A tool wear online monitoring method based on a working condition adaptive implicit semi-Markov model

Abstract

The present application relates to a kind of tool wear online monitoring method based on working condition adaptive implicit semi-Markov model, belong to intelligent monitoring technical field of machining process.First, the sound, acoustic emission and spindle three-way acceleration signal generated in machining process are collected, and features are extracted from its time domain, frequency domain and time-frequency domain respectively for fusion.On this basis, considering the difference between tool life distribution and signal feature distribution under different cutting conditions, working condition adaptive implicit semi-Markov model is established to describe the wear process of tool.Further, tool life data and multi-source signal feature data obtained by orthogonal experiment are used to train the model established;Finally, the wear state and remaining useful life of tool can be evaluated online using the model.The present application has strong working condition adaptability, and can accurately evaluate the wear state and remaining useful life of tool under time-varying milling condition.

Description

Technical Field

[0001] This invention relates to an online tool wear monitoring method based on a working condition adaptive implicit semi-Markov model, belonging to the field of intelligent monitoring technology for machining processes. Background Technology

[0002] Titanium alloys, high-temperature alloys, and other difficult-to-machine materials have poor thermal conductivity and severe work hardening, leading to easy tool wear during milling. This severely affects workpiece machining quality and production efficiency, and can even damage machine tools. Therefore, establishing an accurate online tool wear monitoring system that adapts to different cutting conditions is crucial for achieving intelligent machining. It is of great significance for reducing downtime, ensuring machining quality, and lowering production costs.

[0003] Currently, tool condition monitoring is mainly divided into two categories: direct methods and indirect methods. Direct methods use instruments such as microscopes, industrial cameras, and infrared measuring instruments to accurately measure tool wear; however, this requires interrupting machine processing and is significantly affected by chips and cutting fluid, making real-time online monitoring impossible. Indirect methods, on the other hand, typically collect multi-source signals generated during machining, such as cutting forces, vibrations, sound, and acoustic emissions, and establish a mapping relationship between signal characteristics and tool wear to monitor tool wear. This approach offers advantages such as low cost, high real-time performance, and less susceptibility to the machining environment. The patent "A tool wear monitoring method based on multi-sensor information fusion and deep belief network, CN110000610A" identifies tool wear status by collecting triaxial cutting force signals and vibration signals during the cutting process and using a deep belief network. The literature "Online tool wear monitoring via hidden semi-Markov model with dependent durations[J].IEEE Transactions on Industrial Informatics,2017,14(1):69-78" suggests that the dwell time of adjacent tool wear states follows a normal distribution and is correlated, and establishes a dwell time-related hidden semi-Markov model to describe the tool wear process, further improving the accuracy of tool wear monitoring. However, existing methods ignore the differences between tool life distribution and signal feature distribution under different cutting conditions, resulting in poor model versatility and applicability only to specific cutting conditions. Currently, condition-adaptive hidden semi-Markov models that consider the influence of cutting parameters have not been applied in online tool wear monitoring. Summary of the Invention

[0004] To address the shortcomings of existing tool wear monitoring methods, this invention considers the differences in tool life distribution and signal characteristic distribution under different cutting conditions, and proposes an online tool wear monitoring method based on a condition-adaptive implicit semi-Markov model. Compared with existing methods, this method has the advantages of strong versatility, high monitoring accuracy, and applicability to time-varying milling conditions.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] First, sound, acoustic emission, and spindle three-dimensional acceleration signals generated during machining are collected, and features are extracted and fused from their time, frequency, and time-frequency domains, respectively. Based on this, considering the differences in tool life distribution and signal feature distribution under different cutting conditions, a condition-adaptive implicit semi-Markov model is established to describe the tool wear process. Then, the established model is trained using tool life data obtained from orthogonal experiments and multi-source signal feature data. Finally, this model can be used to evaluate the tool wear state and remaining service life online. The process includes the following steps:

[0007] Step 1. Multi-source signal acquisition and feature fusion. Acquire sound signals, acoustic emission signals, and triaxial acceleration signals of the spindle during the processing, and extract features from their time domain, frequency domain, and time-frequency domain, respectively. Then, use kernel principal component analysis to fuse the extracted multi-source signal features.

[0008] Step 2. Construct a condition-adaptive implicit semi-Markov model. Since tool wear is asymptotic and irreversible, based on the left-right implicit semi-Markov model, considering the differences in tool life distribution and signal characteristic distribution under different cutting conditions, a condition-adaptive implicit semi-Markov model is constructed to describe the tool wear process. This model has N hidden states, denoted as {S1, S2, ..., S...}. N}

[0009] Considering the influence of cutting parameters and state dwell time on the probability of tool wear state transition, we define the cutting state t-th cut as already in state S. i Under the condition of dwell time d, the tool moves from state S i Transition to state S j The probability is:

[0010] a ij (t,d)=P(q t+1 =S j |q t =S i ,d i =d,C t (1)

[0011] In the formula: C t =[V t ,Ft ,Ae t ], where V t F t Ae t Let q represent the cutting speed, feed per tooth, and radial depth of cut at the t-th cut, respectively. t d represents the wear state of the tool during the t-th cut. i For already in state S i The length of stay.

[0012] Based on the evolution law of tool wear, Taylor's empirical formula for tool life is embedded into the accelerated failure time model to describe the influence of cutting parameters and state dwell time on the wear state degradation, and the time-varying state transition probability is calculated.

[0013] Furthermore, considering that Gaussian mixture models trained under specific working conditions may not accurately reflect the signal feature distribution under other cutting parameters, an online unsupervised transfer learning algorithm based on the moment matching method is proposed. Based on this algorithm, the model trained using source domain data can be transferred to any working condition, thereby determining the observation probabilities related to the cutting parameters.

[0014]

[0015] In the formula: O t Here, K represents the signal features extracted during the t-th cut, and K is the number of source domain samples. To train the state S using the kth set of source domain sample data i Gaussian mixture model of observed probability distribution, η k Let be the weight of the k-th source domain model, where Together they form the weight vector η = [η1, η2, ... η K ].

[0016] Assume that the weight vector η follows the parameter η during the t-th cut. The Dirichlet distribution is denoted as η~Dir(η; θ) t The weight vector η is updated in real time based on the moment matching method to determine the observation probability b. i (O t ).

[0017] Step 3. Conduct orthogonal experiments and train the established model using the acquired experimental data. Design orthogonal experiments as the training group, collect multi-source sensor signals as in Step 1, extract multi-domain features and fuse them; measure the flank wear value, classify the tool wear state according to the wear threshold, and thus determine the sample data label.

[0018] Furthermore, using the duration data of each tool state, the parameters of the accelerated failure time model are estimated using the maximum likelihood method. In addition, Gaussian mixture models are trained using the EM algorithm based on the multi-source signal feature data acquired under each working condition.

[0019] Step 4. Online assessment of tool wear status and remaining service life. During online monitoring, firstly, multi-source sensor signals generated during machining are collected according to Step 1, multi-domain features are extracted and fused; then, using the model trained in Step 3, the state transition probability and observation probability are calculated online according to Step 2; finally, the posterior probability of the current tool wear status and remaining service life with respect to the observation sequence is calculated using an improved forward algorithm.

[0020] To eliminate the computational underflow problem in traditional implicit semi-Markov models, the improved forward variable is defined as:

[0021] α t (i,d)=P(q t =S i ,d i =d|O 1:t ,λ 1:t (3)

[0022] In the formula: O 1:t and λ 1:t These represent the observation sequence from the first cut to the t-th cut and the parameters of the established model, respectively.

[0023] Furthermore, at the t-th cutting operation, the posterior probabilities of the tool wear state and the remaining tool life are calculated using equations (4) and (5), respectively:

[0024]

[0025] P(RUL=rul|O 1:t ,λ 1:t )=P(q t+r =S3,q t+r-1 =S2|O 1:t ,λ 1:t (5)

[0026] In the formula: rul=τ t+1 +τ t+2 +…τ m …+τ t+r , where τ m Let m be the machining time for the m-th cut.

[0027] Finally, the current wear state of the tool is obtained through equation (6), and the posterior probability P(RUL|O) is calculated. 1:t ,λ 1:tThe expected value of the mathematical expression is the predicted value of the remaining service life of the tool.

[0028]

[0029] The beneficial effects of this invention are as follows: This invention addresses online monitoring of tool wear under varying milling conditions. For the first time, it considers the differences in tool life distribution and signal characteristic distribution under different cutting conditions within an implicit semi-Markov model, establishing a condition-adaptive implicit semi-Markov model to describe the tool wear process. Compared with existing methods, this invention has strong adaptability to different operating conditions and can accurately assess the tool wear state and remaining service life under time-varying milling conditions. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a schematic diagram of sensor installation.

[0032] Figure 3 This is a schematic diagram of an adaptive implicit semi-Markov model for operating conditions.

[0033] Figure 4 This is a schematic diagram of an online unsupervised transfer learning algorithm.

[0034] Figure 5 This is a description of the tool wear status monitoring under time-varying milling conditions using the method of the present invention.

[0035] Figure 6 This invention provides a method for predicting the remaining tool life under time-varying milling conditions. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0037] A method for online monitoring of tool wear based on a condition-adaptive implicit semi-Markov model is described below. Figure 1 As shown. First, sound, acoustic emission, and spindle three-dimensional acceleration signals generated during machining are collected, and features are extracted and fused from their time domain, frequency domain, and time-frequency domain, respectively. Based on this, considering the differences in tool life distribution and signal feature distribution under different cutting conditions, a condition-adaptive implicit semi-Markov model is established to describe the tool wear process. Then, the established model is trained using tool life data obtained from orthogonal experiments and multi-source signal feature data. Finally, this model can be used to evaluate the tool wear state and remaining service life online. The process includes the following steps:

[0038] Step 1. Multi-source signal acquisition and feature fusion. Acquire sound signals, acoustic emission signals, and three-dimensional acceleration signals of the spindle during the processing. A sensor installation diagram is shown below. Figure 2 As shown.

[0039] 1.1 Time-domain features are extracted from the original signal, including mean, standard deviation, root mean square, peak-to-peak value, skewness, kurtosis, peak factor, margin factor, impulse factor, and waveform factor. Frequency-domain features are extracted from the signal power spectrum, including power spectrum mean, centroid frequency, frequency mean square, and frequency standard deviation. Furthermore, a three-level wavelet packet decomposition is performed on the original signal using db8 as the wavelet basis function. Time-frequency features, including wavelet energy ratio and wavelet energy entropy, are extracted from each sub-band of the third-level wavelet packet decomposition.

[0040] 1.2 Following step 1.1, a total of 115-dimensional feature vectors were extracted from the collected signals. Kernel principal component analysis was used to select principal components with a cumulative contribution rate greater than 80% as the fused features.

[0041] Step 2. Construct a working condition adaptive implicit semi-Markov model. Since tool wear is asymptotic and irreversible, a working condition adaptive implicit semi-Markov model is constructed based on the left-right implicit semi-Markov model to describe the tool wear process. This model has N hidden states, denoted as {S1, S2, ..., S...}. N}.like Figure 3 As shown, in this embodiment, the number of hidden states N = 3, and S1, S2, and S3 represent the initial wear state, normal wear state, and severe wear state, respectively.

[0042] 2.1 Considering the influence of cutting parameters and state dwell time on the probability of tool wear state transition, we define the state transition probability as follows: the cutting operation at time t is already in state S. i Under the condition of dwell time d, the tool moves from state S i Transition to state S j The probability is:

[0043] a ij (t,d)=P(q t+1 =S j |q t =S i ,d i =d,C t (1)

[0044] In the formula: C t =[V t ,F t ,Ae t ], where V t F t Ae tLet q represent the cutting speed, feed per tooth, and radial depth of cut at the t-th cut, respectively. t d represents the wear state of the tool during the t-th cut. i For already in state S i The length of stay.

[0045] The established model is a left-right implicit semi-Markov model, and its state transition probabilities satisfy: And when j i+1, a ij (t,d)=0.

[0046] Based on the characteristics of tool wear, Taylor's empirical formula for tool life can be expressed as:

[0047] L=σV ω F ρ Ae ξ (2)

[0048] In the formula: L is the tool life, V, F, and Ae represent the cutting speed, feed per tooth, and radial depth of cut, respectively, and σ, ω, ρ, and ξ are constants obtained from experimental data.

[0049] Based on equation (2), an accelerated failure time model is established to describe the influence of cutting parameters and state dwell time on wear state degradation, and the state S is characterized by... i The risk function of the accelerated failure time model for the degradation process (i = 1, 2) is expressed as:

[0050]

[0051] In the formula: c i ,β i , For characterizing state S i Parameters of the accelerated failure time model for the degradation process.

[0052] Assume the tool is in state S i The maximum duration under is D i Reliability function R i (t,d)=P(D i >d) indicates that the tool is at least in state S during the t-th cut. i The probability of staying for a duration of d is calculated by the following formula:

[0053]

[0054] The t-th cut is already in state S. i Given a dwell time of d, the probability that the wear state remains unchanged is:

[0055]

[0056] In the formula: τ t Let t be the machining time for the t-th cut.

[0057] Furthermore, from state S i Transition to the next state S i+1 The probability is calculated using the following formula:

[0058] a i(i+1) (t,d)=1-a ii (t,d) (6)

[0059] 2.2 Considering that Gaussian mixture models trained under specific working conditions are difficult to reflect the signal feature distribution under other cutting parameters, an online unsupervised transfer learning algorithm based on the moment matching method is proposed, such as... Figure 4 As shown. Based on this algorithm, the model trained using source domain data can be transferred to any working condition, thereby determining the observation probabilities related to cutting parameters:

[0060]

[0061] In the formula: O t Here, K represents the signal features extracted during the t-th cut, and K is the number of source domain samples. To train the state S using the kth set of source domain sample data i Gaussian mixture model of observed probability distribution, η k Let be the weight of the k-th source domain model, where Together they form the weight vector η = [η1, η2, ... η K ].

[0062] Assume that the weight vector η follows the parameter η during the t-th cut. The Dirichlet distribution is denoted as η~Dir(η; θ) t Initially, each source domain model is assigned the same weights, let in The initial distribution of η can be expressed as:

[0063] P0(η)=Dir(η;θ0) (8)

[0064] Furthermore, the weight vector η with respect to the observation sequence O 1:t The posterior probability is expressed as:

[0065] P t (η)=P(η|O 1:t (9)

[0066] According to Bayes' theorem, equation (9) is calculated as follows:

[0067]

[0068] In the formula: P(O) t |η,q t =S j C t ) represents the state S of the cutting tool during the t-th cutting operation. i The observed value is O t The probability, P i (t-1,d) represents the state of the tool in state S during the (t-1)th cut. i The probability of staying for a time d is calculated using the following formula:

[0069] P i (t-1,d)=P(d<D i <d+τ t ) = R i (t-1,d)-R i (t-1,d+τ t (11)

[0070] P(q t-1 =S i ) represents the state S of the tool during the (t-1)th cut. i The probability of this can be calculated using the following formula:

[0071]

[0072] Substituting equations (7), (11), and (12) into equation (10), we get:

[0073]

[0074] Since the Dirichlet distribution is the conjugate prior of the multinomial distribution, it has the following properties:

[0075]

[0076] In the formula: in,

[0077] Substituting equation (14) into equation (13) and simplifying, we get:

[0078]

[0079] In the formula: The normalization coefficients, z(i,j,k), are obtained by the following formula:

[0080]

[0081] Based on moment matching theory, an exponential distribution family is adopted. As Pt The approximation of (η) is obtained by matching P. t (η) and The sufficient moments determine the parameter θ t P t (η) and Regarding the model weight η of the r-th source domain r The sufficient moments are calculated by equations (17) and (18), respectively:

[0082]

[0083]

[0084] make Find the parameters for:

[0085]

[0086] Furthermore, the online update of the weight vector η is:

[0087]

[0088] Substituting the updated weight vector η into equation (7) yields the observation probability b at the t-th cut. i (O t ).

[0089] Step 3. Conduct orthogonal experiments and use the acquired experimental data to train the established model. Ti6Al4V titanium alloy material was side-milled using a single-flute insert milling cutter, with a cutting length of 100mm per pass. A 3-factor, 3-level orthogonal experiment was designed as the training group, as shown in Table 1.

[0090] Table 1 Orthogonal experimental parameters for the training group

[0091]

[0092] 3.1 Collect multi-source sensor signals generated during the processing according to step 1, extract multi-domain features from them and fuse them.

[0093] 3.2 After each cut, the cutting tool was removed and the flank wear was measured. The maximum wear width of the flank was used as an indicator to classify the tool wear state. A wear value of 0-150μm was considered the initial wear state, a wear value of 150-400μm was considered the normal wear state, and a wear value greater than 400μm was considered the severe wear state, thus determining the sample labels for the dataset.

[0094] 3.3 Using the duration data of each tool state, the parameters of the accelerated failure time model were estimated using the maximum likelihood method, and the results are shown in Table 2. Furthermore, using the multi-source signal feature data acquired under each working condition, nine Gaussian mixture models were trained using the EM algorithm.

[0095] Table 2 Estimated values ​​of accelerated failure time model parameters

[0096]

[0097]

[0098] Step 4. Online assessment of tool wear state and remaining service life. Test experiments were conducted according to the experimental parameters designed in Table 3. During online monitoring, firstly, multi-source sensor signals generated during machining were collected according to Step 1, and multi-domain features were extracted and fused. Then, using the model trained in Step 3, the state transition probability and observation probability were calculated online according to Step 2. Finally, the posterior probability of the current tool wear state and remaining service life with respect to the observation sequence was calculated using an improved forward algorithm.

[0099] Table 3. Experimental cutting parameters for the test group

[0100]

[0101] 4.1 To eliminate the computational underflow problem in traditional implicit semi-Markov models, the improved forward variable is defined as:

[0102] α t (i,d)=P(q t =S i ,d i =d|O 1:t ,λ 1:t ) (twenty one)

[0103] In the formula: O 1:t and λ 1:t These represent the signal feature observation sequence from the first cut to the t-th cut and the parameters of the established model, respectively.

[0104] Define auxiliary variable J t (i,d)=P(q t =S i |O 1:t-1 ,λ 1:t-1 Therefore, equation (21) can be calculated as:

[0105]

[0106] Auxiliary variable J t (i,d) are recursively calculated using the following formula:

[0107]

[0108] 4.2 At the t-th cutting operation, the posterior probabilities of the tool wear state and the remaining tool life are calculated by equations (24) and (25), respectively:

[0109]

[0110]

[0111] In the formula: rul=τ t+1 +τ t+2 +…+τ t+r .

[0112] Since when T > t, the observation sequence O t+1:T Since the context is unknown, the forward variable for T > t is defined as:

[0113]

[0114] Substituting equation (25) into equation (24) yields the current observation sequence O. 1:t The posterior probability of the remaining tool life.

[0115] 4.3 Finally, the current wear state of the tool is obtained through equation (27), and the posterior probability P(RUL|O) is calculated. 1:t ,λ 1:t The expected value of the mathematical expression is the predicted value of the remaining service life of the tool.

[0116]

[0117] Figure 5 The accuracy rate of the tool wear monitoring results during the test experiment was 89.92%. Figure 6 To test the prediction results of the remaining tool life in the experiment, the average prediction error was 2.32 min. These results demonstrate that the method of the present invention has strong adaptability to different working conditions and can accurately assess the wear state and remaining tool life under time-varying milling conditions.

Claims

1. A method for online monitoring of tool wear based on an adaptive implicit semi-Markov model, characterized in that, Includes the following steps: Step 1. Multi-source signal acquisition and feature fusion: Acquire sound signals, acoustic emission signals and spindle triaxial acceleration signals during the processing, and extract features from their time domain, frequency domain and time-frequency domain respectively. Then, use kernel principal component analysis to fuse the extracted multi-source signal features. Step 2. Constructing an Adaptive Implicit Semi-Markov Model: Based on the left-right type implicit semi-Markov model, considering the differences in tool life distribution and signal characteristic distribution under different cutting conditions, an adaptive implicit semi-Markov model is constructed to describe the tool wear process. This model has the following characteristics: There are three hidden states, denoted as follows: ; Considering the influence of cutting parameters and state dwell time on the probability of tool wear state transition, we define the state transition after the t-th cut as follows: Under the condition of dwell time d, the tool moves from state d. Transition to state The probability is: （1） In the formula: ,in , , Let these represent the cutting speed, feed per tooth, and radial depth of cut at the t-th cut, respectively. This indicates the wear state of the tool during the t-th cut. already in the state Duration of stay; Based on the evolution law of tool wear, Taylor's empirical formula for tool life is embedded into the accelerated failure time model to describe the influence of cutting parameters and state dwell time on wear state degradation, and the time-varying state transition probability is calculated. Considering that Gaussian mixture models trained under specific working conditions are difficult to reflect the signal feature distribution under other cutting parameters, an online unsupervised transfer learning algorithm based on the moment matching method is proposed. Based on this algorithm, the model trained using source domain data can be transferred to any working condition, thereby determining the observation probabilities related to cutting parameters: （2） In the formula: Here, K represents the signal features extracted during the t-th cut, and K is the number of source domain samples. To train using the k-th group of source domain sample data to describe the state Gaussian mixture model of observed probability distribution, Let be the weight of the k-th source domain model, where Together they form the weight vector ; Let the weight vector be at the t-th cutting time. Obtain the parameter as The Dirichlet distribution, denoted as ; Real-time updating of weight vector based on moment matching method Determine the observation probability ; Step 3. Conduct orthogonal experiments and use the acquired experimental data to train the model established in Step 2: Design orthogonal experiments as training groups, collect multi-source sensor signals according to Step 1, extract multi-domain features and fuse them; measure the wear value of the back face, classify the tool wear state according to the wear threshold, and thus determine the sample data label; Using the duration data of each tool state, the parameters of the accelerated failure time model are estimated by the maximum likelihood method; in addition, using the multi-source signal feature data obtained under each working condition, Gaussian mixture models are trained by the EM algorithm. Step 4. Online assessment of tool wear condition and remaining service life: During online monitoring, firstly, multi-source sensor signals generated during machining are collected according to Step 1, and multi-domain features are extracted and fused from them; Then, using the model trained in step 3, the state transition probability and observation probability are calculated online according to step 2; finally, the posterior probability of the current tool wear state and remaining service life with respect to the observation sequence is calculated using the improved forward algorithm. To eliminate the computational underflow problem in traditional implicit semi-Markov models, the improved forward variable is defined as: （3） In the formula: and These are represented as the observation sequence from the first cut to the t-th cut and the parameters of the established model, respectively. At the t-th cutting operation, the posterior probabilities of the tool wear state and the remaining tool service life are calculated by equations (4) and (5), respectively: （4） （5） In the formula: ,in Let m be the machining time for the m-th cut; Finally, the current wear state of the tool is obtained through equation (6), and the posterior probability is calculated. The expected value of the result is the predicted value of the remaining tool life: （6）。

Images