Tool wear life prediction method and system based on internet of things perception

Abstract

The application relates to the technical field of intelligent manufacturing and cutting processing state monitoring, and discloses a tool wear life prediction method and system based on Internet of Things sensing, which comprises the following steps: obtaining original signals, carrying out denoising treatment to obtain clean data, carrying out feature extraction and principal component dimension reduction to obtain a key feature combination, calculating a power spectrum density to determine a working condition and extract evolution and load features, using Kalman filtering to carry out multi-source recursive fusion to obtain a feature trend value, classifying and quantifying to obtain a wear category, obtaining a real-time monitoring deviation and calculating a monitoring deviation change rate, combining the monitoring deviation and a fault probability, using a random forest model to estimate a preliminary life, using an exponential smoothing method to dynamically correct to obtain a predicted life, and combining maintenance constraints to execute multi-target decision evaluation to obtain a maintenance guide. The method can realize high-precision self-adaptive prediction of tool wear life and optimal maintenance decision under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and cutting process condition monitoring technology, and in particular to a method and system for predicting tool wear life based on Internet of Things (IoT) sensing. Background Technology

[0002] Currently, in modern manufacturing, the wear condition of cutting tools directly affects machining quality and production efficiency. As a core actuator of machining equipment, real-time monitoring and life prediction of cutting tool operating status are important means to achieve preventive maintenance and reduce unplanned downtime losses.

[0003] In existing technologies, basic operating parameters during machining are typically acquired through monitoring, control, and data acquisition systems. Threshold alarms are then performed based on preset fixed thresholds or single-dimensional sensor signals (such as vibration or cutting force signals alone). This approach often relies on post-processing analysis or simplistic mapping logic, making it difficult to adapt to the dynamic fluctuations and noise interference of physical signals under complex conditions such as high-speed cutting. Because of the complex nonlinear coupling between vibration, temperature, and force changes in the machining environment, single-dimensional monitoring methods are highly susceptible to false alarms or missed alarms due to environmental interference. A deeper technical challenge lies in the fact that even with multi-dimensional monitoring data, existing prediction models often lack the ability to adaptively calibrate the prediction results, leading to a significant discrepancy between the predicted remaining tool life and the actual tool wear state.

[0004] Existing technologies suffer from low fusion of multi-source monitoring information in complex processing environments and a lack of dynamic real-time correction capabilities for lifetime prediction results. Summary of the Invention

[0005] This invention provides a method and system for predicting tool wear life based on Internet of Things (IoT) sensing, in order to solve the problems in the prior art where the fusion of multi-source monitoring information in complex machining environments is low and the life prediction results lack dynamic real-time correction capabilities.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a method for predicting tool wear life based on Internet of Things (IoT) sensing, comprising: The original vibration signal and the original temperature signal are acquired. The original vibration signal is subjected to time-frequency denoising processing to obtain a denoised signal. The denoised signal and the original temperature signal are then filtered and optimized to obtain clean data. Multidimensional feature set is obtained by extracting multidimensional features from the clean data, and principal component dimensionality reduction is performed on the multidimensional feature set to obtain key feature combination. Calculate the average power spectral density of the cleaning data within a preset time window, and determine the current operating condition based on the average power spectral density; Based on the current operating condition, the corresponding evolution features and load features are extracted. The Kalman filter algorithm is then used to perform recursive state estimation on the key feature combination, the evolution features, and the load features to obtain feature trend values. The wear category is obtained by classifying and quantifying the characteristic trend values. Real-time monitoring deviation is obtained and the rate of change of monitoring deviation is calculated. Based on the wear category, the real-time monitoring deviation and the preset failure probability, a pre-trained random forest model is used to estimate the lifespan and obtain the preliminary lifespan. The preliminary lifetime is dynamically corrected by using an adaptive correction algorithm based on exponential smoothing combined with the rate of change of the monitoring deviation, so as to obtain the predicted lifetime. Based on the predicted lifespan and pre-acquired maintenance constraints, a multi-objective decision evaluation is performed to obtain maintenance guidelines.

[0007] Secondly, the present invention provides a tool wear life prediction system based on Internet of Things (IoT) sensing, comprising: The data preprocessing module is used to acquire the original vibration signal and the original temperature signal, perform time-frequency denoising processing on the original vibration signal to obtain a denoised signal, and perform filtering optimization processing on the denoised signal and the original temperature signal to obtain clean data. The feature dimensionality reduction module is used to extract multidimensional features from the clean data to obtain a multidimensional feature set, and to perform principal component dimensionality reduction on the multidimensional feature set to obtain a key feature combination. The working condition identification module is used to calculate the average power spectral density of the cleaning data within a preset time window and determine the current working condition; The feature fusion module is used to extract evolution features and load features, and uses the Kalman filter algorithm to perform recursive state estimation to obtain feature trend values. The classification and quantization module is used to perform classification and quantization processing based on the characteristic trend values ​​to obtain the wear category; The preliminary estimation module is used to obtain the real-time monitoring deviation and calculate the rate of change of the monitoring deviation. Based on the wear category, the real-time monitoring deviation and the preset failure probability, the module uses a pre-trained random forest model to estimate the lifespan and obtain the preliminary lifespan. The dynamic correction module is used to dynamically correct the preliminary lifetime by combining the monitoring deviation change rate with an adaptive correction algorithm based on exponential smoothing to obtain the predicted lifetime. The decision evaluation module is used to perform multi-objective decision evaluation based on the predicted lifespan and pre-acquired maintenance constraints to obtain maintenance guidance.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses a threshold mechanism triggered by judging the vibration evolution characteristics in the key feature combination to extract temperature and load features. It then uses the Kalman filter algorithm with the observed covariance reference value as a constraint to perform multi-source recursive state estimation and fusion, which changes the limitations of traditional monitoring that relies on a single physical dimension or fixed threshold. This mechanism can accurately identify and dynamically respond to multi-source nonlinear signal features under different machining conditions, effectively filtering out interference caused by machine tool vibration and ambient temperature fluctuations. It ensures the physical fidelity of feature extraction from the source and significantly improves the accuracy and noise resistance of tool state recognition in complex and variable cutting environments.

[0009] (2) This invention extracts the rate of change of real-time monitoring deviation within a continuous monitoring period and uses an adaptive correction algorithm based on exponential smoothing to dynamically correct the initial lifetime output by the random forest model, breaking through the technical bottleneck of the traditional prediction model's severe lag in response to high-frequency small mutations. This scheme innovatively uses the deterioration rate of real-time working conditions as a self-healing compensation factor, which can track the nonlinear acceleration trend of tool wear in real time and adaptively calibrate the lifetime prediction, greatly enhancing the anti-interference self-healing ability of the prediction system under abnormal cutting conditions, thereby achieving high-precision convergence of the remaining lifetime prediction results and close alignment with the actual working conditions.

[0010] (3) This invention constructs dynamic maintenance constraints by acquiring equipment downtime loss costs and production scheduling maintenance window times, and combines high-precision predicted lifespan for multi-objective optimization calculations and minimum value solutions, abandoning the traditional passive and extensive maintenance strategy that relies solely on tool life exhaustion. This scheme deeply integrates the underlying physical lifespan prediction results with the macro-level production scheduling and cost models, enabling it to accurately lock the target maintenance window with the optimal overall cost before tool failure, realizing a leap from passive post-event maintenance to proactive and economical preventive maintenance, and significantly reducing the probability of unplanned downtime. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the tool wear life prediction method based on Internet of Things sensing provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the tool wear life prediction system based on Internet of Things sensing provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a tool wear life prediction method based on Internet of Things (IoT) sensing, including the following steps: S11, acquire the original vibration signal and the original temperature signal, perform time-frequency denoising processing on the original vibration signal to obtain a denoised signal, and perform filtering optimization processing on the denoised signal and the original temperature signal to obtain clean data; S12, perform multidimensional feature extraction based on the clean data to obtain a multidimensional feature set, and perform principal component dimensionality reduction on the multidimensional feature set to obtain a key feature combination; S13, calculate the average power spectral density of the cleaning data within a preset time window, and determine the current working condition based on the average power spectral density; S14. Based on the current working condition, extract the corresponding evolution features and load features, and use the Kalman filter algorithm to perform recursive state estimation processing on the key feature combination, the evolution features and the load features to obtain feature trend values. S15, perform classification and quantification processing based on the characteristic trend value to obtain the wear category; S16, acquire real-time monitoring deviation and calculate monitoring deviation change rate, and use a pre-trained random forest model to estimate lifespan based on the wear category, the real-time monitoring deviation and the preset failure probability to obtain preliminary lifespan; S17, The preliminary lifetime is dynamically corrected by using an adaptive correction algorithm based on exponential smoothing combined with the rate of change of the monitoring deviation to obtain the predicted lifetime. S18, perform multi-objective decision evaluation based on the predicted lifespan and pre-acquired maintenance constraints to obtain maintenance guidance.

[0014] In step S11, the original vibration signal and the original temperature signal are acquired, the original vibration signal is subjected to time-frequency denoising processing to obtain a denoised signal, and the denoised signal and the original temperature signal are subjected to filtering optimization processing to obtain clean data.

[0015] The clean data obtained through time-frequency denoising and filtering optimization includes: The original vibration signal is decomposed using a wavelet transform algorithm, and soft thresholding is performed based on a preset high-frequency noise threshold to reconstruct a denoised vibration signal. The energy concentration frequency band of the denoised vibration signal is identified using a short-time Fourier transform algorithm, and a clean vibration signal is obtained by filtering the energy concentration frequency band using a bandpass filter. The original temperature signal is smoothed using a moving average filtering algorithm to obtain a cleaning temperature signal, and the cleaning vibration signal is combined with the cleaning temperature signal to obtain the cleaning data.

[0016] In one implementation, a high-precision accelerometer and an infrared temperature sensor configured on the machine tool spindle collect physical state signals during tool operation, respectively, and extract the corresponding discrete time-series sequences to obtain the original vibration signal and the original temperature signal. This embodiment utilizes a discrete wavelet transform algorithm to decompose the original vibration signal. Specifically, this embodiment selects the Daubechies wavelet basis function to perform multi-scale discrete decomposition on the original vibration signal, extracting a low-frequency profile coefficient sequence and a high-frequency detail coefficient sequence.

[0017] It should be noted that this embodiment uses a no-load reference calibration method to determine the preset high-frequency noise threshold. In this embodiment, the machine tool spindle is controlled to operate under no-load rotation, and the vibration sequence under this reference state is collected and subjected to wavelet decomposition at the same level to extract the corresponding high-frequency detail coefficients. This embodiment calculates the standard deviation of the high-frequency detail coefficients under the no-load state, multiplies the standard deviation by a constant three to obtain a judgment value, and objectively sets this judgment value as the preset high-frequency noise threshold. This embodiment iterates through the high-frequency detail coefficient sequence under the main machining state and compares its absolute value with the preset high-frequency noise threshold. If the absolute value of the high-frequency detail coefficient is greater than the preset high-frequency noise threshold, this embodiment subtracts the preset high-frequency noise threshold from the absolute value of the high-frequency detail coefficient and multiplies it by the initial sign function of the high-frequency detail coefficient to restore its polarity; if the absolute value of the high-frequency detail coefficient is less than or equal to the preset high-frequency noise threshold, this embodiment directly sets the high-frequency detail coefficient to zero, completing the soft thresholding process. Subsequently, in this embodiment, the high-frequency detail coefficients after soft thresholding and the unmodified low-frequency profile coefficients are extracted, and an inverse wavelet transform operation is performed to reconstruct and generate a denoised vibration signal.

[0018] In one implementation, a short-time Fourier transform algorithm is used to identify the energy concentration frequency band of the denoised vibration signal. This embodiment uses a Hanning window as a time-domain sliding window to perform frame-by-frame windowing processing on the denoised vibration signal, and performs a fast Fourier transform on each frame to calculate the power spectral density corresponding to each discrete frequency component. This embodiment sums the power spectral densities of the same frequency components in all time-domain frames to obtain a global spectral energy distribution vector. This embodiment sorts the energy values ​​in the global spectral energy distribution vector in descending order, and sequentially accumulates the sorted energy values ​​until the sum reaches 80% of the total energy value. The frequency range covered by this sum is extracted, and the continuous frequency band enclosed by the upper and lower cutoff frequencies corresponding to this frequency range is determined as the energy concentration frequency band. Based on the determined lower and upper cutoff frequencies, this embodiment configures the passband parameters of a finite-length unit impulse response bandpass filter, inputs the denoised vibration signal into the bandpass filter for filtering, removes interference harmonics outside the passband, and outputs a clean vibration signal.

[0019] It is worth noting that this embodiment uses a moving average filtering algorithm to smooth the original temperature signal to obtain the cleaning temperature signal. Regarding the determination of the moving average filtering algorithm's sliding window size, this embodiment performs autocorrelation analysis on the initial temperature time series signal during the machine preheating stage, calculates the autocorrelation function curve, extracts the coordinates of the intersection point of the autocorrelation function curve when it first drops and crosses zero on the horizontal axis, calculates the time delay interval corresponding to this intersection coordinate, and determines the number of sampling points contained within this time delay interval as the sliding window size. This embodiment performs a point-by-point sliding arithmetic mean calculation on the original temperature signal according to the sliding window size, smoothing local abrupt thermal noise to obtain the cleaning temperature signal. This embodiment extracts the cleaning vibration signal and the cleaning temperature signal, performs linear interpolation alignment using a unified hardware clock timestamp as the time axis reference, and concatenates the two time-aligned one-dimensional time series into a multi-dimensional feature matrix, combining them to obtain the cleaning data.

[0020] For example, an accelerometer acquires the raw vibration signal containing the machine tool's native high-frequency chatter noise at a sampling frequency of 1000 Hz. In this embodiment, an 8-level discrete wavelet decomposition is performed on it. The standard deviation is extracted based on the no-load state and multiplied by a constant of three to calculate a preset high-frequency noise threshold of 0.05 units of gravitational acceleration. Based on this threshold, soft thresholding and inverse transform are performed on the high-frequency detail coefficients to reconstruct the denoised vibration signal. Subsequently, a short-time Fourier transform is used to calculate that the frequency band accounting for 80% of the total energy is concentrated in the 50 Hz to 150 Hz range. A bandpass filter is configured within this range to perform filtering, outputting a clean vibration signal with improved signal-to-noise ratio. Simultaneously, an infrared temperature sensor acquires the raw temperature signal at 10 Hz. Based on the zero-crossing point of the autocorrelation function, the sliding window size is determined to be 5 sampling points, and the clean temperature signal is output point by point. Finally, the system aligns the 1000 Hz vibration signal and the interpolated upsampled temperature signal according to the absolute timestamp, outputting a high-quality clean data matrix for subsequent feature extraction.

[0021] In step S12, multidimensional feature extraction is performed on the clean data to obtain a multidimensional feature set, and principal component dimensionality reduction is performed on the multidimensional feature set to obtain a key feature combination, including: The clean data is analyzed and calculated using the Fast Fourier Transform algorithm and Hilbert Transform to obtain the frequency band energy ratio, kurtosis index and octave peak intensity; Principal component analysis algorithm is used to perform dimensionality reduction mapping on the frequency band energy ratio, the kurtosis index and the peak intensity of the octave, to obtain a new spatial feature vector; The key feature combination is obtained by extracting and filtering the new spatial feature vector based on the preset cumulative variance contribution rate threshold and the preset sliding standard deviation limit.

[0022] In one implementation, the cleaning vibration signal is extracted from the cleaning data, and the cleaning vibration signal is converted from the time domain to the frequency domain using a Fast Fourier Transform algorithm to calculate the power spectral density in the range of 0 Hz to 500 Hz. This embodiment extracts a preset main energy frequency band, namely the 50 Hz to 150 Hz interval, calculates the integral value of the power spectral density within this interval, and divides it by the total integral value of the power spectral density across the entire frequency band to obtain the frequency band energy proportion. Simultaneously, this embodiment extracts the cleaning data from the time domain dimension and calculates the ratio of the fourth moment of the signal sampling point deviating from the mean to the square of the variance to obtain a kurtosis index reflecting the steepness of the signal distribution.

[0023] It is worth noting that this embodiment utilizes Hilbert transform to perform envelope demodulation processing on the clean data, extracting the instantaneous amplitude envelope of the signal. Subsequently, this embodiment performs Fast Fourier Transform on the instantaneous amplitude envelope to extract the envelope spectrum. This embodiment obtains the real-time rotational frequency of the machine tool spindle, identifies the amplitude maxima points in the envelope spectrum that are integer multiples of the rotational frequency, and determines the amplitude value corresponding to the maxima point as the harmonic peak intensity, thereby quantifying the rotational imbalance characteristics caused by tool wear.

[0024] In one implementation, a high-dimensional original feature set is constructed, consisting of the frequency band energy proportion, the kurtosis index, the octave peak intensity, and other preset time-frequency domain indices. To eliminate the influence of different dimensions on the dimensionality reduction result, this embodiment performs zero-mean normalization on the high-dimensional original feature set. This embodiment uses principal component analysis to calculate the covariance matrix of the normalized feature set, and performs eigenvalue decomposition on the covariance matrix to extract each eigenvalue and its corresponding eigenvector. This embodiment projects the original eigenvectors into an orthogonal space constructed by the first k eigenvectors arranged in descending order of eigenvalues, obtaining a new space eigenvector.

[0025] It should be noted that this embodiment uses scree plot analysis to determine the preset cumulative variance contribution rate threshold. This embodiment collects all feature samples from historical processing states, performs preliminary principal component decomposition, calculates the proportion of each feature value to the sum of all feature values ​​to obtain the individual contribution rate, and accumulates them in descending order to obtain the cumulative contribution rate curve. This embodiment extracts the coordinates of the inflection point where the slope of the curve tends to flatten, and objectively sets the value corresponding to this inflection point as the preset cumulative variance contribution rate threshold.

[0026] It is worth noting that this embodiment utilizes a sliding window to perform time-series monitoring on the new spatial feature vector. Regarding the determination of the preset sliding standard deviation limit, this embodiment employs percentile statistics. This embodiment extracts the new spatial feature vector sequence of the tool under normal operating conditions and calculates the local standard deviation distribution set using a sliding window with a length of 100 sampling points. This embodiment sorts this set in ascending order, extracts the standard deviation value corresponding to the 95th percentile, and sets this value as the preset sliding standard deviation limit. This embodiment extracts highly identifiable dimensional components from the new spatial feature vector and combines them to generate a key feature combination consisting of frequency band energy proportion, kurtosis index, and second harmonic peak intensity.

[0027] For example, the original feature set contains statistics in 20 dimensions. In this embodiment, based on the cumulative variance contribution rate threshold of 95% determined by the scree plot, the feature dimensions are mapped from 20 dimensions to 3 dimensions, resulting in a new spatial feature vector containing the first three principal components. The system monitors the fluctuation of the first principal component through a sliding window of length 100. When its standard deviation exceeds the statistically determined threshold of 2.5, the feature extraction logic is triggered. At this time, the system extracts the frequency band energy proportion (50-150 Hz) supported by physical background, the kurtosis index reflecting abrupt shocks (normally about 3.0, rising to 5.0 in anomalies), and the second harmonic peak intensity reflecting mechanical instability as inputs to the subsequent prediction model, thereby eliminating redundant interference while retaining more than 95% of the original information.

[0028] In step S13, the average power spectral density of the cleaning data within a preset time window is calculated, and the current operating condition is determined based on the average power spectral density.

[0029] In one implementation, the cleaning vibration signal is extracted from the cleaning data. Regarding the determination of the preset time window, this embodiment employs a signal stationarity analysis method. This embodiment extracts the vibration signal sequence of the machine tool at different spindle speeds and calculates the evolution curves of the signal mean and variance over time. This embodiment identifies the shortest time span for the evolution curve to enter steady-state fluctuations and determines it as the preset time window, exemplarily set to 10 seconds.

[0030] It should be noted that this embodiment uses the periodogram method to calculate the average power spectral density of the clean data within the preset time window. This embodiment divides the clean vibration signal within the preset time window into multiple discrete data segments of equal length, and performs a Fast Fourier Transform on each discrete data segment to calculate the instantaneous power spectrum of each segment. This embodiment extracts a preset key frequency band, such as the 100 Hz to 200 Hz interval, and calculates the arithmetic mean of the instantaneous power spectra of all segments within the key frequency band at corresponding frequency points to obtain the average power spectral density reflecting the average energy distribution within the time window.

[0031] It is worth noting that this embodiment determines the current operating condition based on the average power spectral density. Regarding the logic for determining the current operating condition, this embodiment pre-constructs an operating condition mapping matrix. This embodiment collects a sample set of average power spectral density from historical machining processes and uses the K-means clustering algorithm to divide the sample set into three categories according to energy characteristic distribution: light-load stable operating condition, standard cutting operating condition, and heavy-load / impact operating condition. This embodiment calculates the centroid vectors of each cluster center and sets the operating condition judgment threshold based on the coverage of each category in the feature space.

[0032] In one implementation, the Euclidean distance of the current real-time average power spectral density relative to the cluster centers of each operating condition category is calculated. In this embodiment, the current average power spectral density is determined to belong to the category with the smallest Euclidean distance, thereby identifying the current operating condition.

[0033] For example, the system extracts cleaning data every 10 seconds. If the average amplitude of the calculated average power spectral density is 0.5g within the critical frequency band of 100 Hz to 200 Hz, and based on the condition determination threshold determined by cluster analysis (e.g., less than 0.6g is considered a stable state), this embodiment determines the current condition as a standard cutting condition. If the average amplitude rises to 0.9g, exceeding the preset threshold range, this embodiment determines the current condition as a heavy load / impact condition and uses it as the logical trigger condition for subsequent step S14 to perform adaptive feature extraction and covariance adjustment.

[0034] In step S14, the corresponding evolution features and load features are extracted according to the current operating conditions. The Kalman filter algorithm is used to perform recursive state estimation processing on the key feature combination, the evolution features and the load features to obtain feature trend values.

[0035] The extraction of corresponding evolution features and load features based on the current operating conditions includes: Determine whether the average power spectral density is greater than a preset power spectral density threshold; If the average power spectral density is greater than the preset power spectral density threshold, the wavelet transform method is used to extract the main periodic fluctuation component from the clean data to obtain the vibration evolution characteristics. The envelope spectrum peak at a specific rotation frequency in the cleaning data is extracted using an envelope analysis algorithm to obtain the load impact characteristics. The vibration evolution characteristic is determined as the evolution characteristic, and the load impact characteristic is determined as the load characteristic.

[0036] Among them, the Kalman filter algorithm is used for recursive state estimation to obtain characteristic trend values, including: Determine the preset observation covariance reference value based on the current operating conditions; Using the Kalman filter algorithm and taking the observation covariance reference value as the observation constraint, multi-source recursive prediction and update processing is performed on the key feature combination, the evolution feature and the load feature to obtain the feature fusion sequence. The slope of the feature fusion sequence over time is extracted to obtain the feature trend value.

[0037] In one implementation, the feature extraction logic is dynamically adjusted based on the current operating condition generated in step S13. This embodiment compares the average power spectral density calculated in step S13 with a preset power spectral density threshold. Regarding the determination of the preset power spectral density threshold, this embodiment extracts historical vibration spectrum data of the machine tool under no-load and standard load conditions, and calculates the power spectral density amplitude distribution of the main energy frequency band (e.g., 100 Hz to 200 Hz). This embodiment uses the percentile method to extract the value corresponding to the 95th percentile of the amplitude distribution, and objectively sets it as the preset power spectral density threshold; for example, it is set to 0.8g.

[0038] It should be noted that if the average power spectral density is greater than the preset power spectral density threshold, it is determined that the current condition is under high load or impact, triggering the extraction process of evolution features and load features. Conversely, if the average power spectral density is less than or equal to the preset power spectral density threshold, it is determined that the current condition is under stable light load or standard cutting conditions. In this case, since the long-cycle thermal deformation and transient mechanical impact effects are not significant, in order to improve the system feature extraction efficiency and avoid introducing background thermal noise, this embodiment skips the extraction process of evolution features and load features and directly sets the values ​​of evolution features and load features to zero to maintain the dimensionality consistency of the observation vector in the subsequent Kalman filter algorithm. When it is determined that the condition is under high load or impact, this embodiment uses the wavelet transform method and selects the Daubechies wavelet basis function to perform multi-level decomposition of the temperature time series signal in the clean data. This embodiment identifies and extracts low-frequency wavelet coefficients with frequencies between 0.01 Hz and 0.1 Hz, and reconstructs the main periodic fluctuation component reflecting the long-cycle temperature rise trend through inverse wavelet transform. In this embodiment, the first-order difference absolute value of the main period fluctuation component is calculated to obtain the vibration evolution characteristics. Long-period temperature rise will cause thermal expansion of the tool and workpiece, thereby changing the equivalent stiffness and damping of the cutting system, which is ultimately manifested as a slow drift of the vibration main frequency. Therefore, although the vibration evolution characteristics are directly calculated from the temperature signal, they are essentially an indirect characterization of the vibration evolution behavior.

[0039] It is worth noting that this embodiment utilizes an envelope analysis algorithm to extract the envelope spectrum peak value at a specific rotational frequency in the cleaning data. This embodiment extracts the rated rotational frequency of the machine tool spindle (e.g., 20 Hz or 30 Hz) and locks this frequency point in the envelope spectrum of the cleaning vibration signal. This embodiment obtains the amplitude maxima at this frequency point and its harmonic locations, defining them as load impact characteristics. This embodiment defines the extracted vibration evolution characteristics and load impact characteristics as the evolution characteristics and load characteristics for subsequent fusion calculations, respectively.

[0040] In one implementation, a Kalman filter algorithm is used to perform recursive state estimation of multi-source features. Regarding the determination of the preset observation covariance reference value, this embodiment establishes a dynamic mapping table based on the current operating conditions. This embodiment pre-models the measurement noise of the sensor under different cutting ratios in a laboratory environment and calculates the covariance matrix of the measurement error. This embodiment extracts the covariance value matching the current operating conditions as an observation constraint to reduce the interference of random environmental noise on feature fusion.

[0041] This embodiment constructs a discrete-time state transition equation, mapping the feature fusion vector from the previous time step to the current time step through a state transition matrix to obtain the predicted state. This embodiment obtains the current observation vector composed of key feature combinations, evolution features, and load features, and uses a Kalman gain matrix to weight and correct the residual between the predicted state and the current observation vector, calculating the optimal state estimate for the current time step, i.e., the current element in the feature fusion sequence. This embodiment uses a recursive algorithm to continuously update the feature fusion sequence.

[0042] It is worth noting that this embodiment extracts the slope of the feature fusion sequence over time to obtain the feature trend value. This embodiment truncates a fixed-length sliding time window at the end of the feature fusion sequence, and uses the least squares method to perform linear regression fitting on the fusion values ​​within this window, calculating the slope of the regression line. This slope objectively reflects the evolution rate of the tool degradation state under the current working conditions.

[0043] For example, if the current operating condition is a heavy load condition and the average power spectral density reaches 0.9g, the system extracts vibration evolution features with an amplitude change rate of 0.5℃ / min and load impact features with an amplitude of 0.7g. The system extracts the corresponding preset observation covariance reference value of 0.02 units and substitutes it into the Kalman filter recursive operator. As the time step progresses, the feature fusion sequence gradually converges from 0.8 to 1.2. The system calculates its growth slope as 0.16 units / hour through linear fitting, that is, the feature trend value is 0.16. This value is passed to step S15 to trigger the subsequent wear grading and life prediction logic.

[0044] The Kalman filter recursive process is a typical prediction-correction system. The following interactive components can be used to understand how the observation covariance (noise control) affects the smoothness and accuracy of the characteristic trend values.

[0045] In step S15, the wear category is obtained by classifying and quantifying the characteristic trend value, including: The wear rate acceleration index is obtained by calculating the derivative of the characteristic trend value using time series analysis methods. The support vector machine algorithm is used to perform nonlinear mapping calculations on the feature trend value, the wear rate acceleration index and the preset historical wear curve to obtain the deterioration rate quantification index. The wear category is obtained by hierarchically dividing the feature trend value and the deterioration rate quantification index using a decision tree classification algorithm.

[0046] In one implementation, a wear rate acceleration index is obtained by differentiating the characteristic trend values ​​using time series analysis. This embodiment extracts a discrete characteristic trend value sequence generated within a continuous monitoring period, calculates the absolute difference between characteristic trend values ​​at two adjacent moments, and divides the absolute difference by the time step of the monitoring period to obtain the characteristic change rate. Subsequently, this embodiment performs a time-dependent difference operation on the characteristic change rate again to calculate the rate of change of the characteristic change rate, which is then determined as the wear rate acceleration index to quantify the intensity of the transition from linear degradation to nonlinear accelerated deterioration of tool wear.

[0047] It should be noted that this embodiment uses a support vector machine (SVM) algorithm to obtain the deterioration rate quantification index. Regarding the feature processing method of the preset historical wear curve involved in the input of the SVM algorithm, this embodiment extracts the standard wear evolution curve stored in the historical database, and uses a geometric feature extraction algorithm to obtain the tangent slope and local curvature of the standard wear evolution curve within the interval corresponding to the current feature trend value. The tangent slope and the local curvature are used as fixed-dimensional feature components characterizing the historical wear curve.

[0048] It should be further explained that the preset historical wear curve uses normalized wear degree as the horizontal axis and characteristic trend value as the vertical axis. A normalized wear degree value of zero represents a new tool, and a value of one represents a complete failure. This curve is obtained by collecting at least fifty characteristic trend value sequences of the same type of tool under the same standard working conditions, from new tool to failure, and dynamically aligning each sequence on the time axis. Then, based on the total duration of each sequence, the time axis is proportionally compressed to the normalized wear degree range of zero to one. At each normalized wear degree sampling point, the median of all sequence characteristic trend values ​​is taken, and the medians are connected sequentially to form a smooth curve. In subsequent use, the characteristic trend value obtained at the current moment is used as the vertical axis. Linear interpolation is used to find points with equal vertical coordinates on the curve. If no perfectly equal vertical coordinates are found, the two nearest points are used for linear interpolation to locate the corresponding normalized wear degree. Then, a wear degree unit interval is taken before and after this point, and the tangent slope and local curvature of the curve within this interval are calculated for input to the support vector machine algorithm.

[0049] It is worth noting that in this embodiment, the characteristic trend value, the wear rate acceleration index, and the fixed-dimensional feature components are combined to form a fixed-dimensional input feature vector. This embodiment uses a pre-trained support vector machine regression model with a radial basis function (RBF) kernel to process the fixed-dimensional input feature vector. Regarding the training process of the support vector machine regression model, this embodiment obtains a full-lifecycle degradation training sample set, uses a grid search algorithm to optimize the penalty coefficient and kernel parameters within a preset parameter space, and uses five-fold cross-validation to extract parameter combinations that minimize the mean squared error to complete model training. This embodiment performs a nonlinear mapping calculation from a multi-dimensional feature space to a scalar score through the support vector machine regression model, outputting a degradation rate quantification index.

[0050] In one implementation, the wear category is obtained using a decision tree classification algorithm. Regarding the method for determining the logical branch threshold of each decision node in the decision tree classification algorithm, this embodiment extracts a historical labeled sample set and calculates the information gain and Gini coefficient at different feature split points. This embodiment extracts the numerical point that results in the fastest decrease in the Gini coefficient and the lowest classification error rate, and objectively labels it as the logical branch threshold. This embodiment inputs the feature trend value and the deterioration rate quantification index into the decision tree classification model and executes hierarchical judgment logic. If the feature trend value is greater than a first preset trend threshold and the deterioration rate quantification index is greater than a first preset deterioration threshold, then the wear category is determined to be severe wear; if the aforementioned conditions are not met and the feature trend value is within a second preset trend range, then the wear category is determined to be moderate wear; if none of the above conditions are met, then the wear category is determined to be slight wear.

[0051] For example, the system extracts the current feature trend value as 0.16, and calculates the wear rate acceleration index as 0.02 by performing a second derivative on the trend sequence. In this embodiment, the slope feature of the preset historical wear curve in the corresponding interval is extracted as 0.12, which is combined to form a vector input RBF kernel support vector machine model, and the output deterioration rate quantification index is calculated to be 4.5. According to the decision tree rule determined by maximizing information gain, the first preset trend threshold is 0.12, and the first preset deterioration threshold is 3.5. Since 0.16 is greater than 0.12 and 4.5 is greater than 3.5, the system matches to the corresponding logical branch and outputs that the current tool wear category is severe wear. The system updates the wear category to the equipment status record in real time, as a key basis for the subsequent life estimation in step S16.

[0052] In step S16, real-time monitoring deviation is acquired and the rate of change of monitoring deviation is calculated. Based on the wear category, the real-time monitoring deviation, and the preset failure probability, a pre-trained random forest model is used to estimate the lifetime and obtain a preliminary lifetime, including: The current operating parameters of the tool are obtained, and the ratio of the difference between the current operating parameters and the preset standard operating parameters is calculated to obtain the real-time monitoring deviation. Extract the difference of the real-time monitoring deviation within a continuous monitoring period, calculate the rate of change of the difference relative to the monitoring period, and obtain the rate of change of the monitoring deviation. The preset failure probability is obtained by performing a mapping query in a preset historical database based on the wear category. The wear category, the real-time monitoring deviation, and the preset failure probability are input into the pre-trained random forest model for voting calculation to obtain the estimated lifespan. Obtain the equipment operating load index, and use the equipment operating load index to perform a weighted correction process on the estimated lifespan to obtain the preliminary lifespan.

[0053] In one implementation, the current operating parameters of the tool are retrieved in real time through the external data interface of the machine tool CNC system. These current operating parameters include the peak value of the spindle vibration frequency at the current moment. This embodiment obtains preset standard operating parameters. The method for determining these preset standard operating parameters is to collect the vibration frequency values ​​of the new tool during the first 100 sampling cycles of stable operation under standard conditions, calculate the arithmetic mean of these values, and set this as the preset standard operating parameter. In this embodiment, the current operating parameters of the tool are subtracted from the preset standard operating parameters to obtain an absolute difference. This absolute difference is then divided by the preset standard operating parameters to obtain a numerical proportion reflecting the degree to which the current operating characteristics deviate from the baseline, which is determined as the real-time monitoring deviation.

[0054] In one implementation, the rate of change of the monitoring deviation is obtained. This embodiment extracts the numerical sequence of the real-time monitoring deviation within a continuous monitoring period. This embodiment calculates the absolute difference between the real-time monitoring deviation in the current monitoring period and the real-time monitoring deviation in the immediately preceding monitoring period. Subsequently, this embodiment retrieves the clock frequency of the CNC system to obtain the duration of the monitoring period. The absolute difference is divided by the duration of the monitoring period to obtain the rate of change of the monitoring deviation, reflecting the rate of evolution of the monitoring deviation. The original calculated value of the rate of change of the monitoring deviation is in percentage per hour, for example, 2.85% per hour. Before substituting it into the adaptive correction formula, this value needs to be divided by 100 to convert it to a dimensionless decimal, i.e., 2.85% per hour corresponds to 0.0285 per hour. If the rate of change of the monitoring deviation is negative, it is also converted according to the above rules and substituted into the formula.

[0055] It should be noted that in this embodiment, a mapping query is performed in a preset historical database based on the wear category to obtain the preset failure probability. Regarding the construction method of the preset historical database, this embodiment collects full-lifecycle monitoring records of more than one thousand tools of the same model, counts the number of functional failure samples under each category of slight wear, moderate wear, and severe wear, and calculates the proportion of failure samples in each category to the total number of samples in that category, obtaining the prior probability distribution corresponding to each category, and storing it as the preset historical database. In this embodiment, the wear category produced in step S15 is extracted as the search key, a mapping query is performed in the preset historical database, and the corresponding percentage value is extracted to obtain the preset failure probability.

[0056] It is worth noting that this embodiment utilizes a pre-trained random forest model to estimate the estimated lifespan. The pre-trained random forest model employs an ensemble regression structure consisting of one hundred decision trees. Regarding the training process of the pre-trained random forest model, this embodiment extracts historical full-lifecycle datasets as training samples. The input features include wear category labels, monitoring deviation values, and failure probabilities, with the output target being the measured remaining lifespan. This embodiment uses a bootstrap method to resample the samples to construct a sub-training set, and randomly selects a subset of features for splitting during the classification process at each decision tree node. The weight parameters of each tree are iteratively optimized with the objective function of minimizing the mean squared error (MSE). During real-time prediction, this embodiment inputs the current wear category, real-time monitoring deviation, and preset failure probability into the pre-trained random forest model. Regression calculations are performed in parallel by each decision tree, and the arithmetic mean of the results is taken to output the theoretical prediction time, thus obtaining the estimated lifespan.

[0057] In one implementation, the equipment operating load index is obtained. The equipment operating load index is determined by retrieving the real-time cutting current from the CNC system and dividing it by the tool's rated cutting current to obtain a value reflecting the current machining intensity ratio. This embodiment uses the equipment operating load index to perform a weighted correction process on the assessed lifespan. Dividing the assessed lifespan by the equipment operating load index compensates for the accelerated lifespan decay caused by high-load machining, resulting in a preliminary lifespan that more closely reflects the actual physical wear and tear process.

[0058] For example, if the current tool wear category is severe wear, the system queries a preset historical database to obtain a corresponding preset failure probability of 0.78. The real-time vibration frequency deviation is calculated to yield a real-time monitoring deviation of 12.5%. Simultaneously, the system extracts deviation data within a continuous monitoring period and calculates the monitoring deviation change rate (e.g., 2.85% per hour) for subsequent dynamic correction. The aforementioned real-time monitoring deviation, wear category, and failure probability data are input into a pre-trained random forest model composed of one hundred decision trees for regression analysis, resulting in an estimated lifespan of 120 hours. At this point, system parameters are retrieved to calculate the current equipment operating load index as 1.3 (standard load is 1.0). In this embodiment, 120 hours are divided by 1.3 to calculate a corrected initial lifespan of approximately 92.3 hours. This result effectively quantifies the reduction in the expected tool lifespan under high-intensity loads, providing a data basis for dynamic correction in subsequent step S17.

[0059] In step S17, the monitoring deviation change rate is obtained, and the preliminary lifetime is dynamically corrected using an adaptive correction algorithm based on exponential smoothing combined with the monitoring deviation change rate, including: Extract a preset smoothing coefficient, use an adaptive correction formula to calculate the sum of the product of the preset smoothing coefficient and the rate of change of the monitoring deviation and a numerical value, and divide the preliminary lifetime by the sum to obtain the predicted lifetime.

[0060] It should be noted that this embodiment utilizes an adaptive correction algorithm based on exponential smoothing combined with the monitoring deviation change rate to dynamically correct the preliminary lifespan to obtain the predicted lifespan. Regarding the method for determining the preset smoothing coefficient, this embodiment adopts the historical prediction error minimization criterion. This embodiment collects tool wear monitoring data with a sample size of over 500 sets from historical machining records. For each sample set, this embodiment sets different candidate smoothing coefficient values ​​in a step size of 0.05 within the range of zero to one, and substitutes the candidate values ​​and the monitoring deviation change rate into the adaptive correction formula to calculate the corresponding simulated predicted lifespan. This embodiment calculates the mean square error between the simulated predicted lifespan and the actual tool failure time. This embodiment extracts the candidate smoothing coefficient value that minimizes the mean square error and objectively sets it as the preset smoothing coefficient, for example, 0.8.

[0061] In one implementation, dynamic adaptive correction of the lifetime is performed. In this embodiment, the preliminary lifetime generated in step S16 is determined as the divisor. This embodiment calculates the product of the preset smoothing coefficient and the rate of change of the monitoring deviation. This embodiment adds the product to the value to obtain the correction denominator. This embodiment divides the preliminary lifetime by the correction denominator to calculate the predicted lifetime after compensation for real-time operating condition degradation rate.

[0062] It is worth noting that this embodiment introduces the monitoring deviation change rate as a correction factor, enabling the life prediction results to respond in real time to the nonlinear accelerated deterioration characteristics of the tool during the cutting process. Compared to the static prediction model, this algorithm smooths out short-term noise through exponential smoothing logic, while retaining the penalty effect of long-term trend deviation on life assessment, ensuring rapid convergence of the prediction results to the actual failure time.

[0063] For example, the system receives the monitoring deviation change rate (e.g., 2.85% per hour, converted to a dimensionless decimal of 0.0285) and preliminary lifetime (e.g., 92.3 hours) generated in step S16. In this embodiment, a preset smoothing coefficient of 0.8 is extracted. The corrected denominator is calculated as the sum of the product of the value 1, 0.8, and 0.0285, which is 1.0228. Dividing 92.3 by 1.0228 yields a predicted lifetime of approximately 90.2 hours. This result is dynamically reduced compared to the preliminary lifetime, reflecting the real-time increase in processing risk.

[0064] In step S18, a multi-objective decision evaluation is performed based on the predicted lifespan and pre-acquired maintenance constraints to obtain maintenance guidance, including: The maintenance constraint is obtained by acquiring the equipment downtime loss cost and the production and maintenance window time. Calculate the sum of downtime loss costs and emergency maintenance costs for each available time point; The total cost is minimized using a multi-objective optimization algorithm to determine the target maintenance window corresponding to the minimum cost. The urgency index is obtained by quantifying the decay error of the predicted lifetime using a fuzzy comprehensive evaluation algorithm. Decision text is generated based on the urgency index and the target maintenance window to obtain the maintenance guidance.

[0065] In one implementation, the enterprise resource planning (ERP) database and the production execution system (MES) database are retrieved to extract the unit cost parameters for unplanned downtime and the baseline cost for equipment status recovery within a preset production cycle. This embodiment multiplies the expected downtime duration by the unit cost parameters for unplanned downtime to obtain the equipment downtime loss cost. Simultaneously, the production schedule list is read, and continuous idle time periods when equipment is in a non-processing state are extracted, with their start and end time coordinates determined as the production maintenance window time. This embodiment combines the equipment downtime loss cost with the production maintenance window time to construct the system's boundary condition matrix, thus obtaining maintenance constraints.

[0066] It is worth noting that, in this invention, cost refers to the technical costs associated with maintenance decisions. Specifically, these include production time losses due to equipment downtime, additional resources consumed by emergency maintenance forced by sudden failures, such as maintenance manpower occupation and spare tool scheduling time, and the risk of secondary equipment damage and processing quality degradation caused by further tool wear due to missing the optimal maintenance opportunity. These costs are measured in terms of time, resource occupation, and risk level, and are used to evaluate the merits of different maintenance windows at the technical level. Among them, the cost of equipment downtime loss is equal to the downtime multiplied by the output lag caused by the production line being idle per unit time, where the output lag is measured by the number of unfinished workpieces. The cost of emergency maintenance is equal to the spare tool scheduling time required for emergency maintenance plus the additional maintenance manpower occupation time, both in minutes. In actual calculations, the sum of the above two costs is calculated for three available maintenance windows before and after the predicted lifespan, and the window with the smallest sum is taken as the target maintenance window.

[0067] It should be noted that this embodiment calculates the sum of downtime loss costs and emergency maintenance costs for each available time node. This embodiment uses the predicted lifespan as the time base and iterates through all available time nodes within the production and maintenance window. For each available time node, the estimated equipment downtime loss cost caused by postponing maintenance to that node is calculated; simultaneously, the emergency maintenance costs corresponding to component replacement and expedited material scheduling at that time node are extracted. This embodiment algebraically adds the equipment downtime loss cost and emergency maintenance cost corresponding to that node to obtain the sum of the total costs for each available time node.

[0068] It is worth noting that the use of a multi-objective optimization algorithm to minimize the total cost means that the optimization process simultaneously considers multiple constraints such as downtime loss cost, emergency maintenance cost, and available time nodes, but the final optimization metric is minimizing the total cost. In this embodiment, the total cost is used as the objective function to perform a minimum value search and determine the available time node with the minimum total cost, which serves as the target maintenance window. This embodiment performs global optimization of the objective function using a heuristic search algorithm within a preset time solution space. This embodiment extracts the specific time point that minimizes the output value of the total cost function and delineates the work interval for performing replacement operations centered on this time point, defining it as the target maintenance window, thereby locking in the optimal maintenance time under the premise of minimizing cost. Specifically, the heuristic search can employ a particle swarm optimization algorithm, using the available maintenance window time point as the particle position, the reciprocal of the total cost as the fitness function, setting the number of particles to thirty, and the number of iterations to one hundred, searching within the production maintenance window time range for the position that minimizes the total cost as the target maintenance window.

[0069] In one implementation, a fuzzy comprehensive evaluation algorithm is used to quantify the tracking error of degradation features to obtain an urgency index. This embodiment extracts actual feature monitoring values ​​(such as vibration signal amplitude) over multiple consecutive monitoring periods and retrieves a pre-established theoretical attenuation model for normal wear. The root mean square error between the actual feature monitoring curve and the theoretical attenuation curve for normal wear is calculated and determined as the tracking error of the degradation features. Regarding the establishment method of the theoretical attenuation model for normal wear, this embodiment collects full lifecycle monitoring data of the same model of new tool under standard working conditions until complete failure, and uses the least squares method to perform polynomial fitting to generate the theoretical attenuation model for normal wear.

[0070] This embodiment constructs a Gaussian membership function to map the tracking error of degenerate features to a continuous interval from zero to one. Regarding the setting of the boundary parameters of the Gaussian membership function, this embodiment performs K-means clustering on the monitoring error dataset from historical fault downtime cases, extracts the centroid values ​​of the severe fault early warning clusters as the expected value of the Gaussian membership function, and uses its variance as the width parameter. This embodiment substitutes the previously calculated tracking error into the Gaussian membership function to calculate the specific membership score, which is then determined as an indicator of urgency.

[0071] Finally, a decision text is generated based on the urgency index and the target maintenance window to obtain maintenance guidance. This embodiment pre-constructs a decision rule matrix. This embodiment determines whether the urgency index is greater than a preset urgency threshold, which is objectively calibrated based on the principle of maximizing the Youden exponent of historical receiver operation characteristic curves. If the urgency index is greater than the preset urgency threshold, this embodiment extracts the specific time point of the target maintenance window, concatenates the parameters into a formatted string using a natural language generation template, generates a decision text containing a forced replacement instruction, and outputs it as maintenance guidance. If the urgency index is less than or equal to the preset urgency threshold, the current tool status is determined to be safe, and the system generates a decision text containing prompts such as "equipment is operating normally" and "continue inspection according to the predetermined cycle," outputting it as regular maintenance guidance.

[0072] For example, the system identifies an idle window that exists after 36 hours from the production scheduling database. The system calculates that if maintenance is performed at the 36-hour node, the sum of the downtime loss cost and the expedited scheduling cost will be the lowest globally. Simultaneously, the system calculates the tracking error (root mean square error) of the current vibration characteristics relative to the theoretical attenuation model to be 4.7. Substituting this into a Gaussian membership function generated based on historical fault data clustering, the urgency index is obtained as 0.82. Combining the high urgency of 0.82 with the optimal 36-hour window, the system generates a decision text through template splicing, rejecting delayed maintenance and forcibly replacing the component after 36 hours. The output is a maintenance guide for the specific degradation scenario of the current tool.

[0073] In summary, this invention constructs a multi-source nonlinear feature recursive fusion mechanism based on Kalman filtering, accurately filtering out multi-dimensional environmental interference in complex machining environments and extracting high-fidelity tool degradation trends. Simultaneously, it innovatively introduces monitoring deviation change rate combined with exponential smoothing to dynamically and adaptively correct the preliminary evaluation results of the random forest model. Finally, it integrates high-precision predicted lifespan with downtime losses and production window depth for multi-objective decision optimization, achieving accurate insight into the actual wear state of tools under extreme and nonlinear deterioration conditions and real-time self-healing calibration of remaining lifespan expectations. This completely breaks through the technical barriers of low information fusion and severely lagging model response in traditional monitoring methods, thereby seamlessly transforming the underlying physical lifespan prediction into a globally cost-optimal proactive preventive maintenance strategy. It effectively solves the technical problems of poor anti-interference capability and lack of dynamic real-time correction closed loop in existing tool lifespan prediction systems, significantly improving the scientific nature of tool operation and maintenance in intelligent manufacturing workshops.

[0074] Reference Figure 2 The second embodiment of the present invention provides a tool wear life prediction system based on Internet of Things (IoT) sensing, comprising: The data preprocessing module is used to acquire the original vibration signal and the original temperature signal, perform time-frequency denoising processing on the original vibration signal to obtain a denoised signal, and perform filtering optimization processing on the denoised signal and the original temperature signal to obtain clean data. The feature dimensionality reduction module is used to extract multidimensional features from the clean data to obtain a multidimensional feature set, and to perform principal component dimensionality reduction on the multidimensional feature set to obtain a key feature combination. The working condition identification module is used to calculate the average power spectral density of the cleaning data within a preset time window and determine the current working condition; The feature fusion module is used to extract evolution features and load features, and uses the Kalman filter algorithm to perform recursive state estimation to obtain feature trend values. The classification and quantization module is used to perform classification and quantization processing based on the characteristic trend values ​​to obtain the wear category; The preliminary estimation module is used to obtain the real-time monitoring deviation and calculate the rate of change of the monitoring deviation. Based on the wear category, the real-time monitoring deviation and the preset failure probability, the module uses a pre-trained random forest model to estimate the lifespan and obtain the preliminary lifespan. The dynamic correction module is used to dynamically correct the preliminary lifetime by combining the monitoring deviation change rate with an adaptive correction algorithm based on exponential smoothing to obtain the predicted lifetime. The decision evaluation module is used to perform multi-objective decision evaluation based on the predicted lifespan and pre-acquired maintenance constraints to obtain maintenance guidance.

[0075] It should be noted that the tool wear life prediction system based on Internet of Things (IoT) sensing provided in this embodiment of the invention is used to execute all the process steps of the tool wear life prediction method based on IoT sensing in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0076] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0077] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for predicting tool wear life based on Internet of Things (IoT) sensing, characterized in that, include: The original vibration signal and the original temperature signal are acquired. The original vibration signal is subjected to time-frequency denoising processing to obtain a denoised signal. The denoised signal and the original temperature signal are then filtered and optimized to obtain clean data. Multidimensional feature set is obtained by extracting multidimensional features from the clean data, and principal component dimensionality reduction is performed on the multidimensional feature set to obtain key feature combination. Calculate the average power spectral density of the cleaning data within a preset time window, and determine the current operating condition based on the average power spectral density; Based on the current operating condition, the corresponding evolution features and load features are extracted. The Kalman filter algorithm is then used to perform recursive state estimation on the key feature combination, the evolution features, and the load features to obtain feature trend values. The wear category is obtained by classifying and quantifying the characteristic trend values. Real-time monitoring deviation is obtained and the rate of change of monitoring deviation is calculated. Based on the wear category, the real-time monitoring deviation and the preset failure probability, a pre-trained random forest model is used to estimate the lifespan and obtain the preliminary lifespan. The preliminary lifetime is dynamically corrected by using an adaptive correction algorithm based on exponential smoothing combined with the rate of change of the monitoring deviation, so as to obtain the predicted lifetime. Based on the predicted lifespan and pre-acquired maintenance constraints, a multi-objective decision evaluation is performed to obtain maintenance guidelines.

2. The tool wear life prediction method based on IoT sensing according to claim 1, characterized in that, The process of performing time-frequency denoising processing on the original vibration signal to obtain a denoised signal, and then filtering and optimizing the denoised signal and the original temperature signal to obtain clean data, includes: The original vibration signal is decomposed using a wavelet transform algorithm, and soft thresholding is performed based on a preset high-frequency noise threshold to reconstruct a denoised vibration signal. The energy concentration frequency band of the denoised vibration signal is identified using a short-time Fourier transform algorithm, and a clean vibration signal is obtained by filtering the energy concentration frequency band using a bandpass filter. The original temperature signal is smoothed using a moving average filtering algorithm to obtain a cleaning temperature signal, and the cleaning vibration signal is combined with the cleaning temperature signal to obtain the cleaning data.

3. The tool wear life prediction method based on IoT sensing according to claim 1, characterized in that, The process involves extracting multidimensional features from the clean data to obtain a multidimensional feature set, and then performing principal component dimensionality reduction on the multidimensional feature set to obtain a key feature combination, including: The clean data is analyzed and calculated using the Fast Fourier Transform algorithm and Hilbert Transform to obtain the frequency band energy ratio, kurtosis index and octave peak intensity; Principal component analysis algorithm is used to perform dimensionality reduction mapping on the frequency band energy ratio, the kurtosis index and the peak intensity of the octave, to obtain a new spatial feature vector; The key feature combination is obtained by extracting and filtering the new spatial feature vector based on the preset cumulative variance contribution rate threshold and the preset sliding standard deviation limit.

4. The tool wear life prediction method based on Internet of Things sensing according to claim 1, characterized in that, The step of extracting the corresponding evolution features and load features based on the current operating conditions includes: Determine whether the average power spectral density is greater than a preset power spectral density threshold; If the average power spectral density is greater than the preset power spectral density threshold, the wavelet transform method is used to extract the main periodic fluctuation component from the clean data to obtain the vibration evolution characteristics. The envelope spectrum peak at a specific rotation frequency in the cleaning data is extracted using an envelope analysis algorithm to obtain the load impact characteristics. The vibration evolution characteristic is determined as the evolution characteristic, and the load impact characteristic is determined as the load characteristic.

5. The tool wear life prediction method based on Internet of Things sensing according to claim 1, characterized in that, The recursive state estimation process using the Kalman filter algorithm to obtain characteristic trend values ​​includes: Determine the preset observation covariance reference value based on the current operating conditions; Using the Kalman filter algorithm and taking the observation covariance reference value as the observation constraint, multi-source recursive prediction and update processing is performed on the key feature combination, the evolution feature and the load feature to obtain the feature fusion sequence. The slope of the feature fusion sequence over time is extracted to obtain the feature trend value.

6. The tool wear life prediction method based on IoT sensing according to claim 1, characterized in that, The step of classifying and quantifying wear categories based on the characteristic trend values ​​includes: The wear rate acceleration index is obtained by calculating the derivative of the characteristic trend value using time series analysis methods. The support vector machine algorithm is used to perform nonlinear mapping calculations on the feature trend value, the wear rate acceleration index and the preset historical wear curve to obtain the deterioration rate quantification index. The wear category is obtained by hierarchically dividing the feature trend value and the deterioration rate quantification index using a decision tree classification algorithm.

7. The tool wear life prediction method based on Internet of Things sensing according to claim 1, characterized in that, The process of acquiring real-time monitoring deviations and calculating the rate of change of monitoring deviations, and then using a pre-trained random forest model to estimate the lifespan based on the wear category, the real-time monitoring deviations, and the preset failure probability to obtain a preliminary lifespan, includes: The current operating parameters of the tool are obtained, and the ratio of the difference between the current operating parameters and the preset standard operating parameters is calculated to obtain the real-time monitoring deviation. Extract the difference of the real-time monitoring deviation within a continuous monitoring period, calculate the rate of change of the difference relative to the monitoring period, and obtain the rate of change of the monitoring deviation. The preset failure probability is obtained by performing a mapping query in a preset historical database based on the wear category. The wear category, the real-time monitoring deviation, and the preset failure probability are input into the pre-trained random forest model for voting calculation to obtain the estimated lifespan. Obtain the equipment operating load index, and use the equipment operating load index to perform a weighted correction process on the estimated lifespan to obtain the preliminary lifespan.

8. The tool wear life prediction method based on Internet of Things sensing according to claim 1, characterized in that, The step of dynamically correcting the preliminary lifetime using an adaptive correction algorithm based on exponential smoothing combined with the rate of change of the monitoring deviation to obtain the predicted lifetime includes: Extract a preset smoothing coefficient, use an adaptive correction formula to calculate the sum of the product of the preset smoothing coefficient and the rate of change of the monitoring deviation and a numerical value, and divide the preliminary lifetime by the sum to obtain the predicted lifetime.

9. The tool wear life prediction method based on Internet of Things sensing according to claim 1, characterized in that, The step of performing multi-objective decision evaluation based on the predicted lifespan and pre-acquired maintenance constraints to obtain maintenance guidance includes: The maintenance constraint is obtained by acquiring the equipment downtime loss cost and the production and maintenance window time. Calculate the sum of downtime loss costs and emergency maintenance costs for each available time point; The total cost is minimized using a multi-objective optimization algorithm to determine the target maintenance window corresponding to the minimum cost. The urgency index is obtained by quantifying the decay error of the predicted lifetime using a fuzzy comprehensive evaluation algorithm. Decision text is generated based on the urgency index and the target maintenance window to obtain the maintenance guidance.

10. A tool wear life prediction system based on Internet of Things (IoT) sensing, characterized in that, include: The data preprocessing module is used to acquire the original vibration signal and the original temperature signal, perform time-frequency denoising processing on the original vibration signal to obtain a denoised signal, and perform filtering optimization processing on the denoised signal and the original temperature signal to obtain clean data. The feature dimensionality reduction module is used to extract multidimensional features from the clean data to obtain a multidimensional feature set, and to perform principal component dimensionality reduction on the multidimensional feature set to obtain a key feature combination. The working condition identification module is used to calculate the average power spectral density of the cleaning data within a preset time window and determine the current working condition; The feature fusion module is used to extract evolution features and load features, and uses the Kalman filter algorithm to perform recursive state estimation to obtain feature trend values. The classification and quantization module is used to perform classification and quantization processing based on the characteristic trend values ​​to obtain the wear category; The preliminary estimation module is used to obtain the real-time monitoring deviation and calculate the rate of change of the monitoring deviation. Based on the wear category, the real-time monitoring deviation and the preset failure probability, the module uses a pre-trained random forest model to estimate the lifespan and obtain the preliminary lifespan. The dynamic correction module is used to dynamically correct the preliminary lifetime by combining the monitoring deviation change rate with an adaptive correction algorithm based on exponential smoothing to obtain the predicted lifetime. The decision evaluation module is used to perform multi-objective decision evaluation based on the predicted lifespan and pre-acquired maintenance constraints to obtain maintenance guidance.

Images