A Milling Tool Wear Detection Method Based on Matching Wavelet Depth Transfer Learning

Abstract

This application provides a milling tool wear detection method based on matched wavelet deep transfer learning. The method includes obtaining original machining signals from the milling process on a machine tool using vibration, acoustic emission, and power sensors; extracting stable cutting process signals from the original machining signals; and then using Morse continuous wavelet to match and upscale the one-dimensional fault signal to capture subtle changes and obtain a visualized enhanced feature image. Secondly, the source domain model of the deep transfer network is effectively transferred. This model has efficient image learning experience, which can reduce the number of training samples in the target domain. Finally, the process parameters are optimized based on limited data during model transfer. This method has strong generalization ability, can detect and locate subtle features under multiple working conditions, has a strong ability to identify subtle vibrations that occur during machining due to tool wear, and effectively reduces dependence on data, greatly improving computational speed and diagnostic accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical cutting tool wear detection, and particularly relates to a milling tool wear detection method based on matched wavelet deep transfer learning. BACKGROUND

[0002] With the development of intelligent manufacturing, intelligent cutting is a powerful guarantee for intelligent manufacturing, and realizing intelligent cutting is the key to improving machining efficiency and machining quality. Accurate monitoring of the machining process is the premise of realizing intelligent machining. As the tool directly involved in cutting, the wear of the milling tool is an inevitable problem in the milling process. The worn tool can easily lead to problems such as a decrease in the quality of the machined surface, a decrease in machining precision, a decrease in tool life, and the like, which seriously affect the machining efficiency and cost.

[0003] In order to detect the wear of the milling tool in a timely manner, many scholars and engineers have conducted a large amount of research and exploration. Traditional tool wear detection methods mainly include visual methods, acoustic methods, vibration methods, and the like. Although these methods are simple and easy to implement, they have problems such as low detection precision and poor reliability.

[0004] In recent years, the rise of machine learning technology has brought new ideas for milling tool wear detection, and tool monitoring models constructed by artificial neural networks, support vector machines, and hidden Markov models have gradually been applied to actual machining. However, the micro-fault signals obtained in real device environments have characteristics such as difficulty in data acquisition, low feature amplitude, and susceptibility to external noise disturbance. Therefore, under the condition of limited data, how to efficiently optimize the diagnosis process of the micro-fault signals is a difficult problem to be solved.

[0005] Therefore, the accurate identification and monitoring of tool wear states by using the organic fusion of information from multiple sensors is a key technology and a key research direction of intelligent machining. SUMMARY

[0006] In view of the actual problems and the deficiencies of the prior art, the main technical problem to be solved by the application is to provide a milling tool wear detection method based on matched wavelet deep transfer learning, which can better identify micro-vibration faults in the early stage of wear. These micro-vibration faults actually represent the development trend and deterioration tendency of tool wear. Accurate diagnosis of micro-faults can effectively improve system reliability and reduce the risk of accidents, thereby ensuring the safe operation of the machine tool.

[0007] To solve the above technical problems, the application provides a milling tool wear detection method based on matched wavelet deep transfer learning, which adopts the following technical scheme:

[0008] The application discloses a milling tool wear detection method based on matched wavelet depth migration learning, and relates to the technical field of milling tool wear detection.

[0009] The milling machine further comprises a power instrument acquisition machine.

[0010] Step A: after setting cutting parameters, single-cutting-parameter milling experiments or multi-cutting-parameter milling experiments are carried out;

[0011] Step B: signal acquisition sample selection, signal samples of the vibration sensor, the acoustic emission sensor and the power acquisition machine are extracted;

[0012] Step C: the extracted signal samples are subjected to signal pretreatment and Morse continuous wavelet transformation to obtain a visualized reinforced feature image;

[0013] Step D: taking a deep migration network model as a migration object, the visualized reinforced feature image is input into a model input layer to construct a migration input sample; the model is migrated from a source domain by sequentially utilizing an ImageNet database and limited target samples, migration learning is completed, and a new signal is formed;

[0014] Step F: the new signal data are classified or predicted by using the DTN model, and a prediction evaluation report is output.

[0015] In a preferred embodiment, in step B, the signal acquisition sample selection comprises obtaining original machining signals, and the original machining signals are obtained by collecting vibration, acoustic emission and cutting force data in multiple same machining processes.

[0016] In a preferred embodiment, in step B, for the original machining signal acquisition, the power acquisition machine collects a total power signal in the single-cutting-parameter milling experiment, and collects a spindle total power signal in the multi-cutting-parameter milling experiment.

[0017] In a preferred embodiment, the signal acquisition sample selection step is as follows:

[0018] Step a: the wear of a tool after single machining is measured and recorded by using an electron microscope, and the wear value is the maximum wear value of a tool flank face;

[0019] Step b: the single-machining data obtained are one-to-one corresponding to tool wear values according to machining times, and each tool wear value corresponds to multiple tool path processes;

[0020] Step c; extract vibration signal samples and acoustic emission signals, first select 3 times of walking tool full process signals corresponding to each tool wear value; secondly, select 1s vibration signal in single walking tool full process signal as sample;

[0021] Step d; extract power signal samples, each tool wear value also corresponds to extract 3 walking tool full process power signals; extract 8s signal as power signal sample in single walking tool full process power signal.

[0022] In a preferred embodiment, in step C, the visual reinforcement feature image is divided into a training set and a test set; the training set is used to input the connection model layer.

[0023] In a preferred embodiment, the ImageNet database is used for model migration, and the specific steps are as follows:

[0024] First, on the ImageNet dataset of the source domain, the pre-trained DTN model is used to extract the feature vector of the image, and according to the category information of the image, the connection layer and the classification layer of the DTN model are trained;

[0025] Then, on the vibration signal dataset of the target domain, the original data is converted into a feature vector similar to the image data in the ImageNet dataset using a preprocessing method, so as to extract the features using the pre-trained DTN model.

[0026] In a preferred embodiment, the model migration is performed using limited target samples, and the parameters in the connection layer and the classification layer of the DTN model are adapted to the vibration signal data of the target domain, and the specific steps are as follows:

[0027] Step (1); on the vibration signal dataset of the target domain, the pre-trained DTN model is used to extract the feature vector, and the feature representation of each sample is obtained;

[0028] Step (2); according to the data distribution of the target domain, the structure of the connection layer and the classification layer of the DTN model is adjusted, the number of connection layers is increased or decreased, the number of neurons in the connection layer is adjusted, and the type of activation function is adjusted;

[0029] Step (3); use the vibration signal dataset of the target domain to train the parameters of the connection layer and the classification layer to adapt to the data distribution of the target domain.

[0030] In a preferred embodiment, in step F, the DTN model is an adjusted DTN model.

[0031] In a preferred embodiment, in step A, in the single cutting parameter milling experiment, the workpiece is connected with a dynamometer; in the multi-cutting parameter milling experiment, the workpiece is clamped by a vice.

[0032] In summary, the present application includes the following advantages:

[0033] The matching wavelet deep transfer learning method provided by the application is a relatively advanced deep learning technology, which can realize the detection and prediction of the milling tool wear condition by training a deep neural network model of a source domain image first, and then migrating it to a target domain data set for fine tuning and training. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a method flowchart of the application;

[0035] Figure 2 is a schematic diagram of the Morse wavelet transform. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application; obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments of the application, and all other embodiments obtained by those skilled in the art without creative labor on the basis of the embodiments in the application belong to the protection scope of the application.

[0037] In the description of the application, it should be noted that the terms "upper", "lower", "inner", "outer", "top / bottom end" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore it cannot be understood as a limitation on the application. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0038] In the description of the application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "provided with", "sleeved / connected", "connected" and the like should be broadly understood, for example, "connected" can be wall-mounted connection, can be detachable connection, or integral connection, can be mechanical connection, can be electrical connection, can be direct connection, can be indirect connection through an intermediate medium, can be the internal communication of two elements, and those skilled in the art can understand the specific meaning of the above terms in the application according to the specific circumstances.

[0039] The following will be described in detail in combination with the accompanying Figure 1 The application will be further described in detail.

[0040] The embodiment of the application discloses a milling tool wear detection method based on matched wavelet deep migration learning, so that the early micro vibration faults of wear are better recognized, and the micro vibration faults actually represent the development trend and deterioration tendency of tool wear. Accurate diagnosis of the micro faults can effectively improve system reliability and reduce accident risk, thereby guaranteeing safe operation of the machine tool.

[0041] The application discloses a milling tool wear detection method based on matched wavelet deep migration learning, which comprises the following steps: obtaining original machining signals from a milling planar machining process of a machine tool through vibration, acoustic emission and power sensors; extracting a stable cutting process signal from the original machining signal; and capturing micro changes and obtaining a visual enhanced feature image through matched dimensionality increase of a one-dimensional fault signal by a Morse continuous wavelet.

[0042] The milling tool wear detection method based on matched wavelet deep migration learning mainly comprises the following steps:

[0043] Step A: sensor arrangement and experimental parameter setting: a complete part machining process needs to be completed by multiple sets of cutting parameters, therefore, after the range of the cutting parameters is determined, single cutting parameter milling experiments and multiple cutting parameter milling experiments need to be performed. When the single cutting parameter milling experiment is performed, the dynamometer is installed on the workbench by using a bolt, the workpiece is connected to the dynamometer by using a bolt, the vibration sensor is installed on the main shaft, the acoustic emission sensor is placed near the workpiece, and the power instrument is used to collect the total power of the machine tool. After the sensor arrangement is completed, the cutting process is performed under the same cutting parameters and machining conditions, and the data of each sensor in the machining process is obtained. When the multiple cutting parameter milling experiment is performed, the workpiece is clamped by a vice; the arrangement modes of the vibration sensor and the acoustic emission sensor are the same as those in the single cutting parameter milling experiment; and the power instrument is used to collect the total power of the main shaft. Because the cutting force signal and the power signal have a relatively strong correlation in the single cutting parameter milling experiment, in addition, the cutting force signal acquisition cost is high, therefore, the cutting force signal is not collected in the multiple cutting parameter milling experiment. During the experiment, the vibration, acoustic emission and cutting force data in multiple same machining processes are collected, the original input data is represented as XS, and the single machining data is represented as XSn, n=1, 2, …, n, and n represents the number of machining times.

[0044] Step B: Selection of signal acquisition samples: Extract the signal samples of the vibration sensor, acoustic emission sensor and power acquisition machine.

[0045] The specific steps are as follows:

[0046] Step a: Measure and record the wear of the single machining tool with an electron microscope, and the wear value is the maximum wear value of the relief surface.

[0047] Step b: According to the number of machining times, the obtained single machining data is one-to-one corresponding to the tool wear value. Each tool wear value corresponds to multiple tool path processes.

[0048] Step c: When extracting vibration signal samples and acoustic emission signals, first select 3 times of tool path whole process signals corresponding to each tool wear value; secondly, select 1s vibration signal in the single tool path whole process signal as the sample, then each tool wear value corresponds to 3 vibration signal samples, and each sample contains XY and Z three direction vibrations. The single cutting parameter experiment collects the total power signal of the machine tool, and the multi-cutting experiment collects the total power signal of the spindle.

[0049] Step d: When extracting power signal samples: each tool wear value also corresponds to 3 tool path whole process power signals; extract 8s signal as power signal sample in the single tool path whole process power signal, then each tool wear value also corresponds to 3 power signal samples, and each signal sample contains 6s cutting time and 2s non-cutting time power.

[0050] Step C: Signal preprocessing and Morse continuous wavelet transform: preprocess the vibration signal, acoustic emission signal and power signal extracted in the above process to remove noise and perform normalization operation to ensure the quality and accuracy of the data. Then, the preprocessed signal is subjected to Morse continuous wavelet transform, so that the local changes of signal characteristics in time-frequency domain are strengthened, and the visualized strengthened feature image is obtained. As shown in Figure 2 The upper picture represents the waveform of a random signal, and the horizontal axis represents time (seconds) and the vertical axis represents amplitude. The lower picture represents the result of continuous wavelet transform of the random signal, the horizontal axis represents time (seconds), and the vertical axis represents the scale size. Different colors represent the amplitude size of different wavelet coefficients, that is, the contribution size of different frequency components in the signal. Dark color represents wavelet coefficient with large amplitude, and light color represents wavelet coefficient with small amplitude. Therefore, by observing the distribution of different colors, we can roughly understand the energy distribution of the signal at different frequency components. Finally, the processed image is divided into training set and test set according to the training ratio. The training set is connected to the model input layer, the migration input sample is constructed, the sample features are operated layer by layer and converted into feature vectors.

[0051] Step D: Migration and diagnostic training: The migration object of this method is the deep migration network model. First, use the image multi-classification task of the source domain as the original task, and use the ImageNet database and limited target samples to perform model migration. First, use the ImageNet database to train the diagnostic ability of the model, and then use the limited target samples to train the model to migrate knowledge from the source domain to realize the diagnosis of the micro-vibration signal. The key to model migration is the migration of the framework layer stage, which migrates the prediction path of the source domain layer stage to the target domain. The main migration objects are the fully connected layer, the Softmax classification layer, the Classification classification layer and the related functions.

[0052] Model migration using the ImageNet database is as follows:

[0053] First, use the pre-trained DTN model to extract the feature vector of the image on the ImageNet dataset of the source domain, and train the connection layer and classification layer of the DTN model according to the class information of the image.

[0054] Then, use the pre-processing method to convert the original data into a feature vector similar to the image data in the ImageNet dataset on the vibration signal dataset of the target domain, so as to extract the features using the pre-trained DTN model.

[0055] Model migration using limited target samples: By migrating the framework parameters using limited target samples, the parameters in the connection layer and classification layer of the DTN model can be adapted to the vibration signal data of the target domain. The specific method includes:

[0056] Step (1) Use the pre-trained DTN model to extract the feature vector on the vibration signal dataset of the target domain to obtain the feature representation of each sample;

[0057] Step (2) According to the data distribution of the target domain, adjust the structure of the connection layer and classification layer of the DTN model, increase or decrease the number of connection layers, adjust the number of neurons in the connection layer and the type of activation function;

[0058] Step (3) Use the vibration signal dataset of the target domain to train the parameters of the connection layer and classification layer to adapt to the data distribution of the target domain.

[0059] After the migration learning is completed, use the adjusted DTN model to classify or predict new signal data. In the prediction process, use the softmax function of the classification layer to output the prediction result to obtain the probability value of each class.

[0060] Milling tool wear detection: input the milling signal to be detected into the fine-tuned and trained model, predict the wear condition of the milling tool, and give the corresponding evaluation result.

[0061] The above are all preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, and thus: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for detecting tool wear in milling based on matching wavelet deep transfer learning, characterized in that: The milling machine tool body, force gauge installed on the milling table, vibration sensor and acoustic emission sensor installed on the milling spindle; the acoustic emission sensor is placed near the workpiece; The milling machine also includes a power instrument acquisition machine; the detection method includes the following steps: Step A: After setting the cutting parameters, single cutting parameter milling experiment or multiple cutting parameter milling experiment is carried out; Step B: Signal acquisition sample selection, extract the signal samples of the vibration sensor, acoustic emission sensor and power acquisition machine; Step C: The extracted signal samples are preprocessed and Morse continuous wavelet transformed to obtain a visualized reinforced feature image; Step D: Take the deep migration network model as the migration object, input the visualized reinforced feature image to the model input layer, and construct the migration input sample; ImageNet database and limited target samples are used in sequence to migrate the model, so that the model migrates knowledge from the source domain, completes migration learning, and forms new signals; Step F: Classify or predict the new signal data using the DTN model, and output a prediction evaluation report; The signal acquisition sample selection step is: Step a: Measure and record the tool wear after single machining using an electron microscope, and the wear value is the maximum wear value of the tool face; Step b: Correspond the single machining data obtained with the tool wear value according to the number of machining times; each tool wear value corresponds to multiple tool path processes; Step c: Extract the vibration signal sample and the acoustic emission signal; first, select the 3-time tool path whole process signal corresponding to each tool wear value; second, select the 1s vibration signal in the single tool path whole process signal as the sample; Step d: Extract the power signal sample; each tool wear value also corresponds to the extraction of 3 tool path whole process power signals; extract the 8s signal in the single tool path whole process power signal as the power signal sample; The specific steps of model migration using ImageNet database are as follows: First, on the ImageNet dataset of the source domain, use the pre-trained DTN model to extract the feature vector of the image, and according to the category information of the image, train the connection layer and classification layer of the DTN model; Then, on the vibration signal dataset of the target domain, use the preprocessing method to convert the original data into a feature vector similar to the image data in the ImageNet dataset, so as to use the pre-trained DTN model to extract the feature; The specific steps of model migration using limited target samples are as follows: Step (1): On the vibration signal dataset of the target domain, use the pre-trained DTN model to extract the feature vector to obtain the feature representation of each sample; Step (2): According to the data distribution of the target domain, adjust the structure of the connection layer and classification layer of the DTN model, increase or decrease the number of connection layers, adjust the number of neurons in the connection layer and the type of activation function; Step (3): Use the vibration signal dataset of the target domain to train the parameters of the connection layer and the classification layer to adapt to the data distribution of the target domain.

2. The method according to claim 1, characterized in that: In step B, the signal acquisition sample selection includes obtaining the original machining signal, and the original machining signal is obtained by collecting vibration, acoustic emission and cutting force data in multiple same machining processes.

3. The method of claim 2, wherein the method is characterized by: In step B, for the original machining signal, in the single cutting parameter milling experiment, the power acquisition machine collects the signal as the total power of the machine tool; in the multi cutting parameter milling experiment, the power acquisition machine collects the signal as the total power signal of the main shaft.

4. The method of claim 1, wherein the method is characterized by: In step C, the visualized reinforced feature image is divided into a training set and a test set; and the training set is used to input the model.

5. The method of claim 1, wherein the method is characterized by: In step F, the DTN model is an adjusted DTN model.

6. The method of claim 1, wherein the method is characterized by: In step A, in the single cutting parameter milling experiment, the workpiece is connected with a dynamometer; and in the multi cutting parameter milling experiment, the workpiece is clamped by a vice.

Images