A method for monitoring tool wear status using data augmentation and multi-feature fusion under small sample sizes

Abstract

This invention provides a method for monitoring tool wear status using data augmentation and multi-feature fusion under small sample conditions, designed for the field of metal cutting processing technology. This invention converts one-dimensional signals into two-dimensional grayscale images and time-frequency images respectively. An improved deep convolutional generative adversarial network (MDCGAN) is created to augment a small number of two-dimensional image samples, and its generation effect is evaluated using evaluation metrics. Then, the generated image samples, after evaluation and screening, and real image samples are input into a multi-source feature fusion network (MFFNet). By extracting and fusing features from different modalities, tool status recognition is achieved. By transferring the parameters of the MFFNet model, cross-domain diagnosis between different datasets under small sample conditions is achieved. Experiments are conducted on two different datasets and compared with other deep learning network models. Experiments show that training the MFFNet model on the dataset augmented with MDCGAN achieves a classification accuracy of over 99%, effectively improving the accuracy of tool status monitoring.

Description

Technical Field

[0001] This invention relates to the field of metal cutting technology, and in particular to a method for monitoring tool wear status using data augmentation and multi-feature fusion with small sample sizes. Background Technology

[0002] As an indispensable part of the milling process, tool wear directly affects the quality and precision of the machined products. Continuous tool wear causes changes in the contact state between the cutting edge and the workpiece. If the tool is not replaced in time, it will lead to workpiece scrap and equipment downtime, ultimately causing serious economic losses.

[0003] With the development of artificial intelligence, more and more scholars are applying machine learning, especially deep learning, to the monitoring of machining conditions, such as tool condition monitoring and equipment health status monitoring. However, unlike other application areas, most machining processes operate under normal conditions, and the amount of sample data under abnormal conditions is relatively small, which easily leads to data imbalance. If imbalanced sample data is used as training samples, the network model will overemphasize the characteristics of samples with a large number of labeled samples, resulting in a decline in the classification characteristics of imbalanced class labels, which seriously affects the performance of tool wear monitoring models. Therefore, how to solve the data imbalance problem has become the key to improving the accuracy of machining condition monitoring.

[0004] Generative Adversarial Networks (GANs), as an unsupervised learning model, have broad application prospects in augmenting datasets and monitoring machining conditions. By training GANs to generate new samples, the problem of scarce abnormal samples can be solved, further improving the accuracy of tool monitoring and identification.

[0005] Traditional monitoring and recognition methods extract features from deep learning models using a single signal, and the input state information is often unimodal, lacking the complementary characteristics between multiple modalities. Therefore, this paper proposes a one-dimensional signal-to-image method, using the converted grayscale image and time-frequency image as input to the network model. This achieves feature extraction and fusion of different modal information, fully exploring the model's learning potential. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for monitoring tool wear status using data augmentation and multi-feature fusion with small sample sizes.

[0007] A method for monitoring tool wear status using data augmentation and multi-feature fusion under small sample conditions includes the following steps:

[0008] Step 1, Data Acquisition and Preprocessing: Select the cutting force signal, convert the one-dimensional cutting force signal into a two-dimensional grayscale image, and convert the one-dimensional cutting force signal into a two-dimensional time-frequency image through continuous wavelet transform;

[0009] The conversion formula for the two-dimensional grayscale image is as follows:

[0010]

[0011] Where, round{·} represents the rounding function, which normalizes all pixel values ​​to between 0 and 255, P(j,k),j=1......M,k=1......M represents the pixel value of each pixel after transformation, M is the side length of the two-dimensional image, L is an arbitrarily selected length in the one-dimensional cutting force signal, and L is a continuous M×M sampling point;

[0012] The continuous wavelet transform (CWT) is used to transform each time-domain waveform into a corresponding time spectrum. The Morlet wavelet function is selected, and the transformation principle is shown in Equations 2 to 4.

[0013]

[0014]

[0015]

[0016] Where ψ(t) is the basic mother wavelet function, ψ a,τ ψ(t) is a set of time-scaled wavelets, where a and τ are the scaling factor and translation factor, respectively. For any signal function ψ(t)∈L 2 (R), the corresponding continuous wavelet function is defined by formula (3), where CWT is the complex conjugate function of the mother wavelet function ψ(t). f (a,τ) represents f(t) and ψ a,τ The inner product of (t). Morlet is selected as the mother function of the wavelet, and its expression is Equation (4), where w0 is the center frequency of the wavelet.

[0017] Step 2: Use a generative adversarial network to augment the original data; build an improved deep convolutional generative adversarial network MDCGAN and train it. Input the original grayscale image and time-frequency image into MDCGAN to obtain the generated grayscale image and time-frequency image. Use two evaluation metrics, FID and MMD, to measure and filter the original image and the generated image. Balance the original image and the filtered image to obtain the augmented dataset.

[0018] Step 2.1: Build an improved deep convolutional generative adversarial network (MDCGAN) and train the improved deep convolutional generative adversarial network (MDCGAN).

[0019] Step 2.1.1: Set the model structure parameters and establish an improved deep convolutional generative adversarial network (MDCGAN); on the framework of deep convolutional generative adversarial network, modify its generator and discriminator.

[0020] The generator consists of fully connected layers and two-dimensional transposed convolutional layers; the discriminator consists of fully connected layers and two-dimensional convolutional layers; batch normalized (BN) is added to each convolutional layer of the discriminator, and ReLU and Leaky ReLU are selected as the activation functions for the convolutional and transposed convolutional layers; the optimizers of both the generator and the discriminator use adaptive moment estimation (Adam) to update the model weights; the Nadam algorithm is introduced, which uses second-order moment estimation of the gradient to achieve adaptive adjustment of the learning rate α. The update rule of Nadam is described as follows:

[0021]

[0022]

[0023]

[0024] Where θ is the weight or threshold to be optimized, and m t and These are the first-order momentum deviation and the corrected first-order momentum deviation, respectively, V t and denoted as the second-order momentum deviation and the corrected second-order momentum deviation, respectively, where t is the current iteration, and β1, β2, and ε are the correction coefficients. t Let α be the gradient vector and α be the learning rate.

[0025] Step 2.1.2: Set the model training parameters, including the number of training iterations, the number of iterations, the learning rate, the batch size, and the dimension of the latent space. At the same time, set the discriminator's loss for distinguishing true from false to binary cross-entropy loss.

[0026] The loss functions for the generator and discriminator are as follows:

[0027]

[0028]

[0029] Where x represents the actual sample data, This represents the sample generated by the generator. data (x) represents the distribution from the real data, p g The distribution derived from the generated data is formed by having prior p z The simple random variable z of (z) is obtained through the G mapping, i.e., z ~ p z (z),x=G(z).

[0030] Step 2.1.3: Start training. First train the discriminator, then train the generator. Input the training set containing real sample data into the discriminator. At the same time, input random noise into the generator to generate pseudo samples. Then input the pseudo samples into the discriminator to implement the adversarial behavior between the generator and the discriminator to train the generator.

[0031] Step 2.2: Input the two-dimensional grayscale image and the two-dimensional time-frequency image into the improved deep convolutional generative adversarial network MDCGAN to generate sample images;

[0032] Step 2.3: Use two metrics, FID and MMD, to evaluate the similarity and diversity of the generated sample images, and determine whether the distribution of the generated sample data is similar to that of the real tool condition sample data; use FID and MMD to evaluate the quality of the artificially generated data, and discard data that does not meet the criteria.

[0033] Specifically, artificially generated data of the same type are sequentially compared with real datasets using FID and MMD measurements, their average values ​​are calculated, and a threshold M is set. If the value is less than the threshold M, the artificially generated data is considered to meet the standard for similarity with real data, and the data that meets the conditions are used to synthesize a training dataset.

[0034] The FID represents the diversity and quality of the generated image. The smaller the FID, the better the image diversity and the better the quality. FID is defined as:

[0035]

[0036] Where Tr represents the sum of the elements on the diagonal of the matrix, referred to as the trace of the matrix; x and g represent the real image and the generated sample image, respectively; μ x and μ g ∑ represents the mean features of the real image and the generated image. x and ∑ g It is the covariance matrix of the real image and the generated image.

[0037] The MMD (Mean Distance Distance) is used to measure the distance between two data distributions, i.e., the similarity between the two data points. The smaller the MMD value, the closer the two distributions are, and the better the quality of the generated image. MMD is defined as:

[0038]

[0039] In the formula, F represents the chosen kernel function, k(x,y) is the kernel mapping result of x and y, and m and n are the distributions P and N, respectively. r(x) and P g(y) The number of actual samples and generated samples in the dataset, x i and y iThese are the real sample dataset and the generated sample dataset, respectively.

[0040] Step 3, Monitoring and Identification: Construct a multi-source feature fusion network MFFNet, select a set proportion of original real images and generated images after evaluation and screening to form an enhanced dataset, and input it into the multi-source feature fusion network MFFNet for feature extraction and identification classification.

[0041] Step 3.1: Construct the multi-source feature fusion network MFFNet;

[0042] The original real images and the generated images after evaluation and screening are combined to form an enhanced dataset, which is then input into the multi-source feature fusion network MFFNet for feature extraction and fusion.

[0043] The multi-source feature fusion network MFFNet includes an encoding denoising module and a parameterless attention mechanism module.

[0044] The encoding and noise reduction module structure includes: setting an adaptive threshold to remove noise information; wherein the formula for calculating the adaptive threshold is:

[0045]

[0046]

[0047] Where τ is the threshold, x is the input, and y is the output. α∈[0,1] is the contraction factor, and ρ is the adjustment factor.

[0048] The parameterless attention mechanism module structure utilizes an attention mechanism to redistribute the weights of feature information;

[0049] An energy function E is defined. t Then, its analytical solution was obtained.

[0050]

[0051]

[0052]

[0053] Where λ is the regularization term, and t and x i These are the characteristics of the target neuron and other neurons in each channel. The mean of all neurons in a single channel. Let H be the variance of all neurons on a single channel, and H and W be the height and width of the feature map.

[0054] The entire process of refining the parameterless attention mechanism module is as follows:

[0055]

[0056] Where X is the input feature tensor, Sigmoid(·) is the activation function, and ⊙ is the Hadamard product.

[0057] Step 3.2: Tool detection and identification using the multi-source feature fusion network MFFNet:

[0058] The enhanced dataset is input into MFFNet for tool wear condition identification and classification.

[0059] Step 4, Transfer Learning: By transferring the parameters of the multi-source feature fusion network MFFNet, monitoring and diagnosis can be achieved across different datasets;

[0060] First, the network layer structure of the multi-source feature fusion network (MFFNet) is frozen using source domain data, and the parameters of MFFNet are extracted. Then, the parameters of the classifier of MFFNet are fine-tuned using target domain data to achieve parameter transfer of MFFNet.

[0061] The fine-tuning specifically involves saving the trained weight parameters from the source domain and, when training in the target domain, not starting from scratch, but continuing to train based on the weights of the source feature fusion network MFFNet trained in the source domain.

[0062] The beneficial effects of adopting the above technical solution are as follows:

[0063] This invention provides a method for monitoring tool wear conditions using data augmentation and multi-feature fusion under small sample sizes. Experimental results show that the accuracy of the neural network is higher when the data-augmented samples are input into the neural network for training than when the original images are directly input into the neural network. This solves the problem of the limited number of tool wear samples and verifies the effectiveness of the proposed method in tool wear condition monitoring. Simultaneously, the multi-feature fusion network achieves the extraction and fusion of different features, resulting in higher classification and recognition performance. Finally, a parameter transfer method is used to achieve monitoring and diagnosis between different datasets, providing a new approach for tool monitoring under various working conditions. Attached Figure Description

[0064] Figure 1 This is a flowchart of the tool wear condition monitoring method in an embodiment of the present invention;

[0065] Figure 2 This is the converted grayscale image in an embodiment of the present invention;

[0066] Among them, (a) - initial wear, (b) - normal wear, and (c) - severe wear;

[0067] Figure 3 This is the converted time-frequency graph image in an embodiment of the present invention;

[0068] Figure 4 This is a diagram of the improved deep convolutional generative adversarial network MDCGAN in an embodiment of the present invention.

[0069] Figure 5 This is a structural diagram of the encoding and noise reduction module in an embodiment of the present invention;

[0070] Figure 6 This is a structural diagram of the parameterless attention mechanism module in an embodiment of the present invention;

[0071] Figure 7 This is a diagram showing the monitoring and identification results in an embodiment of the present invention;

[0072] Figure 8 This is a graph showing the monitoring and identification accuracy and loss in an embodiment of the present invention;

[0073] Among them, Figure (a) is the monitoring and identification accuracy map, and (b) is the loss map;

[0074] Figure 9 This is a framework diagram for migration monitoring and diagnosis in an embodiment of the present invention. Detailed Implementation

[0075] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0076] A method for monitoring tool wear condition using data augmentation and multi-feature fusion under small sample sizes, such as... Figure 1 As shown, it includes the following steps:

[0077] Step 1, Data Acquisition and Preprocessing: Select the cutting force signal, convert the one-dimensional cutting force signal into a two-dimensional grayscale image, and convert the one-dimensional cutting force signal into a two-dimensional time-frequency image through continuous wavelet transform;

[0078] The grayscale image conversion process requires no human intervention; the entire input-output process is completed using a conversion formula, eliminating the impact of manual intervention on feature extraction and making the entire process more intelligent. The conversion formula for the two-dimensional grayscale image is as follows:

[0079]

[0080] Wherein, round{·} represents the rounding function, which normalizes all pixel values ​​to between 0 and 255, which is exactly the pixel value range of the grayscale image. P(j,k),j=1......M,k=1......M represents the pixel value of each pixel after transformation, M is the side length of the two-dimensional image, and L is an arbitrarily selected length in the one-dimensional cutting force signal, which is a continuous M×M sampling point; in this embodiment, L is 4096 and M is 64.

[0081] In this embodiment, based on actual machining conditions, tool wear is divided into three states: initial wear, normal wear, and severe wear (with flank wear amounts of [0, 0.07), [0.07, 0.120), and [0.120, +∞) respectively). From each state, 4096 points are randomly selected 300 times as usable samples. 50 samples are used in MACGAN to generate new samples, and 100 samples are used as the test set for monitoring and identification. The original one-dimensional tool signal is reconstructed into a two-dimensional matrix, and the transformed tool wear image data is shown below. Figure 2 As shown, tool images in different wear states have distinct characteristics that make them relatively easy to distinguish.

[0082] The time-frequency plot can clearly describe the fluctuation of signal frequency over time, and also express the signal's energy distribution. Continuous wavelet transform (CWT) is used to transform each time-domain waveform into a corresponding time-frequency spectrum. The Morlet wavelet function is selected, and the transformation principle is as follows:

[0083] As shown in equations 2 to 4.

[0084]

[0085]

[0086]

[0087] Where ψ(t) is the basic mother wavelet function, ψ a,τ ψ(t) is a set of time-scaled wavelets, where a and τ are the scaling factor and translation factor, respectively. For any signal function ψ(t)∈L 2 (R), the corresponding continuous wavelet function is defined by formula (3), where CWT is the complex conjugate function of the mother wavelet function ψ(t). f (a,τ) represents f(t) and ψ a,τ The inner product of (t). Morlet is selected as the mother function of the wavelet, and its expression is Equation (4), where w0 is the center frequency of the wavelet.

[0088] To ensure consistent image input for network training, the time-frequency spectrum is visualized as a 64x64 pixel RGB three-channel time-frequency graph. The converted time-frequency graph is shown below. Figure 3 As shown.

[0089] Step 2: Insufficient training data often hinders the performance of classification algorithms. Generative adversarial networks (GANs) are used to augment the original data. An improved deep convolutional generative adversarial network (MDCGAN) is built and trained. The original grayscale image and time-frequency image are input into MDCGAN to obtain the generated grayscale image and time-frequency image. The original image and the generated image are measured and filtered using two evaluation metrics, FID and MMD. The original image and the filtered image are then balanced to obtain the augmented dataset, which is used for subsequent monitoring and recognition networks.

[0090] Step 2.1: Build an improved deep convolutional generative adversarial network (MDCGAN) and train the improved deep convolutional generative adversarial network (MDCGAN).

[0091] In the basic GAN structure, both the generator and discriminator are multilayer perceptron structures, and they are trained adversarially to complete the process of building the generative adversarial network model. We create a new generative adversarial network, MDCGAN, by improving the model's convolutional structure and loss function.

[0092] Step 2.1.1: Set the model structure parameters and establish an improved deep convolutional generative adversarial network (MDCGAN); on the framework of deep convolutional generative adversarial network, modify its generator and discriminator.

[0093] The generator consists of fully connected layers and two-dimensional transposed convolutional layers; the discriminator consists of fully connected layers and two-dimensional convolutional layers. To accelerate training convergence and avoid overfitting, batch normalization (BN) is added to each convolutional layer of the discriminator. Rectified linear unit (ReLU) and Leaky ReLU are chosen as the activation functions for the convolutional and transposed convolutional layers, respectively. The specific structure is as follows: Figure 4 As shown, both the generator and discriminator optimizers use the Adam adaptive moment estimation algorithm to update the model weights. The Nadam algorithm is introduced, employing second-order moment estimation of the gradient to adaptively adjust the learning rate α. Combining the Adam adaptive moment estimation algorithm and the Nesterov acceleration gradient algorithm, Nadam can iterate and update faster than other optimizers, obtaining the optimal parameter updates. The Nadam update rule is described as follows:

[0094]

[0095]

[0096]

[0097] Where θ is the weight or threshold to be optimized, and m t and These are the first-order momentum deviation and the corrected first-order momentum deviation, respectively, V t and V^ t denoted as the second-order momentum deviation and the corrected second-order momentum deviation, respectively, where t is the current iteration, and β1, β2, and ε are the correction coefficients. t Let be the gradient vector and α be the learning rate. In this training and testing, the learning rates for the generator and discriminator are set to 0.002, and β1 and β2 are set to 0.5 and 0.99, respectively.

[0098] Step 2.1.2: Set the model training parameters, including the number of training iterations, the number of iterations, the learning rate, the batch size, and the dimension of the latent space. At the same time, set the discriminator's loss for distinguishing true from false to binary cross-entropy loss.

[0099] The loss functions for the generator and discriminator are as follows:

[0100]

[0101]

[0102] Where x represents the actual sample data, This represents the sample generated by the generator. data (x) represents the distribution from the real data, p g The distribution derived from the generated data is formed by having prior p z The simple random variable z of (z) is obtained through the G mapping, i.e., z ~ p z (z),x=G(z).

[0103] Step 2.1.3: Start training. First train the discriminator, then train the generator. Input the training set containing real sample data into the discriminator. At the same time, input random noise into the generator to generate pseudo samples. Then input the pseudo samples into the discriminator to implement the adversarial behavior between the generator and the discriminator to train the generator.

[0104] Step 2.2: Input the two-dimensional grayscale image and the two-dimensional time-frequency image into the improved deep convolutional generative adversarial network MDCGAN to generate sample images;

[0105] Step 2.3: Use two metrics, FID and MMD, to evaluate the similarity and diversity of the generated sample images, and determine whether the distribution of the generated sample data is similar to that of the real tool condition sample data; use FID and MMD to evaluate the quality of the artificially generated data, and discard data that does not meet the criteria.

[0106] Specifically, artificially generated data of the same type are sequentially compared with real datasets using FID and MMD measurements, and their average values ​​are calculated. A threshold M is set, and if the value is less than the threshold M, the artificially generated data is considered to meet the standard of similarity with real data. The threshold can be set by calculating the FID and MMD of real data. Data that meet the conditions are used to synthesize a training dataset.

[0107] The generated samples obtained after training a certain number of times are evaluated and compared with real images. The evaluation metrics are FID and MMD; images with better performance on these metrics are saved. After training is complete, the real samples and the selected generated samples are balanced to obtain the augmented dataset, which is used for subsequent monitoring and recognition network models.

[0108] The FID (Fréchet Inception Distance) represents the diversity and quality of the generated image. The smaller the FID, the better the image diversity and quality. FID is defined as:

[0109]

[0110] Where Tr represents the sum of the elements on the diagonal of the matrix, referred to as the trace of the matrix; x and g represent the real image and the generated sample image, respectively; μ x and μ g ∑ represents the mean features of the real image and the generated image. x and ∑ g It is the covariance matrix of the real image and the generated image.

[0111] The MMD (Mean Distance Distance) is used to measure the distance between two data distributions, i.e., the similarity between the two data points. The smaller the MMD value, the closer the two distributions are, and the better the quality of the generated image. MMD is defined as:

[0112]

[0113] In the formula, F represents the chosen kernel function, typically a Gaussian kernel function, k(x,y) is the kernel mapping result between x and y, and m and n are the distributions P and N, respectively. r(x) and P g(y) The number of actual samples and generated samples. x i and y i These are the real sample dataset and the generated sample dataset, respectively.

[0114] Step 2.4: Compare with other GAN models to verify the effectiveness of the data augmentation of the proposed MDCGAN.

[0115] Step 3, Monitoring and Recognition: Construct a multi-source feature fusion network (MFFNet). Select a set proportion of original real images and generated images after evaluation and filtering to form an enhanced dataset. Input this dataset into the MFFNet network for feature extraction and classification. Compare it with other deep network models.

[0116] Step 3.1: Construct the multi-source feature fusion network MFFNet;

[0117] The Multi-Source Feature Fusion Network (MFFNet) refers to a neural network model for tool wear monitoring and identification where grayscale images and time-frequency images are input into the model via dual channels. Features from different modalities are extracted, and the feature vectors are concatenated to achieve feature fusion. Finally, the loss is minimized to achieve identification and classification.

[0118] The original real images and the generated images after evaluation and screening are combined to form an enhanced dataset, which is then input into the multi-source feature fusion network MFFNet for feature extraction and fusion.

[0119] The multi-source feature fusion network MFFNet includes an encoding denoising module and a parameterless attention mechanism module.

[0120] The proposed encoding denoising module structure addresses the issue of noise present in images input into neural network models. Noise reduction is necessary to eliminate its interference with training. This paper proposes an encoding denoising module that uses an adaptive threshold to remove noise. The network structure of the encoding denoising module is as follows: Figure 5 As shown. The formula for calculating the adaptive threshold is:

[0121]

[0122]

[0123] Where τ is the threshold, x is the input, and y is the output. α∈[0,1] is the contraction factor, and ρ is the adjustment factor.

[0124] The parameterless attention mechanism module structure (SimAM) involves mining the information of interest through the parameterless attention mechanism module. The attention mechanism is used to redistribute the weights of feature information to weaken redundant features. The multimodal feature extraction module based on the parameterless attention mechanism is as follows: Figure 6 As shown.

[0125] Existing attention-based modules suffer from two problems. First, they can only refine features in one dimension—either channel or spatial—lacking flexibility in situations where both channels and spatial dimensions change simultaneously. Second, their structures often require a series of complex operations, such as pooling. The module proposed in this paper, based on well-developed neuroscience theory, effectively addresses these two issues. Specifically, to enable the network to learn more discriminative neurons, it proposes directly inferring three-dimensional weights from the current neurons and then using this inverse to optimize those neurons. To effectively infer the three-dimensional weights, an energy function E is defined based on neuroscience knowledge. t Then, its analytical solution was obtained.

[0126]

[0127]

[0128]

[0129] Where λ is the regularization term, and t and x i These are the characteristics of the target neuron and other neurons in each channel. The mean of all neurons in a single channel. Let H be the variance of all neurons on a single channel, and H and W be the height and width of the feature map.

[0130] The entire process of refining the parameterless attention mechanism module is as follows:

[0131]

[0132] Where X is the input feature tensor, Sigmoid(·) is used to constrain excessively large values, it does not affect the relative importance of each neuron because it is a monotonic function, and ⊙ is the Hadamard product.

[0133] Step 3.2: Tool detection and identification using the multi-source feature fusion network MFFNet:

[0134] The enhanced dataset is input into MFFNet for tool wear state identification and classification. To verify the effectiveness of the proposed MFFNet model, it is compared with other deep learning network models, such as CNN, LSTM, and GRU. Experimental results show that MFFNet achieves a classification accuracy of 99%, which is superior to other deep learning models.

[0135] Step 4, Transfer Learning: By transferring the parameters of the multi-source feature fusion network MFFNet, monitoring and diagnosis can be achieved across different datasets;

[0136] First, training is performed using source domain data. In this embodiment, the source domain data uses the PHM2010 tool dataset. The network layer structure of the Multi-Source Feature Fusion Network (MFFNet) is frozen, and the parameters of MFFNet are extracted. Then, the parameters of the MFFNet classifier are fine-tuned using target domain data to achieve parameter transfer of MFFNet. In this embodiment, the target domain data is the tool data collected experimentally. Figure 7 This is a diagram showing the monitoring and identification results in an embodiment of the present invention; Figure 8 This is a graph showing the monitoring and identification accuracy and loss in an embodiment of the present invention;

[0137] Specifically, the fine-tuning involves saving the pre-trained weight parameters from the source domain and, during training in the target domain, not starting from scratch, but continuing training based on the weights of the pre-trained multi-source feature fusion network (MFFNet) from the source domain. The migration detection and diagnostic framework is as follows: Figure 9 As shown.

[0138] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for monitoring tool wear status using data augmentation and multi-feature fusion under small sample conditions, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: Select the cutting force signal, convert the one-dimensional cutting force signal into a two-dimensional grayscale image, and convert the one-dimensional cutting force signal into a two-dimensional time-frequency image through continuous wavelet transform; Step 2: Use a generative adversarial network to augment the original data; build an improved deep convolutional generative adversarial network MDCGAN and train it. Input the original grayscale image and time-frequency image into MDCGAN to obtain the generated grayscale image and time-frequency image. Use two evaluation metrics, FID and MMD, to measure and filter the original image and the generated image. Balance the original image and the filtered image to obtain the augmented dataset. Step 3, Monitoring and Identification: Construct a multi-source feature fusion network MFFNet, select a set proportion of original real images and generated images after evaluation and screening to form an enhanced dataset, and input it into the multi-source feature fusion network MFFNet for feature extraction and identification classification; Step 4, Transfer Learning: By transferring the parameters of the multi-source feature fusion network MFFNet, monitoring and diagnosis can be achieved across different datasets; The conversion formula for the two-dimensional grayscale image mentioned in step 1 is as follows: （1）； in, This represents the rounding function, which normalizes all pixel values ​​to a range of 0 to 255. This represents the pixel value of each pixel after conversion, M is the side length of the two-dimensional image, L is an arbitrarily selected length in the one-dimensional cutting force signal, and L is a continuous M×M sampling point; The continuous wavelet transform (CWT) is used to transform each time-domain waveform into the corresponding time spectrum. The Morlet wavelet function is selected, and the transformation principle is shown in formulas (2)-(4). （3）； （4）； in, For the basic mother wavelet function, A set of time-scale wavelets, and These are the scaling factor and the translation factor, respectively, for any signal function. The corresponding continuous wavelet function is defined by formula (3), where Mother wavelet function The complex conjugate function, for and The inner product; Morlet is selected as the wavelet mother function, and its expression is given by formula (4), where is the center frequency of the wavelet.

2. The tool wear state monitoring method with data augmentation and multi-feature fusion under small sample conditions according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Build an improved deep convolutional generative adversarial network (MDCGAN) and train the improved deep convolutional generative adversarial network (MDCGAN). Step 2.1.1: Set the model structure parameters and establish an improved deep convolutional generative adversarial network (MDCGAN); on the framework of deep convolutional generative adversarial network, modify its generator and discriminator. The generator consists of fully connected layers and two-dimensional transposed convolutional layers; the discriminator consists of fully connected layers and two-dimensional convolutional layers; batch normalized (BN) is added to each convolutional layer of the discriminator, and ReLU and Leaky ReLU are selected as the activation functions for the convolutional and transposed convolutional layers; the optimizers of both the generator and the discriminator use adaptive moment estimation (Adam) to update the model weights; the Nadam algorithm is introduced, which uses second-order moment estimation of the gradient to achieve adaptive adjustment of the learning rate α. The update rule of Nadam is described as follows: （5）； （6）； （7）； in, These are the weights or thresholds to be optimized. and These are the first-order momentum deviation and the corrected first-order momentum deviation, respectively. and These are the second-order momentum deviation and the corrected second-order momentum deviation, respectively, where t is the current iteration. and It is a correction factor; The gradient vector, The learning rate; Step 2.1.2: Set the model training parameters, including the number of training iterations, the number of iterations, the learning rate, the batch size, and the dimension of the latent space. At the same time, set the discriminator's loss for distinguishing true from false to binary cross-entropy loss. The loss functions for the generator and discriminator are as follows: （8）； （9）； Where x represents the actual sample data, This represents the samples generated by the generator; For the distribution derived from real data, The distribution derived from the generated data is based on prior knowledge. Simple random variables pass The result of the mapping, i.e. ; Step 2.1.3: Start training. First train the discriminator, then train the generator. Input the training set containing real sample data into the discriminator. At the same time, input random noise into the generator to generate pseudo samples. Then input the pseudo samples into the discriminator to implement the adversarial behavior between the generator and the discriminator to train the generator. Step 2.2: Input the two-dimensional grayscale image and the two-dimensional time-frequency image into the improved deep convolutional generative adversarial network MDCGAN to generate sample images; Step 2.3: Use two metrics, FID and MMD, to evaluate the similarity and diversity of the generated sample images, and determine whether the distribution of the generated sample data is similar to that of the real tool condition sample data; use FID and MMD to evaluate the quality of the artificially generated data, and discard data that does not meet the criteria. Specifically, the same type of artificially generated data is compared with the real dataset in turn by FID and MMD, and the average value is calculated. At the same time, a threshold M is set. If the value is less than the threshold M, the artificially generated data is considered to meet the standard of similarity with the real data. Data that meet the conditions are used to synthesize a training dataset. The FID represents the diversity and quality of the generated image. The smaller the FID, the better the image diversity and the better the quality. FID is defined as: （10）； in, The sum of the elements on the diagonal of the matrix is ​​called the trace of the matrix; x and g represent the real image and the generated sample image, respectively. and This represents the mean of features for the real image and the generated image. and It is the covariance matrix between the real image and the generated image; The MMD is used to measure the distance between two data distributions, i.e., the similarity between the two data points; the smaller the MMD value, the closer the two distributions are, and the better the quality of the generated image. MMD is defined as: （11）； In the formula, F represents the chosen kernel function, k(x, y) is the kernel mapping result of x and y, and m and n are the distributions respectively. and The number of actual samples and generated samples in the sample. and These are the real sample dataset and the generated sample dataset, respectively.

3. The tool wear state monitoring method with data augmentation and multi-feature fusion under small sample conditions according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Construct the multi-source feature fusion network MFFNet; The original real images and the generated images after evaluation and screening are combined to form an enhanced dataset, which is then input into the multi-source feature fusion network MFFNet for feature extraction and fusion. The multi-source feature fusion network MFFNet includes an encoding noise reduction module and a parameterless attention mechanism module; The encoding and noise reduction module: sets an adaptive threshold to remove noise information; The parameterless attention mechanism module: uses the attention mechanism to redistribute the weights of feature information; An energy function is defined. Then, its analytical solution was obtained; （14）； （15）； （16）； Among them, For regularization terms, t and These are the characteristics of the target neuron and other neurons in each channel; The mean of all neurons in a single channel. The variance of all neurons on a single channel. The height and width of the feature map; The entire process of refining the parameterless attention mechanism module is as follows: （17）； in, Given the input feature tensor, For activation function, The product of Hadamard; Step 3.2: Multi-source feature fusion network MFFNet for tool monitoring and identification: The enhanced dataset is input into MFFNet to identify and classify tool wear status.

4. The tool wear state monitoring method with data augmentation and multi-feature fusion under small sample conditions according to claim 1, characterized in that, Step 4 specifically includes: First, training is performed using source domain data, the network layer structure of the multi-source feature fusion network MFFNet is frozen, and the parameters of the multi-source feature fusion network MFFNet are extracted. Then, the parameters of the classifier of the multi-source feature fusion network MFFNet are fine-tuned using target domain data to achieve parameter transfer of the multi-source feature fusion network MFFNet. The fine-tuning specifically involves saving the trained weight parameters from the source domain and, when training in the target domain, not starting from scratch, but continuing to train based on the weights of the source feature fusion network MFFNet trained in the source domain.

Images