A Machine Tool Feed Axis Health Monitoring Method Based on Optimized Conditional Autoencoder

Abstract

The present invention belongs to the field of machine tool health status monitoring, and discloses a machine tool feed axis health monitoring method based on an optimized conditional autoencoder. Sensors collect vibration signals and power signals under the healthy state and some fault states of the machine tool feed axis; the vibration signals and power signals are preprocessed and denoised by wavelet packet decomposition; after denoising, time-domain and frequency-domain eigenvalues are calculated to form a machine tool feature vector, a time-frequency matrix is generated and converted into an RGB image, and a machine tool multi-source data set is constructed; the machine tool multi-source data set is divided to obtain a training set, a validation set, and a test set, and the training set only contains healthy data; the machine tool feature vector and the training set are input into a machine tool feed axis health status monitoring model based on an optimized conditional variational autoencoder for training, and the trained model is used for machine tool feed axis health status monitoring. The present invention realizes the health status monitoring of the feed axis only by using healthy data, and reduces the influence of the imbalance between fault data and healthy data on the data-driven monitoring method.

Description

Technical Field

[0001] The present invention relates to the field of machine tool health status monitoring, and specifically to a method for monitoring the health of a machine tool feed axis based on an optimized conditional autoencoder. Background Art

[0002] The feed axis of a numerically controlled machine tool is an important part of the numerically controlled machine tool. During long-term processing operations, under the action of external loads, the performance of the feed axis gradually deteriorates over time, resulting in a reduction in part processing accuracy, processing quality, and qualification rate. In severe cases, it may even cause machine tool failures and affect normal production operations. Therefore, it is of great significance to realize real-time monitoring of the health status of the machine tool feed axis.

[0003] In recent years, with the rise of neural networks, more and more health status diagnosis methods for key components of machine tools based on neural networks have been proposed. In the patent "A Method for Monitoring the Health of a Numerically Controlled Machine Tool Based on Digital Twin" (CN118192423A), a multi-sensor dataset of a numerically controlled machine tool is constructed using a cutting force sensor, a vibration sensor, a single temperature sensor, and a temperature and humidity sensor, and a CNN-LSTM neural network model is used to infer and judge the current health status and milling status of the numerically controlled machine tool, and output a health prediction result. In the patent "A Method for Monitoring the Health of a Tool for an Aeronautical Numerically Controlled Machine Tool Based on OPCUA" (CN116372665A), discrete wavelet transform is used to perform symmetric wavelet transform on the force signal, acceleration signal, and acoustic emission signal during the machining process of the numerically controlled machine tool to construct a model dataset, and an LSTM-CNN neural network model is used to predict the tool wear amount. In the patent "A Method for Predicting the RUL of a Rolling Bearing Based on Deep Learning" (CN117634300A), wavelet transform is performed on the acquired bearing vibration signal to obtain a wavelet power spectrum diagram to characterize the bearing degradation state, and a CNN-Transformer is used to predict the remaining service life of the bearing.

[0004] The above research shows that neural networks have achieved good diagnostic effects in the field of health status monitoring. However, in actual engineering applications, it takes a long time for the feed axis to run from a healthy state to a faulty state, and the time cost of obtaining faulty data is high, resulting in a serious imbalance between faulty data and healthy data. It is difficult for traditional data-driven neural network detection methods to achieve their due effects. The present invention realizes the health status monitoring of the machine tool feed axis only by using healthy data, which is beneficial for enterprises to maintain and manage equipment in a timely manner and has practical significance. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for monitoring the health of a machine tool feed axis based on an optimized conditional autoencoder, which is used to solve the problem that it is difficult to obtain faulty data of the machine tool feed axis, resulting in difficult diagnosis of the health status, and only uses healthy data for training to realize the health status monitoring of the machine tool feed axis.

[0006] The technical solution of the present invention is as follows: A method for monitoring the health of a machine tool feed axis based on an optimized conditional autoencoder, including the following steps:

[0007] S1. Collect the power signal and vibration signal of the machine tool feed axis in a healthy state and the vibration signal and power signal in some fault states through sensors;

[0008] S2. Perform signal preprocessing and denoising on the vibration signal and the power signal through wavelet packet decomposition to obtain a denoised vibration signal and a denoised power signal;

[0009] S3. Calculate the time-domain eigenvalues and frequency-domain eigenvalues of the denoised vibration signal and the denoised power signal to form a machine tool feature vector; perform wavelet transform on the denoised vibration signal and the denoised power signal to generate a vibration signal time-frequency matrix and a power signal time-frequency matrix, obtain a vibration time-frequency grayscale image and a power time-frequency grayscale image; draw a grayscale image of the denoised vibration signal, input the vibration time-frequency grayscale image into the red channel, the power time-frequency grayscale image into the green channel, and the grayscale image of the denoised vibration signal into the blue channel to generate a machine tool state RGB image, and construct a machine tool multi-source dataset;

[0010] S4. Divide the machine tool multi-source dataset to obtain a training set, a validation set, and a test set; only healthy data is included in the training set and the validation set, and both healthy data and fault data are included in the test set; build a machine tool feed axis health status monitoring model based on an optimized conditional variational autoencoder, input the machine tool feature vector and the training set into the machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder for training, and iteratively update the network parameters by minimizing the loss function to obtain a trained machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder;

[0011] Further, the machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder takes the conditional variational autoencoder as the main framework and includes two parts: an encoder and a decoder;

[0012] The machine tool feature vector and the training set are input into the encoder; the machine tool feature vector is used as the conditional input of the conditional variational autoencoder, and the latent space is constrained by the machine tool feature vector, so that the machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder generates a corresponding mapping space with the machine tool feature vector as the label during the training process, and in each subsequent health diagnosis, the corresponding mapping space is found through the machine tool feature vector for data reconstruction; the machine tool state RGB image is used as the feature input of the conditional variational autoencoder;

[0013] The encoder mainly consists of a convolutional layer, a residual block, and a SENet attention module. The input RGB image of the machine tool state passes through convolutional layers in sequence in the encoder for preliminary feature extraction, and residual blocks are added between convolutional layers; the SENet attention module is used to increase the weights of important features and suppress irrelevant features. Finally, the features are compressed into a one-dimensional vector through a flattening layer , and is concatenated with the machine tool feature vector and then connected to the distributed calculation layer;

[0014] The distributed calculation layer consists of two fully connected layers and is used to calculate the mean and the log variance of the data in the latent space. The latent variable is sampled through the mean and the log variance ;

[0015] is merged with the machine tool feature vector to obtain the output of the encoder . Subsequently, the encoded features are reconstructed into data with the same size as the original input through a decoder, where the structure of the decoder is symmetric to that of the encoder.

[0016] S5. The validation set is input to the trained machine tool feed axis health state monitoring model based on the optimized conditional variational autoencoder to calculate the reconstruction error; according to the 3σ principle, the health state judgment threshold is set through the reconstruction error of the validation set; when the data reconstruction error exceeds the health state judgment threshold, the machine tool deviates from the health state and the machine tool failure is determined. When the data reconstruction error is less than or equal to the health state judgment threshold, it is determined that the machine tool is still in a healthy state.

[0017] Furthermore, the sensor includes an acceleration sensor and a power sensor; the acceleration sensor is arranged at the proximal bearing of the machine tool feed axis and is used to collect vibration signals of the machine tool feed axis under different health states; the power sensor is connected to the feed axis servo system and is used to collect power signals of the machine tool feed axis under different health states; the vibration signal is expressed as , and the power signal is expressed as , represents the number of sampling points, the signal acquisition time each time is min, and the sampling frequency is .

[0018] Furthermore, the signal preprocessing for denoising is specifically to segment the vibration signal and the power signal respectively. The length of each sub-signal after segmentation is , and each sub-signal is subjected to layers of wavelet packet decomposition, and the wavelet basis function is​ , a wavelet packet tree is obtained; the energy at the -th layer and the -th node in the wavelet packet tree is expressed as , where is the wavelet coefficient under the -th layer and the -th node of the vibration signal or the wavelet coefficient under the -th layer and the -th node of the power signal. Summing the energies of all nodes in the -th layer, the total energy of the nodes in the -th layer is obtained . Calculate the wavelet denoising threshold . Define the nodes with energy greater than or equal to the wavelet denoising threshold as effective nodes, and the nodes with energy less than the wavelet denoising threshold as noise nodes, where is the threshold adjustment coefficient, which is used to adjust the sensitivity of noise signal judgment; Reconstruct the vibration signal or the power signal respectively according to the remaining effective nodes to obtain the denoised vibration signal and the denoised power signal .

[0019] Further, the machine tool feature vector is expressed as , where , represents the time-domain eigenvalue and frequency-domain eigenvalue of different signals; the time-domain eigenvalue includes rectified average, root mean square, standard deviation, kurtosis, and the frequency-domain eigenvalue includes spectral amplitude mean, spectral amplitude root mean square, spectral amplitude standard deviation, spectral frequency centroid, spectral frequency standard deviation, spectral frequency root mean square, spectral frequency skewness, spectral variation coefficient.

[0020] Further, the specific calculation method of the vibration signal time-frequency matrix and the power signal time-frequency matrix is to perform continuous wavelet transform on the denoised vibration signal and the denoised power signal respectively:

[0021]

[0022] In the formula, is the vibration wavelet transform coefficient or the power wavelet transform coefficient , is the denoised vibration signal or the denoised power signal , is the scale parameter of the wavelet function, is the translation parameter of the wavelet function, is the conjugate function of the wavelet basis function;

[0023] For and Take the second norm respectively to obtain the time-frequency matrix of the vibration signal and the time-frequency matrix of the power signal .

[0024] Furthermore, the method for generating the RGB image of the machine tool state is as follows: convert the denoised vibration signal into a two-dimensional matrix with n rows and n columns: ; normalize the values in the power signal time-frequency matrix, the vibration signal time-frequency matrix and the two-dimensional matrix to the range of [0, 255], and then map the matrices into grayscale images respectively to obtain the vibration time-frequency grayscale image, the power time-frequency grayscale image, and the denoised vibration signal grayscale image. Set the size of each grayscale image to W*H:

[0025]

[0026] In the formula, is the vibration time-frequency matrix , the power time-frequency matrix or the two-dimensional matrix , is the converted grayscale image; input the vibration time-frequency grayscale image, the power time-frequency grayscale image, and the denoised vibration signal grayscale image into the red, green, and blue color channels in sequence to generate the RGB image of the machine tool state.

[0027] Furthermore, divide the multi-source dataset of the machine tool, where P1% of the healthy data is used as the training set for model training, P2% of the healthy data is used as the validation set for calculating the healthy state judgment threshold, and all the fault data and the remaining P3% of the healthy data are used as the test set for evaluating the ability of the machine tool feed axis health state monitoring model based on the optimized conditional variational autoencoder to distinguish between the healthy state and the fault state of the machine tool feed axis.

[0028] Furthermore, the loss function is:

[0029]

[0030] where is the KL divergence of the mean and variance of the encoder output from the standard normal distribution, , , are the red, green, and blue channels of the original RGB image of the machine tool state respectively; , , are the red, green, and blue channels of the reconstructed RGB image respectively, and are the width and height of the image.

[0031] Furthermore, the healthy state judgment threshold is the sum of the mean of the reconstruction errors of the validation set and 3 times the standard deviation of the reconstruction errors of the validation set. The specific formula is:

[0032]

[0033] wherein is the mean value of the reconstruction error of the validation set, and is the standard deviation of the reconstruction error of the validation set. If the data reconstruction error is greater than the health state judgment threshold, it indicates that the machine tool deviates from the healthy state and is determined to be faulty. If the data reconstruction is less than or equal to the health state judgment threshold, it indicates that the machine tool is still in a healthy state.

[0034] Advantages of the present invention:

[0035] (1) Using the time-domain eigenvalues and frequency-domain eigenvalues of the vibration signal and power signal of the machine tool feed axis to generate the machine tool state feature vector, which is used as the conditional input of the conditional variational autoencoder, enabling the model to generate the corresponding mapping space with the machine tool feature vector as the label during the training process. At the same time, during data reconstruction, the model can find the corresponding mapping space according to the machine tool feature vector for data reconstruction, thereby reducing the probability of misjudgment of the model;

[0036] (2) The monitoring method of the present invention arranges a variety of sensors for monitoring the abnormal state of the machine tool feed axis, converts the time-domain signal and time-frequency domain signal of the data into grayscale images and synthesizes the machine tool state RGB image, realizing the multi-source fusion of machine tool signals, enabling the model to simultaneously capture the time-domain features and frequency-domain features of the signals, thereby improving the model's understanding and recognition ability of the machine tool state and enhancing its diagnostic effect under complex working conditions;

[0037] (3) The prior art not only requires manual annotation of training data but also requires the number of normal samples to be similar to the number of abnormal samples. However, in actual engineering applications, it takes a long time for the feed axis to go from healthy to faulty, and the time cost of obtaining fault data is high, resulting in a serious imbalance between fault data and healthy data. This method uses an unsupervised conditional variational autoencoder, which only requires normal samples for training, without abnormal data and manual annotation, eliminating the need for manual annotation work, thereby effectively overcoming the training difficulties brought about by the scarcity of fault samples;

[0038] (4) This method has a certain universality, and the health state monitoring method provided by this method can be extended to other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a schematic flow diagram of the machine tool feed axis health monitoring method based on the optimized conditional autoencoder.

[0040] Figure 2 is a schematic signal processing flow diagram of the machine tool feed axis health monitoring method based on the optimized conditional autoencoder.

[0041] Figure 3 It is a schematic structural diagram of a machine tool feed axis health monitoring model based on an optimized conditional variational autoencoder. Specific implementation manners

[0042] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be described in detail below with reference to the accompanying drawings.

[0043] The provided solution, a machine tool feed axis health monitoring method based on an optimized conditional autoencoder, has the following steps:

[0044] Taking the Y axis of a certain type of five-axis vertical machining center as an example, the implementation manner of the present invention will be described in detail. The stroke range of this five-axis vertical machining center is 0 to 500 mm (machine tool coordinate value), and the maximum feed speed is 40,000 mm / min. The flow of the machine tool feed axis health monitoring method based on an optimized conditional autoencoder is as Figure 1 shown.

[0045] S1. Collect the power signal and vibration signal of the machine tool feed axis in a healthy state through sensors, and at the same time obtain the vibration signal and power signal in some fault states for verifying the effectiveness of the model;

[0046] The sensors include an acceleration sensor and a power sensor; the acceleration sensor is arranged on the proximal bearing of the machine tool feed axis, the power sensor is connected to the feed axis servo system, the sampling frequency is 10,000 Hz, the machine tool movement stroke is 0 to 420 mm, the feed speed is 10,000 mm / min, and the acquisition time is 180 min.

[0047] S2: Perform signal preprocessing and denoising on the vibration signal and the power signal through wavelet packet decomposition to obtain a denoised vibration signal and a denoised power signal;

[0048] The schematic diagram of the signal preprocessing and denoising process is as Figure 2 shown. The specific method is as follows: First, the collected vibration signal and power signal are respectively segmented, and the length of each sub-signal after segmentation is 4096. Subsequently, each sub-signal is subjected to 4-layer wavelet packet decomposition, and the wavelet basis function is selected as the "dmey" wavelet to obtain a wavelet packet tree. The energy at the th node of the 4th layer in the wavelet packet tree can be expressed as , where is the wavelet coefficient at the th node of the 4th layer of the vibration signal or the wavelet coefficient at the th node of the 4th layer of the power signal. The energies of all nodes in the 4th layer are summed to obtain the total energy of the 4th layer nodes, and the wavelet denoising threshold , nodes with energy greater than or equal to the wavelet denoising threshold are defined as valid nodes, and nodes less than the wavelet denoising threshold are defined as noise nodes; where is the threshold adjustment coefficient, used to adjust the sensitivity of noise signal judgment, and takes as 0.5 in this example; reconstruct the vibration signal or power signal respectively according to the remaining valid nodes to obtain the denoised vibration signal and the denoised power signal .

[0049] S3. Calculate the time-domain eigenvalue and frequency-domain eigenvalue of the denoised vibration signal and the denoised power signal to form the machine tool feature vector; perform wavelet transform on the denoised vibration signal and the denoised power signal to generate the vibration signal time-frequency matrix and the power signal time-frequency matrix, and obtain the vibration time-frequency grayscale image and the power time-frequency grayscale image; draw the grayscale image of the denoised vibration signal, input the vibration time-frequency grayscale image into the red channel, the power time-frequency grayscale image into the green channel, and the grayscale image of the denoised vibration signal into the blue channel to generate the machine tool state RGB image and construct the machine tool multi-source dataset;

[0050] The machine tool feature vector is expressed as , where , are the time-domain eigenvalues of the denoised vibration signal: rectified average, root mean square, standard deviation, kurtosis; are the time-domain eigenvalues of the denoised power signal: rectified average, root mean square, standard deviation, kurtosis; are the frequency-domain eigenvalues of the denoised vibration signal in turn: spectral amplitude mean, spectral amplitude root mean square, spectral amplitude standard deviation, spectral frequency centroid, spectral frequency standard deviation, spectral frequency root mean square, spectral frequency skewness, spectral coefficient of variation, are the frequency-domain eigenvalues of the denoised power signal in turn: spectral coefficient of variation and spectral frequency centroid. The mathematical expressions of each eigenvalue are as follows:

[0051] (1) Rectified average

[0052]

[0053] (2) Root mean square

[0054]

[0055] (3) Standard deviation

[0056]

[0057] (4) Kurtosis

[0058]

[0059] (5) Spectral amplitude mean

[0060]

[0061] (6) Root mean square of spectral amplitude

[0062]

[0063] (7) Standard deviation of spectral amplitude

[0064]

[0065] (8) Center of gravity of spectral frequency

[0066]

[0067] (9) Standard deviation of spectral frequency

[0068]

[0069] (10) Root mean square of spectral frequency

[0070]

[0071] (11) Skewness of spectral frequency

[0072]

[0073] (12) Coefficient of variation of spectrum

[0074]

[0075] Wherein is the denoised vibration signal or denoised power signal, is the frequency domain amplitude obtained by performing fast Fourier transform on the denoised vibration signal or the frequency domain amplitude obtained by performing fast Fourier transform on the denoised power signal, is related to corresponding frequency.

[0076] The specific calculation method of the vibration signal time-frequency matrix and the power signal time-frequency matrix is to perform continuous wavelet transform on the denoised vibration signal and the denoised power signal respectively:

[0077]

[0078] In the formula, is the vibration wavelet transform coefficient or the power wavelet transform coefficient , is the denoised vibration signal or denoised power signal, is the scale parameter of the wavelet function, is the translation parameter of the wavelet function, is the conjugate function of the wavelet basis function;

[0079] For and take the two-norm respectively to obtain the time-frequency matrix of the vibration signal and the time-frequency matrix of the power signal .

[0080] The method for generating the RGB image of the machine tool state is as follows: First, convert the denoised vibration signal into a two-dimensional matrix of 64 rows and 64 columns: ; Normalize the values in the power signal time-frequency matrix, the vibration signal time-frequency matrix, and the two-dimensional matrix to the range of [0, 255], and then map the matrices to grayscale images respectively to obtain the vibration time-frequency grayscale image, the power time-frequency grayscale image, and the denoised vibration signal grayscale image. At the same time, set the size of each grayscale image to 64*64 uniformly:

[0081]

[0082] In the formula, is the vibration time-frequency matrix , the power time-frequency matrix or the two-dimensional matrix , is the converted grayscale image. Input the vibration time-frequency grayscale image, the power time-frequency grayscale image, and the denoised vibration signal grayscale image into the red, green, and blue color channels in sequence to generate the RGB image of the machine tool state.

[0083] S4. Divide the multi-source dataset of the machine tool to obtain a training set, a validation set, and a test set; only healthy data is included in the training set and the validation set, and both healthy data and faulty data are included in the test set; Build a health status monitoring model for the machine tool feed axis based on the optimized conditional variational autoencoder, input the machine tool feature vector and the training set into the health status monitoring model for the machine tool feed axis based on the optimized conditional variational autoencoder for training, and iteratively update the network parameters by minimizing the loss function to obtain the trained health status monitoring model for the machine tool feed axis based on the optimized conditional variational autoencoder;

[0084] Divide the multi-source dataset of the machine tool, where 60% of the healthy data is used as the training set for model training, 20% of the healthy data is used as the validation set for calculating the health status judgment threshold, and all faulty data and the remaining 20% of the healthy data are used as the test set for evaluating the ability of the health status monitoring model for the machine tool feed axis based on the optimized conditional variational autoencoder to distinguish between the healthy state and the faulty state of the machine tool feed axis.

[0085] The schematic diagram of the health monitoring model for the machine tool feed axis based on the optimized conditional variational autoencoder is as shown in Figure 3As shown in the figure, the machine tool feed axis health monitoring model based on the optimized conditional variational autoencoder takes the conditional variational autoencoder as the main framework, including an encoder and a decoder;

[0086] The machine tool feature vector and the training set are input into the encoder; the machine tool feature vector is used as the conditional input of the conditional variational autoencoder, and the machine tool state RGB image is used as the feature input of the conditional variational autoencoder;

[0087] The encoder is mainly composed of a convolutional layer, a residual block, and a SENet attention module. The input machine tool state RGB image passes through 3 convolutional layers in the encoder in sequence. The size of the convolutional kernel in each layer is 3*3, and the number of channels is 16, 32, and 64 in sequence for preliminary feature extraction. Residual blocks are introduced after the second and third convolutional layers to prevent gradient explosion or disappearance in the deep network; the SENet attention module is used to increase the weights of important features and suppress irrelevant features. Finally, the features are compressed into a one-dimensional vector through a flattening layer , and is concatenated with the machine tool feature vector and then connected to the distribution calculation layer;

[0088] The distribution calculation layer is composed of two fully connected layers, which are used to calculate the mean and the log variance of the data in the latent space. The latent variable is sampled through the mean and the log variance ;

[0089] is concatenated with the machine tool feature vector to obtain the encoder output . The encoded features are reconstructed into data with the same size as the original input through the decoder. In the decoder module, a symmetric structure is used with the encoder to reconstruct the data. The transposed convolutional layers Tconv1, Tconv2, and TConv3 have a convolutional kernel size of 3*3, and the number of channels is 64, 32, and 16 in sequence. Finally, a transposed convolutional layer TConv4 with a size of 3*3 and a channel number of 3 is used to output the reconstructed image. The maximum pooling in the network has a pooling window size of 2 and a stride of 1.

[0090] The loss function is:

[0091]

[0092] where is the KL divergence of the mean and variance of the encoder output with respect to the standard normal distribution, , , are the red, green, and blue channels of the original machine tool state RGB image respectively;​ , , They are the red, green, and blue channels of the reconstructed RGB image respectively. In this example, both the width and height of the image are 64.

[0093] S5. Input the validation set into the trained machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder to calculate the reconstruction error; according to the 3σ principle, set the health status judgment threshold through the reconstruction error of the validation set; input the data to be detected into the trained machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder. When the data reconstruction error exceeds the health status judgment threshold, the machine tool deviates from the healthy state and is determined to be faulty. When the data reconstruction error is less than or equal to the health status judgment threshold, it is determined that the machine tool is still in a healthy state.

[0094] The health status judgment threshold is the sum of the mean value of the reconstruction error of the validation set and 3 times the standard deviation of the reconstruction error of the validation set. The specific formula is:

[0095]

[0096] In the formula is the mean value of the reconstruction error of the validation set, and is the standard deviation of the reconstruction error of the validation set. If the data reconstruction error is greater than the health status judgment threshold, it indicates that the machine tool deviates from the healthy state and is determined to be faulty. If it is less than or equal to the health status judgment threshold, it indicates that the machine tool is still in a healthy state.

[0097] It should be noted that the above specific embodiments of the present invention are only used to exemplarily illustrate the principles and processes of the present invention and do not constitute a limitation to the present invention. Therefore, any modifications and equivalent replacements made without departing from the spirit and scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine tool feed axis health monitoring method based on an optimized condition autoencoder, characterized in that: The steps include: S1. Collecting the power signal and vibration signal of the machine tool feed axis in a healthy state and the vibration signal and power signal in a partial fault state through sensors; S2. Perform signal preprocessing and denoising on the vibration signal and the power signal by wavelet packet decomposition to obtain a denoised vibration signal and a denoised power signal; The signal preprocessing denoising is specifically as follows: the vibration signal and the power signal are segmented respectively, each segment of the sub-signal has a length of L, each segment of the sub-signal is decomposed by an i-layer wavelet packet, the wavelet basis function is ψ(t), and a wavelet packet tree is obtained; the energy at the k-th node of the i-th layer in the wavelet packet tree is expressed as Among them C i,k (h) is the wavelet coefficient of the kth node in the i-th layer of the vibration signal or the wavelet coefficient of the kth node in the i-th layer of the power signal. The energy of all nodes in the i-th layer is summed to obtain the total energy of the nodes in the i-th layer. Calculate wavelet denoising threshold The nodes with energy greater than or equal to the wavelet denoising threshold are defined as valid nodes, and the nodes with energy less than the wavelet denoising threshold are defined as noise nodes, where γ is the threshold adjustment coefficient, which is used to adjust the sensitivity of noise signal judgment; the vibration signal or power signal is reconstructed according to the remaining valid nodes to obtain the denoised vibration signal X(n) and the denoised power signal Y(n); S3. Calculate the time domain eigenvalues ​​and frequency domain eigenvalues ​​of the denoised vibration signal and the denoised power signal to form a machine tool feature vector; the machine tool feature vector is represented by T = {t1, t2, ..., t q }, where q = 1, 2, ..., Q, t q Representing the time domain eigenvalues ​​and frequency domain eigenvalues ​​of different signals; the time domain eigenvalues ​​include rectified mean, root mean square, standard deviation, and kurtosis; the frequency domain eigenvalues ​​include spectrum amplitude mean, spectrum amplitude root mean square, spectrum amplitude standard deviation, spectrum frequency centroid, spectrum frequency standard deviation, spectrum frequency root mean square, spectrum frequency skewness, and spectrum variation coefficient; Perform wavelet transform on the denoised vibration signal and the denoised power signal to generate a vibration signal time-frequency matrix and a power signal time-frequency matrix, and obtain a vibration time-frequency grayscale image and a power time-frequency grayscale image; draw a denoised vibration signal grayscale image, input the vibration time-frequency grayscale image into the red channel, input the power time-frequency grayscale image into the green channel, and input the denoised vibration signal grayscale image into the blue channel, generate an RGB image of the machine tool status, and construct a multi-source data set for the machine tool; S4. Divide the multi-source data set of the machine tool into a training set, a validation set, and a test set; The training set and the validation set only contain healthy data, and the test set contains both healthy data and fault data; a machine tool feed axis health status monitoring model based on an optimized conditional variational autoencoder is constructed, and the machine tool feature vector and the training set are input into the machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder for training, and the network parameters are iteratively updated by minimizing the loss function to obtain a trained machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder; The machine tool feed axis health status monitoring model based on optimized conditional variational autoencoder takes the conditional variational autoencoder as the main framework, including an encoder and a decoder; The machine tool feature vector and the training set are input to the encoder; the machine tool feature vector is used as the conditional input of the conditional variational autoencoder, and the latent space is constrained by the machine tool feature vector, so that the machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder generates a corresponding mapping space with the machine tool feature vector as the label during the training process, and each subsequent health diagnosis finds the corresponding mapping space through the machine tool feature vector to reconstruct the data; the machine tool status RGB image is used as the feature input of the conditional variational autoencoder; The encoder is mainly composed of a convolutional layer, a residual block, and a SENet attention module. The input machine tool state RGB image is sequentially passed through K con The convolutional layers are used for preliminary feature extraction, and residual blocks are added between the convolutional layers. The SENet attention module is used to increase the weight of important features and suppress irrelevant features. Finally, the flattening layer is used to compress the features into a one-dimensional vector e. x , e x After being spliced ​​with the machine tool feature vector T, it is connected to the distributed computing layer; The distribution calculation layer consists of two fully connected layers, which are used to calculate the mean μ and logarithmic variance logσ of the data in the latent space. 2 , the latent variable e is analyzed by the mean and log variance z Take samples, E z Combined with the machine tool feature vector T, the encoder output e is obtained z * , then the decoder reconstructs the encoded features into data of the same size as the original input, where the structure of the decoder is symmetrical with that of the encoder; S5. The validation set is input into the trained machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder to calculate the reconstruction error; according to the 3σ principle, the health status judgment threshold is set by the validation set reconstruction error; When the data reconstruction error exceeds the health status judgment threshold, the machine tool deviates from the healthy state and is judged to be faulty. When the data reconstruction error is less than or equal to the health status judgment threshold, the machine tool is judged to be still in a healthy state.

2. The machine tool feed axis health monitoring method based on the optimized condition autoencoder according to claim 1 is characterized in that: The sensor includes an acceleration sensor and a power sensor; the acceleration sensor is arranged at the proximal bearing of the machine tool feed shaft, and is used to collect vibration signals of the machine tool feed shaft under different health conditions; the power sensor is connected to the feed shaft servo system, and is used to collect power signals of the machine tool feed shaft under different health conditions; the vibration signal is expressed as x(n)=(x1,x2,…,x N ), the power signal is expressed as y(n)=(y1,y2,…,y N ), n=1,2…,N represents the number of sampling points, each signal acquisition time is Tmin, and the sampling frequency is KHZ.

3. The machine tool feed axis health monitoring method based on the optimized condition autoencoder according to claim 1 is characterized in that: The specific calculation method of the vibration signal time-frequency matrix and the power signal time-frequency matrix is ​​to perform continuous wavelet transform on the denoised vibration signal and the denoised power signal respectively: Where W ψ (a, b) are the vibration wavelet transform coefficients X ψ Or power wavelet transform coefficient Y ψ , f(n) is the denoised vibration signal X(n) or the denoised power signal Y(n), a is the scale parameter of the wavelet function, b is the translation parameter of the wavelet function, ψ * is the conjugate function of the wavelet basis function; X ψ With Y ψ Take the two norms respectively and get the vibration signal time-frequency matrix M X =||X ψ || 2 And the power signal time-frequency matrix M Y =||Y ψ || 2 .

4. The machine tool feed axis health monitoring method based on the optimized condition autoencoder according to claim 1 is characterized in that: The method for generating the machine tool status RGB image is to convert the denoised vibration signal into a two-dimensional matrix with n rows and n columns: z (i, j) = X(n*(i-1)+j), i = 1, 2, 3…, n, j = 1, 2, 3…, n; normalize the values ​​in the power signal time-frequency matrix, the vibration signal time-frequency matrix and the two-dimensional matrix to the range of [0, 255], and then map the matrices to grayscale images respectively to obtain the vibration time-frequency grayscale image, the power time-frequency grayscale image, and the denoised vibration signal grayscale image, and uniformly set the size of each grayscale image to W*H: Where M is the vibration time-frequency matrix M X , power time-frequency matrix M Y Or a two-dimensional matrix M Z , Gray(i,j) is the converted grayscale image; the vibration time-frequency grayscale image, power time-frequency grayscale image and denoised vibration signal grayscale image are input into the red, green and blue color channels in sequence to generate the machine tool status RGB image.

5. The machine tool feed axis health monitoring method based on the optimized condition autoencoder according to claim 1 is characterized in that: The machine tool multi-source data set is divided, wherein P1% healthy data is used as a training set for model training, P2% healthy data is used as a validation set for calculating the health status judgment threshold, and all fault data and the remaining P3% healthy data are used as a test set to judge the ability of the machine tool feed axis health status monitoring model based on the optimized conditional variational autoencoder to distinguish between the health status and fault status of the machine tool feed axis.

6. The machine tool feed axis health monitoring method based on the optimized condition autoencoder according to claim 1 is characterized in that: The loss function is: Where KL(μ,σ 2 ) is the KL divergence of the encoder output mean and variance for the standard normal distribution, I R ,I G ,I B They are the red, green and blue channels of the original machine tool status RGB image; I * R ,I * G ,I * B They are the red, green and blue channels of the reconstructed RGB image respectively, and W and H are the width and height of the image.

7. The machine tool feed axis health monitoring method based on the optimized condition autoencoder according to claim 1 is characterized in that: The health status judgment threshold is the sum of the mean of the validation set reconstruction error and 3 times the standard deviation of the validation set reconstruction error. The specific formula is: Threshold=μ c +3s c Where μ c Reconstruct the mean error of the validation set, σ c The standard deviation of the reconstruction error of the validation set. If the data reconstruction error is greater than the health status judgment threshold, it means that the machine tool deviates from the healthy state and is judged to be faulty. If the data reconstruction error is less than or equal to the health status judgment threshold, it means that the machine tool is still in a healthy state.

Images