Multi-source signal tool wear amount prediction method and system based on one-dimensional residual network

Abstract

The application belongs to the technical field of industrial monitoring, and specifically discloses a multi-source signal tool wear amount prediction method and system based on a one-dimensional residual network, which realizes full-process monitoring from signal collection to tool wear amount prediction; a signal collection module is arranged, the precision of signal analysis is improved through multi-sensor signal conversion and wavelet transformation; a data processing module is arranged, the data imbalance and small sample problems are solved through data balancing operation, and the effectiveness of data processing is improved; a complex relationship processing module is arranged, the complex relationship between multi-source signals is processed through the construction of a one-dimensional residual network, and comprehensive signal analysis is realized; a feature enhancement module is arranged, the feature extraction capability is enhanced through the introduction of a multi-head attention mechanism, and the accuracy of tool wear amount prediction is improved; a task execution module is arranged, tool wear amount prediction is carried out through the model combining the one-dimensional residual network and the multi-head attention mechanism, and the detailed analysis capability of fault detection is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial monitoring, in particular to a multi-source signal tool wear amount prediction method and system based on a one-dimensional residual network. BACKGROUND

[0002] High-precision numerical control machine tools are one of the important equipment indispensable in the development of advanced manufacturing industry. The tool of the numerical control machine tool inevitably produces wear during cutting process and part friction, and the wear state of the tool directly affects the machining quality of the part surface. Once the tool is seriously worn, the machined parts will be unqualified or even scrapped, causing significant losses to the enterprise. Therefore, it is necessary to monitor the wear state of the tool of the numerical control machine tool in real time during the machining process. The current monitoring method is mainly to predict the tool wear value by using the vibration signal generated during the machining process. This method has great limitations and inaccuracy because the signal source is single and is easily affected by noise and environmental disturbances in complex cutting environments. At the same time, most of the collected data is the data of the normal wear state of the tool, and the data of the serious wear is rare. The unbalanced data under the single signal causes insufficient prediction accuracy, which is not conducive to accurately predicting the tool wear value. With the continuous development of deep learning theory, the multi-signal fusion method has obvious advantages in feature extraction and pattern recognition, and has great application potential for tool wear prediction under different working conditions of unbalanced data sets. SUMMARY

[0003] The purpose of the present application is to solve the technical problems existing in the background art. To this end, a multi-source signal tool wear amount prediction method and system based on a one-dimensional residual network are provided, which improves the accuracy and reliability of tool wear amount prediction by fusing multi-sensor signals and introducing complex relationship processing and feature enhancement technology.

[0004] In order to achieve the above purpose, the technical solutions adopted by the present application are as follows:

[0005] The multi-source signal tool wear amount prediction method based on the one-dimensional residual network comprises:

[0006] Step 1: Construct a signal acquisition module to convert the collected multi-sensor signals into a frequency spectrum and perform wavelet transform;

[0007] Step 2: Construct a data processing module to balance the data and solve the problems of unbalanced data and small sample data;

[0008] Step 3: Construct a complex relationship processing module to build a one-dimensional residual network structure and process the complex relationship between multi-source signals;

[0009] Step 4: Construct a feature enhancement module to introduce a multi-head attention mechanism and enhance the feature extraction capability of the model;

[0010] Step 5: Build a task execution module and use a model that combines a one-dimensional residual network and a multi-head attention mechanism to perform the tool wear prediction task.

[0011] The following is a technical solution further defined in the method of the present invention. The construction process of the signal acquisition module in step 1 includes:

[0012] Step 11: Collect signals through multiple sensors, filter the audio information, and then perform a fast Fourier transform (FFT) to obtain a spectrum diagram;

[0013] Step 12: Use the db4 wavelet function to perform wavelet transform on the spectrum graph, so that the signal is decomposed and analyzed at different scales and a wavelet scale graph is obtained;

[0014] Wherein, the wavelet transform in step 12 includes:

[0015] Step 121: performing signal decomposition. First, the signal to be processed is subjected to multi-scale decomposition. The signal is decomposed into components at multiple scales. The low-frequency component reflects the overall trend of the signal, while the high-frequency component reflects the details and mutations of the signal.

[0016] Step 122: Use wavelet filters, one is a low-pass filter and the other is a high-pass filter. The two sets of filters are determined by the coefficients of the db4 wavelet function. The low-pass filter is used to extract the low-frequency components of the signal, while the high-pass filter is used to extract the high-frequency components of the signal.

[0017] Step 123: Convolution operation, convolving the signal with low-pass and high-pass filters to obtain low-frequency and high-frequency coefficients, where the low-frequency coefficients represent the low-frequency components of the signal, and the high-frequency coefficients represent the high-frequency components of the signal;

[0018] Step 124: downsampling, performing a downsampling operation on the convolution result to reduce the frequency so that the scale of the signal is reduced by half, which can also be called subsampling;

[0019] Step 125: Repeat steps 121-125 to decompose the signal into lower-scale components until the required number of decomposition levels is reached; if necessary, the decomposed signal can be reconstructed to restore the original signal. This operation usually requires the use of wavelet coefficients and inverse wavelet filters for inverse transform operations.

[0020] The following is a technical solution further defined in the method of the present invention. The construction process of the data processing module in step 2 includes:

[0021] Step 21: Data preparation, ensuring that the multi-sensor signal dataset contains various characteristics and distribution of the original data;

[0022] Step 22: Design the autoencoder and build the encoder and decoder parts of the autoencoder. The encoder maps the original data to a low-dimensional representation, and the decoder attempts to reconstruct the original data from the low-dimensional representation.

[0023] Step 23: Build a generator and a discriminator to adapt to the characteristics of the sensor signal data. The generator is responsible for generating sample data from random noise, and the discriminator attempts to distinguish between generated fake samples and real samples.

[0024] Step 24: Define the loss function, which consists of the reconstruction loss of the autoencoder and the adversarial loss of the generative adversarial network. The reconstruction loss is used to measure the accuracy of data reconstruction, and the adversarial loss is used to encourage the generator to generate real data samples.

[0025] Step 25: Network training, optimize the model by alternately training the autoencoder and the generative adversarial network, first using the reconstruction loss of the autoencoder to train the autoencoder, and then using the adversarial loss to train the generator and discriminator; this process requires multiple iterations until the model converges or reaches the predetermined training rounds.

[0026] Step 26: Evaluate the data samples generated by the generator to check whether they conform to the distribution characteristics of the original data, and then optimize and adjust the model based on the evaluation results, including adjusting the network structure, loss function and training parameters to improve the quality fidelity of the generated results.

[0027] The following is a technical solution further defined by the method of the present invention, wherein the process of constructing the autoencoder in step 22 includes:

[0028] Step 221: Define the encoder architecture, select a multi-layer perceptron (MLP) structure, and use the ReLU rectified linear unit (RLU) as the activation function. The ReLU function can help the model converge better during training while avoiding the gradient vanishing problem. During the forward propagation process, the input signal data passes through the various hidden layers of the encoder and is gradually converted into a lower-dimensional representation. This process is expressed as follows:

[0029] h=ReLU(W1x+b1)

[0030] z=W2h+b2

[0031] Where x is the input signal data, h is the hidden layer representation, z is the encoded representation, W1 and W2 are the weight matrices of the hidden layer and the encoding layer, and b1 and b2 are the corresponding bias vectors;

[0032] Step 222: Define the decoder framework. The decoder structure, activation function selection, and weight matrix initialization are similar to those of the encoder. During the forward propagation process of the decoder, the decoder receives the representation generated by the encoder as input and attempts to decode it into the original input data. The forward propagation process of the decoder is expressed as follows:

[0033] h′=ReLU(W′1x+b′1)

[0034]

[0035] Where h′ is the hidden layer representation of the decoder, is the output of the decoder, representing the reconstructed input data;

[0036] Step 223: Define the loss function and optimizer. The goal of the autoencoder is to minimize the reconstruction error between the input data and the decoder output data. The mean square error (MSE) is used as the loss function. MSE is defined by the following formula:

[0037]

[0038] Where N is the number of samples, xi is the original input data, is the output data of the decoder;

[0039] Step 224: Backpropagation and parameter update, calculate the gradient of the loss with respect to the parameters through backpropagation, and update the parameters using the gradient descent optimization algorithm.

[0040] The construction process of the generator and discriminator in step 23 includes:

[0041] Step 231: GAN consists of two parts: the generator and the discriminator. The generator receives a random noise vector z and outputs the generated data sample G(z; θ g ):

[0042]

[0043] Where G is the generator function with parameter θ g , used to map the random noise z dimension to generate data samples The discriminator receives real data samples x or generated data samples By parameter θ disc Output the probability D(x;θ) that the sample is real data disc )and Where D is the function of the discriminator, and the parameter is θ disc ;

[0044] Step 232: Discriminator and generator loss function definition, the goal of the discriminator is to maximize the log-likelihood estimate, so that the probability of real samples is close to 1, and the probability of generated samples is close to 0:

[0045]

[0046] The goal of the generator is to minimize the log-likelihood estimate, so that the probability of generated samples is close to 1, and the probability of generated samples being considered real samples by the discriminator is maximized:

[0047]

[0048] The following is a further defined technical solution of the method in the present application, the construction process of network training and loss function definition in step 24 includes:

[0049] Step 241: initialization, randomly initialize the parameters θ e 、θ d of the encoder and decoder, and the parameters θ g 、θ disc of the generator and discriminator;

[0050] Step 242: forward propagation of autoencoder and reconstruction loss calculation, the encoder maps the input data x to the latent space representation z:

[0051] z = E(x; θ e )

[0052] The decoder reconstructs the latent space representation z into the input data

[0053]

[0054] Calculate the reconstruction loss L recon :

[0055]

[0056] Step 243: forward propagation of GAN and calculation of adversarial loss, calculate the discriminator loss L D :

[0057]

[0058] Calculate the generator loss L G :

[0059]

[0060] Step 244: back propagation and parameter update, update the parameters of the encoder and decoder, and minimize the reconstruction loss:

[0061]

[0062]

[0063] Update the discriminator parameters to minimize the discriminator loss:

[0064]

[0065] Update the generator parameters to minimize the generator loss:

[0066]

[0067] Step 245: Iterate training and repeat steps 241-245 until the model converges. Each iteration includes an update of the encoder and decoder, and an update of the generator and discriminator.

[0068] The following is a technical solution further defined in the method of the present invention. The construction process of the complex relationship processing module in step 3 includes:

[0069] Step 31: Collect lathe signal data from multiple sensors and use filters to extract and classify signal features;

[0070] Step 32: Convert the processed time domain signal into a frequency domain signal through Fourier transform, and further extract multi-scale features through wavelet transform;

[0071] Step 33: Construct a one-dimensional residual network (1D ResNet) to capture and process the complex relationships between multi-source signals. By introducing skip connections, the residual network makes the training of deep networks more stable and can effectively integrate and extract the correlation features of multi-source signals.

[0072] In step 31, the process of using filters to extract and classify signal features includes:

[0073] Step 311: Signal preprocessing, which removes noise and highlights the main features of the signal. Use a low-pass filter to remove high-frequency noise:

[0074]

[0075] The signals are then normalized to keep their amplitudes within a standard range:

[0076]

[0077] Where μ and σ are the mean and standard deviation of the signal, respectively.

[0078] Step 312: Extract time domain features from the signal. These features reflect the statistical properties of the signal. The time domain features include: maximum value, minimum value, peak-to-peak value, mean value, variance, standard deviation, mean square value, kurtosis, skewness, shape factor, crest factor, impulse factor, margin factor, and clearance factor.

[0079] Maximum: The maximum value in the data set, representing the largest observation in the data.

[0080] Maximum=max(x1,x2,...,xn)

[0081] Minimum: The smallest value in the data set, representing the smallest observed value in the data.

[0082] Minimum=min(x1,x2,...,xn)

[0083] Peak-to-peak value: The peak-to-peak value is the difference between the maximum and minimum values ​​and represents the range of the data.

[0084] Peak-to-Peak=max(x1,x2,...,xn)-min(x1,x2,...,xn)

[0085] Mean: The mean is the arithmetic average of all the observations in a data set and is used to indicate the central tendency of the data.

[0086]

[0087] Variance: Variance is the average of the squares of the differences between the observations and the mean. It indicates the dispersion or spread of the data.

[0088]

[0089] Standard Deviation: Standard deviation is the square root of the variance. It measures the dispersion of the data relative to the units of the mean.

[0090]

[0091] Mean Squared Value: The mean squared value is the average of the squares of all the observations in a data set, then taking the square root. It is often used to represent the root mean square magnitude of a signal or waveform, similar to the RMS value.

[0092]

[0093] Kurtosis: Kurtosis is used to describe the tail and peak of a data distribution. Positive kurtosis indicates that the data distribution has a heavier tail and more extreme values, while negative kurtosis indicates that the data distribution has a lighter tail and more data is concentrated around the mean.

[0094] Kurtosis=(1 / n)*∑((xi-Mean) / Standard Deviation)^4-3

[0095] Skewness: Skewness is used to describe the asymmetry of the data distribution. Positive skewness indicates that the data distribution is skewed to the right, negative skewness indicates that the data distribution is skewed to the left, and zero skewness indicates that the data distribution is symmetrical.

[0096] Skewness=(1 / n)*∑((xi-Mean) / Standard Deviation)^3

[0097] Form Factor: Form factor measures the relationship between the peak value and the RMS value of a signal waveform. It is often used to analyze the peak properties of a signal. A larger form factor indicates a signal with a sharper peak.

[0098] Crest Factor=Peak Value / RMS Value

[0099] Crest Factor: Crest factor is also used to measure the relationship between the peak value and the RMS value of a signal waveform. A larger crest factor indicates that the signal has a larger peak value.

[0100] Peak Factor=Peak Value / RMS Value

[0101] Pulse factor: The pulse factor is used to describe the impulse nature of the signal, that is, the relationship between the short-term peak value and the root mean square value of the signal. It is usually used to analyze transient signals. A larger pulse factor indicates that the signal contains more transient peaks.

[0102] Pulse Factor=Peak Value / RMS Value

[0103] Margin factor: Margin factor is used to describe the relationship between the peak value and the mean value of a signal. It is often used to analyze the margin of a signal. A larger margin factor indicates that the signal has a larger peak value relative to the mean value.

[0104] Crest Factor = Peak Value / Mean Value Crest Factor: The crest factor describes the relationship between a signal's peak value and its range. It is often used to analyze a signal's dynamic range; a larger crest factor indicates a signal with a larger peak value relative to its range.

[0105] Margin Factor=Peak Value / Range

[0106] In step 32, the process of converting the time domain signal into the frequency domain signal through Fourier transform includes:

[0107] Step 321: Convert the time series signal to the frequency domain through Fourier transform and extract the frequency domain features. The frequency domain feature extraction formula is as follows:

[0108] Fourier transform: Convert time domain signals into frequency domain signals through Fourier transform.

[0109]

[0110] Spectral energy: represents the energy distribution of the signal in different frequency bands.

[0111]

[0112] Center frequency: Indicates the center of gravity of the spectrum.

[0113]

[0114] Spectral entropy: represents the complexity of the spectrum.

[0115]

[0116] in,

[0117] Step 322: Time-frequency domain feature extraction, combining the time domain and frequency domain, decomposing the signal into different frequency band energies through wavelet transform, and extracting local features;

[0118]

[0119] Step 323: Combine all extracted features into a feature vector.

[0120] In step 33, the process of constructing the one-dimensional residual network includes:

[0121] Step 331: Input data preparation, using the pre-processed multi-sensor signal data (cutting force signal, vibration signal and acoustic emission signal, etc.) as the input of the network. Assume that the input signal is

[0122] X∈R N×d

[0123] Where N is the time step and d is the dimension of the signal. This step ensures that the model receives processed, high-quality data ready for subsequent feature extraction and analysis.

[0124] Step 332: Construct an initial convolution layer to perform initial convolution on the input signal to extract low-level features. The convolution operation captures local temporal patterns and frequency features. The specific formula is:

[0125] X conv1 =Conv1D(X, W1, b1)

[0126] where Conv1D represents a one-dimensional convolution operation, and W1 and b1 are the convolution kernel weight and bias, respectively. The initial convolution layer serves to reduce the complexity of the input signal and extract preliminary local features, laying the foundation for subsequent deep feature extraction.

[0127] Step 333: Construction of residual blocks, residual blocks are the core part of ResNet, each residual block includes two convolution layers and a shortcut connection. This structure allows the signal to pass through the jump connection directly, thereby alleviating the problem of gradient disappearance and promoting the training of deep networks.

[0128] 1) First convolution layer of residual block:

[0129] X conv2 = ReLU(BatchNorm(Conv1D(X conv1 , W2, b2))

[0130] where BatchNorm represents the batch normalization operation, and ReLU is the activation function. This convolution layer serves to further extract features from the signal and stabilize the training process through batch normalization.

[0131] 2) Second convolution layer of residual block:

[0132] X conv3 = BatchNorm(Conv1D(X conv2 , W3, b3))

[0133] 3) Shortcut connection:

[0134] X res = X conv1 + X conv3

[0135] If the dimensions of the input and output are inconsistent, a linear transformation is needed to match the dimensions:

[0136] X res = X conv1 + Conv1D(X conv1 , W skip , b skip )

[0137] The role of the shortcut connection is to directly add the input signal to the output, ensuring direct information transmission, reducing gradient disappearance, and improving the training effect of the model.

[0138] 4) Activation output:

[0139] X out=ReLu(X res )

[0140] Through the ReLU activation function, nonlinearity is introduced into the model to improve the expressiveness of the model.

[0141] Repeating multiple residual blocks to build a deep network. These residual blocks are used to extract and integrate features layer by layer to capture complex patterns and relationships in the signal.

[0142] Step 334: Global average pooling layer, which reduces feature dimensions and aggregates information. The specific formula is:

[0143]

[0144] The role of the global average pooling layer is to average the features of all time steps to generate a fixed-size feature vector, reducing the dimension while retaining important information.

[0145] Step 335: Construct a fully connected layer and an output layer, input the result of global average pooling into the fully connected layer to obtain the final output. The specific formula is:

[0146] X fc =ReLU(W fc X gap +b fc )

[0147] Y pred =Softmax(W out X fc +b out )

[0148] Among them, W fc and b fc are the weights and biases of the fully connected layer, W out and b out are the weights and biases of the output layer. The fully connected layer performs nonlinear transformation on the pooled features, and the output layer performs classification through the Softmax function to generate the final prediction result.

[0149] The following is a technical solution further defined in the method of the present invention. The process of constructing the feature enhancement module in step 4 includes:

[0150] Step 41: Preprocess the signals, standardize and normalize the signals collected by multiple sensors to ensure that all dimensions of the signals are on the same scale, which helps the model better process the input data;

[0151] Step 42: Linear transformation: input the preprocessed signal into the linear layer to generate query, key and value matrices;

[0152] Q=XW Q , K=XW K ,V=XW V

[0153] Where X is the input signal matrix; W Q , W K , W V is a learnable weight matrix; Q, K, V are query, key, and value matrices respectively.

[0154] Step 43: Calculate the attention score between the query and the key by dot product:

[0155]

[0156] Among them, d k is the dimension of the key, used to scale the dot product result.

[0157] After scaling and normalization, weighted summation is performed. The attention scores are scaled and softmax normalized to ensure the stability and interpretability of the scores. The values ​​are weighted summed using the normalized attention weights to obtain the output of each attention head:

[0158]

[0159] Where h is the number of attention heads. i ,K i ,V i is the query, key, and value matrix of the i-th head.

[0160] Step 44: After concatenating the outputs of the attention heads, perform residual connections and layer normalization. The outputs of all attention heads are concatenated and linearly transformed to generate the final output feature representation:

[0161] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0162] Among them: head i =Attention(Q i ,K i ,V i ),W O is the output weight matrix.

[0163] Then the output and input features of the multi-head attention mechanism are residually connected and layer normalized:

[0164] Y=LayerNorm(X+MultiHead(Q, K, V)).

[0165] The following is a technical solution further defined in the method of the present invention. The construction process of the task execution module in step 5 includes:

[0166] Step 51: Combine the one-dimensional residual network and the multi-head attention mechanism;

[0167] Step 52: Define the loss function of the model, use an optimization algorithm to minimize the loss function, and then iteratively train the model until convergence;

[0168] In step 51, the process of combining the one-dimensional residual network and the multi-head attention mechanism includes:

[0169] Step 511: The one-dimensional residual network is formed by stacking multiple residual blocks in depth to form a deep one-dimensional residual network to enhance the model's ability to model complex signals. The residual block is the basic structure of the network. Each residual block contains two or more one-dimensional convolutional layers. Each convolutional layer is followed by a ReLU activation function and batch normalization, and a skip connection is introduced between the input and output.

[0170] y=F(x,{W i})+x

[0171] Where F(x,{Wi}) represents the convolution and activation operations in the residual block, x is the input, and y is the output. In the one-dimensional convolution layer, multi-sensor signals are preprocessed and then enter feature extraction.

[0172] Step 512: Input the output features of the one-dimensional residual network (1DResNet) into the multi-head attention mechanism to enhance the feature representation through the attention mechanism. The features output by the attention mechanism are residually connected and layer normalized to form the final feature representation:

[0173] Y=LayerNorm(X+MulitiHead(Q, K, V)) The following is a technical solution further defined in the method of the present invention, which includes, in step 52:

[0174] Step 521: Define the loss function and use the mean square error as the loss function to measure the difference between the model prediction value and the actual value:

[0175]

[0176] MSE is more sensitive to large errors and can effectively punish prediction results with large deviations, which helps to improve the accuracy of the model.

[0177] Step 522: Design an optimization algorithm using the Adam optimizer, which is an optimization algorithm that combines momentum and RMSProp. It can adaptively adjust the learning rate of each parameter and is suitable for processing sparse gradients and high-dimensional data.

[0178] m t =β1m t-1 +(1-β1)g t

[0179] v t =β2v t-1 +(1-β2)g 2 t

[0180]

[0181] Among them, mt is the first-order moment estimate, vt is the second-order moment estimate, β1, β2 are the exponential decay rates of the momentum term, g t is the gradient, α is the learning rate, and θt is the model parameter. The Adam optimizer combines the advantages of momentum and RMSProp, can adaptively adjust the learning rate, performs well in processing sparse gradients and high-dimensional data, and can quickly converge to the optimal solution.

[0182] The multi-source signal tool wear prediction system based on one-dimensional residual network is used to implement the above-mentioned multi-source signal tool wear prediction method based on one-dimensional residual network, including:

[0183] Signal acquisition module, used to convert the collected multi-sensor signals into a spectrum and perform wavelet transform;

[0184] Data processing module, used for performing balancing operations on data;

[0185] Complex relationship processing module, used to build a one-dimensional residual network structure to handle complex relationships between multi-source signals;

[0186] Feature enhancement module, which is used to introduce a multi-head attention mechanism to enhance the model's feature extraction capabilities;

[0187] The task execution module uses a model that combines a one-dimensional residual network and a multi-head attention mechanism to perform tool wear prediction tasks.

[0188] Compared with the prior art, the present invention has the following technical effects:

[0189] The present invention realizes full-process monitoring from signal acquisition to tool wear prediction; a signal acquisition module is set up to improve the accuracy of signal analysis through multi-sensor signal conversion and wavelet transform; a data processing module is set up to solve data imbalance and small sample problems through data balancing operations, thereby improving the effectiveness of data processing; a complex relationship processing module is set up to process the complex relationship between multi-source signals by constructing a one-dimensional residual network, thereby realizing comprehensive signal analysis; a feature enhancement module is set up to enhance the feature extraction capability and improve the accuracy of tool wear prediction by introducing a multi-head attention mechanism; a task execution module is set up to predict tool wear by combining a model of a one-dimensional residual network and a multi-head attention mechanism, thereby improving the detailed analysis capability of fault detection.

[0190] The present invention will be further described below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0191] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0192] Figure 1 It is a schematic flow chart of the overall method of the present invention;

[0193] Figure 2 This is the AE-GAN network model structure diagram of the present invention;

[0194] Figure 3 This is the structural diagram of the improved 1DResNet-MHSA model in the present invention;

[0195] Figure 4 This is an example diagram of the data collection process in the present invention. DETAILED DESCRIPTION

[0196] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0197] This embodiment provides a tool wear prediction system based on a multi-source signal of a one-dimensional residual network and a tool wear prediction method based on a multi-source signal of a one-dimensional residual network.

[0198] The system consists of a signal acquisition module, a data processing module, a complex relationship processing module, a feature enhancement module, and a task execution module. The system developed these three modules. The signal acquisition module converts the collected multi-sensor signals into a spectrum and performs a wavelet transform. The data processing module balances the data to address issues such as data imbalance and small sample sizes. The complex relationship processing module constructs a one-dimensional residual network to handle the complex relationships between multi-source signals. The feature enhancement module introduces a multi-head attention mechanism to enhance the model's feature extraction capabilities. The task execution module combines a one-dimensional residual network and a multi-head attention mechanism to predict tool wear. This system addresses the difficulties faced by traditional tool wear prediction technologies in extracting complex relationships between signals from multi-sensor inputs and inefficiently removing highly correlated redundant information. It significantly improves the ability to capture complex relationships between multi-sensor signals, enhancing the accuracy and reliability of wear prediction, and has broad engineering application prospects.

[0199] The method includes the construction of signal acquisition module, data processing module, complex relationship processing module, feature enhancement module and task execution module. The specific steps are as follows: Figure 1 As shown:

[0200] S1: Construct a signal acquisition module to convert the signals collected by multiple sensors into a spectrum and perform wavelet transform.

[0201] S11, such as Figure 4 As shown in Figure 1, a triaxial dynamometer is installed between the workpiece and the worktable to measure the three-dimensional cutting forces. Three piezoelectric accelerometers and an acoustic emission sensor are mounted on the workpiece to measure the cutting force signals, vibration signals, and acoustic emission signals generated in the X, Y, and Z directions during high-speed CNC machine tool cutting.

[0202] S12. The signals collected by the multiple sensors are first filtered to remove noise and interference to ensure signal clarity and accuracy. The filtered signals are then subjected to a fast Fourier transform (FFT) to convert the time domain signals into frequency domain signals, resulting in a spectrum.

[0203] S13. Perform a wavelet transform on the spectrum using the db4 wavelet function, decomposing and analyzing the signal at different scales to produce a wavelet scalogram. Specifically, the multi-scale decomposition technique decomposes the signal into components at different scales. Low-frequency components reflect the overall trend of the signal, while high-frequency components reflect details and mutations.

[0204] S14. The filtered signal is then convolved to obtain low-frequency coefficients that represent the overall trend of the signal and high-frequency coefficients that reflect signal details and mutations. Finally, the result is downsampled to reduce the scale of the signal and thus the frequency. This decomposition step is repeated until the signal is decomposed to the desired number of decomposition levels, i.e., the target scale.

[0205] S2: If Figure 2 As shown in Figure 2, a data processing module is constructed to solve the problems of small samples and sample imbalance that occur in most offline training tasks.

[0206] S21. Organize the multi-sensor signal data to ensure that the dataset contains the relevant features. Then clean the data to remove outliers and fill in missing values ​​to ensure the integrity and consistency of the data points.

[0207] S22. Build an autoencoder module, which includes an encoder and decoder. First, the high-order input data (the processed features) is fed into the encoder's input layer. After the input layer receives the data, it is passed to the hidden layer for linear transformation and ReLU activation. The hidden layer extracts the input data's features and compresses them into a low-dimensional representation, which is then passed to the output layer. The encoder's output layer passes the low-dimensional representation to the decoder. The first layer of the decoder uses linear transformation and ReLU activation to convert the low-dimensional representation into a hidden representation. The hidden representation is then further converted to a high-dimensional representation, and the original input is finally reconstructed through the output layer.

[0208] S23. The generator and discriminator are described below. The generator is fed a random noise vector (a vector that follows a standard normal distribution) as input into its first hidden layer. The generator's first hidden layer transforms the random noise into a preliminary hidden representation through a linear transformation and a ReLU activation function. This hidden representation is then processed through a further linear transformation and a ReLU activation function to generate a more complex hidden representation. Finally, the output layer uses a tanh activation function to convert the hidden representation into generated sample data, ensuring that the generated data values ​​are within a reasonable range. The discriminator is then trained with real data, and the generated data is then fed into the discriminator. The discriminator's first hidden layer transforms the input data into a hidden representation through a linear transformation and a ReLU activation function. The hidden representation is further processed through a linear transformation and a ReLU activation function to extract deeper features. The output layer uses a sigmoid activation function to calculate the probability that the sample is real data. The generator and discriminator are trained to improve the quality of the generated data through adversarial training. The discriminator attempts to distinguish between real and generated data, while the generator attempts to trick the discriminator into misidentifying the generated data as real.

[0209] S24. Next, define the loss functions for both. The reconstruction loss of the autoencoder measures the difference between the reconstructed data and the original data, typically using the mean squared error (MSE). The loss function for the generative adversarial network includes the discriminator loss and the generator loss. The discriminator loss measures the discriminator's ability to distinguish between real data and generated data, while the generator loss encourages the generator to produce samples that are closer to real data.

[0210] S25. The model parameters are then continuously optimized by alternately training the autoencoder and the generative adversarial network. First, the encoder and decoder are trained using the autoencoder's reconstruction loss to minimize the reconstruction error. Next, the generator and discriminator are trained using the adversarial loss. By minimizing the discriminator loss and the generator loss, the parameters of the generator and discriminator are alternately optimized until the model converges or the predetermined number of training rounds is reached. Finally, the data samples generated by the generator are evaluated to check whether they conform to the distribution characteristics of the original data. Based on the evaluation results, the model is optimized and adjusted to ensure the quality and fidelity of the generated data. If necessary, the model structure, training parameters, and loss function can be adjusted to improve the generation results.

[0211] S3: Build a complex relationship processing module that can capture the complex relationships between signals and improve the model's feature extraction capabilities, thereby providing high-quality feature input for subsequent feature enhancement and task execution modules.

[0212] S31. First, preprocess the multi-sensor signals. Use a bandpass filter to remove high-frequency noise and low-frequency drift to ensure signal purity. Normalize the filtered signal to scale the signal amplitude to a uniform range, improving the stability of subsequent feature extraction and model training. Then, extract time-domain features such as maximum and minimum values ​​and variance. Perform a Fourier transform to convert the time-domain signal into a frequency-domain signal, generating a spectrogram. Calculate spectral energy to analyze the signal's energy distribution across different frequency bands. Finally, extract frequency-domain features such as center frequency and spectral entropy.

[0213] S32, such as Figure 3 As shown in Figure 2, the preprocessed signal is fed into a one-dimensional residual network to capture and process the complex relationships between multiple source signals. By introducing skip connections, the residual network makes the training of deep networks more stable and can effectively integrate and extract the correlation features of multiple source signals.

[0214] S33. Construction of a one-dimensional residual network: By stacking multiple residual blocks, each residual block contains two convolutional layers and a shortcut connection. The first convolutional layer extracts the local features of the input signal through a one-dimensional convolution operation, and uses batch normalization and ReLU activation functions for nonlinear transformation. The second convolutional layer further extracts the deep features of the signal, also using batch normalization and ReLU activation functions. The shortcut connection adds the input signal directly to the output of the second convolutional layer to form a residual output. By stacking multiple residual blocks, a deep one-dimensional residual network is constructed. Each residual block can further extract more complex signal features based on the previous residual block.

[0215] S34. Use the global average pooling layer to perform pooling on the output of the residual block. The function of the global average pooling layer is to average the features of all time steps to generate a fixed-size feature vector, reducing the dimension while retaining important information.

[0216] S4: Build a feature enhancement module and introduce a multi-head attention mechanism to enhance the model's feature extraction capabilities.

[0217] S41. Based on the preprocessed signal, the signal is first standardized and normalized to ensure that all dimensions of the signal are on the same scale, which helps the model better process the input data. The standardized and normalized signal is input into the multi-head attention mechanism.

[0218] S42. In the multi-head attention mechanism, the input signal is passed through multiple self-attention layers. Each self-attention layer generates different feature representations that capture the relevance of the signal in different contexts.

[0219] S43. Perform a linear transformation on the output of each self-attention layer to adjust the feature dimension. Then, concatenate the outputs of all self-attention layers to generate a comprehensive feature vector.

[0220] S44. The concatenated feature vector is passed through a fully connected layer to further process and transform the comprehensive features. The output of this fully connected layer is the enhanced feature representation, which can better reflect the complex relationships and important features of the signal.

[0221] S45. After the fully connected layer is processed, the enhanced feature representation is added to the original input signal through a residual connection. Then, layer normalization is performed. This step helps stabilize and accelerate the network training process while preventing gradients from vanishing or exploding.

[0222] Finally, the enhanced feature representation is used as the model input for the subsequent task execution module. By introducing a multi-head attention mechanism, the feature enhancement module can capture the correlation of signals in different contexts, improve the model's feature extraction capability and prediction accuracy, and thus improve the accuracy and reliability of tool wear prediction.

[0223] S5: Build a task execution module and use a model that combines a one-dimensional residual network and a multi-head attention mechanism to perform tool wear prediction tasks.

[0224] S51. First, after the signal goes through preprocessing, data processing, complex relationship processing, and feature enhancement modules, high-quality feature representations are generated. These feature representations are used as input to the task execution module.

[0225] S52. Input the features into the one-dimensional residual network. A one-dimensional residual network is constructed by stacking multiple residual blocks in depth, forming a deep one-dimensional residual network to enhance the model's ability to model complex signals. Each residual block contains two or more one-dimensional convolutional layers, each followed by a ReLU activation function and batch normalization. By introducing skip connections, information is effectively transferred within the deep network, mitigating the vanishing gradient problem and improving the stability and effectiveness of model training.

[0226] S53. Input the output features of the one-dimensional residual network into the multi-head attention mechanism. The multi-head attention mechanism increases the model's focus on key features by assigning weights to different time steps. The specific steps are: first, perform a linear transformation to generate query, key, and value matrices; then, calculate the attention score between the query and key using a dot product, and scale and normalize the attention score; finally, use the normalized attention weights to weight the sum of the values ​​to obtain the output of each attention head.

[0227] S54: Concatenate the features output by the multi-head attention mechanism, perform residual connections, and perform layer normalization to form the final feature representation. This step ensures the comprehensive expression of features at different levels and improves the model's predictive ability.

[0228] S55: Input the final feature representation into the fully connected layer to predict tool wear. The fully connected layer maps the feature vector to the prediction result and outputs the predicted value of wear through the activation function.

[0229] S56. Define the model's loss function, using mean squared error (MSE) as the loss function to measure the difference between the predicted value and the actual wear amount. Use the Adam optimizer to minimize the loss function. Use the backpropagation algorithm to calculate the gradient of the loss with respect to the model parameters and update the model parameters. Iterate the model training until the loss function converges or the predetermined number of training rounds is reached. During training, monitor the model's performance metrics to ensure good performance on the training and validation sets.

[0230] The present invention provides a tool wear prediction model based on a one-dimensional residual network and a multi-sensor signal (multi-source signal) model, which integrates multi-dimensional signals such as cutting force, vibration, acoustic emission and temperature field. Through feature dimensionality reduction, the multimodal input is converted into two-dimensional data, thereby accelerating the data processing speed and providing efficient input for offline training and online prediction. By combining the residual network with the multi-head attention mechanism, the model accurately removes complex redundant information in the multi-dimensional signal and extracts deeper dependency patterns. In order to solve the common problems of small samples and sample imbalance in offline training, feature enhancement technology is introduced, and automatic encoding adversarial generative network is used for optimization. This model covers the entire process from signal acquisition, data processing to wear prediction, significantly improving the ability to capture complex relationships between multi-sensor signals, enhancing the accuracy and reliability of wear prediction, and has broad engineering application prospects.

[0231] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any person skilled in the art can utilize the methods and technical contents disclosed above to make many possible variations and modifications to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, or modify them into equivalent embodiments with equivalent variations. Therefore, any equivalent variations made in accordance with the shape, structure, and principles of the present invention without departing from the content of the technical solutions of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-source signal tool wear prediction method based on one-dimensional residual network, characterized by: include: Step 1: Build a signal acquisition module to convert the collected multi-sensor signals into a spectrum and perform wavelet transform; Step 2: Build a data processing module to perform balancing operations on the data; Step 3: Construct a complex relationship processing module to build a one-dimensional residual network structure to handle the complex relationship between multi-source signals; Step 4: Construct a feature enhancement module and introduce a multi-head attention mechanism to enhance the model's feature extraction capabilities; The construction process of the feature enhancement module includes: Step 41: Preprocess the signal, standardize and normalize the signals collected by multiple sensors to ensure that all dimensions of the signal are on the same scale; Step 42: Linear transformation: input the preprocessed signal into the linear layer to generate query, key and value matrices; Step 43: Calculate the attention score. Calculate the attention score between the query and the key by dot product. Scale and softmax normalize the attention score to ensure the stability and interpretability of the score. Use the normalized attention weight to weight the sum of the values ​​to get the output of each attention head. Step 44: Concatenate the outputs of all attention heads and perform a linear transformation to generate the final output feature representation. Then, perform a residual connection between the output and input features of the multi-head attention mechanism and perform layer normalization. Step 5: Build a task execution module and use a model that combines a one-dimensional residual network and a multi-head attention mechanism to perform the tool wear prediction task.

2. The multi-source signal tool wear prediction method based on one-dimensional residual network according to claim 1 is characterized in that: The construction process of the signal acquisition module includes: Step 11: The signals are collected by multiple sensors and filtered to remove noise and interference. The filtered signals are then fast Fourier transformed to convert the time domain signals into frequency domain signals to obtain a spectrum diagram. Step 12: Use the db4 wavelet function to perform wavelet transform on the spectrum graph, so that the signal is decomposed and analyzed at different scales and a wavelet scale graph is obtained; Among them, wavelet transform includes: Step 121: performing signal decomposition. First, the signal to be processed is subjected to multi-scale decomposition. The signal is decomposed into components at multiple scales. The low-frequency component reflects the overall trend of the signal, while the high-frequency component reflects the details and mutations of the signal. Step 122: Use wavelet filters, one is a low-pass filter and the other is a high-pass filter. The two sets of filters are determined by the coefficients of the db4 wavelet function. The low-pass filter is used to extract the low-frequency components of the signal, and the high-pass filter is used to extract the high-frequency components of the signal. Step 123: Convolve the signal with low-pass and high-pass filters to obtain low-frequency and high-frequency coefficients; Step 124: downsampling the convolution result to reduce the frequency so that the scale of the signal is reduced by half; Step 125: Repeat steps 121-125 to decompose the signal into lower-scale components until the desired number of decomposition levels is reached.

3. The tool wear prediction method based on multi-source signals of one-dimensional residual network according to claim 1, characterized in that: The construction process of the data processing module includes: Step 21: Data preparation, ensuring that the multi-sensor signal dataset contains various characteristics and distribution of the original data; Step 22: Design the autoencoder and build the encoder and decoder parts of the autoencoder. The encoder maps the original data to a low-dimensional representation, and the decoder attempts to reconstruct the original data from the low-dimensional representation. Step 23: Construct a generator and a discriminator to adapt to the characteristics of sensor signal data. The generator is responsible for generating sample data from random noise, and the discriminator distinguishes between generated fake samples and real samples. Step 24: Define the loss function, which consists of the reconstruction loss of the autoencoder and the adversarial loss of the generative adversarial network. The reconstruction loss is used to measure the accuracy of data reconstruction, and the adversarial loss is used to encourage the generator to generate real data samples. Step 25: Network training, optimize the model by alternately training the autoencoder and the generative adversarial network, first using the reconstruction loss of the autoencoder to train the autoencoder, and then using the adversarial loss to train the generator and discriminator; Step 26: Evaluate the data samples generated by the generator to check whether they conform to the distribution characteristics of the original data, and then optimize and adjust the model based on the evaluation results, including adjusting the network structure, loss function and training parameters to improve the quality fidelity of the generated results.

4. The tool wear prediction method based on multi-source signals of one-dimensional residual network according to claim 3 is characterized in that: The construction process of the autoencoder includes: Step 221: Define the encoder architecture. Select a multi-layer perceptron (MLP) structure and use the ReLU rectified linear unit (ReLU) as the activation function. During the forward propagation process, the input signal data passes through the various hidden layers of the encoder and is gradually converted to a lower-dimensional representation. This process is expressed as follows: h=ReLU(W1x+b1) z=W2h+b2 Where x is the input signal data, h is the hidden layer representation, z is the encoded representation, W1 and W2 are the weight matrices of the hidden layer and the encoding layer, and b1 and b2 are the corresponding bias vectors; Step 222: Define the decoder framework, decoder structure, and activation function selection. During the forward propagation process of the decoder, the decoder receives the representation generated by the encoder as input and decodes it into the original input data. The forward propagation process of the decoder is expressed as follows: h′=ReLU(W′1x+b′1) Where h′ is the hidden layer representation of the decoder, is the output of the decoder, representing the reconstructed input data; Step 223: Define the loss function and optimizer. The goal of the autoencoder is to minimize the reconstruction error between the input data and the decoder output data. The mean square error (MSE) is used as the loss function. MSE is defined by the following formula: Where N is the number of samples, xi is the original input data, is the output data of the decoder; Step 224: Backpropagation and parameter update, calculating the gradient of the loss with respect to the parameters through backpropagation and updating the parameters using the gradient descent optimization algorithm; The construction process of the generator and discriminator includes: Step 231: GAN consists of two parts: the generator and the discriminator. The generator receives a random noise vector z and outputs the generated data sample G(x; θ g ): Where G is the generator function with parameter θ g , used to map the random noise z dimension to generate data samples The discriminator receives real data samples x or generated data samples By parameter θ disc Output the probability D(x;θ) that the sample is real data disc )and Where D is the function of the discriminator, and the parameter is θ disc ; Step 232: The loss functions of the discriminator and the generator are defined. The goal of the discriminator is to maximize the log-likelihood estimate so that the probability of the real sample is close to 1 and the probability of the generated sample is close to 0. The goal of the generator is to minimize the log-likelihood estimate so that the probability of the generated sample is close to 1 and maximize the probability that the generated sample is considered to be a real sample by the discriminator.

5. The tool wear prediction method based on multi-source signals of one-dimensional residual network according to claim 3, characterized in that: The construction process of network training and defining loss function includes: Step 241: Randomly initialize the parameters θ of the encoder and decoder e ,θ d , the parameters θ of the generator and discriminator g ,θ disc ; Step 242: After forward propagation and reconstruction loss calculation from the encoder, the encoder maps the input data x to the latent space representation z: z=E(x;θ e ) The decoder reconstructs the latent space representation z into the input data Calculate the reconstruction loss L recon： Step 243: GAN forward propagation and adversarial loss calculation, calculate the discriminator loss L D : Calculate the generator loss L G : Step 244: Backpropagation and parameter update, update the encoder and decoder parameters to minimize the reconstruction loss: Update the discriminator parameters to minimize the discriminator loss: Update the generator parameters to minimize the generator loss: Step 245: Iterate training and repeat steps 241-245 until the model converges. Each iteration includes an update of the encoder and decoder, and an update of the generator and discriminator.

6. The multi-source signal tool wear prediction method based on one-dimensional residual network according to claim 1, characterized in that: The construction process of the complex relationship processing module includes: Step 31: Collect lathe signal data from multiple sensors and use filters to extract and classify signal features; Step 32: Convert the processed time domain signal into a frequency domain signal through Fourier transform, and further extract multi-scale features through wavelet transform; Step 33: Construct a one-dimensional residual network to capture and process the complex relationship between multi-source signals. The residual network stabilizes the training of the deep network by introducing skip connections, and integrates and extracts the correlation features of multi-source signals. In step 31, the process of using filters to extract and classify signal features includes: Step 311: Signal preprocessing, using a low-pass filter to remove high-frequency noise, and then normalizing the signals to control their amplitudes within a standard range; Step 312: Extract time domain features from the signal. These features reflect the statistical properties of the signal. The time domain features include: maximum value, minimum value, peak-to-peak value, mean value, variance, standard deviation, mean square value, kurtosis, skewness, shape factor, crest factor, impulse factor, margin factor, and clearance factor. In step 32, the process of converting the time domain signal into the frequency domain signal through Fourier transform includes: Step 321: Convert the time series signal to the frequency domain through Fourier transform and extract frequency domain features; Step 322: Time-frequency domain feature extraction, combining the time domain and frequency domain, decomposing the signal into different frequency band energies through wavelet transform, and extracting local features; Step 323: combining all extracted features into a feature vector; In step 33, the process of constructing the one-dimensional residual network includes: Step 331: input data preparation, using the pre-processed multi-sensor signal data as the input of the network; Step 332: Construct an initial convolution layer to perform initial convolution on the input signal to extract low-level features. The convolution operation captures local temporal patterns and frequency characteristics. Step 333: Construction of residual blocks. Each residual block includes two convolutional layers and a shortcut connection, allowing signals to be directly transmitted through skip connections. Step 334: Global average pooling layer, which reduces feature dimensions and aggregates information; Step 335: Construct a fully connected layer and an output layer, input the result of global average pooling into the fully connected layer, and obtain the final output.

7. The tool wear prediction method based on multi-source signals of one-dimensional residual network according to claim 1, characterized in that: The construction process of the task execution module includes: Step 51: Combine the one-dimensional residual network and the multi-head attention mechanism; Step 52: Define the loss function of the model, use an optimization algorithm to minimize the loss function, and then iteratively train the model until convergence; In step 51, the process of combining the one-dimensional residual network and the multi-head attention mechanism includes: Step 511: The one-dimensional residual network is formed by stacking multiple residual blocks in depth to form a deep one-dimensional residual network to enhance the model's ability to model complex signals. The residual block is the basic structure of the network. Each residual block contains two or more one-dimensional convolutional layers. Each convolutional layer is followed by a ReLU activation function and batch normalization, and a skip connection is introduced between the input and output. Step 512: Input the output features of the one-dimensional residual network into the multi-head attention mechanism to enhance the feature representation through the attention mechanism.

8. The tool wear prediction method based on multi-source signals of one-dimensional residual network according to claim 7, characterized in that: In step 52, it includes: Step 521: Define a loss function and use the mean square error as the loss function to measure the difference between the model prediction value and the actual value; Step 522: Optimize the algorithm design, use the Adam optimizer, and adaptively adjust the learning rate of each parameter, which is suitable for processing sparse gradients and high-dimensional data.

9. A tool wear prediction system based on a multi-source signal of a one-dimensional residual network, used to implement the tool wear prediction method based on a multi-source signal of a one-dimensional residual network according to any one of claims 1 to 8, characterized in that: include: Signal acquisition module, used to convert the collected multi-sensor signals into a spectrum and perform wavelet transform; Data processing module, used for performing balancing operations on data; Complex relationship processing module, used to build a one-dimensional residual network structure to handle complex relationships between multi-source signals; Feature enhancement module, which is used to introduce a multi-head attention mechanism to enhance the model's feature extraction capabilities; The task execution module uses a model that combines a one-dimensional residual network and a multi-head attention mechanism to perform tool wear prediction tasks.

Images