Generative machine tool feed shaft state detection method based on Markov migration field

Abstract

The invention provides a Markov migration field-based generative machine tool feed shaft state detection method, which comprises the following steps of: s1, acquiring a current data set of a feed shaft under each working condition for preprocessing, and establishing a one-dimensional feed shaft current data set; s2, converting the one-dimensional signal into an MTF image by using Markov migration field transformation, cutting the original image to generate a plurality of sub-images, and making a two-dimensional image data set; and s3, dividing the cut image data into real data and a verification set according to a proportion, and training an improved adversarial generation neural network by using the real data so as to generate high-quality data. The objective of the embodiment of the invention is to provide a machine tool feed shaft state detection method based on generative data assistance. The machine tool feed shaft state detection method based on generative data assistance aims to solve the problems that a classifier is low in machine tool feed shaft state recognition rate and training is difficult to converge under the condition of small samples by utilizing the ability of Markov migration field transformation to retain dynamic characteristics and adversarial neural network expansion data.

Description

Technical Field

[0001] The present invention relates to the field of state detection of a machine tool feed shaft, a key component of rotating mechanical equipment, and particularly relates to a generative machine tool feed shaft state detection method based on a Markov migration field. Background Art

[0002] In modern manufacturing, CNC machine tools are core equipment for high-precision machining, and their performance directly determines machining efficiency and quality. The feed axis, a key moving component in CNC machine tools, is responsible for driving the tool or workpiece along a predetermined trajectory. Its operating status is crucial to machining accuracy and stability. However, with the increasing demand for high-speed, high-precision machining in the manufacturing industry, feed axes often need to operate at high speeds and frequently start and stop to accommodate complex machining tasks. These high-speed, high-load, and frequent start-stop operating conditions can easily lead to mechanical wear, thermal deformation, and increased vibration in the feed axis, resulting in reduced accuracy, increased noise, and even sudden failure. For example, high-speed operation increases wear on bearings and guideways, while frequent starts and stops can cause fatigue damage to the drive motor and transmission components. If these issues are not detected and addressed promptly, they can not only affect machining quality but can also cause equipment downtime and production losses. Therefore, real-time monitoring of feed axis operating status and fault warnings are crucial for improving machine tool reliability, reducing maintenance costs, and ensuring production efficiency. Through condition detection technology, potential faults can be detected early to avoid sudden downtime, while providing data support for predictive maintenance, thereby improving the overall performance and service life of machine tools.

[0003] In recent years, deep learning technology has made significant progress in the field of feed shaft condition detection and fault diagnosis, becoming a mainstream research direction. Its powerful feature extraction and pattern recognition capabilities can effectively process complex nonlinear fault signals, achieving high-precision condition monitoring and fault prediction. However, collecting feed shaft operating data under real-world conditions faces many challenges, such as the complex equipment operating environment, the scarcity of fault samples, and the high cost of data annotation. This results in a limited amount of data available for training, making it difficult to meet the large-scale data requirements of deep learning models. Using small sample data collected in the laboratory for data augmentation and then generating synthetic data through adversarial generative neural networks (GANs) has become an effective approach. By learning the distribution of real data, GANs can generate fault data that is highly similar to real operating conditions, thereby expanding the training sample and improving the model's generalization ability and diagnostic accuracy. This method provides a new solution for feed shaft fault diagnosis and has important application prospects.

[0004] Generative adversarial networks (GANs) have achieved remarkable results in the field of data generation, especially in the generation of two-dimensional images. However, there are relatively few studies on the generation of one-dimensional time series data (such as the feed shaft current signal), mainly because the dynamic characteristics and complex patterns of one-dimensional data are difficult to capture effectively directly through GAN. In order to solve this problem, the present invention proposes to convert the one-dimensional current time domain signal of the feed shaft into a Markov transition field (MTF) image. The Markov transformation method not only retains the time series characteristics of the signal but also enhances the visualization expression ability of the data by mapping the state transition probability of the time domain signal into a two-dimensional image. In particular, MTF has unique advantages in processing non-stationary signals. Motor current signals usually exhibit obvious non-stationarity. For example, when starting, stopping or the load suddenly changes, the signal characteristics will change dramatically. MTF can effectively capture this non-stationarity. By modeling the state transition probability, the dynamic changes of the signal are converted into spatial patterns in the image, thereby better reflecting the time series dependency of the signal. Building on this, using GANs to augment the generated Markov images not only overcomes the technical bottleneck of one-dimensional data generation but also provides an innovative solution to the small sample size problem in fault diagnosis. This approach provides richer training samples for feed axis state detection, significantly improving the model's generalization and detection accuracy. Summary of the Invention

[0005] The purpose of the embodiment of the present invention is to provide a generative machine tool feed axis state detection method based on Markov migration field, and to propose a solution to the problems that the classifier has low recognition rate of machine tool feed axis state and difficult training convergence in the case of small samples.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A generative machine tool feed axis state detection method based on a Markov migration field comprises the following steps: step s1, collecting one-dimensional current signal data of the machine tool feed axis under normal working conditions, overload conditions, and fault conditions, and preprocessing the collected signals to generate a one-dimensional feed axis current data set;

[0008] In step s2, the one-dimensional current signal is converted into an MTF image using the Markov migration field transform, and the original image is cropped into sub-images of equal size to produce a two-dimensional image dataset of the feed axis, thereby preliminarily achieving the purpose of data enhancement.

[0009] Step s3: divide the cropped image data into real data and validation sets in proportion, and use the real data to train the improved adversarial generative neural network to generate high-quality data;

[0010] In step s4, the generated data images and the real data images are combined to form a mixed data set, which is divided into a training set and a test set in proportion. The classifier model is trained and the accuracy of the model state detection is verified by testing.

[0011] In step s1, the current of the feed shaft during start-up, shutdown, and stable operation under normal operating conditions, overload conditions, and fault conditions is collected. The preprocessing method specifically uses a sliding window function with a fixed length of 1024 to intercept the current time domain signal, normalizes the data, and produces a one-dimensional feed shaft current data set.

[0012] In step s2, the data is converted into an image of size 1024×1024 using Markov migration field transformation, and the image is cut into 64 sub-images of size 128×128 without overlapping.

[0013] In step s3, the improved adversarial generative neural network includes two modules: a generator G and a discriminator D, wherein:

[0014] The generator G takes as input a concatenated vector of random noise z and conditional side information y. Sub-pixel convolution gradually transforms the low-resolution feature map into a high-resolution image, creating a more detailed image. The input features are processed and transformed through a series of convolutional layers, batch normalization layers, and Parametric ReLU activation functions. A self-attention mechanism is introduced in the final convolutional layer to capture long-range dependencies in the current's temporal domain. Finally, the generated data is normalized using the Tanh activation function. The discriminator D takes as input the discriminator-generated image G(x) and the ground-truth MTF image x. Feature extraction and spatial dimensionality reduction are performed through a series of convolutional layers and Parametric ReLU activation functions. Max-pooling is then applied to the extracted features to reduce computational complexity and the number of parameters. Similarly, a self-attention mechanism is introduced in the final layer. Finally, the concatenated vector of the convolution result and conditional side information y is fed into a fully connected layer, ultimately outputting a scalar value representing the probability that the generated image is ground-truth.

[0015] In step s3, the main improvement of the improved adversarial generative neural network is to reduce the training instability problem of the traditional GAN ​​by introducing Wasserstein distance and gradient penalty. On this basis, the ability of the GAN network to extract features is improved by deepening the convolution layer and introducing the self-attention mechanism. Finally, the auxiliary information y is introduced to enable the generator to generate samples of different categories under specific conditions. The objective function of the improved adversarial generative neural network is:

[0016] In step s4, the real image and the generated image are mixed and input into the mainstream deep learning classification model for classification training, and the gradient backpropagation is used to update the classification model parameters again according to the cross entropy loss function; the test set image is input into the optimal classification model to verify the accuracy of the feed axis state detection.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0018] The innovation of the present invention is to process the current time domain information using the Markov migration field. MTF can retain the dynamic characteristics of the signal by calculating the probability of state transition in the time series, and is suitable for capturing the time evolution law of the motor current signal. The motor current signal is usually non-stationary (such as when starting, stopping or the load suddenly changes), and MTF can effectively handle this non-stationarity. MTF converts one-dimensional signals into two-dimensional images, which is more suitable for convolutional neural network learning and classification. MTF can retain the time dependence of the motor current signal, is suitable for non-stationary signals, and the generated image is suitable for deep learning and has a certain robustness to noise.

[0019] This paper introduces conditional auxiliary information and a self-attention mechanism into the WGAN-GP network. While retaining the advantages of Wasserstein distance and gradient penalty training stability, the inclusion of conditional auxiliary information y enables the generator to generate samples of different categories based on specific conditions, thereby increasing the diversity of generated samples. Finally, the self-attention mechanism allows the model to focus on different parts of the input data, capturing long-range dependencies, making it suitable for processing Markov angle field data with complex structures.

[0020] This invention provides a method for assisting machine tool feed axis status detection. Specifically, it applies a Markov angular field transform to the feed axis's one-dimensional current signal and utilizes an adversarial neural network to generate high-quality data for auxiliary detection. To address the instability of GAN data, a Markov migration field transform is proposed to convert the one-dimensional current time domain signal. This method retains the superior performance of GAN networks in generating two-dimensional images while incorporating the unique advantages of Markov migration field transform in processing non-stationary signals.

[0021] This paper proposes an improved generative adversarial neural network (GAN) that improves the model's ability to generate high-quality data and its stability by introducing Wasserstein distance and gradient penalty, deepening convolutional layers, a self-attention mechanism, and conditional auxiliary information. This approach addresses the low recognition rate of machine tool feed axis states and the difficulty in training convergence in small sample sizes. This approach further addresses the challenges of complex equipment operating environments, a scarcity of fault samples, and the high cost of data annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is an overall implementation flow chart of the generative data-assisted feed axis state detection method of the present invention.

[0023] Figure 2 The improved adversarial neural network structure diagram of the present invention is generated. DETAILED DESCRIPTION

[0024] In order to make the technical solutions and advantages of the present invention more clear, the present invention is described in more detail below with reference to specific embodiments and related drawings.

[0025] Example 1.

[0026] Refer to the attached Figure 1 The overall flow chart of a generative data-assisted machine tool feed axis state detection method includes the following steps:

[0027] Step s1, collecting one-dimensional current signal data of the machine tool feed axis under normal working conditions, overload working conditions and fault working conditions, and preprocessing the collected signals to produce a one-dimensional feed axis current data set;

[0028] In step s1.1, the collected current signal is pre-processed using a sliding window function, and the current signals under the three working conditions are processed into 200 one-dimensional data with a length of 1024. The current data is then normalized to facilitate the subsequent training and operation of the adversarial generative neural network and classifier.

[0029] Among them, the formula for normalizing one-dimensional current data is: Step s2, using Markov migration field transformation to convert the one-dimensional current signal into a two-dimensional image is divided into two processes, namely image conversion and image cropping. The specific steps are as follows:

[0030] Step s2.1 uses the Markov migration field to convert the one-dimensional current data sample described in step 1 into 200 images of 1024 pixels;

[0031] The Markov Transition Field (MTF) is a time series image coding method based on the Markov transition matrix. This method treats the passage of time in a time series as a Markov process. That is, given the current state, its future evolution is independent of its past evolution. This method constructs a Markov transition matrix, which is then expanded to a Markov transition field, enabling image coding.

[0032] For the time series X=(x t ,t=1,2,...,T), the image encoding steps are as follows:

[0033] (1) Divide the time series X(t) into Q bins (labeled 1, 2, ..., Q, with the same amount of data in each bin);

[0034] (2) Change each data in the time series to the serial number of its corresponding quantile bin;

[0035] (3) Construct the transfer matrix W(w ij represents the frequency of transfer from bin i to bin j):

[0036]

[0037] (4) Construct the Markov migration field M:

[0038]

[0039] In step s2.2, in order to meet the training requirements of the GAN network, in order to ensure the generation of high-quality sample data in the adversarial generation of small samples, the 1024×1024 image is processed by cropping. A 1024-pixel MTF image is cropped into 64 128-pixel sub-images without overlapping. These images are classified according to working conditions and divided into a training sample set and a verification sample set in a ratio of 7:3.

[0040] Step s3, the cropped image data is divided into real data and validation set according to the proportion, and the real data is used to train the improved adversarial generative neural network to generate high-quality data. The input of random noise to generate high-quality generated data is mainly divided into the following four stages, which are combined with Figure 2 Elaborate on the process of generating high-quality data images using MTF images;

[0041] In step 3.1, a new vector consisting of 1024-length random noise z and the one-hot label encoding vector y is input into the generator. After two sub-pixel convolutions, the channels of the low-resolution feature map are rearranged to generate a high-resolution image. The specific steps include convolution and pixel shuffle operations. The convolution operation generates feature maps with multiple channels, which are then rearranged to produce a high-resolution image. Compared to traditional upsampling techniques that use interpolation to increase resolution, sub-pixel convolution has learnable parameters, and the convolution kernel is continuously optimized during training, which improves image generation while reducing the checkerboard effect.

[0042] The features transformed by sub-pixel convolution in step 3.2 are then repeatedly processed and transformed through two 3×3 convolutional layers, followed by batch normalization and a LeakyReLU activation function. A self-attention mechanism is introduced before the last convolutional layer to enhance the network's ability to capture long-range dependencies in sequential data and improve its ability to model the dynamic characteristics of time-domain signals (such as motor current signals), such as start-up, stop-down, and sudden load changes. Motor current signals are typically non-stationary, and the self-attention mechanism dynamically adjusts the level of attention paid to different time steps, thereby better handling the signal's non-stationary nature. The Markov transition field image generated using this method can more accurately capture non-stationary changes in the signal (such as transient processes), improving the accuracy of machine tool feed axis state recognition. Finally, the generated image is normalized using the Tanh activation function to a pixel size of 128×128.

[0043] In step 3.3, the generated image G(z) and the real image x are input to the discriminator. Feature extraction and spatial dimensionality reduction are performed on the input data through a series of convolutional layers, LeakyReLU activation functions, regularization, and max pooling layers. A self-attention mechanism is introduced before the last convolutional layer. A max pooling layer is used to reduce the size of the feature map, thereby reducing computational complexity and the number of parameters. Finally, the transformed result and the conditional auxiliary information are concatenated, flattened through a fully connected layer, and then passed through a softmax function to output a scalar value representing the probability that the generated image is classified as a real image.

[0044] In step 3.4, the probability output by the discriminator is used as a parameter to calculate the loss function of the generator G and the discriminator D, using SGDM (momentum gradient descent) as the optimizer. The parameter weights of the generator G and the discriminator D are iteratively optimized through the loss function and backpropagation. Ultimately, the generator reaches the level of generating realistic images.

[0045] The loss function of the generator is:

[0046]

[0047] The loss function of the discriminator is:

[0048]

[0049] Where: p(x) is the probability distribution of the real data x, p(z) is the probability distribution of the random noise z, D(x) is the probability of the discriminator judging x as true or false, G(z) is the sample output by the generator, E x～p(x) represents the expectation of x sampled from the true data distribution p(x), E z～p(x) represents the expectation of z sampled from the random noise distribution p(x), λ is the gradient penalty coefficient, is the gradient operator, Random interpolation for real and generated data;

[0050] In step s4, the generated data images and the real data images are combined to create a mixed data set, which is divided into a training set and a test set in proportion. The classifier model is trained and the accuracy of the model state detection is verified by testing. Step s4 is mainly divided into two stages, including:

[0051] Step s4.1 merges the high-quality Markov migration field image generated in step s3 and the real one-dimensional current data conversion image according to the three working condition labels, and divides the mixed data set into a training set and a test set in a ratio of 7:3.

[0052] In step s4.2, select a mainstream deep learning model classifier, such as 2DCNN, Resnet34, RNN, or Transformer. Use the training set from the mixed dataset to train the classifier for state recognition and classification. Backpropagate the gradients based on the classification loss function to update the parameters of the machine tool feed axis state detection model. After training and testing, save the optimized detection model. Compare the accuracy of the optimal model from different classifiers, select the optimal classifier, and test its accuracy using the test set.

Claims

1. A generative machine tool feed axis state detection method based on Markov migration field, characterized in that It includes the following steps: Step s1, collecting one-dimensional current signal data of the machine tool feed axis under normal working conditions, overload working conditions and fault working conditions, and preprocessing the collected signals to produce a one-dimensional feed axis current data set; In step s2, the one-dimensional current signal is converted into an MTF image using the Markov migration field transform, and the original image is cropped into sub-images of equal size to produce a two-dimensional image dataset of the feed axis, thereby preliminarily achieving the purpose of data enhancement. Step s3: divide the cropped image data into real data and validation sets in proportion, and use the real data to train the improved adversarial generative neural network to generate high-quality data; In step s4, the generated data images and the real data images are combined to form a mixed data set, which is divided into a training set and a test set in proportion. The classifier model is trained and the accuracy of the model state detection is verified by testing.

2. The method for detecting the feed axis state of a machine tool based on generative data assistance according to claim 1, characterized in that: In step s1, the current of the feed shaft during start-up, shutdown, and stable operation under normal operating conditions, overload conditions, and fault conditions is collected. The preprocessing method specifically uses a sliding window function with a fixed length of 1024 to intercept the current time domain signal, normalizes the data, and produces a one-dimensional feed shaft current data set.

3. The method for detecting the feed axis state of a machine tool based on generative data assistance according to claim 1, characterized in that: In step s2, the data is converted into an image of size 1024×1024 using Markov migration field transformation, and the image is cut into 64 sub-images of size 128×128 without overlapping.

4. The method for detecting the feed axis state of a machine tool based on generative data assistance according to claim 1, wherein: In step s3, the improved adversarial generative neural network includes two modules: a generator G and a discriminator D, wherein: The generator G takes as input a concatenated vector of random noise z and conditional side information y. Sub-pixel convolution gradually transforms the low-resolution feature map into a high-resolution image, creating a more detailed image. The input features are processed and transformed through a series of convolutional layers, batch normalization layers, and Parametric ReLU activation functions. A self-attention mechanism is introduced in the final convolutional layer to capture long-range dependencies in the current's temporal domain. Finally, the generated data is normalized using the Tanh activation function. The discriminator D takes as input the discriminator-generated image G(x) and the ground-truth MTF image x. Feature extraction and spatial dimensionality reduction are performed through a series of convolutional layers and Parametric ReLU activation functions. Max-pooling is then applied to the extracted features to reduce computational complexity and the number of parameters. Similarly, a self-attention mechanism is introduced in the final layer. Finally, the concatenated vector of the convolution result and conditional side information y is fed into a fully connected layer, ultimately outputting a scalar value representing the probability that the generated image is ground-truth.

5. The method for detecting the feed axis state of a machine tool based on generative data assistance according to claim 1, characterized in that: In step s3, the main improvement of the improved adversarial generative neural network is to reduce the training instability problem of the traditional GAN ​​by introducing Wasserstein distance and gradient penalty. On this basis, the ability of the GAN network to extract features is improved by deepening the convolution layer and introducing the self-attention mechanism. Finally, the auxiliary information y is introduced to enable the generator to generate samples of different categories under specific conditions. The objective function of the improved adversarial generative neural network is:

6. The method for detecting the feed axis state of a machine tool based on generative data assistance according to claim 1, characterized in that: In step s4, the real image and the generated image are mixed and input into the mainstream deep learning classification model for classification training, and the gradient backpropagation is used to update the classification model parameters again according to the cross entropy loss function; the test set image is input into the optimal classification model to verify the accuracy of the feed axis state detection.

Images