring results and model prediction.

[0033] Taking the laser displacement sensor as an example, the displacement sensor satisfies the sampling frequency f s ≥2f max , where f max is the highest frequency of the signal; the laser displacement sensor works based on the laser triangulation principle, and its measurement formula is: , where L is the displacement amount, d is the distance between the laser emission point and the receiving point, α is the laser emission angle, and β is the laser reflection angle.

[0034] Taking the piezoelectric acceleration sensor as an example, the acceleration sensor collects the cutting vibration signal, and its working principle is based on the piezoelectric effect. The relationship between the generated electric charge Q and the acceleration a is Q = d ij F (d ij is the piezoelectric coefficient, and F is the force generated by the acceleration), and the cutting vibration signal can be obtained through conversion.

[0035] The power sensor collects the cutting energy consumption signal. The measurement basis is the definition of power P = UI (where U is the voltage and I is the current), and the energy consumption signal is obtained by integrating the power over time. Data is collected under different combinations of cutting speed, feed rate, and cutting depth, while recording the influence of factors such as ambient temperature and humidity.

[0036] Generally speaking, step S1 adopts multi-dimensional data acquisition: the cutting displacement signal, cutting vibration signal, and cutting energy consumption signal are synchronously collected by a laser displacement sensor, an acceleration sensor, and a power sensor, covering the geometric changes, mechanical abnormalities, and energy loss characteristics of tool wear. The displacement signal reflects the relative position error between the tool and the workpiece, the vibration signal characterizes the cutting stability, and the energy consumption signal reflects the machining load. The three together constitute the multi-physical representation of the wear state.

[0037] S2. Multi-domain feature extraction: Feature extraction is performed on the collected cutting displacement signal, cutting vibration signal, and cutting energy consumption signal respectively. Displacement-sensitive features highly correlated with displacement error are selected to form a displacement-sensitive feature set, vibration-sensitive features highly correlated with vibration abnormalities are selected to form a vibration-sensitive feature set, and energy consumption-sensitive features highly correlated with energy consumption abnormalities are selected to form an energy consumption-sensitive feature set.

[0038] In step S2, the collected cutting displacement signal is processed by the displacement domain index method, and the extracted feature set F d = {displacement mean , displacement variance , displacement peak value P d , skewness S d , kurtosis K d}, and a correlation function is established between the displacement signals of two different tools and workpieces; ; ; ; ; ; where represents the value of the cutting displacement signal at the i-th sampling point, where i is the index of the sampling point, i = 1, 2,... N , N is the total number of sampling points.

[0039] The displacement mean is the result obtained by averaging the values of the cutting displacement signal at all sampling points, which reflects the average level of the cutting displacement signal in a certain interval.

[0040] Displacement variance It is used to measure the degree to which the displacement signal deviates from the mean. The larger the variance, the greater the fluctuation of the displacement signal.

[0041] Displacement peak value P d It is the maximum value of the cutting displacement signal among the sample points, which can reflect the extreme change of the displacement signal at a certain moment.

[0042] Skewness S d It is used to describe the asymmetry of the displacement distribution. When, S d = 0, the distribution is symmetric; S d >0 indicates a positively skewed distribution, that is, the tail on the right (the direction of larger values) is longer; S d <0 indicates a negatively skewed distribution, that is, the tail on the left (the direction of smaller values) is longer.

[0043] Kurtosis K d It is used to measure the peakedness of the distribution. Compared with the normal distribution, K d The larger it is, the sharper the peak of the distribution, and the greater the probability of extreme values on both sides; K d The smaller it is, the flatter the distribution.

[0044] Using wavelet transform Extract wavelet coefficients at different scales as features, and establish the correlation function between two different displacement signals D1(t) and D2(t) of the tool and the workpiece , ; is the result of wavelet transform, indicating the value of wavelet transform of the cutting displacement signal when the scale parameter is a and the translation parameter is b.

[0045] a controls the stretching of the wavelet function. A larger value of a corresponds to a wider wavelet function, which is used to analyze the low-frequency and global features of the signal; a smaller value of a corresponds to a narrower wavelet function, which is used to capture the high-frequency and local features of the signal.

[0046] b is the translation parameter, which is used to translate the wavelet function on the time axis, so as to analyze the signal at different positions.

[0047] is the wavelet basis function, which is the core of wavelet transform. Different wavelet basis functions have different characteristics, which will affect the result of wavelet transform and the ability to extract signal features.

[0048] ; where , are the means of D1(t) and D2(t) in the time period (1, T) respectively.

[0049] is the time delay.

[0050] In step S2, the collected cutting vibration signal is processed by the frequency-domain index method. The spectrum V(f) is obtained through fast Fourier transform, and the feature set F v ={vibration energy E v , main frequency f v , frequency band energy ratio R v}, where: ; where f is the frequency point, f s is the sampling frequency, and its value range is 0 ≤ f ≤ f s / 2, f s / 2 is the Nyquist frequency, representing the highest effective frequency for signal analysis; ; where argmax f is the frequency value corresponding to the maximum spectrum amplitude, and |V(f)| is the absolute value of the spectrum amplitude; ; where Δf is the frequency band width centered on the main frequency f 𝑣 , the numerator is the total energy sum within the frequency band near the main frequency, and the denominator is the total vibration energy E v .

[0051] In step S2, the acquisition and feature extraction of the cutting energy consumption signal are realized by a power sensor, which specifically includes the following steps: (a) Real-time acquisition of power signal: Install a power sensor at the input end of the machine tool spindle motor to measure the three-phase voltages V a (t), V b (t), V c (t) and the currents I a (t), I b (t), I c (t), and calculate the instantaneous power: , where k = a, b, c: identifiers of three-phase electricity, P(t): instantaneous power, representing the total input power of the machine tool spindle motor at time t; by directly measuring the sum of the products of the three-phase voltages and currents, the instantaneous power of the machine tool is obtained in real time, providing a basis for subsequent energy consumption calculation.

[0052] And the cutting energy consumption signal is obtained through integration: .

[0053] where the integration variable τ is the time variable; t0: integration start time (such as the machining start time); t: current time; E(t): cumulative cutting energy consumption from t0 to t. The total energy consumption is obtained by integrating the instantaneous power, which directly reflects the energy consumption of the tool during the machining process and is positively correlated with the tool wear degree.

[0054] Among them, 1. Power sensor selection and installation: A high-precision three-phase power sensor (such as Yokogawa WT1800) is adopted, with a bandwidth ≥ 5 kHz and an accuracy class of 0.1%; The power sensor is installed at the power input terminal of the spindle motor, and synchronous sampling is carried out through a current clamp and a voltage probe to avoid phase deviation; (b) Signal preprocessing and noise reduction: Perform moving average filtering on the original power signal to filter out high-frequency noise; Adaptive calibration is adopted to eliminate the interference of grid voltage fluctuations: ; where V nominal is the rated voltage, and V actual (t) is the real-time grid voltage.

[0055] Signal synchronization and calibration: Collect the no-load power P idle of the motor under no-load conditions, and the actual cutting energy consumption is corrected as: ; Calibrate the linearity of the sensor through a standard resistance load to ensure that the full-scale error ≤ 1%; (c) Feature extraction: Mean energy consumption: ; Among them, T: analysis time window length (unit: second, s); : mean energy consumption (unit: joule, J), which characterizes the average energy consumption level; Energy consumption volatility: ; Among them : energy consumption variance, which reflects the fluctuation intensity of energy consumption; Peak energy consumption ratio: .

[0056] Among them, max: maximum value function; R peak : peak energy consumption ratio, which characterizes the energy impact under extreme working conditions.

[0057] Feature engineering optimization: Harmonic order selection: According to the characteristics of the spindle motor, focus on monitoring the 3rd, 5th, and 7th harmonics (H = 7); Frequency band adaptive division: Dynamically set the characteristic frequency band based on the spindle speed n (rpm).

[0058] Furthermore, through the above, the power harmonic distortion rate THD can be further calculated P where Ph is the power amplitude of the h-th harmonic (unit: watt, W); P1 is the fundamental wave power (unit: watt, W); THD P : the total harmonic distortion (unit: %), which reflects the interference degree of the power grid quality on the power signal.

[0059] Generally speaking, step S2 adopts multi-domain sensitive feature screening: Displacement domain: extract features such as mean, variance, peak value, wavelet coefficient, etc., to capture the displacement fluctuation trend and spatial correlation (such as the correlation function of the tool-workpiece displacement signal).

[0060] Vibration frequency domain: extract vibration energy, main frequency, and frequency band energy ratio through Fourier transform to identify abnormal frequency components of cutting vibration (such as sudden change of the main frequency indicating the risk of resonance).

[0061] Energy consumption domain: calculate the mean energy consumption, volatility, and peak energy consumption ratio to quantify the energy consumption level and impact intensity, and reflect the tool load state.

[0062] Screen features that are highly correlated with displacement error, vibration abnormality, and energy consumption abnormality, eliminate redundant information, and improve the mapping accuracy between features and wear states.

[0063] S3. Multi-modal feature fusion: Based on the selected displacement-sensitive features, vibration-sensitive features, and energy consumption-sensitive features, use the SVM model to establish the non-linear mapping relationship between the displacement-sensitive features and the vibration-sensitive features, and the non-linear mapping relationship between the displacement-sensitive features and the energy consumption-sensitive features respectively; further through the Bayesian optimization algorithm, establish the three-dimensional non-linear mapping relationship between the displacement-sensitive features, vibration-sensitive features, and energy consumption-sensitive features.

[0064] The specific process of step S3 is as follows: (a) Displacement signal segmentation: Divide the displacement error signal D(t) into N displacement interval segments {S k} k-1 N according to the machining stage; the segmentation rule is: δ D : the displacement change threshold, which is dynamically adjusted according to the machining material. For example: δ D = 0.1 mm (aluminum alloy), δ D = 0.05 mm (titanium alloy).

[0065] t k-1 ,t k: The start and end times of the k-th segment, triggered by ΔD(t) ≥ δ D Trigger the end of the segment.

[0066] Divide the machining stage by a significant jump in displacement change (exceeding δ D ). The displacement change within each segment is small, reflecting the motion state of the tool during a certain stable cutting stage, and avoiding over-segmentation caused by minor fluctuations.

[0067] (b) Displacement-vibration feature correlation modeling: For each segment S k , construct the displacement-vibration feature pair (F d (k) , F v (k) ), where F d (k) = {μ d (k) , σ d (k) , P d (k)} are the mean, variance, and peak value within the displacement segment.

[0068] Use the kernel function SVM to establish a non-linear mapping Φ dv : F d (k) → F v (k) ). The objective function is: .

[0069] : Radial basis kernel function mapping; C: Penalty factor, balancing the maximization of the margin and the classification error; y k is a binary label based on vibration feature anomalies (y k = 1 indicates vibration or energy consumption anomalies, and y k = -1 otherwise).

[0070] (c) Displacement-energy consumption dynamic weighted fusion: For each segment S k , calculate the dynamic weights of the displacement and energy consumption features: where Δ D k = max t∈Sk D ( t ) - min t∈Sk D ( t ), β is the scaling factor; Establish a mapping through weighted SVM , the loss function is: .

[0071] For the feature set F extracted from the cutting displacement signal d , perform feature correlation modeling optimization, and the selected kernel function is as follows: Adopt a composite kernel , where α ∈ [0, 1] is determined by cross-validation; Anomaly label generation: Define the vibration anomaly criterion: where w th , b th is a preset threshold vector.

[0072] (d) Displacement-guided Bayesian fusion: Use the outputs of Φ dv and Φ dv as inputs to construct the joint feature space Z = [Φ dv , Φ dv ; Maximize the posterior probability through Bayesian optimization: where is a hyperparameter, and the optimization goal is to minimize the mean square error (MSE) between the predicted wear value and the true value.

[0073] Adopt incremental SVM (Online SVM) to update the model parameters, and the processing time of each segment S k ≤ 2ms; Bayesian optimization is completed offline, and the optimal hyperparameter table is directly loaded in the online stage to ensure real-time performance.

[0074] Using the above method, the technical effects are as follows: Asynchronous signal processing: By means of displacement segmentation and event triggering mechanism, avoid the problem of multi-sensor time synchronization.

[0075] Physics-driven modeling: The displacement change threshold δ D is associated with material properties to enhance the adaptability to working conditions.

[0076] Dynamic weight allocation: Automatically adjust the weight of the energy consumption feature based on the displacement amplitude to improve the mapping robustness.

[0077] Lightweight calculation: Combine incremental SVM and offline Bayesian optimization to meet the real-time requirements.

[0078] Generally speaking: Step S3 adopts non-linear mapping modeling: Use SVM to establish the non-linear mapping relationships between displacement-vibration and displacement-energy consumption, and capture the implicit correlations between multiple signals (such as when the displacement fluctuation intensifies, the vibration energy increases and the energy consumption increases).

[0079] Dynamic weight and three-dimensional fusion: Construct a three-dimensional non-linear mapping of displacement-vibration-energy dissipation through Bayesian optimization, adaptively adjust the feature weights (such as assigning higher weights to energy dissipation features in high displacement change segments), and enhance the robustness of feature fusion.

[0080] Process adaptability: Based on segmented processing in the machining stage (such as dividing intervals according to displacement change thresholds), adapt to the cutting characteristics of different materials (such as aluminum alloy, titanium alloy), and improve the generalization ability of the model.

[0081] S4. Construction of tool wear prediction model: Using the processed cutting displacement signal as the input and the tool wear value measured by a hand-held microscope as the label, establish a tool wear prediction model based on the GRU network, and incorporate an attention mechanism in the dimension of cutting displacement for model training; use the tool wear monitoring model to monitor and output the real-time wear value of the tool.

[0082] In step S4, a tool wear prediction model is constructed based on the GRU network, and an attention mechanism is incorporated in the dimension of cutting displacement. The displacement-sensitive feature set is constructed into an input matrix according to the spatial sequence. The specific steps are as follows: (a) Spatial sequence definition and feature mapping: Using the processed cutting displacement signal D(t) as the input and the tool wear value as W real label; According to the geometric coordinates (x, y, z) of the tool processing path, the displacement-sensitive feature set F d is sorted according to the spatial position to construct a spatial sequence input matrix Xs ∈ R M×5 , where M is the total number of path points; The spatial sorting rule is: along the tool feed direction, arrange in ascending order of the Euclidean distance between adjacent path points to ensure spatial continuity.

[0083] Spatial sequence construction: Use the G code provided by the numerical control system (such as FANUC) to parse the machining path coordinates, and the sampling interval is 0.1 mm; for non-continuous paths (such as reverse cutting), generate a smooth spatial sequence through B-spline interpolation.

[0084] (b) Design of spatial attention mechanism: Introduce spatial position encoding (Positional Encoding) before the GRU network, and the encoding formula is: , where pos is the spatial position index and d model = 32 is the encoding dimension; Define the spatial attention weight β ij , and calculate the correlation between different positions: , where q i , k j are the query vector and the key vector, which are linearly transformed from the input features.

[0085] Spatial attention optimization: Introduce a local attention window (window size L = 5), and only calculate the correlation of adjacent L positions to reduce the computational complexity; For high-curvature path regions (curvature > 0.1 mm -1 ), dynamically increase the window to L = 10 to enhance feature correlation.

[0086] (c) Spatial-temporal hybrid GRU network: Input the spatial sequence matrix X S and the time series signal are processed in parallel. The network structure includes two branches: Temporal branch: Standard GRU layer, which processes the features in the time dimension; Spatial branch: Spatial GRU layer, cell state: The outputs of the two branches are merged through a gated fusion module: , where W g is a trainable parameter matrix.

[0087] Hybrid network training strategy: Adopt the alternating training method: First, freeze the temporal branch and train the spatial branch until convergence; then jointly fine-tune the parameters of the two branches; (d) Loss function with enhanced spatial features: Based on the Huber loss, add a spatial consistency constraint term: where is the attenuation coefficient, which is used to punish the mutation of the predicted values of adjacent spatial points; the attenuation coefficient is adaptively adjusted according to the material hardness.

[0088] The total loss function is: . The weight of the constraint term is determined by grid search on the validation set, and the range is [0.01, 0.1].[[]END]

[0089] The main technical effects of the above method are as follows: Spatial-temporal bimodal fusion: Simultaneously capture the time dynamics and spatial distribution rules through the hybrid GRU network to improve the prediction accuracy; Physically guided attention mechanism: Combine the geometric characteristics of the processing path to enhance the sensitivity of the model to key regions (such as high-curvature segments); Anti-mutation design: The loss of spatial consistency effectively suppresses the jump of predicted values caused by measurement noise and improves output stability; Engineering adaptability: Compatible with the data interface of the numerical control system and can be directly deployed on existing processing equipment.

[0090] Generally speaking: A tool wear prediction model is constructed based on the GRU network, and an attention mechanism is incorporated in the dimension of cutting displacement. The displacement-sensitive feature set is constructed as an input matrix according to the spatial sequence. A spatio-temporal hybrid model is constructed based on the GRU network. The spatial branch processes the path coordinate sequence of displacement features, the temporal branch captures the time dynamics of vibration / energy consumption, and the gated fusion module synthesizes spatio-temporal information.

[0091] Enhanced attention mechanism: Spatial attention focuses on key areas such as high-curvature paths; the loss function introduces spatial consistency constraints to ensure the smoothness of predicted values at adjacent positions and improve prediction stability.

[0092] High-precision modeling: The GRU network combined with the attention mechanism effectively captures the long-distance dependence and local anomalies of wear, and the prediction error is reduced by more than 30% compared with traditional models.

[0093] S5. Real-time monitoring: Input the signals collected in real time into the trained tool wear prediction model to monitor the tool wear status in real time.

[0094] Refinement of specific implementation methods: 1. Hardware deployment architecture: An edge computing device (such as NVIDIA Jetson Xavier NX) is adopted to run the real-time Linux kernel (PREEMPT-RT); Sensor interface: The cutting displacement signal is accessed through a PCIe data acquisition card (NI PCIe-6321), and the cutting vibration signal and cutting energy consumption signal are synchronized through the EtherCAT bus.

[0095] 2. Streaming processing optimization: Zero-copy data transfer: Use shared memory (SHM) to transfer preprocessed features to avoid memory copy overhead; Priority scheduling: Set the model inference thread to the highest real-time priority (SCHED_FIFO) to ensure cycle stability.

[0096] 3. Fault tolerance mechanism design: Data loss compensation: If the packet loss rate of sensor data > 5%, start the Kalman filter to predict missing values: , , where A and B are state transition matrices, and Q is the process noise covariance.

[0097] Model health monitoring: Periodically calculate the prediction confidence , If Confidence < 0.8, trigger model reloading; 4. Real-time verification: Use an oscilloscope (Keysight DSOX1204G) to measure the end-to-end delay (from signal acquisition to HMI display), required to be ≤ 100 ms; Verify the system throughput through stress testing (simulating 100-channel concurrency).

[0098] In step S5, the specific implementation of the real-time monitoring includes the following steps: (a) Real-time signal synchronous acquisition and preprocessing: Synchronously acquire the displacement signal D(t), vibration signal V(t), and energy consumption signal E(t) through a multi-threaded architecture, with the sampling frequencies not less than 10 kHz, 20 kHz, and 5 kHz respectively; Perform online filtering and standardization on the original signal; , where , is the mean and standard deviation within a sliding window (length = 1 s); (b) Streaming feature extraction and caching: Intercept the signal according to the time window T win = 0.5 s, and real-time extract the displacement-sensitive feature F d , vibration-sensitive feature F V , and energy consumption-sensitive feature F e , and store them in a circular buffer (capacity N buf = 10 windows); The feature extraction delay is constrained to be ≤ 5 ms, and accelerated through parallel (OpenMP) computing; (c) Model inference and state update: Input the feature set X = [F d , F v , F e into the pre-trained GRU-attention model, and output the wear value ; The model inference period Tinf = 50 ms, adopt the TensorRT optimized inference engine, and the single inference time ≤ 10 m; Adopt exponential weighted moving average to smooth the prediction result: , α = 0.7 (d) Anomaly detection and alarm: Define the dynamic wear rate threshold: ; where , is the mean and standard deviation of the historical wear rate; If the wear rates in three consecutive cycles satisfy , trigger a level-three alarm (warning, severe alarm, shutdown); (e) Data persistence and visualization: Store real-time data in a time-series database, supporting time-range queries and playback; Push the wear value curve to the HMI interface via WebSocket, with a refresh rate ≥ 30Hz.

[0099] The technical effects of adopting the above method are as follows: Full-link low latency: The end-to-end latency from signal acquisition to prediction output is controlled within 100ms, meeting the requirements of high-speed machining scenarios.

[0100] Adaptive dynamic threshold: A threshold update mechanism based on historical data to avoid false alarms / missing alarms caused by fixed thresholds.

[0101] Edge-cloud collaboration: Key data is synchronously uploaded to the cloud for long-term trend analysis, and only the real-time inference module is retained locally.

[0102] Visualization and traceability: The HMI interface supports wear process playback and alarm log export, facilitating fault tracing.

[0103] Generally speaking: Step S5 adopts low-latency real-time processing: Multithreaded synchronous signal acquisition, with a streaming feature extraction latency ≤ 5ms, a model inference cycle of 50ms (single inference ≤ 10ms after TensorRT optimization), and an end-to-end latency ≤ 100ms, meeting the requirements of high-speed machining.

[0104] Dynamic anomaly detection: Based on the dynamic wear rate threshold of historical data, combined with a three-level alarm mechanism, reducing false alarms / missing alarms caused by fixed thresholds.

[0105] Data visualization and traceability: Push the real-time wear curve to the HMI interface (refresh rate ≥ 30Hz), store the data in a time-series database for playback, facilitating fault tracing and process optimization.

[0106] S6. Wear prediction: Predict the remaining service life of the tool based on the monitoring results and model prediction. The specific implementation of the wear prediction includes the following steps: Based on the exponential degradation model W(t) = W0 * exp(-λt), fit the parameter λ by the least squares method to predict the remaining service life , where W th is the wear threshold.

[0107] Generally speaking: In step S5, remaining useful life (RUL) quantification prediction is adopted. Based on the exponential degradation model, the wear trend is fitted by the least square method, and combined with real-time monitoring data calculation, it provides a quantitative basis for tool replacement.

[0108] Preventive maintenance: Predict the tool failure time in advance, avoid unexpected shutdowns, reduce tool waste (such as the waste of un-worn tools caused by traditional fixed-time tool changes), and reduce production costs.

[0109] Overall method summary: The tool wear monitoring and prediction model based on multi-signal fusion integrates the cutting displacement signal, cutting vibration signal, and cutting energy consumption signal. Starting from signal acquisition, through multi-domain feature extraction and multi-modal feature fusion, a prediction model is constructed to achieve real-time monitoring and wear prediction, comprehensively improving the accuracy and reliability of tool wear monitoring and prediction.

[0110] 1. Signal acquisition: Use a laser displacement sensor to measure the position change between the tool cutting point and the workpiece to obtain the cutting displacement signal D(t), and its measurement principle is based on laser triangulation. A piezoelectric acceleration sensor collects the cutting vibration signal V(t), which works based on the piezoelectric effect. Install a power sensor at the input end of the machine tool spindle motor to measure the three-phase voltage and current, and calculate the instantaneous power and integrate it to obtain the cutting energy consumption signal E(t). At the same time, measure the wear value Wreal of the tool flank face to provide a real data reference for subsequent model training.

[0111] 2. Multi-domain feature extraction Processing of cutting displacement signal: Use the displacement domain index method to calculate features such as displacement mean, variance, peak value, skewness, and kurtosis, which reflect the average level, fluctuation degree, extreme change, distribution symmetry, and peak sharpness of the displacement signal. Use wavelet transform to extract wavelet coefficients at different scales, establish a correlation function between two different displacement signals of the tool and the workpiece, and analyze the signal features from different scales.

[0112] Processing of cutting vibration signal: Convert the cutting vibration signal in the time domain to the frequency domain through fast Fourier transform to obtain the frequency spectrum V(f). Extract features such as vibration energy, main frequency, and frequency band energy ratio. Vibration energy reflects the intensity of vibration, the main frequency is the frequency component with the largest energy in the vibration, and the frequency band energy ratio reflects the proportion of the energy in the frequency band near the main frequency in the total vibration energy.

[0113] Cutting energy consumption signal processing: After the power sensor collects the power signal, the instantaneous power is calculated and integrated to obtain the cutting energy consumption signal. The original power signal is subjected to moving average filtering and adaptive calibration to eliminate high-frequency noise and grid voltage fluctuation interference. Features such as energy consumption mean value, energy consumption volatility, and peak energy consumption ratio are extracted to respectively characterize the average energy consumption level, energy consumption fluctuation intensity, and energy impact under extreme working conditions.

[0114] 3. Multi-modal feature fusion Displacement signal segmentation: The displacement error signal D(t) is divided into multiple displacement interval segments according to the machining stage. Segmentation is triggered based on the displacement change threshold. The displacement change within each segment is small, representing the motion state of the tool during the stable cutting stage.

[0115] Displacement-vibration feature correlation modeling: For each displacement interval segment, a displacement-vibration feature pair is constructed. The kernel function SVM is used to establish a non-linear mapping relationship between displacement-sensitive features and vibration-sensitive features to find the potential connection between the two.

[0116] Displacement-energy consumption dynamic weighted fusion: The dynamic weights of displacement and energy consumption features are calculated. The energy consumption feature weight is automatically adjusted based on the displacement amplitude. A mapping relationship is established through weighted SVM to improve the robustness of feature fusion.

[0117] Displacement-guided Bayesian fusion: Taking the outputs of the above two mappings as inputs, a joint feature space is constructed. The posterior probability is maximized through Bayesian optimization to determine the optimal hyperparameters, and a three-dimensional non-linear mapping relationship is established to explore the deep connection between multi-modal features.

[0118] 4. Tool wear prediction model construction Spatial sequence definition and feature mapping: Using the processed cutting displacement signal as input, according to the geometric coordinates of the tool machining path, the displacement-sensitive feature set is sorted by spatial position to construct a spatial sequence input matrix. Ensure spatial continuity to provide input data with spatial structure for the model.

[0119] Spatial attention mechanism design: Introduce spatial position encoding before the GRU network, define spatial attention weights, and calculate the correlation between different positions. By means of a local attention window and dynamically adjusting the window size, the computational complexity is reduced, and the sensitivity of the model to key regions is enhanced.

[0120] Spatial-temporal hybrid GRU network: Process the spatial sequence input matrix and the time series signal in parallel. The outputs of the two branches are merged through a gated fusion module to capture both time dynamics and spatial distribution rules simultaneously, improving the prediction accuracy.

[0121] Loss function with enhanced spatial features: Based on the Huber loss, a spatial consistency constraint term is added to penalize the sudden change in the predicted values of adjacent spatial points, improve the output stability, and ensure the reliability of the model prediction results.

[0122] 5. Real-time Monitoring and Wear Prediction Real-time monitoring: The displacement, vibration, and energy consumption signals are synchronously collected through a multi-threaded architecture, and online filtering and normalization processing are performed. The signals are intercepted according to time windows, sensitive features are extracted in real time and stored in a circular buffer. The feature set is input into a pre-trained GRU-attention model to output the wear value, and the exponential weighted moving average is used to smooth the prediction results. A dynamic wear rate threshold is defined for anomaly detection and alarm. The real-time data is stored in a time series database, and the wear value curve is pushed to the HMI interface through WebSocket.

[0123] Wear prediction: Based on the exponential degradation model, the parameters are fitted by the least squares method. According to the currently monitored tool wear value and the set wear threshold, the remaining service life of the tool is predicted.

[0124] The tool wear monitoring and prediction model with multi-signal fusion has significant technical effects in many aspects, covering dimensions such as monitoring accuracy, prediction reliability, working condition adaptability, real-time performance, and stability. It effectively solves the defects of traditional methods and meets the needs of modern manufacturing.

[0125] Improve monitoring accuracy: Integrate the cutting displacement, vibration, and energy consumption signals, and use a variety of feature extraction methods to comprehensively reflect the tool wear state. For example, features such as displacement mean and variance are extracted from the cutting displacement signal to capture the displacement change trend and fluctuations; features such as vibration energy and main frequency are obtained from the cutting vibration signal to reflect the vibration characteristics. The combination of multi-signals and multi-features is more accurate than single-signal monitoring, improving the monitoring accuracy.

[0126] Enhance prediction reliability: Adopt advanced models and algorithms such as GRU networks and Bayesian optimization algorithms, which have strong capabilities in processing complex data. Through Bayesian optimization, a three-dimensional non-linear mapping relationship between multi-modal features is established to explore the deep connections between signal features, making the prediction results closer to the actual situation, reducing the risk of unexpected tool wear, and providing a reliable basis for early tool replacement and maintenance decisions.

[0127] Strong adaptability: Considering various working conditions and environmental factors, the model has good generalization ability. When segmenting the displacement signal, the displacement change threshold is dynamically adjusted according to the processed material. For example, the threshold is set to 0.1 mm for aluminum alloy processing and 0.05 mm for titanium alloy processing, enabling the model to adapt to the processing characteristics of different materials and be applicable to different processing scenarios.

[0128] Full-link low latency: The end-to-end latency of the real-time monitoring system from signal acquisition to prediction output is controlled within 100 ms. The multi-threaded architecture synchronously acquires signals with a high sampling frequency, and uses optimization technologies such as zero-copy data transmission and priority scheduling to ensure fast data processing and timely model inference, meeting the stringent real-time requirements of high-speed machining scenarios.

[0129] Adaptive dynamic threshold: In real-time monitoring, a dynamic wear rate threshold is defined based on the mean and standard deviation of historical wear rates to avoid false alarms or missed alarms caused by fixed thresholds. The wear rate must meet specific conditions for 3 consecutive cycles to trigger an alarm, ensuring the accuracy of the alarm and timely indicating the abnormal state of the tool.

[0130] Edge-cloud collaboration: Key data is synchronously uploaded to the cloud for long-term trend analysis, and a real-time inference module is retained locally to achieve complementary advantages of edge computing and cloud computing. This not only reduces the local data storage pressure but also enables the long-term value of data to be mined with the help of the powerful computing power of the cloud.

[0131] Visualization and traceability: The HMI interface displays information such as the real-time curve of wear values, alarm status, and prediction confidence, and supports playback of the wear process and export of alarm logs. This facilitates operators to grasp the tool status in real time, trace the cause of faults when problems occur, and optimize the machining process and tool management.

[0132] The above are only specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any simple changes, equivalent replacements, or modifications made based on the present invention to solve substantially the same technical problems and achieve substantially the same technical effects are all covered by the protection scope of the present invention.

Claims

1. A method for tool wear monitoring and prediction, characterized in that Including the following steps: S1. Signal acquisition: Acquire the cutting displacement signals of the tool and the workpiece during machining, denoted as D(t), and measure the displacement error data by obtaining the position change between the cutting point of the tool and the workpiece; acquire the cutting signals during machining, where the cutting signals include cutting vibration signals and cutting energy consumption signals, denoted as V(t) and E(t) respectively; measure the wear value W of the flank face of the tool real ; S2. Multi-domain feature extraction: Feature extraction is respectively performed on the collected cutting displacement signal, cutting vibration signal, and cutting energy consumption signal, and displacement-sensitive features highly correlated with displacement error are screened out to form a displacement-sensitive feature set, vibration-sensitive features highly correlated with vibration anomalies are screened out to form a vibration-sensitive feature set, and energy consumption-sensitive features highly correlated with energy consumption anomalies are screened out to form an energy consumption-sensitive feature set; S3. Multi-modal feature fusion: Based on the screened displacement-sensitive features, vibration-sensitive features, and energy consumption-sensitive features, an SVM model is used to establish a non-linear mapping relationship between the displacement-sensitive features and the vibration-sensitive features, and a non-linear mapping relationship between the displacement-sensitive features and the energy consumption-sensitive features respectively; further, through the Bayesian optimization algorithm, a three-dimensional non-linear mapping relationship between the displacement-sensitive features, vibration-sensitive features, and energy consumption-sensitive features is established; S4. Tool wear prediction model construction: Using the processed cutting displacement signal as the input and the tool wear value measured by a hand-held microscope as the label, a tool wear prediction model is established based on the GRU network, and an attention mechanism is incorporated in the cutting displacement dimension for model training; the tool wear monitoring model is used to monitor and output the real-time wear value of the tool; S5. Real-time monitoring: The real-time collected signal is input into the trained tool wear prediction model to monitor the tool wear status in real time; S6. Wear prediction: According to the monitoring results and model prediction, the remaining service life of the tool is predicted.

2. The tool wear monitoring and prediction method according to claim 1, characterized in that: In step S2, the collected cutting displacement signal is processed by the displacement domain index method, and the extracted feature set F d ={mean displacement , displacement variance , displacement peak value P d , skewness S d , kurtosis K d}, and a correlation function between two different displacement signals of the tool and the workpiece is established; ； ； ； ； ； wherein represents the value of the cutting displacement signal at the i-th sampling point, where i is the index of the sampling point, i = 1, 2, …… N , N is the total number of sampling points; Using wavelet transform Extract wavelet coefficients at different scales as features, and establish a correlation function between two different displacement signals D1(t) and D2(t) of the tool and the workpiece , ； ； where a is the scale parameter and b is the translation parameter, is the wavelet basis function; , is the mean value, is the time delay.

3. The tool wear monitoring and prediction method according to claim 1, wherein: In step S2, the collected cutting vibration signals are processed by the frequency domain index method. The spectrum V(f) is obtained through fast Fourier transform, and the feature set F is extracted. v ={vibration energy E v , main frequency f v , frequency band energy ratio R v}, where: ； where f is the frequency point, and f s is the sampling frequency, with a value range of 0 ≤ f ≤ f s / 2 f s / 2 is the Nyquist frequency, representing the highest significant frequency for signal analysis; ； where argmax f is the frequency value corresponding to the maximum spectral amplitude, and ∣V(f)∣ is the absolute value of the spectral amplitude; ； where Δf is the bandwidth centered at the main frequency f v The numerator is the total energy within the frequency band near the main frequency, and the denominator is the total vibration energy E v .

4. The tool wear monitoring and prediction method according to claim 1, wherein: In step S2, the acquisition and feature extraction of the cutting energy consumption signal are realized through a power sensor, which specifically includes the following steps: (a) Real-time acquisition of power signal: Install a power sensor at the input end of the machine tool spindle motor to measure the three-phase voltages V a (t), V b (t), V c (t) and currents I a (t), I b (t), I c (t), and calculate the instantaneous power: ， where k = a, b, c, identifiers of three-phase electricity P ( t ): Instantaneous power, representing the total input power of the machine tool spindle motor at time t And obtain the cutting energy consumption signal through integration: ; Among them, the integral variable τ is a time variable; t0: the starting time of integration; P(τ) is the instantaneous power function; t: the current time; E(t): the cumulative cutting energy consumption from t0 to t; (b) Signal preprocessing and noise reduction: A moving average filter is performed on the original power signal to filter out high-frequency noise; adaptive calibration is used to eliminate the interference of grid voltage fluctuations: ; where V nominal is the rated voltage, V actual (t) is the real-time grid voltage; (c) Feature extraction: Mean energy consumption: ; where T is the length of the analysis time window; : the average energy consumption, representing the average energy consumption level; Energy consumption volatility: ; Among them : The energy consumption variance, which reflects the fluctuation intensity of energy consumption; Peak energy consumption ratio: ; Among them, max: the maximum value function; R peak : Peak energy consumption ratio, representing the energy impact under extreme working conditions.

5. The tool wear monitoring and prediction method according to claim 1, characterized in that: The specific process of step S3 is as follows: (a)Displacement signal segmentation: The displacement error signal D(t) is divided into N displacement interval segments {S k} k-1 N ; (b) Displacement-vibration characteristic correlation modeling: For each segment S k , construct a displacement-vibration characteristic pair (F d (k) , F v (k) ), where F d (k) = {μ d (k) , σ d (k) , P d (k)} is the mean, variance, and peak value within the displacement segment; Establish a non - linear mapping Φ using the kernel function SVM dv :F d (k) →F v (k) ); (c) Displacement - energy dissipation dynamic weighted fusion: For each segment S k , calculate the dynamic weights of displacement and energy dissipation characteristics: establish a mapping through weighted SVM ; (d)Displacement-guided Bayesian fusion: Taking the outputs of Φ dv and Φ dv as inputs, construct the joint feature space Z = [Φ dv , Φ dv of displacement-vibration-energy dissipation; maximize the posterior probability through Bayesian optimization.

6. The tool wear monitoring and prediction method according to claim 1, characterized in that: In step S4, a tool wear prediction model is constructed based on the GRU network, and an attention mechanism is incorporated in the cutting displacement dimension. The displacement-sensitive feature set is constructed as an input matrix according to the spatial sequence, and the specific steps are as follows: (a) Spatial sequence definition and feature mapping: Taking the processed cutting displacement signal D(t) as the input and the tool wear value as W real Label; According to the geometric coordinates (x, y, z) of the tool processing path, the displacement-sensitive feature set F d is sorted by spatial position to construct a spatial sequence input matrix Xs ∈ R M×5 , where M is the total number of path points; The spatial sorting rule is: along the tool feed direction, arranged in ascending order of the Euclidean distance between adjacent path points to ensure spatial continuity; (b) Spatial attention mechanism design: A spatial position encoding is introduced before the GRU network, the spatial attention weight is defined, and the correlation between different positions is calculated: (c) Spatial-temporal hybrid GRU network: Input matrix X of the spatial sequence S is processed in parallel with the time series signal, and the outputs of the two branches are merged through the gated fusion module: (d) Loss function for spatial feature enhancement: Based on the Huber loss, a spatial consistency constraint term is added.

7. The tool wear monitoring and prediction method according to claim 1, characterized in that: In step S5, the specific implementation of the real-time monitoring includes the following steps: (a) Real-time signal synchronous acquisition and preprocessing: The displacement signal D(t), vibration signal V(t), and energy consumption signal E(t) are synchronously acquired through a multi-threaded architecture, and the sampling frequencies are not less than 10 kHz, 20 kHz, and 5 kHz respectively; Online filtering and standardization are performed on the original signal; (b) Streaming feature extraction and caching: Intercept the signal according to the time window T win = 0.5 s, and extract the displacement-sensitive feature F in real time d , vibration-sensitive feature F v , energy consumption-sensitive feature F e , and store them in the circular buffer; The feature extraction delay is constrained to ≤5ms and accelerated through parallel computing; (c) Model inference and state update: Input the feature set X = [F d , F v , F e into the pre-trained GRU-attention model to output the wear value ; The model inference period Tinf = 50ms. The TensorRT optimized inference engine is adopted, and the single inference time ≤10m; The exponential weighted moving average is used to smooth the prediction results; (d) Anomaly detection and alarm: Define the dynamic wear rate threshold: ; wherein , are the mean and standard deviation of the historical wear rate; If the wear rates for three consecutive cycles satisfy , a level 3 alarm is triggered; (e) Data persistence and visualization: The real-time data is stored in the time series database, supporting time range query and playback; The wear value curve is pushed to the HMI interface through WebSocket, and the refresh rate ≥30Hz.

8. The tool wear monitoring and prediction method according to claim 1, wherein: In step S6, the specific implementation of the wear prediction includes the following steps: Based on the exponential degradation model W(t)=W0*exp (-λt), the parameter λ is fitted by the least squares method to predict the remaining service life , where W th is the wear threshold.

9. The tool wear monitoring and prediction method according to claim 1, characterized in that: In step S1, the displacement error signal between the cutting point of the tool and the workpiece is collected by the laser displacement sensor, the cutting vibration signal is collected by the acceleration sensor, and the cutting energy consumption signal is collected by the current sensor.

Images