1 Introduction

With the sustained development of industry 4.0 and intelligent manufacturing, the demand for advanced manufacturing technology and advanced machining processes is increasing in practice. While in the pursuit of higher machining quality and efficiency, tool wear is ever present and affects the entire machining process. Tool wear has a direct impact on the surface quality of the workpiece, while tool breakage can cause machine downtime and reduce machining efficiency. Studies have shown that the time lost to machining discontinuities caused by tool breakage accounts for 20% of total downtime [4.1 Signal acquisition

Currently, the signals are mainly acquired by force sensors, acceleration sensors, eddy current sensors, acoustic emission sensors, microphones and other devices. The force signal is the preferred signal for online tool wear monitoring as it best reflects changes in tool condition. Lubis.et al. [20] found that tool wear causes changes in cutting forces. The cutting force signal is more sensitive to tool wear than other signals. However, the installation of the force sensor is very demanding in practice. Different installation methods have a significant impact on the measurement results and the force sensor is very expensive, which can significantly increase the machining costs.

The motor current signal and power signal are also very reliable. When the cutting force changes, the spindle motor current and output power will also change, and the tool wear status can be obtained by analyzing the changes in motor current and output power. Wu [52] based on the current signal, wavelet decomposition is used to decompose and extract features, and then a feature filtering method is used to classify the sensitive features to obtain the tool wear status. Masahiro Uekita 53 combined motor current and power signals with acoustic emission signals, projected the two signals onto a two-dimensional map, and implemented online monitoring of tool wear using a split line approach. The current detection method is simple, low cost and less destructive to the system. However, it has the disadvantage that it is not very sensitive and subtle tool condition changes its not monitored, especially in the case of small tools, its detection rate is low.

During the machining process, the interaction between the tool and the workpiece will make noise. Using the microphone to collect the machining noise can also analyze the tool wear. F-Huda [17] uses a microphone to record the sound data of tool wear during machining. The resulting sound signal is then analyzed in the time–frequency domain using wavelet transform. The results show that the acoustic signal in the time–frequency domain significantly increases in amplitude under the wear condition. Pan [18] Tool wear monitoring was implemented based on cutting sound signal analysis and processing. It was found that the cutting speed has the greatest influence on the cutting sound signal, while the depth of cut and feed have the least influence. However, it only studied the identification of the tool wear state, which has not yet reached the requirement of real-time tool wear monitoring. Moreover, the range of cutting parameters chosen is small and difficult to adapt to the complex working conditions of actual production. The use of microphones for acoustic signal acquisition is cheap and simple to install, while not affecting the stiffness of the system. However, the microphone acquires a complex signal, which creates a large workload for post-processing. The Fig. 6 shows the process of tool wear monitoring based on acoustic signals.

Fig. 6
figure 6

Tool wear monitoring process based on sound signal [19]

Full size image

In addition to the above signals, Zhang [54] extracted time-domain-frequency-domain features from vibration signals and obtained an average accuracy of 99.1% in a multiscale principal element modal. Bhuiyan [55]. found that the amplitude of the AE signal became larger with increasing tool wear as the material removal speed increased. In addition, cutting temperature and surface roughness can also be used as tool wear monitoring signals. The Table 1 shows the principles, advantages and disadvantages of the various signals.

Table 1 Working principles and advantages and disadvantages of various monitoring signals [56]
Full size table

Single sensors can be flawed to varying degrees, making online monitoring accuracy often less than ideal. Therefore, multi-sensor fusion techniques have been proposed by some scholars. Dimla [57] The dynamic and static forces of the cutting process were decomposed and combined with vibration signals to build a tool wear status monitoring system, and through signal analysis, it was found that the measurement results had a higher accuracy compared to a single sensor. Figure 7a shows the structure of the multi-sensor fusion system, and (b) shows the flow chart of the multi-sensor fusion.

Silva et al. [58] fusion of cutting power signals with AE signals and combination of probabilistic neural networks successfully implemented tool wear monitoring during machining. Heng et al. [59] fused vibration signals, cutting force signals and acoustic emission signals to build a tool RUL model, which has the advantage of improving the prediction accuracy by at least 4% compared with single signal methods.

Multi-signal fusion technology can effectively improve tool wear monitoring accuracy, but the number of sensors needs to be studied, too many sensors will increase production costs and maintenance costs, the machine tool will also become more modified, making machining uncertainties increase, and too many sensors will cause a greater burden on the signal processing system, causing information processing blockages, reducing the system real-time. Too few sensors will not meet the monitoring accuracy requirements. Therefore, the selection of the number and type of sensors is particularly important for the performance of multi-sensor fusion systems.

Fig. 7
figure 7

a Structure model of multi-sensor fusion system [60] b Flow chart of multi-sensor fusion

Full size image

4.2 Signal processing and feature extraction

The signals collected by the sensors contain signals related to the tool wear state, but also inevitably contain signals from environmental noise and changes in tool parameters. Therefore, it is necessary to decompose the collected signals and extract the highly sensitive signals to replace the original signals in order to finally monitor tool wear. Currently the commonly used signal processing methods are time-domain, frequency-domain and time–frequency-domain methods. The Table 2 shows the signal processing methods and their characteristics.

Table 2 Signal processing methods and characteristics
Full size table

Time domain analysis, also known as waveform analysis, is the direct raw sequence analysis of the acquired machining signals. Using time domain analysis, the characteristic parameters of the tool wear state during machining can be effectively obtained. The Table 3 shows the commonly used time domain characteristic parameters and their calculation formulas.

Table 3 Commonly used time-domain statistical parameters and calculation formulas
Full size table

Duo et al. [61] analyzed the predictive power of time-domain statistical features of internal and external signals on tool wear. The most sensitive signals to tool wear were also identified from the time domain features based on an automatic learning algorithm and the predictive power of the method was demonstrated experimentally. Time domain analysis is very widely used, but the drawbacks of time domain analysis are also obvious. During actual machining, the dynamic characteristics of the tool and workpiece can lead to non-linear and non-smooth signals. It is therefore difficult to use a limited number of parameters for stable monitoring of changing machining conditions, and the time domain method is easily disturbed by external signals, resulting in misjudgment of the tool wear state.

Frequency domain analysis, also known as spectrum analysis, uses the frequency characteristics of signals to express the dynamic characteristics of linear systems, using the frequency characteristics of signal analysis method to make up for the shortcomings of time domain analysis. As the basis of frequency domain analysis, the frequency and amplitude of each data point are extracted as eigenvalues. Fast Fourier transform (FFT) is widely used in milling tool wear determination. Jiang [62] The FFT is used to extract features from the vibration signal in the time domain to the frequency domain. Alexandre et al. [63] analyzed the acoustic emission signal in the frequency domain, selected the best frequency band, and processed the frequency band and input it into the fuzzy system to realize the online monitoring of tool wear. Klaic et al. [64] used FFT to transform the vibration signal from time domain to frequency domain, and then selected the best bandwidth sample from the transformed signal power spectrum as the tool wear characteristic value. Frequency domain analysis expresses the steady-state characteristics of the system along with the transient characteristics; it can be a good solution for systems that are difficult to start from physical laws.

When it comes to signal processing, the FFT does not express information about the frequency of a signal at a given moment in time in the frequency spectrum. Furthermore, Fourier analysis essentially uses a set of sine basis functions or cosine basis functions to approximate the signal [65], the sine and cosine functions are both long periodic functions with no wires. Therefore, for non-smooth signals, they cannot be approximated effectively. In practice, however, most signals are non-stationary. In recent years, time–frequency domain analysis has been proposed by some scholars and become a research hotspot. In online tool wear monitoring, the commonly used time–frequency analysis methods are wavelet transform, empirical modal decomposition, variational modal decomposition and various algorithms. Table 4 for the commonly used time–frequency domain features calculation formula.

Table 4 Common time–frequency domain characteristics and calculation formulas
Full size table

As the most widely used time–frequency analysis method, wavelet transform was first proposed by French engineer J. Morlet in 1974. The wavelet transform has a good decomposition function for the low frequency part of the signal, but the effect is not good for the high frequency part, so the wavelet transform cannot completely meet the requirements of real-time monitoring. In addition, it is also difficult to select the optimal wavelet base and signal decomposition layers, and the selection of these parameters will affect the feature extraction results [80] used a Bayesian optimized regression support vector machine for determining tool wear under various machining parameters. Chen et al. [81] proposed a prediction model for drill wear identification based on the adaptive particle swarm optimization (APSO) algorithm and the least squares support vector machine (LS-SVM) algorithm. Kong et al. [82] proposed a WOA-SVM model integrating Support Vector Machine (SVM) and Whale Optimization Algorithm (WOA) for accurate estimation of tool wear of titanium alloy Ti-6Al-4 V end mills under variable cutting conditions. Support vector machines also have limitations and there is still much scope for research in the processing of large amounts of data, the handling of multi-classification problems and the selection of kernel functions.

Since its creation in 70th last century, the hidden Markov Model has been used in various fields such as computer text recognition, mobile communication technology, bio-information technology and fault diagnosis technology. The Fig. 8 shows a monitoring system based on the hidden Markov model.

Fig. 8
figure 8

Hidden Markov model monitoring system [83]

Full size image

Zhang et al. [23] proposed a SSAE-PHMM model based on Hidden Markov Model. Firstly, a DT that can reflect the real state of the tool is established, and the tool wear state is predicted by visual display and analysis in the virtual space; secondly, a tool wear state recognition model based on SSAE-PHMM is established, which can adaptively complete the time-domain feature extraction. And for each tool wear state, multiple HMM models are combined into one PHMM model to achieve accurate recognition of tool wear states. Hidden Markov models are capable of modelling non-smooth physical processes, but most studies have been carried out under fixed cutting conditions. To address this problem, Li et al. [84] proposed an improved HMM to construct a risk model to describe the time-varying and conditional adaptive state transfer probabilities, and a multilayer perceptron (MLP) was used to approximate the non-linear function and calculate the observation probabilities. The state transfer probabilities and observation probabilities are then combined to estimate the tool wear state using a forward algorithm. Jiang [85] in studying the continuous hidden Markov model, it was found that it has the structural deficiency of being difficult to handle multi-channel communication problems, and then proposed to optimize the model using a coupled hidden Markov model.

In addition to the above methods, Zhao et al. [28] combined random forest and principal component analysis models to establish a non-linear map** relationship between spindle motor current and tool wear under different machining conditions, effectively solving the sample imbalance problem. Wu et al. [24] proposed a random forest-based tool wear prediction model, which has better performance compared with feedforward neural networks and support vector regression. The structure of the random forest evaluator is shown in Fig. 9. Martinez et al. [86] proposed a new tool wear classification method based on signal imaging, deep learning and big data. By combining these two techniques, the method is able to process the raw data directly, avoiding the use of statistical pre-processing or filtering methods. He et al. [87] proposed a BP neural network and stacked sparse autoencoder model based on BP neural networks and used an improved loss function with sparse and weight penalty terms to enhance the robustness and generalization of the stacked sparse autoencoder model.

Fig. 9
figure 9

Random forest estimator structure [28]

Full size image

4.4 Summary

This section describes the process of tool wear monitoring from three aspects: data acquisition, signal processing and feature extraction, and pattern recognition. The direct measurement method is no longer suitable for the development trend because it cannot achieve real-time monitoring. This paper therefore focuses on indirect monitoring and systematically describes the commonly used monitoring signals, signal processing techniques and pattern recognition methods from the perspective of the monitoring process.

In the signal acquisition phase, commonly used signals include cutting force signals.

Host current and power signals, acoustic emission signals, sound signals, etc.. From the current research progress, the use of individual signals does not fully handle the complex situations that arise during processing. Therefore, multi-sensor fusion will become the future development direction.

The second describes the signal processing and feature extraction. The signal processing method is to extract the signal feature values by time domain, frequency domain and time–frequency domain methods. This part is the key to realizing online monitoring of tool wear, and the quality of the extracted feature values will directly affect the monitoring accuracy, so the feature extraction stage is the core of the whole monitoring process. In addition, in the process of pattern recognition, due to the complexity of the processing process, most of the current models are not quite ideal in terms of generalization ability and accuracy.

5 Tool RUL prediction

The tool RUL is defined as "the length from the current time to the end of the service life" [88], which can be identified as \({l}_{k}={t}_{Eol}-{t}_{k}\), where \({t}_{Eol}\) is the cut-off time, \({t}_{k}\) is the current time, \({l}_{k}\) is the RUL at the current time [89].

RUL prediction is the prediction of the time remaining before the machine loses its motion capacity based on condition monitoring signals [89], the prediction process is shown in the Fig. 10. Tool RUL can be divided into two categories: (1) RUL based on tool availability, (2) RUL based on machining quality judgement, either by the time taken from the tool not being started to be used until it is damaged, or by the time taken from the start of machining until the product machining quality deteriorates is. Depending on the classification method. Tool RUL predictions can be divided into those based on physical models and those based on data-driven predictions.

Fig. 10
figure 10

The RUL prediction process of the tool [90]

Full size image

5.1 RUL prediction based on physical models

Predicting tool RUL based on physical models is the process of develo** mathematical models through fault diagnosis or damage principles to describe the remaining useful life of machinery [91], the physical model parameters are related to material properties and stress levels, which are usually determined by using specific experiments, finite element analysis or other suitable techniques [89].

The Paris-Erdogan (PE) model is one of the most widely used models for mechanical RUL prediction and was first proposed by Paris et al. [92] was first proposed as a method to describe crack growth. As research has progressed, the model has been refined in a number of ways [93,94,95,96], Wang et al. [97] transformed the PE model into an empirical model for predicting mechanical RUL, which is usually expressed as a series of permanent ordinary differential equations or partial differential equations. Liao [98] and Sun et al. [99] enhanced the PE model into a state-space model and applied it to tool RUL prediction. In addition, Yang et al. [100] described the creep evolution of a turbine using Norton's law and combined the Kalman filter (KF) and particle filter (PF) to predict the tool RUL.

In addition to the physical models mentioned above, there are other methods based on physical models, such as the Archard model proposed by John F. Archard, which is used to predict future changes in tool wear values. The remaining tool life can be expressed as

$$W=\int K\frac{{P}^{a}{V}^{b}}{{H}^{c}}dt$$
(1)

where W: wear depth; P: interface pressure; V: relative sliding velocity; H: material hardness; time t, K, a, b and c are correction factors. This method is used to determine the remaining tool life by predicting the rate of tool wear over time.

Physical models are developed with a full understanding of the failure mechanism and a valid estimation of the model parameters, and are effective in predicting RUL, but in practice machining processes are often influenced by a variety of factors, such as shear stresses and thermal influences when cutting workpieces. Therefore, the physical model-based approach is only suitable for predicting RUL for simple mechanical systems or processes, and for complex machining conditions, the physical model-based approach to predicting RUL is subject to large errors.

5.2 Data-driven tool RUL prediction based on

The data-driven approach derives models from online and offline historical measurements from artificial intelligence techniques without the need for quantitative mathematical models. The data-driven approach does not need to take into account the physical processes of tool wear and extracts information related to tool life directly from sensor signals. Compared to methods based on physical models, data-driven methods are easier to collect data than building accurate physical models [97]. The main data-driven RUL prediction methods based on data are neural networks, support vector machines and deep learning. Figure 11 shows the tool wear prediction based on data-driven techniques.

BP neural networks have also been used in RUL prediction. Cong et al. [101] proposed to use the evolutionary thinking algorithm (MEA) to optimize the BP neural network for the random assignment of initial weights and thresholds in the traditional BP neural network. Zhang et al. [102] Wavelet analysis and Pearson correlation coefficient (PCC) were used to analyze the monitoring signals and select key features, and then a fuzzy neural network was used to achieve the RUL prediction.

In the support vector machine based method for predicting tool RUL, Yang et al. [103] proposed a least squares support vector machine based tool wear prediction model, which used the leave-one-out method to adjust the least squares support vector machine regularization factor and radial basis function kernel parameters to enhance the global search capability. Brezak et al. [104] proposed a tool wear prediction model based on two modules: classification and prediction. The prediction module consists of a support vector machine non-linear algorithm and the classification module is designed by a fuzzy logic concept, which can improve the robustness of the module by using fuzzy logic decisions without the limitation of the number of tool wear features. In response to the problem that traditional support vector machines are prone to complex algorithms and low efficiency, Shan et al. [105] used the MF-DFA method to noise reduce the acoustic emission signal and extract the features, combined with multiple fractal detrended fluctuation analysis and least squares support vector machine to construct a tool RUL prediction model, which has higher prediction accuracy compared with the traditional SVM performed.

Fig. 11
figure 11

Tool wear prediction process based on data-driven technology [106]

Full size image

With the development of deep learning technology, deep learning techniques have been gradually applied to the study of tool RUL prediction. Similar to neural networks, the principle of deep learning is also based on multiple hidden layers, which are gradually expressed from low latitude features to high latitude abstraction [107]. Guofeng et al. [108] based on the principles of noise reduction self-encoder (DAE) and hybrid trend particle filtering (HTPF), using trending features as particle filtering observations and parameter optimization through improved particle filtering algorithms. The process enables automatic selection of feature values related to tool wear, and this method of combining multiple state equations can be used to characterize tool wear trends with stochastic properties compared to a single state equation. Wang [109] extracted tool wear eigenvalues using time–frequency analysis and the principle of restricted Boltzmann machines, and implemented automatic assessment of tool wear and RUL based on machine vision, deep learning methods and improved long and short term memory networks. Zhang et al. [13, intelligent spindle integrates perception, decision making and execution, which has a significant impact on improving processing efficiency and quality. With the development of technology, the intelligent spindle will develop towards the direction of diversified diagnosis range, quantitative diagnosis form, high decision timeliness and integration with manufacturing big data. Combined with more advanced CNC machine tool system, will make CNC machine tool more intelligent, greatly improve the processing efficiency and processing quality.

Fig. 13
figure 13

aSmart spindle system structure and workflow; b Smart spindle prototype [116]

Full size image

(4) Application of data-driven algorithm in tool wear mechanism, online monitoring and RUL prediction under big data background.

On the basis of big data technology, the tool wear mechanism is fused with data-driven algorithm, and the judgment of tool wear mechanism is realized by using methods such as photographing and identifying pictures, which will save a lot of time and save complicated detection links in the research process of tool wear mechanism.

In addition, the use of big data platform of the advantages of large capacity data storage and deep learning algorithm to learn the characteristics of the inherent law of large amounts of data in the database, and then combined with the corresponding hardware and software, build a big data and deep learning technology based on data collection and analysis of the library, the tool wear monitoring system would make the process more intelligent and facilitation. Therefore, data-driven algorithms will have a wide application prospect.

With the arrival of THE 5G era, the application of 5G technology is becoming more and more mature and has been put into practical application (unmanned driving). By combining 5G technology and multi-sensor fusion technology, signals can be obtained comprehensively, analyzed and processed quickly and transmitted to the decision-making system, so as to achieve the purpose of rapid decision-making. The signal processing and feature extraction technology in the process of tool wear on-line monitoring and RUL prediction is combined with the decision method of unmanned driving technology to achieve fast, accurate and efficient tool wear processing.

In addition, 5G technology is featured with high speed, low latency and large connections. Rely on big data, establish tool wear database. The database requires the experimental data between different cutters-workpiece materials and fixtures; It is required to collect a wide range of data, which has strong robustness and updating ability. On this basis, a perfect tool wear monitoring and prediction training model is established. Through continuous training and updating in the application of the model, make it more perfect. Offline training model, online prediction and update.