A tool life prediction method based on multi-source state perception

Abstract

The present application relates to the technical field of turning tool life prediction, and particularly relates to a turning tool life prediction method based on multi-source state perception, which takes a cutting time window as a basic calculation unit, synchronously collects numerical control working conditions, cutting force, vibration, cutting temperature, acoustic emission and spindle power data, forms multi-source state features after preprocessing, calculates a physical reference life according to cutting speed, feed rate, cutting depth and material hardness, constructs a sensing correction factor with a reference working condition health window as an anchor point, recursively calculates flank wear and its uncertainty under monotonic constraint, corrects it by using microscopic measurement or virtual observation meeting a confidence threshold, solves residual life in combination with a future working condition sequence, obtains a confidence interval through parameter, observation noise and working condition disturbance sampling, and outputs a decision according to a preset priority. The present application forms a closed loop among physical life, online state, sparse observation and future plan, and the prediction process is traceable, calibratable and convenient for engineering deployment.

Description

Technical Field

[0001] This invention relates to the field of cutting tool life prediction technology, and in particular to a cutting tool life prediction method based on multi-source state perception. Background Technology

[0002] In CNC turning, the cutting tool typically serves as the direct actuator responsible for material removal. Its cutting edge is in continuous contact with the workpiece material and bears the combined effects of mechanical loads, thermal loads, and frictional wear. As machining time increases, the tool's service life inevitably deteriorates. This degradation process is influenced by factors such as cutting speed, feed rate, depth of cut, workpiece material hardness, cooling and lubrication conditions, clamping rigidity, and batch variations. Its manifestations are not fixed; common degradation forms include flank wear, rake face crater wear, edge chipping, built-up edge, hot cracks, and coating peeling. Under certain actual working conditions, these wear or failure phenomena may occur alternately, overlap, or exhibit abrupt development at a particular machining stage. When the tool wear approaches or exceeds the allowable range, the stability of the machining system decreases. For example, workpiece dimensions may drift, surface roughness may increase, and cutting forces, vibrations, and cutting temperatures may continuously rise. Furthermore, it may lead to decreased machining accuracy, poor product consistency, reduced yield, and even workpiece scrap. If wear conditions are not identified in time, under heavy loads or abnormal cutting conditions, it may induce tool breakage, unplanned downtime, or have a cascading effect on the machine tool spindle, fixture, and workpiece. Therefore, estimating the remaining usable life of the cutting tool online during turning and providing a relatively reasonable tool change prompt before the risk of failure increases significantly is of practical significance for stabilizing machining quality, reducing tool waste, and improving equipment utilization.

[0003] Among existing methods for managing cutting tool life, the most common are timed tool changing, condition judgment based on empirical thresholds, and life estimation based on the Taylor life formula. Timed tool changing typically sets a uniform tool changing cycle based on conservative operating conditions, which is relatively simple to implement, but it is prone to premature tool replacement before the tool is fully utilized, leading to increased tool costs. Empirical threshold methods generally select one of the following features as the judgment criterion: cutting force, vibration, acoustic emission, spindle power, or temperature. However, this method is sensitive to batch variations in materials, differences in workpiece clamping, fluctuations in cutting fluid conditions, and ambient noise, resulting in insufficient stability of the judgment results. The traditional Taylor life formula can describe the empirical relationship between cutting speed and tool life, but it is more inclined towards macroscopic average life estimation and cannot reflect the real-time degradation state of the same cutting tool under varying operating conditions, nor can it incorporate multi-source online monitoring signals into the individualized correction process. In recent years, machine learning methods have been increasingly used for tool wear identification and life prediction. These methods can extract degradation features from multi-source signals such as cutting force, vibration, acoustic emission, spindle power, and temperature, and establish a mapping relationship between monitoring signals and tool state. However, in actual production settings, relying solely on data-driven models still has several limitations. On the one hand, model training typically requires a large number of full-life samples labeled with actual wear or failure times. However, tool wear often needs to be obtained through microscopic measurement after machine shutdown, resulting in high acquisition costs, poor continuity, and sparse and unevenly distributed labeled data. On the other hand, changes in machining batches, material hardness, tool grade, machine tool condition, and cutting parameters can cause feature distribution drift, making existing models prone to prediction bias under new working conditions.

[0004] For the reasons mentioned above, there is an urgent need for a tool life prediction method that can integrate empirical physical life models, online sensing characteristics, sparse wear observation data, and future machining plans to improve the accuracy, interpretability, and engineering feasibility of tool life prediction results. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the remaining life of a lathe tool based on multi-source state perception, which can improve the accuracy, interpretability, and engineering deployment feasibility of the prediction results.

[0006] To achieve the above objectives, this invention provides a method for predicting the life of a cutting tool based on multi-source state perception, comprising: During the turning process, S1 synchronously collects multi-source status data according to a unified clock. The multi-source status data includes at least CNC operating condition data, cutting force data, vibration data, and cutting temperature data, and may also include acoustic emission data and spindle power data. The CNC operating condition data includes cutting speed, feed rate, depth of cut, tool number, workpiece material number, and machining program segment number. S2 sequentially performs timestamp alignment, resampling, filtering and noise reduction, outlier removal, and sliding window sampling on multi-source state data, extracts time-domain, frequency-domain, and time-frequency-domain features of each time window, and standardizes the features using the fixed mean and standard deviation from the training phase to obtain standardized sensing features. S3 calculates the physical reference life that is strictly greater than zero under the current working conditions based on the cutting speed, feed rate, depth of cut and workpiece material hardness of the current time window; S4 concatenates standardized sensing features and normalized operating condition features into a comprehensive feature vector to construct a sensing correction factor that is strictly greater than zero; the sensing correction factor is anchored using a reference operating condition health window so that its central value is 1 under the reference state, and the actual damage rate is calculated in combination with the physical reference life. S5 recursively derives the prior wear amount and state uncertainty of the cutting tool flank face based on the actual damage rate and process increment, and limits the prior wear amount to the wear amount corrected in the previous time window and the allowable wear limit, thus obtaining the prior damage degree that satisfies the monotonically non-decreasing constraint. S6 When microscopic measurements are available, microscopic wear is preferred as the observation; when microscopic measurements are not available, virtual wear is used as the observation only when the virtual observation confidence level reaches the minimum confidence threshold; the observation noise variance is determined according to the observation source, the prior wear is corrected for gain, and the correction result is subject to monotonic interval constraints. S7 uses the current correction damage degree as the initial value, and recursively calculates the remaining tool life based on the future working condition sequence in the CNC program or production schedule until the damage degree reaches 1. When the future sensing characteristics are unknown, the statistical value or safety quantile of the historical correction factor of the same working condition is used to determine the future sensing correction factor. S8 samples model parameters, process increments, observation noise, and future operating condition disturbances to obtain multiple remaining lifetime samples, and calculates the mean and confidence interval of the remaining lifetime. Based on the current safe termination status, the lower limit of the confidence interval, the mean of the remaining lifetime, the corrected damage degree, and the process threshold, it outputs the warning level according to the priority of "immediate tool change, warning, and normal".

[0007] In step S3, the physical reference lifetime is calculated according to the following formula: ; in, The reference life is under reference operating conditions. , , These represent the cutting speed, feed rate, and depth of cut for the current time window. This refers to the material's hardness factor. For reference operating parameters, All are non-negative lifetime sensitivity indices; in, For the first The physical reference life under the current operating conditions for each cutting time window. Number the cutting time window. For the first The Brinell hardness of the workpiece material corresponding to each cutting time window; The reference working condition is a calibration working condition that is consistent with the tool grade, workpiece material and cooling conditions to be predicted. These are the cutting speed, feed per revolution, depth of cut, and Brinell hardness of the workpiece under the specified calibration conditions. The speed is 120–220 m / min. The value is 0.10–0.25 mm / r. The diameter is 0.5–2.0 mm. The range is 180–260 HB; the calibration range of the lifetime sensitivity index is... , , , .

[0008] In step S4, the sensing correction factor Constructed using an exponential function: ,in Let be the damage rate parameter, and , The composite feature vector; actual damage rate ,in, , where is the baseline damage rate; the dimension of the weight vector in the damage rate parameter is consistent with the comprehensive feature vector, the bias is a scalar, and to avoid scale coupling between the reference lifetime and the bias, the logarithmic center value of the correction factor is constrained to 0 on the health window set of the reference operating condition, making the center value of the correction factor 1; where Adjust the weight vector to reflect the damage rate. express transpose, For biased scalars, This represents the natural exponential function.

[0009] In step S5, the prior wear amount of the cutting tool flank is... Calculate using the following recursive formula: ; in, This refers to the wear on the flank face. This is the initial wear amount. To allow limits, This represents the length of the k-th window. The damage rate is represented by the superscript pri, which indicates the prior estimate before observation correction, and the superscript cor, which indicates the corrected estimate after observation correction. This is the process increment, used to characterize unmodeled wear fluctuations; the deterministic point estimate takes the expected value of this increment, and during random sampling, the recursive result is projected between the previous window's corrected wear amount and the allowable wear limit to ensure that the wear amount does not reverse; the corresponding state uncertainty recursively is: ; in, Indicates the first The variance of wear state estimation for each cutting time window; its superscripts pri and cor represent prior estimate and corrected estimate, respectively; The process incremental variance; the prior damage degree .

[0010] Specifically, step S6 includes obtaining the actual wear amount using a microscope during a stop-and-spot inspection. ; without shutting down the system, utilize the virtual observation model trained from sensor features Wear observations are given, and both are collectively referred to as observations. : ; in Indicates whether microscopic measurement exists. Indicates the confidence level of virtual observations. The lowest confidence threshold, This indicates that there are no reliable observations in the current window. This represents the observation noise variance; microscopic observations use the measurement variance obtained from calibration, while the variance of virtual observations increases as the confidence level decreases; among which, This is the parameter set for the virtual observation model; when reliable observations exist, a one-dimensional gain correction method is used to fuse prior wear and observed wear. ; ; ; in, For the first The observation correction gain for each cutting time window; after gain correction, the corrected wear amount is limited to the corrected wear amount of the previous time window and the allowable wear limit. This limit is used to suppress non-physical wear regression caused by observation noise; when there are no reliable observations in the current window, let , , The corrected damage and health levels are: .

[0011] Specifically, step S7 includes: To obtain the corrected damage level at the current moment. Then, read the sequence of future cutting conditions from the CNC program or production schedule. ,make For the future Each window calculates the corresponding baseline lifetime, sensing correction factor, and damage rate. And perform recursion: When there are no real-time sensing features for the future time window, the correction factor is retrieved according to the historical window that is closest to the future operating conditions, and the mean, median or upper percentile is selected according to risk preference. ; in, For the first A vector of planned cutting conditions for a future time window, where the superscript p indicates the planned or predicted quantity for the future. For the current number The future time window sequence number starting from the current time window. To advance towards the future Predicted damage level after a time window For the first The duration of the planned future time window. The corresponding predicted damage rate; Define the future window number for the first time reaching the end of life as: ; If the end of life occurs in the first Within a future window, linear interpolation is performed on that window to obtain the remaining tool life: .

[0012] In the formula, To avoid division by zero caused by zero damage rate, a very small positive number is set. When the damage level of 1 is not reached within the given prediction range, the future operating condition sequence is continued to be extended; if it cannot be extended, the remaining lifetime is output as a right censored lower bound greater than the current prediction range.

[0013] In step S8, the lower limit of the confidence interval is set to L. k The early warning time threshold is The forced tool change time threshold is The damage warning threshold is η1, where And 0 < η1 < 1; the warning judgment is executed according to the following priority: current safety termination criterion trigger, Mean remaining lifespan ≤ Or, if the damage correction is ≥1, output "immediate tool change"; if the immediate tool change condition is not met and Mean remaining lifespan < If the damage level is ≥η1, an early warning will be output; otherwise, a normal output will be output.

[0014] This invention presents a multi-source state-sensing-based tool life prediction method that combines the generalized Taylor physical life model with multi-source sensing feature correction factors. This method can calculate the equivalent life consumption rate for each time window even when cutting speed, feed rate, depth of cut, and material hardness change, overcoming the problem that traditional fixed life or a single Taylor formula cannot reflect real-time conditions. This invention derives a discrete window life consumption model from damage integrals, establishing a clear mathematical relationship between the baseline life, damage rate, flank wear, and remaining life. Under constant operating conditions, the model can degenerate into a traditional life formula, while under varying operating conditions, it can perform piecewise cumulative calculations. This invention uses an exponential sensing correction factor to ensure the damage rate is strictly positive and uses damage degree... and the amount of wear on the back face The recursive state prediction satisfies the physical laws of irreversible wear and monotonically increasing tool life consumption, improving the interpretability of the prediction results. This invention establishes a unified correction mechanism for microscopic and virtual observations, enabling real-time signal correction of wear states even under sparse measured labels, reducing the dependence of pure data-driven models on a large number of full-lifetime labeled samples. This invention recursively predicts the tool life to the end of the tool life based on future planned working conditions and calculates the remaining life (RUL) through intersection interpolation, outputting the remaining life in minutes, seconds, or remaining workable parts; it also provides confidence intervals and warning levels, providing a quantitative basis for tool change decisions. This invention can be deployed at the edge of the workshop, outputting health status, remaining life, and warning levels, reducing tool waste caused by premature tool changes, lowering the risk of chipping and quality abnormalities, and improving the stability and automation level of CNC turning. This invention improves the accuracy, interpretability, and engineering deployment feasibility of turning tool remaining life prediction results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0016] Figure 1 This is a flowchart of the tool life prediction method based on multi-source state perception of the present invention.

[0017] Figure 2 This is an architecture diagram of the cutting tool life prediction system based on multi-source state perception of the present invention.

[0018] Figure 3 This is a schematic diagram of the tool life prediction system based on multi-source state perception of the present invention.

[0019] 101-Multi-source signal acquisition module, 102-Signal preprocessing module, 103-Feature construction module, 104-Physical benchmark lifetime calculation module, 105-Degradation state recursion module, 106-Observation correction module, 107-Remaining lifetime solution module, 108-Early warning output module. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0021] Please see Figures 1-3 This invention provides a method for predicting the life of a cutting tool based on multi-source state perception, comprising: During the turning process, S1 synchronously collects multi-source status data according to a unified clock. The multi-source status data includes at least CNC operating condition data, cutting force data, vibration data, and cutting temperature data, and may also include acoustic emission data and spindle power data. The CNC operating condition data includes cutting speed, feed rate, depth of cut, tool number, workpiece material number, and machining program segment number. S2 sequentially performs timestamp alignment, resampling, filtering and noise reduction, outlier removal, and sliding window sampling on multi-source state data, extracts time-domain, frequency-domain, and time-frequency-domain features of each time window, and standardizes the features using the fixed mean and standard deviation from the training phase to obtain standardized sensing features. S3 calculates the physical reference life that is strictly greater than zero under the current working conditions based on the cutting speed, feed rate, depth of cut and workpiece material hardness of the current time window; In this step, the physical reference lifetime is calculated according to the following formula: in, The reference life is under reference operating conditions. , , These represent the cutting speed, feed rate, and depth of cut for the current time window. This refers to the material's hardness factor. For reference operating parameters, All are non-negative lifetime sensitivity indices.

[0022] in, For the first The physical reference life under the current operating conditions for each cutting time window. Number the cutting time window. For the first The Brinell hardness of the workpiece material corresponding to each cutting time window;

[0023] The reference working condition is a calibration working condition that is consistent with the tool grade, workpiece material and cooling conditions to be predicted. These are the cutting speed, feed per revolution, depth of cut, and Brinell hardness of the workpiece under the specified calibration conditions. Preferably, The speed is 120–220 m / min. The value is 0.10–0.25 mm / r. The diameter is 0.5–2.0 mm. The range is 180–260 HB; in one embodiment, the values ​​are 160 m / min, 0.15 mm / r, 1.0 mm, and 220 HB. The calibration range of the lifetime sensitivity index is... , , , The specific value is determined through full-life test data and validation set constraint optimization. In the formula, all operating parameters adopt the same unit system and are normalized with positive reference values; the material hardness factor is greater than 0, and the reference life is greater than 0, thereby ensuring that the physical reference life is always greater than 0.

[0024] S4 concatenates standardized sensing features and normalized operating condition features into a comprehensive feature vector to construct a sensing correction factor that is strictly greater than zero; the sensing correction factor is anchored using a reference operating condition health window so that its central value is 1 under the reference state, and the actual damage rate is calculated in combination with the physical reference life. In this step, the sensing correction factor Constructed using an exponential function: ,in Let be the damage rate parameter, and , The composite feature vector; actual damage rate ,in, , where is the baseline damage rate; the dimension of the weight vector in the damage rate parameter is consistent with the comprehensive feature vector, the bias is a scalar, and to avoid scale coupling between the reference lifetime and the bias, the logarithmic center value of the correction factor is constrained to 0 on the health window set of the reference operating condition, making the center value of the correction factor 1; where Adjust the weight vector to reflect the damage rate. express transpose, For biased scalars, This represents the natural exponential function. To ensure parameter identifiability, the bias is centered on the reference operating condition health window, making the logarithmic mean of the sensing correction factor within this window set zero. Therefore, the physical reference lifetime is responsible for describing the main effects of the operating condition, while the sensing correction factor only describes the residual effect of the online state relative to the reference state.

[0025] S5 recursively derives the prior wear amount and state uncertainty of the cutting tool flank face based on the actual damage rate and process increment, and limits the prior wear amount to the wear amount corrected in the previous time window and the allowable wear limit, thus obtaining the prior damage degree that satisfies the monotonically non-decreasing constraint. In this step, the prior wear amount of the cutting tool's flank face is... Calculate using the following recursive formula: ; in, This refers to the wear on the flank face. This is the initial wear amount. To allow limits, Indicates the first Window length, The damage rate is represented by the superscript pri, which indicates the prior estimate before observation correction, and the superscript cor, which indicates the corrected estimate after observation correction. The process increment is used to characterize unmodeled wear fluctuations; the deterministic point estimate takes its expected value, and the sampling result participates in uncertainty propagation after being projected onto the monotonic interval (the deterministic point estimate takes the expected value of this increment, and during random sampling, the recursive result is projected onto the area between the previous window's corrected wear amount and the allowable wear limit to ensure that the wear amount does not reverse); the corresponding state uncertainty is recursively calculated as follows: ; in, Indicates the first The variance of wear condition estimation for each cutting time window; its superscripts pri and cor represent prior estimate and corrected estimate, respectively; The process incremental variance; the prior damage degree .

[0026] S6 When microscopic measurements are available, microscopic wear is preferred as the observation; when microscopic measurements are not available, virtual wear is used as the observation only when the virtual observation confidence level reaches the minimum confidence threshold; the observation noise variance is determined according to the observation source, the prior wear is corrected for gain, and the correction result is subject to monotonic interval constraints. Specifically, it includes: The actual wear amount was obtained using a microscope during random inspections after machine shutdown. ; without shutting down the system, utilize the virtual observation model trained from sensor features Wear observations are given, and both are collectively referred to as observations. : The virtual observation model employs a four-layer feedforward fully connected network with an input dimension of d=Q+4, corresponding to Q-dimensional standardized sensing features and 4-dimensional normalized working condition features. The first hidden layer contains 64 neurons, using ReLU activation, batch normalization, and a Dropout of 0.2. The second hidden layer contains 32 neurons, using ReLU activation, batch normalization, and a Dropout of 0.1. The output layer contains one neuron, using Softplus activation to output non-negative flank wear. The model uses microscopic wear measurement as the supervision label and is trained using Huber loss and Adam optimizer. The virtual observation confidence is jointly calibrated by the validation set residuals, Monte Carlo Dropout variance, and Mahalanobis distance from the sample to the training distribution.

[0027]

[0028] in Indicates whether microscopic measurement exists. Indicates the confidence level of virtual observations. The lowest confidence threshold, This indicates that there are no reliable observations in the current window. This represents the observation noise variance; the variance of microscopic observations is obtained through calibration, and the variance of microscopic measurements is determined by instrument calibration and repeated measurements. The variance of virtual observations increases as the confidence level decreases; among which... This is the parameter set for the virtual observation model; when reliable observations exist, a one-dimensional gain correction method is used to fuse prior wear and observed wear.

[0029] in, For the first The observation correction gain for each cutting time window is calculated; after gain correction, the corrected wear amount is limited to the corrected wear amount of the previous time window and the allowable wear limit. This limit is used to suppress non-physical wear regression caused by observation noise. Then, the corrected damage degree and health degree are calculated. When there are no reliable observations in the current window, let... , , The corrected damage and health levels are: .

[0030] S7 uses the current correction damage degree as the initial value, and recursively calculates the remaining tool life based on the future working condition sequence in the CNC program or production schedule until the damage degree reaches 1. When the future sensing characteristics are unknown, the statistical value or safety quantile of the historical correction factor of the same working condition is used to determine the future sensing correction factor. Specifically, it includes: To obtain the corrected damage level at the current moment. Then, read the sequence of future cutting conditions from the CNC program or production schedule. ,make For the future Each window calculates the corresponding baseline lifetime, sensing correction factor, and damage rate. And perform recursion: When there are no real-time sensing features for the future time window, the correction factor is retrieved according to the historical window closest to the future working condition, and the mean, median, or upper quantile is selected based on risk preference; the safety mode prioritizes the upper quantile to obtain a conservative remaining life. The historical window is limited to a valid window where the tool brand, workpiece material, and cooling conditions are consistent with the future working condition. The cutting speed, feed rate, depth of cut, and material hardness are normalized according to the standard deviation fixed in the training set, and the weighted Euclidean distance of the four normalized differences is calculated. All four weights are non-negative and their sum is 1. The windows are sorted from smallest to largest distance, and the top K windows with distances not exceeding a preset similarity threshold δ are selected as similar working conditions. K, δ, and the weights are determined through the validation set. If there are fewer than K candidate windows, all windows that meet the threshold are used. If no window meets the similarity threshold, it is marked. For extrapolation of operating conditions, a safety quantile is used with the same tool brand and workpiece material historical correction factor. Risk preferences are divided into normal mode, robust mode and safe mode, respectively taking the mean, median and q upper quantile of similar operating condition correction factors, where q is 0.75 to 0.95 and preset according to the validation set or process safety requirements. The safe mode is used for high-value workpieces, difficult-to-cut materials, unattended machining or scenarios where the control system triggers conservative prediction flags. Since the larger the sensing correction factor, the higher the damage rate, using the q upper quantile can obtain a shorter and more conservative remaining life.

[0031] , in, For the first A vector of planned cutting conditions for a future time window, where the superscript p indicates the planned or predicted quantity for the future. For the current number The future time window sequence number starting from the current time window. To advance towards the future Predicted damage level after a time window For the first The duration of the planned future time window. The corresponding predicted damage rate; Define the future window number for the first time reaching the end of life as: ; If the end of life occurs in the first Within a future window, linear interpolation is performed on that window to obtain the remaining tool life: .

[0032] In the formula, A very small positive number is set to avoid division by zero caused by zero damage rate. The very small positive number in the denominator is used to avoid division by zero caused by zero damage rate. If the end of the lifetime does not occur within the prediction range, the future operating condition sequence is extended. If it cannot be extended, the cumulative duration of the current prediction range is used as the lower bound of the remaining lifetime. Specifically, when the damage degree of 1 is not reached within the given prediction range, the future operating condition sequence is extended. If it cannot be extended, the right censored lower bound of "remaining lifetime is greater than the current prediction range" is output.

[0033] S8 samples model parameters, process increments, observation noise, and future operating condition disturbances to obtain multiple remaining lifetime samples, and calculates the mean and confidence interval of the remaining lifetime. Based on the current safe termination status, the lower limit of the confidence interval, the mean of the remaining lifetime, the corrected damage degree, and the process threshold, it outputs the warning level according to the priority of "immediate tool change, warning, and normal".

[0034] In this step, the process safety threshold includes an early warning time threshold. Forced tool change time threshold The damage warning threshold η1 is defined; the warning branches are determined in the order of immediate tool change, warning, and normal operation to eliminate branch ambiguity when multiple conditions are met simultaneously. Specifically, the lower limit of the confidence interval is denoted as L. k The early warning time threshold is The forced tool change time threshold is The damage warning threshold is η1, where And 0 < η1 < 1; the warning judgment is executed according to the following priority: current safety termination criterion trigger, Mean remaining lifespan ≤ Or, if the damage correction is ≥1, output "immediate tool change"; if the immediate tool change condition is not met and Mean remaining lifespan < If the damage level is ≥η1, an early warning will be output; otherwise, a normal output will be output.

[0035] To better understand the multi-source state perception-based tool life prediction method of the present invention, the principle and specific implementation process of the method are described in detail below.

[0036] First, define the symbols and tool life, using the first... Each cutting time window is used as the basic calculation unit, and the end time of the window is... Window length is Within the same window, the cutting speed, feed rate, depth of cut, and material hardness remain approximately constant, denoted as follows: , , and The amount of wear on the back face is recorded as... The initial wear amount is recorded as The limit can be denoted as This invention uses the point at which the wear on the cutting edge reaches its allowable limit as the physical endpoint for the recursive wear life; when the process simultaneously limits surface roughness... Dimensional deviation or main cutting force At that time, set an independent online security termination criterion: (1) Formula (1) is the safety coverage rule: when any quality or load index reaches the upper limit, the machine is immediately stopped and the remaining life is set to 0; when the safety coverage rule is not triggered, the remaining life is still decremented to 1 according to the wear damage degree of the flank face. Thus, the meanings of safe shutdown and wear life prediction are distinguished from each other.

[0037] in, Indicates the first The lifespan of the step is marked as terminated. This indicates that the tool has reached the end of its lifespan or the tool needs to be replaced. This indicates that the person has not reached the end of their lifespan. Indicates the first Wear on the back face of the lathe tool. This indicates the maximum permissible flank wear threshold. Indicates the first Step-by-step surface roughness processing. This indicates the upper limit of permissible surface roughness. Indicates the first The sign of the step processing dimension relative to the target dimension is taken by formula (1) and the absolute value is used; this sign is distinguished from the lowercase e of the state prediction error in the following text. Indicates the first The main cutting force of the step. This represents the allowable threshold for the main cutting force. To facilitate online calculations, the tool wear condition is normalized to a damage degree. and define health : (2) In formula (2) This indicates that the damage is limited to the closed interval [0,1]. When When indicates a new or nearly new knife; This indicates that the set lifespan has been reached; The closer it is to 0, the lower the remaining available capacity.

[0038] Table 1

[0039]

[0040]

[0041] Table 1 is a table defining the symbols of the main parameters of this method. The specific process and principle of steps S1 to S8 of the multi-source state perception-based tool life prediction method of this invention are as follows: During the turning process, S1 synchronously collects multi-source status data according to a unified clock. The multi-source status data includes at least CNC operating condition data, cutting force data, vibration data, and cutting temperature data, and may also include acoustic emission data and spindle power data. The CNC operating condition data includes cutting speed, feed rate, depth of cut, tool number, workpiece material number, and machining program segment number. Step S2: Data Windowing and Feature Standardization: Divide the continuous cutting process into segments of length [missing information]. Time window Within each time window, the cutting force signal... Vibration signals Acoustic emission signals Spindle power and cutting temperature Features such as mean, root mean square value, peak-to-peak value, kurtosis, skewness, spectral energy, dominant frequency amplitude, wavelet packet energy, and temperature rise slope are extracted to form the original feature vector: Each channel is aligned via the CNC system clock or hardware trigger signal. High-frequency channels undergo anti-aliasing filtering and are downsampled to the feature calculation frequency, while low-frequency channels are mapped to the same time window using hold-or interpolation. Abnormal samples are determined jointly by physical range, mutation rate, and consecutive missing length; windows exceeding the allowable missing proportion are marked as invalid and do not proceed to state recursion. When acoustic emission or spindle power channels are not configured, the feature extraction operator deletes the corresponding components and recalibrates the model according to the actual input dimension, preventing misjudgment of unconfigured channels as missing data.

[0042] (3) To eliminate the influence of dimensions and sampling scale, the first [unit / scale] is saved during the training phase. The mean of each feature and standard deviation In the online prediction phase, the same parameters are used for standardization: (4) in For feature extraction operators, For the number of features, To prevent extremely small positive numbers with a denominator of zero.

[0043] Step S3: Derivation of the Generalized Taylor Reference Life Model: The traditional Taylor formula describes the empirical relationship between cutting speed and tool life under fixed feed rate and depth of cut conditions. To adapt to varying working conditions in the workshop, this invention incorporates cutting speed, feed rate, depth of cut, and material hardness into the life model simultaneously. Let the reference working condition be... The reference lifespan is The material hardness factor is In logarithmic coordinates, assuming the sensitivity of the reference life to each operating condition factor is a linear superposition, then: (5) in All are non-negative lifetime sensitivity indices. Taking the indices on both sides of formula (5), we obtain the physical reference lifetime under the current operating conditions: (6) Formula (6) can be converted to and from the generalized Taylor relation. Let , , , Then we get: (7) From formula (7), we can obtain: (8) Therefore, both formula (6) and formula (8) can be used as equivalent expressions for physical reference lifetime. Formula (6) is convenient for normalization interpretation of different operating conditions, while formula (8) is convenient for continuing the traditional Taylor parameter calibration method.

[0044] Step S4: Equivalent life consumption and damage derivation: Under constant current operating conditions and without abnormal degradation, if the tool life is... Then the baseline damage rate is: (9) The life loss of a cutting tool under varying working conditions can be expressed as the integral of the damage rate over time: (10) The physical meaning of formula (10) is: when the cumulative damage reaches 1, the cutting tool reaches the end of its lifespan. If the operating conditions remain constant throughout the entire lifespan... Then we have: (11) Therefore, the baseline damage rate It strictly matches the physical reference life under constant operating conditions. Numerical integration over the discrete window yields the... The baseline damage increment for each window: (12) Derivation of Sensing Correction Factor and Positive Damage Rate: Due to factors such as individual tool differences, clamping stiffness, cooling status, edge chipping, and local hardness fluctuations in the workpiece material, the actual wear rate under the same working condition can vary. This invention introduces a sensing correction factor to adjust the baseline damage rate. First, the standardized sensing features and the normalized features of the working condition are concatenated into a comprehensive feature vector: (13) Let the effect of sensing characteristics on the residual degradation rate be... To ensure the damage rate remains positive and to meet the physical requirement of irreversible lifetime depletion, an exponential function is used to construct a correction factor: (14) This yields the actual damage rate after sensor correction: (15) To eliminate reference lifetime With bias Due to the indistinguishability of scales between them, a set of windows in the healthy phase under the reference operating condition is selected. Let the set contain The average value is 0; equivalently, this average value can be used to reset the training time after each round of training. This anchoring ensures that the sensing correction factor of the reference window is centered at 1, preventing the main effects of the operating conditions from being repeatedly absorbed into the data correction term.

[0045] If degradation-sensitive characteristics such as vibration spectrum energy, acoustic emission energy, cutting temperature, root mean square value of cutting force, or power fluctuation increase abnormally, then Increase, and thus make and An increase indicates a faster depletion of the sensor's lifespan; when the sensor state is stable... Approaching 1, the life consumption is mainly determined by the physical reference life. Converting the normalized damage rate to the flank wear rate, we get: (16) Step S5: Construction of the wear state recursive model: with Assuming the tool's degenerate state, a one-dimensional state recursive model is established. Let the superscript p... ri Let represent the prior estimate before observation correction, and the superscript cor represent the corrected estimate after observation correction. Then the ... The prior wear amount for each window is: (17) in The process increment is used to characterize unmodeled wear fluctuations; the deterministic point estimate takes its expected value of 0, while the uncertainty sampling generates the original disturbance and performs a monotonic interval projection on the state result. Let Qk be its variance. The corresponding state uncertainty is recursively given as: (18) The prior wear amount can be converted into the prior damage degree using formula (2): (19) The deterministic point estimate of formula (17) takes the expected value of the process increment as 0, and performs monotonic interval projection on the recursive result. Since formula (15) guarantees The lower limit of the projection is taken from the previous window to correct the wear amount, thus ensuring... This ensures that the degree of damage satisfies the physical law of monotonicity.

[0046] Step S6: Unified calibration of microscopic observation and virtual observation: In actual production, the actual wear amount can be obtained through a microscope during downtime spot checks. Without shutting down the system, a virtual observation model trained from sensor features can be used. Wear observations are provided. Both are collectively referred to as observations. : The virtual observation model employs a four-layer feedforward fully connected network with an input dimension of d=Q+4, corresponding to Q-dimensional standardized sensing features and 4-dimensional normalized working condition features. The first hidden layer contains 64 neurons, using ReLU activation, batch normalization, and a Dropout of 0.2. The second hidden layer contains 32 neurons, using ReLU activation, batch normalization, and a Dropout of 0.1. The output layer contains one neuron, using Softplus activation to output non-negative flank wear. The model uses microscopic wear measurement as the supervision label and is trained using Huber loss and Adam optimizer. The virtual observation confidence is jointly calibrated by the validation set residuals, Monte Carlo Dropout variance, and Mahalanobis distance from the sample to the training distribution.

[0047] (20) in Indicates whether microscopic measurement exists. Indicates the confidence level of virtual observations. The lowest confidence threshold, This indicates that there are no reliable observations in the current window. Observation noise variance. The value is determined based on the observation method: a smaller value is taken for microscopic measurements, and an increase is taken as the confidence level decreases for virtual observations. For example: The virtual observation confidence is calculated by formula (44) from the validation error, Monte Carlo Dropout prediction variance and Mahalanobis distance from the sample to the training distribution.

[0048] (twenty one) Among them, the fundamental variance of microscopic measurement and the fundamental variance of virtual observation are both greater than 0. The smallest positive number in the denominator is used to avoid division by zero when the confidence level is 0. Therefore, the lower the confidence level of virtual observation, the larger its observation variance, and the smaller its impact on the correction results.

[0049] When reliable observations exist, a one-dimensional gain correction method is used to fuse prior wear and observed wear: (twenty two) (twenty three) (twenty four) When there are no reliable observations in the current window, let , , The corrected damage and health levels are: Regardless of whether microscopic observation or virtual observation is used, the wear amount after gain correction is further limited to the wear amount corrected in the previous time window and the wear limit; this conservative projection is used to suppress wear regression caused by measurement noise.

[0050] (25) Step S7: Operating condition recursion and remaining life calculation: Obtain the corrected damage degree at the current moment. Then, read the sequence of future cutting conditions from the CNC program or production schedule. .make For the future Each window calculates the corresponding baseline lifetime, sensing correction factor, and damage rate. And perform recursion: The sensing correction factor corresponding to the future working condition is determined in the following order: if online predictable features exist, predictive features are used; otherwise, the statistics of the historical correction factor for similar working conditions are used; in the safety mode, its upper quantile is taken. Similar working conditions are retrieved based on the normalized distance of cutting speed, feed rate, depth of cut, and material hardness. The historical window is limited to a valid window whose tool brand, workpiece material, and cooling conditions are consistent with the future working condition. The cutting speed, feed rate, depth of cut, and material hardness are normalized according to the fixed standard deviation of the training set, and the weighted Euclidean distance of the four normalized differences is calculated. All four weights are non-negative and their sum is 1. The windows are sorted from smallest to largest distance, and the top K windows with distances not exceeding the preset similarity threshold δ are selected as similar working conditions. K, δ, and the weights are determined through the validation set. If there are fewer than K candidate windows, all windows that meet the threshold are used; if no window meets the similarity threshold, it is marked. For extrapolation of operating conditions, a safety quantile is used with the same tool brand and workpiece material historical correction factor. Risk preferences are divided into normal mode, robust mode and safe mode, respectively taking the mean, median and q upper quantile of similar operating condition correction factors, where q is 0.75 to 0.95 and preset according to the validation set or process safety requirements. The safe mode is used for high-value workpieces, difficult-to-cut materials, unattended machining or scenarios where the control system triggers conservative prediction flags. Since the larger the sensing correction factor, the higher the damage rate, using the q upper quantile can obtain a shorter and more conservative remaining life.

[0051] (26) Define the future window number for the first time reaching the end of life as: (27) If the end of life occurs in the first Within a future window, linear interpolation is performed on that window to obtain the remaining lifetime: (28) In the formula, the smallest positive number is used to avoid division by zero. If the end of the lifespan is not found within the maximum prediction window, the subsequent CNC program or production schedule is read. If the subsequent plan still cannot be obtained, the covered future processing time is used as the right censoring lower bound of the RUL and marked as "insufficient prediction range", and no false finite lifespan point estimates are output.

[0052] When the operating conditions are approximately constant in the near future, a closed-form approximation can be used: (29) The remaining tool life is the number of remaining machinable parts. Let the actual cutting time per part be... Then we have: (30) Step S8: Calculation of Confidence Interval and Warning Level: To output reliable tool change suggestions, this invention performs calculations on model parameters, measurement noise, and future operating condition fluctuations. After random sampling, we obtain One remaining lifetime sample: (31) in The recursive solution operators for formulas (26) to (28) are given. The mean remaining lifetime and the two-sided confidence interval are: (32) (33) in and These represent the quantile operators. The warning level is output based on the mean remaining lifetime, lower confidence interval, damage level, and process safety factor. : (34) To ensure that the branches of formula (34) are mutually exclusive and the result is unique, the actual execution follows the order of Replace, Warning, and Normal. Once a high-priority condition is met, the low-priority branch is no longer entered. The lower limit of the confidence interval and the mean of RUL are used in the judgment simultaneously. If either indicator enters a higher risk interval, a higher risk level is adopted.

[0053] Among them, each threshold satisfies and ; To provide early warning time thresholds, To enforce the tool change time threshold, This is the damage level warning threshold. The threshold should meet the following requirements. When RUL is timed in minutes, The preferred time is 10–30 min. The preferred time is 2 to 10 minutes. The preferred value is 0.70 to 0.90.

[0054] The model training and parameter optimization process is explained below.

[0055] (1) Parameter set and positive constraints: model parameters Including generalized Taylor parameters Damage rate parameter Virtual observation model parameters Process incremental variance and observation noise variance To ensure a positive lifetime sensitivity index, this invention employs softplus parameterization: (35) in For unconstrained learnable parameters, It is a very small positive number. This parameterization avoids the non-physical result that the lifespan increases with the increase of cutting speed during training.

[0056] (2) Loss function and model training: The training data consists of historical cutting processes, and each tool life trajectory is denoted as... This includes operating condition sequences, multi-source sensing characteristics, partial microscopic wear measurements, virtual observation tags, and the final failure time. For the first... The trajectory, the first The data for each window is: (36) To ensure physical consistency in the prediction process, this invention constructs a total loss function that includes observation error, lifetime end error, state recursion error, monotonicity constraints, and parameter regularization terms. The observation error term is used to constrain the model's output wear amount to be consistent with microscopic measurements or high-confidence virtual observations. (37) in For the first There exists a set of reliable observation windows within each trajectory. For observation weights, for trajectories with complete failure records, the lifetime end error term constrains the cumulative damage of the entire trajectory to be close to 1 at the time of failure. (38) The state recursion error term is used to constrain the model's predicted wear amount to be consistent with the damage rate integral: (39) Monotonicity constraints are used to penalize non-physical phenomena such as damage decreasing over time. (40) Taking all the above constraints into account, the total loss function is: (41) in The weighting coefficients can be determined by the validation set or process safety requirements. Parameter optimization can employ gradient descent, Adam, or constrained least squares methods. The general update formula is: (42) (3) Online update of model parameters: When running online, when new microscopic measurements are obtained or a tool change event occurs, the prediction error can be updated. Used for online updates, allowing the model to gradually adapt to the actual degradation patterns of the machine tool, material batch, and tool grade: (43) in, For online learning rate, The observation confidence level; when the observation comes from artificial microscopy. Higher, when observations come from a virtual model and the signal-to-noise ratio is low. The values ​​are relatively low. To prevent online learning from corrupting the general model, an upper limit can be set on the magnitude of parameter updates, or only the bias term can be updated. Security policies.

[0057] After each online update, the parameters are projected back to the preset feasible domain: the lifetime sensitivity index, reference lifetime, and noise variance remain positive, and the single relative change does not exceed the safety limit; when the observation is unreliable, the error is abnormal, or the parameter projection fails, the update is canceled and the previous version is retained.

[0058] The online learning rate is determined using a method that decays and is limited with the number of trusted updates: ,in , , , , The number of trusted updates performed; in one embodiment , Observation confidence Normalization to [0,1]: Determined during microscopy based on instrument calibration error and the standard deviation of repeated measurements, preferably 0.95~1.00; the formula for virtual observation is calculated as follows: (44) in, The root mean square error is normalized to the validation set truncated to [0,1]. The predicted variance of the current sample obtained by Monte Carlo Dropout; To measure the variance of the predictions on the validation set, the 95th percentile is preferred. The Mahalanobis distance of the current sample relative to the training distribution; As a reference distance scale, the 95th percentile of the Mahalanobis distance in the training set is preferred.

[0059] Implementation Case: A Set of Implementable Tool Life Prediction Configurations This implementation case uses a CNC lathe, a CNMG120408 carbide cutting tool, and 50 mm diameter 45 steel bar. The workpiece hardness is 180–260 HB, the cutting speed is 120–220 m / min, the feed rate is 0.10–0.25 mm / r, and the depth of cut is 0.5–2.0 mm. The reference working conditions are V0 = 160 m / min, F0 = 0.15 mm / r, and a... p,0 =1.0 mm, H0=220 HB; the reference working condition full life calibration result of 5 cutting tools in the same batch is taken as T ref =45 min, initial wear amount VB0=0.02 mm, flank wear limit VB lim =0.30 mm.

[0060] Training data was collected using 12 combinations of cutting parameters, with 5 cutting tools per combination, resulting in 60 complete life trajectories. The sampling frequency for cutting force and vibration signals was 10 kHz, for temperature signals it was 10 Hz, and for CNC operating condition data it was 20 Hz. The data was segmented into 5-second windows with 50% overlap, yielding approximately 65,000 window samples. Microscopic wear measurements were performed every 3 minutes. The data was further divided into 42 training trajectories, 9 verification trajectories, and 9 test trajectories based on the tool path to prevent data leakage caused by adjacent windows of the same tool crossing sets.

[0061] 30-dimensional degradation features (Q=30) are extracted from each window, and the input dimension of the virtual observation model is d=34. The network structure consists of a 34-dimensional input layer, a 64-neuron hidden layer, a 32-neuron hidden layer, and a 1-neuron Softplus output layer. Huber loss, Adam optimizer, and an initial training learning rate of 1×10⁻⁶ are used. -3 The batch size is 128 and the maximum number of training epochs is 300. Training is stopped early if the validation set loss does not decrease for 20 consecutive epochs. A set of calibration values ​​for the Generalized Taylor Lifetime Sensitivity Index is provided. , , , .

[0062] In the online phase, the standard deviation of microscopic measurement noise is 0.005 mm, the standard deviation of virtual observation baseline noise is 0.03 mm, and the lowest confidence threshold is used. Initial online learning rate The relative change in parameters per instance is limited to within 1%. Uncertainty calculations use M=500 and α=0.05; the warning threshold is set to... , , The above values ​​are used to illustrate specific embodiments of the present invention and can be re-determined based on the tool grade, workpiece material, machine tool rigidity, and quality requirements using calibration data.

[0063] The lifetime prediction calculation process of this invention is as follows: Figure 2 As shown, the process includes eight main steps: data windowing, feature standardization, baseline life calculation, damage rate calculation, wear state recursion, observation correction, future working condition recursion, and early warning output. The pseudocode for the overall tool life prediction calculation process is shown in Table 2. Table 2

[0064]

[0065]

[0066]

[0067] This invention presents a method for predicting the life of lathe tools based on multi-source state perception. The method constructs a tool life prediction calculation process, which includes steps such as life end point definition, baseline life conversion, equivalent damage accumulation, damage rate correction based on sensor features, wear state recursion, observation data fusion correction, life extrapolation under future working conditions, and early warning result output. This process transforms the traditional tool life management method, which relies on fixed cycles, single empirical rules, or isolated thresholds, into an online state estimation method that is continuously updated with the machining process. Therefore, this invention, while retaining empirical physical laws such as the Taylor life model, incorporates multi-source monitoring information from the field and limits the prediction results from changes that do not conform to the wear mechanism through a physical constraint degradation model. This makes the remaining tool life, health status, and tool replacement recommendations more engineering interpretable and valuable. This invention combines the generalized Taylor physical life model with multi-source sensor feature correction factors, enabling the calculation of the equivalent life consumption rate for each time window even when cutting speed, feed rate, depth of cut, and material hardness change, overcoming the problem that traditional fixed life or single Taylor formulas cannot reflect real-time conditions. This invention derives a discrete window life consumption model based on damage integrals, establishing a clear mathematical relationship between the baseline life, damage rate, flank wear, and remaining life. Under constant operating conditions, the model degenerates into a traditional life formula, while under varying operating conditions, it allows for piecewise cumulative calculations. This invention employs an exponential sensing correction factor to ensure the damage rate is strictly positive, and uses damage degree... and the amount of wear on the back face The recursive state prediction satisfies the physical laws of irreversible wear and monotonically increasing tool life consumption, improving the interpretability of the prediction results. This invention establishes a unified correction mechanism for microscopic and virtual observations, enabling real-time signal correction of wear states even under sparse measured labels, reducing the dependence of pure data-driven models on a large number of full-lifetime labeled samples. This invention recursively predicts the tool life to the end of the tool life based on future planned working conditions and calculates the remaining life (RUL) through intersection interpolation, outputting the remaining life in minutes, seconds, or remaining workable parts; it also provides confidence intervals and warning levels, providing a quantitative basis for tool change decisions. This invention can be deployed at the edge of the workshop, outputting health status, remaining life, and warning levels, reducing tool waste caused by premature tool changes, lowering the risk of chipping and quality abnormalities, and improving the stability and automation level of CNC turning. This invention improves the accuracy, interpretability, and engineering deployment feasibility of turning tool remaining life prediction results.

[0068] Furthermore, this invention also provides a tool life prediction system based on multi-source state perception, employing a multi-source state perception-based tool life prediction method. The system includes a multi-source signal acquisition module 101, a signal preprocessing module 102, a feature construction module 103, a physical baseline life calculation module 104, a degradation state recursion module 105, an observation and correction module 106, a remaining life solution module 107, and an early warning output module 108, connected sequentially. The multi-source signal acquisition module 101 is used to synchronously acquire multiple types of data from the CNC machine tool machining site, including CNC operating condition data, cutting force data, vibration data, cutting temperature data, acoustic emission data, and spindle power data. The CNC operating condition data may include information such as cutting speed, feed rate, depth of cut, tool number, workpiece material number, and machining program segment number. The signal preprocessing module 102 is used to uniformly process the acquired multi-source data, sequentially performing timestamp alignment, resampling, filtering and noise reduction, outlier removal, sliding window sampling, and fixed parameter standardization to form standardized samples that can be used for life prediction. The tool life prediction module includes a feature construction module 103, a physical baseline life calculation module 104, a degradation state recursion module 105, an observation and correction module 106, a remaining life solution module 107, and an early warning output module 108. The feature construction module 103 fuses sensor features with operating condition features to generate comprehensive degradation features; the physical baseline life calculation module 104 calculates the baseline life of the cutting tool based on the current machining conditions; the degradation state recursion module 105 recursively calculates the tool wear state and damage degree based on the life consumption process; the observation and correction module 106 corrects the degradation state using online observation, offline measurement, or virtual observation results; the remaining life solution module 107 predicts the remaining tool life based on the current degradation state and subsequent operating conditions; and the early warning output module 108 outputs health status, remaining life, life trajectory, early warning level, and tool replacement suggestions. This system can be deployed collaboratively using edge computing terminals, industrial servers, and a cloud platform. The edge computing terminal is deployed on the CNC machine tool site, responsible for real-time sampling, feature calculation, early warning output, and windowed processing of high-frequency signals, enabling real-time monitoring of the cutting tool status. Industrial servers or cloud platforms are responsible for parameter training, model version management, and cross-machine tool data aggregation, providing training support for the life prediction model. For high-sampling-frequency signals such as vibration, window feature extraction can be completed at the edge, uploading only low-dimensional features, window features, and process data such as predicted life trajectories, thereby compressing the amount of data transmitted and reducing storage pressure. After the server or cloud platform completes model training and parameter optimization based on the aggregated data, it distributes the parameters, thresholds, and model version to the field terminals, enabling continuous updates to the prediction model and early warning strategies.

[0069] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for predicting the life of a lathe tool based on multi-source state perception, characterized in that, include: During the turning process, S1 synchronously collects multi-source status data according to a unified clock. The multi-source status data includes at least CNC operating condition data, cutting force data, vibration data, and cutting temperature data, and may also include acoustic emission data and spindle power data. The CNC operating condition data includes cutting speed, feed rate, depth of cut, tool number, workpiece material number, and machining program segment number. S2 sequentially performs timestamp alignment, resampling, filtering and noise reduction, outlier removal, and sliding window sampling on multi-source state data to extract time-domain, frequency-domain, and time-frequency-domain features of each time window. The features are then standardized using the mean and standard deviation fixed during the training phase to obtain standardized sensing features. S3 calculates the physical reference life that is strictly greater than zero under the current working conditions based on the cutting speed, feed rate, depth of cut and workpiece material hardness of the current time window; S4 concatenates standardized sensing features with normalized operating condition features into a comprehensive feature vector, and constructs a sensing correction factor that is strictly greater than zero. The sensor correction factor is anchored using the reference operating condition health window so that its center value is 1 under the reference state, and the actual damage rate is calculated in combination with the physical reference lifetime. S5 recursively derives the prior wear amount and state uncertainty of the cutting tool flank face based on the actual damage rate and process increment, and limits the prior wear amount to the wear amount corrected in the previous time window and the allowable wear limit, thus obtaining the prior damage degree that satisfies the monotonically non-decreasing constraint. S6 When microscopic measurements are available, microscopic wear is preferred as the observation; when microscopic measurements are not available, virtual wear is used as the observation only when the virtual observation confidence level reaches the minimum confidence threshold; the observation noise variance is determined according to the observation source, the prior wear is corrected for gain, and the correction result is subject to monotonic interval constraints. S7 uses the current correction damage degree as the initial value, and recursively calculates the remaining tool life based on the future working condition sequence in the CNC program or production schedule until the damage degree reaches 1. When the future sensing characteristics are unknown, the statistical value or safety quantile of the historical correction factor of the same working condition is used to determine the future sensing correction factor. S8 samples model parameters, process increments, observation noise, and future operating condition disturbances to obtain multiple remaining lifetime samples and calculates the mean and confidence interval of the remaining lifetime. Based on the current safe termination status, lower limit of the confidence interval, average remaining lifetime, corrected damage degree, and process threshold, the warning level is output according to the priority of immediate tool change, early warning, and normal.

2. The tool life prediction method based on multi-source state perception as described in claim 1, characterized in that, In step S3, the physical reference lifetime is calculated according to the following formula: ； in, The reference life is under reference operating conditions. , , These represent the cutting speed, feed rate, and depth of cut for the current time window. This refers to the material's hardness factor. For reference operating parameters, All are non-negative lifetime sensitivity indices; in, For the first The physical reference life under the current operating conditions for each cutting time window. Number the cutting time window. For the first Brinell hardness of the workpiece material corresponding to each cutting time window; The reference working condition is a calibration working condition that is consistent with the tool grade, workpiece material and cooling conditions to be predicted. These are the cutting speed, feed per revolution, depth of cut, and Brinell hardness of the workpiece under the specified calibration conditions. The speed is 120–220 m / min, and f0 is 0.10–0.25 mm / r. It ranges from 0.5 to 2.0 mm. The range is 180–260 HB; the calibration range of the lifetime sensitivity index is... , , , .

3. The tool life prediction method based on multi-source state perception as described in claim 2, characterized in that, In step S4, The sensing correction factor Constructed using an exponential function: ,in Let be the damage rate parameter, and , The composite feature vector; actual damage rate ,in, , where is the baseline damage rate; the dimension of the weight vector in the damage rate parameter is consistent with the comprehensive feature vector, the bias is a scalar, and to avoid scale coupling between the reference lifetime and the bias, the logarithmic center value of the correction factor is constrained to 0 on the health window set of the reference operating condition, making the center value of the correction factor 1; where Adjust the weight vector to reflect the damage rate. express transpose, For biased scalars, This represents the natural exponential function.

4. The tool life prediction method based on multi-source state perception as described in claim 3, characterized in that, In step S5, the prior wear amount of the cutting tool flank is determined. Calculate using the following recursive formula: ； in, This refers to the wear on the flank face. This is the initial wear amount. To allow limits, This represents the length of the k-th window. The damage rate is represented by the superscript pri, which indicates the prior estimate before observation correction, and the superscript cor, which indicates the corrected estimate after observation correction. This is the process increment, used to characterize unmodeled wear fluctuations; the deterministic point estimate takes the expected value of this increment, and during random sampling, the recursive result is projected between the previous window's corrected wear amount and the allowable wear limit to ensure that the wear amount does not reverse; the corresponding state uncertainty recursively is: ； in, Indicates the first The variance of wear state estimation for each cutting time window; its superscripts pri and cor represent prior estimate and corrected estimate, respectively; The process incremental variance; the prior damage degree .

5. The tool life prediction method based on multi-source state perception as described in claim 4, characterized in that, Step S6 specifically includes: The actual wear amount was obtained using a microscope during random inspections after machine shutdown. ; without shutting down the system, utilize the virtual observation model trained from sensor features Wear observations are given, and both are collectively referred to as observations. : ； in This indicates whether microscopic measurement is present. Indicates the confidence level of virtual observations. The lowest confidence threshold, This indicates that there are no reliable observations in the current window. This represents the variance of observation noise; the variance of microscopic observations is obtained from calibration, while the variance of virtual observations increases as the confidence level decreases. in, This is the set of parameters for the virtual observation model; When reliable observations exist, a one-dimensional gain correction method is used to fuse prior wear and observed wear: ； ； ； in, For the first The observation correction gain for each cutting time window; after the gain correction is completed, the corrected wear amount is limited between the corrected wear amount of the previous time window and the allowable wear limit. This limit is used to suppress non-physical wear regression caused by observation noise. When there are no reliable observations in the current window, let , , The corrected damage and health levels are: 。 6. The tool life prediction method based on multi-source state perception as described in claim 5, characterized in that, Step S7 specifically includes: To obtain the corrected damage level at the current moment. Then, read the sequence of future cutting conditions from the CNC program or production schedule. ,make For the future Each window calculates the corresponding baseline lifetime, sensing correction factor, and damage rate. And perform recursion: When there are no real-time sensing features for the future time window, the correction factor is retrieved according to the historical window that is closest to the future operating conditions, and the mean, median or upper percentile is selected according to risk preference. ； in, For the first A vector of planned cutting conditions for a future time window, where the superscript p indicates the planned or predicted quantity for the future. For the current number The future time window sequence number starting from the current time window. To advance towards the future Predicted damage level after a time window For the first The duration of the planned future time window. The corresponding predicted damage rate; Define the future window number for the first time reaching the end of life as: ； If the end of life occurs in the first Within a future window, linear interpolation is performed on that window to obtain the remaining tool life: ； In the formula, To avoid division by zero caused by zero damage rate, a very small positive number is set. When the damage level of 1 is not reached within the given prediction range, the future operating condition sequence is continued to be extended; if it cannot be extended, the remaining lifetime is output as a right censored lower bound greater than the current prediction range.

7. The tool life prediction method based on multi-source state perception as described in claim 6, characterized in that, In step S8, the lower limit of the confidence interval is denoted as L. k The early warning time threshold is The forced tool change time threshold is The damage warning threshold is η1, where And 0 < η1 < 1; Warning decisions are made according to the following priority: current safety termination criteria triggered, Mean remaining lifespan ≤ Or, if the damage correction is ≥1, output "immediate tool change"; if the immediate tool change condition is not met and Mean remaining lifespan < If the damage level is ≥η1, an early warning will be output; otherwise, a normal output will be output.

Images