A tool wear online monitoring method based on long short-term memory neural network-mechanism hybrid driving

Abstract

A tool wear online monitoring method based on long short-term memory neural network-mechanism hybrid driving is provided.The spindle vibration signal is collected, denoised, and divided into training set and test set; the long short-term memory neural network is used to construct a tool wear data-driven monitoring network; the training set data is used to train the data-driven monitoring network, the tool wear data-driven monitoring model is obtained, and the model related parameters are saved. Based on the tool wear law, a mechanism-driven tool wear state misclassification judgment model is constructed. The vibration signal collected in the actual machining process is input into the tool wear data-driven monitoring model after data preprocessing, the tool wear state corresponding to the signal is obtained, and the obtained continuous three tool states are input into the mechanism-driven misclassification judgment model as a group, and the final tool wear state is obtained. The method can effectively reduce the misclassification of the data-driven monitoring model caused by random interference of the monitoring data, and significantly improve the reliability of the monitoring result.

Description

Technical Field

[0001] This invention belongs to the field of tool condition monitoring technology, specifically a method for online monitoring of tool wear based on a hybrid mechanism driven by a long short-term memory neural network. Background Technology

[0002] In machining, tool condition is one of the key factors affecting part machining quality and production efficiency. Tool condition monitoring technology identifies tool breakage, damage, and wear by collecting and analyzing information related to tool condition, thereby ensuring stable and efficient production. Indirect tool condition monitoring methods can achieve real-time monitoring during the cutting process without stopping the machine or detecting during cutting intervals, making them easier to apply in actual production.

[0003] Machine learning, as a commonly used indirect monitoring method, has gained widespread attention from researchers by establishing a mapping relationship between dynamic monitoring signals and tool status to obtain tool status. In the patent "A Tool Wear Monitoring Method Based on Composite Current and Acoustic Emission Signals" (CN104723171B), the current signal of the spindle motor and the acoustic emission signal of the tool wear state during cutting are acquired. Wavelet packet analysis, correlation analysis, and principal component analysis are used to extract feature information of the tool wear state, and the degree of tool wear is determined by analyzing the correlation between the feature information and the initial state. In the patent "A Real-time Monitoring Method for CNC Machine Tool Milling Wear Based on Deep Convolutional Neural Networks" (CN202110976220.8), machine tool milling process status data and tool wear data are collected, a deep learning network is constructed to achieve accurate regression prediction of tool wear results, and a deep convolutional neural network is constructed to effectively identify the critical state of tool wear. The real-time tool wear is compared with the corresponding critical state threshold, and measures such as tool replacement or parameter changes are taken in a timely manner to achieve real-time monitoring of tool wear status. In the patent "A Tool Wear State Identification Method Based on Vibration and Acoustic Emission" (CN201811375637.3), vibration and acoustic emission signals during the machining process are collected in real time. After feature extraction in the time domain, frequency domain, and time-frequency domain, a BP neural network is used to identify the tool state and finally output the current tool wear status. In the patent "A Tool Wear State Monitoring Method Based on Machine Learning" (CN201811297325.5), cutting force and vibration signals during the machining process are collected as monitoring information for the tool state. After manual feature extraction and feature selection, a neural network model is used to perform regression analysis on the tool wear amount, and a particle swarm optimization algorithm is used to improve the predictive performance of the network.

[0004] In summary, existing tool wear condition monitoring methods still have some problems, such as: (1) the monitoring model has isolated misclassified samples, and the misclassified samples are relatively scattered; (2) they rely solely on data-driven models and have not achieved a hybrid driving of data and mechanism to obtain a highly robust monitoring model. This invention addresses the problems existing in current tool wear condition monitoring models by proposing a hybrid driving method for online tool wear condition monitoring based on a long short-term memory neural network and mechanism. Summary of the Invention

[0005] The purpose of this invention is to provide an online monitoring method for tool wear status driven by a hybrid long short-term memory neural network and mechanism, which solves the problems of poor robustness to data fluctuations and isolated misclassified samples in existing tool condition monitoring models, and realizes reliable monitoring of tool wear status.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A method for online tool wear monitoring based on a hybrid mechanism driven by a long short-term memory neural network is proposed. First, the spindle vibration signal during machine tool machining is selected as the monitoring signal for tool wear status, and a triaxial accelerometer is used to collect the data. Mean filtering and normalization are performed on the signal for noise reduction and data preprocessing. A labeled dataset is constructed using the preprocessed monitoring data and corresponding tool wear status labels, and divided into training and testing sets. Second, a tool wear mechanism discrimination model is constructed based on the distribution characteristics of the tool wear status labels. Then, an online tool wear monitoring model is constructed based on a long short-term memory neural network, and the model is trained using the training set data to optimize its performance. The relevant parameters of the model are then saved. Next, the real-time vibration signal from the actual machining process, after data preprocessing, is input into the saved monitoring model to obtain preliminary tool status labels for the corresponding monitoring signals. Finally, the obtained preliminary tool status labels are input in groups of three into a pre-constructed misclassification mechanism discrimination model based on tool wear patterns to obtain the final tool status labels. The specific steps are as follows:

[0008] The first step is the acquisition and preprocessing of monitoring signals during the processing.

[0009] (1) Monitoring signal acquisition

[0010] Vibration signals of the spindle in the X, Y, and Z directions during CNC machine tool machining are collected using a piezoelectric triaxial accelerometer. The tool wear is then observed to determine whether the tool wear has reached a critical value, and the tool condition is divided into healthy wear state and unhealthy wear state.

[0011] (2) Preprocessing of monitoring signals

[0012] First, in order to eliminate the adverse effects of environmental noise on the monitoring signal when the piezoelectric triaxial accelerometer collects data, the mean filtering algorithm of Equation (1) is used to denoise the monitoring signal; then, the denoised monitoring signal is normalized by Equation (2) so that the normalized data amplitude range is [0,1]. The normalized monitoring signal and the corresponding tool status label are used to construct a labeled dataset, and the labeled dataset is divided into a training set and a test set.

[0013]

[0014]

[0015] The second step is to construct a tool wear mechanism discrimination model.

[0016] Tool wear is a continuous and gradual process. The distribution of tool wear status labels strictly follows the time sequence. That is, a new tool wear status label starts from a healthy tool wear state until the tool wear value reaches a preset threshold T. After that, the tool wear status label changes to an unhealthy tool wear state, and this change happens only once. Based on the above rules, a tool wear mechanism discrimination model is constructed. The final tool wear status label at the intermediate time is determined according to the tool wear status labels at the adjacent time points, as shown in formula (3). If the tool wear status label at the intermediate time point is inconsistent with the tool wear status labels at the adjacent time points, the tool status label at the intermediate time point needs to be adjusted. If the tool wear status labels at the adjacent time points are different, the tool wear status label at the intermediate time point remains unchanged.

[0017]

[0018] In the formula, Label (k) Label is the tool status label at time k. (k-1) and Label (k+1) These represent the tool status labels at times k-1 and k+1, respectively.

[0019] The third step is the construction and training of a tool condition monitoring model based on a long short-term memory neural network.

[0020] Assuming the memory cells of an LSTM are updated once at each time step t, then the value of its input gate i t and memory cell candidate state values They are respectively:

[0021] i t =σ(W i ·[h t-1 ,x t ]+b i (4)

[0022]

[0023] Then the value of the forget gate f at time t t for:

[0024] f t =σ(W f ·[h t-1 ,x t ]+b f (6)

[0025] The current value C of the memory cell at time t is calculated from the values ​​of the input gate, the forget gate, and the candidate states of the memory cell. t :

[0026]

[0027] Finally, the memory unit output value h is obtained. t for:

[0028] h t =σ(W o [h t-1 ,x t ]+b o )*tanh(C t (8)

[0029] In the formula, x t is the input to the memory unit at time t; W is the model's weight parameters; b is the model's bias; σ and tanh are the Sigmoid activation function and the Tanh activation function, respectively.

[0030] Stack multiple LSTM networks to construct a deep LSTM network; train the deep LSTM network using the training set constructed in the first step, and test the performance of the trained deep LSTM network using the test set constructed in the first step; save the parameters of the optimal deep LSTM network as the parameters of the tool wear condition monitoring model.

[0031] The fourth step is to reliably adjust the tool monitoring results based on the tool wear mechanism discrimination model through self-supervision.

[0032] Input the monitoring data corresponding to the test set into the tool wear state monitoring model to obtain the preliminary tool state monitoring label; then input the preliminary tool state monitoring label into the tool wear mechanism discrimination model in groups of three, and realize the self-supervised reliable adjustment of the tool monitoring results based on formula (3) to obtain the final tool state label;

[0033] Step 5: Monitoring the wear condition of the cutting tool

[0034] In the actual machining process, the collected spindle vibration signals are first saved. When the data storage reaches the threshold of a single sample, the stored data is denoised and normalized, and then input into the constructed tool wear condition monitoring model to obtain the preliminary tool condition monitoring label. Then, the preliminary tool condition monitoring labels are input into the constructed tool wear mechanism discrimination model in groups of three to realize unsupervised and reliable adjustment of the tool monitoring results and obtain the final tool condition label.

[0035] The beneficial effects of this invention are: intelligent monitoring of tool condition can be achieved through this method, and unsupervised reliable adjustment of monitoring results driven by tool wear data can be realized based on the tool wear mechanism, thereby obtaining more reliable tool condition monitoring results.

[0036] Compared with the prior art, the advantages of this invention are:

[0037] (1) A tool wear state data-driven model was constructed using a long short-term memory neural network, which avoided human intervention and established the dependency relationship between adjacent sample monitoring data.

[0038] (2) Unsupervised reliable adjustment of monitoring results was achieved, and the constructed hybrid driving model has higher robustness. Attached Figure Description

[0039] Figure 1 This is a flowchart of the construction process for a hybrid LSTM and mechanism-driven tool wear monitoring model.

[0040] Figure 2 This is a flowchart of the tool wear mechanism discrimination model construction process.

[0041] Figure 3 This is an accuracy confusion matrix diagram of a tool wear monitoring model driven by LSTM data.

[0042] Figure 4 This is a confusion matrix diagram showing the accuracy of LSTM and mechanism-driven tool wear monitoring models. Detailed Implementation

[0043] To make the technical solution and beneficial effects of the present invention clearer, the present invention will be described in detail below with reference to the specific embodiments of milling tool condition monitoring and the accompanying drawings. This embodiment is based on the technical solution of the present invention and provides detailed implementation methods and specific operating procedures, but the scope of protection of the present invention is not limited to the following embodiments.

[0044] Taking the milling process of a domestically produced vertical machining center as an example, the implementation method of the present invention will be described in detail.

[0045] Cutting experiments were conducted on a three-axis vertical machining center using a vertical end mill. The basic information of the three-axis vertical machining center is as follows: maximum travel of the X, Y, and Z axes are 710 mm, 500 mm, and 350 mm, respectively; the maximum spindle speed is 15000 r / min. The basic information of the cutting tool is as follows: tool type is vertical end mill; tool material is carbide; tool diameter is 10 mm; number of cutting edges is 4. The basic information of the workpiece is as follows: workpiece material is 45# steel; workpiece shape is 200 mm x 100 mm x 10 mm.

[0046] The construction and training process of the hybrid drive tool wear condition monitoring model is as follows: Figure 1 As shown, the specific implementation method is as follows:

[0047] The first step is the acquisition and preprocessing of monitoring signals during the processing.

[0048] (1) Spindle vibration data acquisition during machining process

[0049] A cutting test was conducted on the aforementioned three-axis vertical machining center. The three-axis accelerometer was installed on the spindle near the tool holder. Spindle vibration data in the X, Y, and Z directions during machining were collected at a sampling frequency of 1000Hz and saved.

[0050] (2) Vibration data preprocessing

[0051] First, to eliminate the adverse effects of environmental noise on the monitoring signal during data acquisition, the mean filtering algorithm of Equation (1) is used to denoise the dynamic signal, where N is 5. Then, the denoised dynamic signal is normalized using Equation (2) so that the normalized data amplitude range is [0,1]. Second, a dedicated measuring device is used to monitor tool wear. Tool wear is detected every 100mm of cutting distance, and the tool wear state is divided into healthy and unhealthy states based on whether the tool wear value reaches 0.3mm. Third, the collected cutting process vibration data is divided into a sample monitoring data segment of 250 data points, and labeled with the corresponding tool status label, containing a total of 410 labeled samples. Finally, the labeled sample set is randomly divided into a training set and a test set in a 7:3 ratio, with 287 and 123 samples in the training set and test set, respectively.

[0052] The second step is to construct a tool wear mechanism discrimination model.

[0053] The tool wear mechanism discrimination model determines whether the tool monitoring result at time t is correct based on the tool condition monitoring results at times (t-1) and (t+1), and makes corresponding adjustments.

[0054] The third step involves constructing and training a tool condition monitoring model based on a long short-term memory neural network.

[0055] A tool wear monitoring model was constructed based on a Long Short-Term Memory (LSTM) neural network. This model consists of two LSM layers and one BP network layer. The model was trained using training data and tested using test data. The hyperparameters were adjusted based on the test results until optimal performance was achieved, and the model was then saved. The final model was determined to have 250, 50, and 2 neurons per layer; a learning rate of 0.01; and 5000 iterations. Testing showed that the accuracy of the monitoring model based on the LSM neural network was 95.9%.

[0056] A deep LSTM network is constructed by stacking multiple LSTM networks. The deep LSTM network is trained using training data and tested using test data. The parameters of the optimal deep LSTM network are saved as the parameters of the tool wear condition monitoring model.

[0057] The fourth step is to reliably adjust the tool monitoring results based on the mechanism discrimination model without supervision.

[0058] A total of 123 vibration data segments corresponding to the test set were input into the constructed tool wear condition monitoring model to obtain preliminary tool condition monitoring labels. Then, these preliminary tool condition monitoring labels were input in groups of three into a constructed mechanism discrimination model to achieve unsupervised and reliable adjustment of the tool monitoring results, obtaining the final tool condition labels. The final results show that after unsupervised adjustment by the mechanism discrimination model, the accuracy rate of the tool wear monitoring results is 97.6%, which is higher than the accuracy rate of tool wear monitoring based solely on long short-term memory networks.

[0059] Step 5: Monitoring the wear condition of the cutting tool

[0060] In the actual machining process, the collected spindle vibration signals are first saved. When the data storage reaches the threshold of a single sample, the stored data is denoised and normalized. The data is then input into the constructed tool wear condition monitoring model to obtain the preliminary tool condition monitoring label. Then, the preliminary tool condition monitoring labels are input into the constructed mechanism discrimination model in groups of three to achieve unsupervised and reliable adjustment of the tool monitoring results, and finally obtain the tool condition label.

Claims

1. A method for online monitoring of tool wear based on a hybrid long short-term memory neural network-mechanism drive, characterized in that, The steps are as follows: The first step is the acquisition and preprocessing of monitoring signals during the processing. (1) Monitoring signal acquisition Vibration signals of the spindle in the X, Y, and Z directions during CNC machine tool machining are collected using a piezoelectric triaxial accelerometer. The tool wear is then observed to determine whether the tool wear has reached a critical value, and the tool condition is divided into healthy wear state and unhealthy wear state. (2) Preprocessing of monitoring signals First, in order to eliminate the adverse effects of environmental noise on the monitoring signal when the piezoelectric triaxial accelerometer collects data, the mean filtering algorithm of Equation (1) is used to denoise the monitoring signal; then, the denoised monitoring signal is normalized by Equation (2) so that the normalized data amplitude range is [0,1]. The normalized monitoring signal and the corresponding tool status label are used to construct a labeled dataset, and the labeled dataset is divided into a training set and a test set. (1) (2) in, ; The second step is to construct a tool wear mechanism discrimination model. Tool wear is a continuous and gradual process. The distribution of tool wear status labels strictly follows the time sequence. That is, a new tool wear status label starts from a healthy tool wear state until the tool wear value reaches a preset threshold T. After that, the tool wear status label changes to an unhealthy tool wear state, and this change happens only once. Based on the above rules, a tool wear mechanism discrimination model is constructed. The final tool wear status label at the intermediate time is determined according to the tool wear status labels at the previous and next times, as shown in formula (3). If the tool wear status label at the intermediate time is inconsistent with the tool wear status labels at the previous and next times, the tool status label at the intermediate time needs to be adjusted. If the tool wear status labels at the previous and next times are different, the tool wear status label at the intermediate time remains unchanged. （3） In the formula, Let k be the tool status label at time k. and These represent the tool status labels at times k-1 and k+1, respectively. The third step is the construction and training of a tool condition monitoring model based on a long short-term memory neural network. Assuming the memory cells of an LSTM are updated once at each time step t, then the value of its input gate... and memory cell candidate state values They are respectively: （4） （5） Then the value of the forget gate at time t for: （6） The current value of the memory cell at time t is calculated from the values ​​of the input gate, the forget gate, and the candidate states of the memory cell. : （7） Finally, the memory unit output value is obtained. for: （8） In the formula, is the input to the memory unit at time t; W is the model's weight parameters; b is the model's bias; σ and tanh are the Sigmoid activation function and the Tanh activation function, respectively. Stack multiple LSTM networks to construct a deep LSTM network; train the deep LSTM network using the training set constructed in the first step, and test the performance of the trained deep LSTM network using the test set constructed in the first step; save the parameters of the optimal deep LSTM network as the parameters of the tool wear condition monitoring model. The fourth step is to reliably adjust the tool monitoring results based on the tool wear mechanism discrimination model through self-supervision. Input the monitoring data corresponding to the test set into the tool wear state monitoring model to obtain the preliminary tool state monitoring label; then input the preliminary tool state monitoring label into the tool wear mechanism discrimination model in groups of three, and realize the self-supervised reliable adjustment of the tool monitoring results based on formula (3) to obtain the final tool state label; Step 5: Monitoring the wear condition of the cutting tool In the actual machining process, the collected spindle vibration signals are first saved. When the data storage reaches the threshold of a single sample, the stored data is denoised and normalized, and then input into the constructed tool wear condition monitoring model to obtain the preliminary tool condition monitoring label. Then, the preliminary tool condition monitoring labels are input into the constructed tool wear mechanism discrimination model in groups of three to realize unsupervised and reliable adjustment of the tool monitoring results and obtain the final tool condition label.

Images