A ball screw health assessment method based on the difference between DCAE and the maximum mean.
By using a denoising convolutional autoencoder and the maximum mean difference method, the signal features of ball screws are automatically extracted, which solves the problem of dependence on expert knowledge and large amounts of data in existing technologies and realizes the automation and efficiency of ball screw health assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-05-13
- Publication Date
- 2026-03-10
AI Technical Summary
Existing technologies rely on expert knowledge and a large amount of experimental data in ball screw health assessment, making it difficult to effectively extract features under different conditions, and feature selection and dimensionality reduction are also challenging.
A denoising convolutional autoencoder is used to automatically process the ball screw sensing signal. The maximum mean difference measure feature distribution difference is mapped to a healthy value, and only signal data under normal performance is needed to build the model.
It enables ball screw health assessment without requiring multiple degradation state data, improves the automation and industrial applicability of feature extraction, and can intuitively reflect the health status of the ball screw.
Smart Images
Figure CN115344950B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of mechanical processing, and particularly relates to a ball screw health evaluation method based on a denoising convolutional autoencoder (DCAE) and a maximum mean difference. BACKGROUND
[0002] In a numerical control machine tool, the precision of a ball screw plays a crucial role in the machining precision of the machine tool. In a long machining process, the performance of the ball screw will decline to different degrees, resulting in a decrease in machining precision. Therefore, it is of great significance to evaluate the health state of the ball screw. At present, a data-driven method is still the most commonly used method for ball screw health evaluation technology, that is, a health value in a current state is obtained by extracting and analyzing certain features of a sensing signal of the ball screw in operation. However, these methods generally need expert knowledge for feature selection, and the effectiveness of the features often needs to be verified by a large amount of experimental data, and the effectiveness in different situations is questionable. In addition, the selection and dimensionality reduction of the features often need to analyze the feature correlation based on experimental data in different health states, and it is often difficult to obtain sufficient data in actual application. SUMMARY
[0003] The present application proposes a ball screw health evaluation method based on a DCAE and a maximum mean difference to address the above-mentioned deficiencies in the prior art. The method uses a denoising convolutional autoencoder to automatically process a sensing signal of the ball screw, extract useful information, and map the difference between feature distributions into a health value by a maximum mean difference. The method only needs to obtain a signal of the ball screw in a very healthy state to establish a model, and does not need data in multiple degradation states to establish key features.
[0004] The present application is implemented by the following technical solutions:
[0005] The present application relates to a ball screw health evaluation method based on a DCAE and a maximum mean difference. A signal of the ball screw in normal performance is collected to generate a training sample to train a denoising convolutional autoencoder. The training sample is input into the trained denoising convolutional autoencoder to obtain a baseline feature. A working signal of a ball screw to be evaluated is collected to generate an evaluation sample. The sample feature is extracted by the trained denoising convolutional autoencoder. The maximum mean difference between the baseline feature and the sample feature is mapped into a health value of the ball screw to be evaluated.
[0006] The signal collected in the normal performance refers to a vibration signal of the ball screw in stable and continuous operation in the normal performance, which is collected by a vibration sensor installed on the ball screw and further divided into multiple signal samples for model training.
[0007] The vibration sensor is arranged at the screw nut and the bearing seat at the two ends of the ball screw according to actual conditions; and the collected vibration signal contains at least one vibration direction.
[0008] The normal performance refers to that the ball screw is newly installed or has not been put into use for a large amount of machining.
[0009] The division refers to that the collected vibration signals of multiple channels are combined to form a multi-channel signal for subsequent processing, and the vibration signal is divided into multiple signal samples according to a specified sample length N, and the setting principle of N is that at least 0.1 seconds of signal is contained to ensure sufficient information, but not more than 1 second of signal to avoid repeated information to reduce the calculation efficiency.
[0010] The encoder sequentially includes a noise layer and a plurality of one-dimensional convolution layers and batch normalization layers connected in alternation, wherein the noise layer adds random noise to the input signal to improve the denoising ability of the encoder. The noise layer randomly sets a certain proportion of points in the signal to zero, and the training autoencoder reconstructs the original signal, which can improve the robustness of the encoder learning features; the one-dimensional convolution layer performs one-dimensional convolution operation on the input sequence to obtain an output sequence, and the batch normalization layer performs normalization operation on each batch of samples in the training process to speed up the training.
[0011] The one-dimensional convolution operation is specifically: when the time step of a sliding convolution kernel is l, and the input sequence has d channels, then a convolution kernel of the one-dimensional convolution layer is a one-dimensional vector v∈R ld×1 In each step, the input sequence [x i , x i+1 ,..., x i+l-1 ] T ∈R l×d is multiplied by the convolution kernel at the corresponding position, and the result is obtained by concatenating the results of the multiplication, and the specific operation is: , wherein: X i ∈R ld×1 is a one-dimensional vector formed after concatenation, denotes the concatenation operation. The result of the one-dimensional convolution operation can be obtained by dot multiplication, specifically: o=v T X i +b, wherein b is a bias parameter vector.
[0012] The batch normalization operation is specifically: wherein x (k)E[·] denotes the mean, Var[·] denotes the variance, γ (k) and β (k) are parameters to be learned by the network.
[0013] The decoder comprises a plurality of one-dimensional deconvolution layers and batch normalization layers connected alternately, wherein: the one-dimensional deconvolution layer first adds s-1 zero elements between adjacent elements in the input, and then performs normal one-dimensional convolution operation; the batch normalization layer in the decoder performs the same operation as the batch normalization layer in the encoder, wherein: s is the deconvolution step length.
[0014] The training is performed by a mean square error loss function, and the loss function value is iteratively optimized, and the training is stopped after a preset iteration step number is reached to obtain the trained denoising convolutional autoencoder.
[0015] The training specifically comprises layer-by-layer training of the denoising convolutional autoencoder and overall fine-tuning training of the denoising convolutional autoencoder, wherein: in the layer-by-layer training process, the first one-dimensional convolution layer and the batch normalization layer of the encoder and the last one-dimensional convolution layer and the batch normalization layer of the decoder are sequentially connected in the original order first, and the new denoising convolutional autoencoder is iteratively trained; then the first two one-dimensional convolution layers and the batch normalization layers of the encoder and the last two one-dimensional convolution layers and the batch normalization layers of the decoder are sequentially connected in the original order, and the new denoising convolutional autoencoder is iteratively trained; and the process is repeated until all layers are included; the overall fine-tuning training process is to train the complete denoising convolutional autoencoder after the layer-by-layer training.
[0016] The to-be-evaluated sample refers to: collecting the vibration signal of the ball screw currently needing to be evaluated when it is stably and continuously working, and dividing it into a plurality of signal samples with the same length N as the training sample.
[0017] The maximum mean difference value Wherein: is a mapping from the data space to a reproducing Hilbert space, x i and y i are the i-th and j-th samples in the two distributions, respectively, and m and n are the number of samples in the two distributions, respectively, and ||·|| H is the norm in the reproducing Hilbert space, and the maximum mean difference is used to measure the distribution difference between the two data distributions, and when the feature distribution of the collected to-be-evaluated sample is significantly different from the baseline feature distribution of the healthy performance sample, it is considered that the degradation is more serious.
[0018] The maximum mean difference can be calculated by specifying a kernel function, specifically: Where: k(·,·) is the selected kernel function.
[0019] The mapping is a nonlinear mapping, specifically H=a-exp(b*h), wherein: h is the maximum mean difference, H is the mapped health value, and a and b are selected parameters for adjusting the health value distribution in the [0, 1] interval according to the actual situation, so as to more intuitively understand the current health condition of the ball screw.
[0020] The present application relates to a system for implementing the above method, comprising: a vibration signal acquisition unit, a denoising convolutional autoencoder training unit, a baseline feature unit, and a health value calculation unit, wherein: the vibration signal acquisition unit is connected to the vibration sensor arranged on the ball screw and acquires the vibration signal during stable continuous operation, the denoising convolutional autoencoder training unit is trained using the samples obtained by dividing the collected ball screw vibration signal under normal performance as input, the baseline feature unit inputs the samples obtained by dividing the collected ball screw vibration signal under normal performance into the trained denoising convolutional autoencoder and constructs the baseline feature set from all outputs of the encoder, and the health value calculation unit inputs the samples obtained by dividing the collected ball screw vibration signal under the state to be evaluated into the trained denoising convolutional autoencoder and calculates the maximum mean difference from all outputs of the encoder and the baseline feature and maps it as a health value.
[0021] Technical effects
[0022] The present application solves the problems of dependence on empirical knowledge in the feature extraction stage and difficulty in obtaining sufficient data of different health stages in the feature establishment stage of the prior art; compared with the prior art, the present application realizes automatic feature extraction of the original vibration signal through the denoising convolutional autoencoder, thereby eliminating the need for prior knowledge in the feature extraction and feature selection operations; the model can be established using only data under the normal performance state of the ball screw without the need for multiple data of different health stages, and the present application has stronger industrial practicability. BRIEF DESCRIPTION OF DRAWINGS
[0023] Figure 1 The present application is a method flowchart.
[0024] Figure 2 The present application is a test device schematic diagram of the embodiment.
[0025] In the figure: guide rail 1, vibration sensor 2, ball screw 3, motor 4, bearing 5, screw nut 6, gear 7, bearing 8, electromagnetic brake 9. DETAILED DESCRIPTION
[0026] As shown in Figure 1 The present application relates to a ball screw health evaluation method based on a denoising convolutional autoencoder and a maximum mean difference, comprising the following steps:
[0027] Step 1: as shown in Figure 2 vibration signals of the ball screw under normal performance during working are collected by vibration sensors installed on the ball screw;
[0028] In this embodiment, FFZD4005-3 type ball screw is used for accelerated degradation experiment, and various loads and rotating speeds are used in the experiment to simulate variable working conditions in actual machining process, and signals are collected every 30 minutes for about 5 minutes. One vibration sensor is installed on the screw nut and the bearing seat at both ends of the ball screw. The two vibration sensors installed on the bearing seat only collect radial vibration signals, and the vibration sensor installed on the screw nut collects x, y and z direction vibration signals. The frequency of the collected signals is 5000 Hz.
[0029] The vibration sensor can be arranged at multiple sites according to actual conditions, including the screw nut and the bearing seat at both ends of the ball screw; and the collected vibration signals can include multiple vibration directions.
[0030] Step 2: dividing the collected vibration signals into multiple signal samples for model training;
[0031] In this embodiment, the collected vibration signals of 5 channels are combined to form 5-channel signals for subsequent processing, and the vibration signals are divided into multiple signal samples according to a specified sample length N = 1000.
[0032] Step 3: constructing a denoising convolutional autoencoder including an encoder and a decoder, and taking the signal samples for training as network input: the constructed denoising convolutional autoencoder includes an encoder and a decoder, which can perform feature learning on the input information. The input space is mapped to the feature space by the encoder, and then mapped back to the input space by the decoder, so that the reconstruction error is as small as possible, thereby learning the most effective information contained in the input.
[0033] The encoder first adds random noise to the input signal using a noise layer to improve the denoising ability of the encoder. The noise layer randomly sets a certain proportion of points in the signal to zero, and the autoencoder is trained to reconstruct the original signal, which can improve the robustness of the encoder in learning features;
[0034] In this embodiment, multiple one-dimensional convolutional layers and batch normalization layers are used after the noise layer.
[0035] The one-dimensional convolutional layer performs one-dimensional convolution operation on the input sequence to obtain the output sequence. Assuming that the time step of a sliding convolution kernel is l, and the input sequence has d channels, then a convolution kernel is a one-dimensional vector v ∈ R ld×1 In each step, the input sequence [x i , x i+1x i+l-1 ] T ∈R l×d First, the splicing is performed, and the specific operation is: Wherein: X i ∈R ld×1 is a one-dimensional vector formed after splicing, Indicates the splicing operation. The result of the one-dimensional convolution operation can be obtained by dot multiplication operation, specifically: o=v T X i +b, wherein: b is a bias parameter vector;
[0036] The batch normalization layer normalizes each batch of samples in the training process to speed up the training, specifically: Wherein: x (k) is the output of the previous layer, E[·] represents the mean, Var[·] represents the variance, γ (k) and β (k) are parameters to be learned by the network.
[0037] In this embodiment, the method of randomly setting a certain proportion of points in the signal to zero by the noise layer is specifically: a normal distribution random signal with a mean of 0 and a standard deviation of 0.1 is generated, and the points in the original signal corresponding to the positions with a signal value less than-0.1 are set to zero.
[0038] In this embodiment, the encoder includes 4 one-dimensional convolution layers and 4 batch normalization layers, the number of convolution kernels in each convolution layer is 32, 64, 64, and 64 respectively, the size of the convolution kernel is 64, 3, 3, and 3 respectively, and the sliding step of the convolution kernel is 16, 2, 2, and 2 respectively.
[0039] The decoder in step 3 is: using a plurality of one-dimensional deconvolution layers and batch normalization layers connected alternately;
[0040] The one-dimensional deconvolution layer first adds s-1 zero elements between adjacent elements in the input, wherein: s is the deconvolution step, and then performs normal one-dimensional convolution operation; The batch normalization layer in the decoder performs the same operation as the batch normalization layer in the encoder.
[0041] In this embodiment, the decoder includes 4 one-dimensional deconvolution layers and 4 batch normalization layers, the number of convolution kernels in each deconvolution layer is 64, 64, 32, and 5 respectively, the size of the convolution kernel is 3, 3, 3, and 64 respectively, and the sliding step of the convolution kernel is 2, 2, and 16 respectively.
[0042] Step 4: using the input signal sample and the output of the decoder to establish a loss function, iteratively optimizing the loss function value, stopping after reaching the preset iteration number, and obtaining the trained denoising convolution autoencoder;
[0043] The loss function is used to measure the reconstruction error of the decoder reconstructing the input signal, and the mean square error function is selected, and the expression is as follows: Wherein: M represents the number of samples, z j And The original input signal and the reconstructed signal output by the denoising convolutional autoencoder are represented by and respectively, and ||·||2 represents the 2-norm of the vector.
[0044] The iterative training process includes two stages: layer-by-layer training of the denoising convolutional autoencoder, and overall fine-tuning training of the denoising convolutional autoencoder, specifically: the layer-by-layer training process first connects the first one-dimensional convolutional layer and the batch normalization layer of the encoder with the last one-dimensional convolutional layer and the batch normalization layer of the decoder in the original order, and iteratively trains this new denoising convolutional autoencoder; then the first two one-dimensional convolutional layers and the batch normalization layers of the encoder are connected with the last two one-dimensional convolutional layers and the batch normalization layers of the decoder in the original order, and iteratively train this new denoising convolutional autoencoder; and so on, until all layers are included.
[0045] The overall fine-tuning training process is to train the complete denoising convolutional autoencoder after layer-by-layer training.
[0046] In this embodiment, the layer-by-layer training process uses the Adam optimization algorithm, and the learning rate is set to 0.1; the overall fine-tuning process also uses the Adam optimization algorithm, and the learning rate is set to 0.001; the training process is stopped after 100 iterations.
[0047] Step 5: Obtain the signal sample for training
[0048] Step 6: Collect the vibration signal of the ball screw to be evaluated when working, divide it into multiple signal samples, and input it into the trained denoising convolutional autoencoder to obtain the output of the encoder as the feature of the sample to be evaluated;
[0049] The division divides the vibration signal into multiple signal samples according to the same sample length N in step 2.
[0050] Step 7: Calculate the maximum mean difference between the baseline feature obtained in step 5 and the sample to be evaluated obtained in step 6, which is used to measure the distribution difference between the two data distributions. When the feature distribution of the collected sample to be evaluated is significantly different from the baseline feature distribution of the healthy performance sample, it is considered that the degradation is more serious, and the health degree is lower. The expression of the maximum mean difference is:
[0051] Wherein: denotes the mapping from the data space to a reproducing Hilbert space, x i and y i are the i-th and j-th samples in the two distributions respectively, m and n are the number of samples in the two distributions respectively, and ||·|| denotes the norm in the reproducing Hilbert space. H
[0052] Further, the maximum mean discrepancy can be calculated by specifying a kernel function, specifically: wherein k(·,·) denotes the selected kernel function. In this embodiment, a Gaussian kernel function is selected as the kernel function for calculating the maximum mean discrepancy.
[0053] Step 8: mapping the obtained maximum mean discrepancy value to a health value, specifically: using a nonlinear mapping to map the obtained maximum mean discrepancy value to a health value, specifically: establishing a nonlinear mapping: H = a-exp(b*h), wherein h is the maximum mean discrepancy obtained in step 7, H is the mapped health value, and a and b are selected parameters for adjusting the health value distribution in the [0, 1] interval according to the actual situation, so as to more intuitively understand the current health condition of the ball screw. In this embodiment, the values of a and b in the established nonlinear mapping are both 2.
[0054] In this embodiment, the vibration signal of a brand new ball screw at the beginning of work for 5 minutes is collected as the normal health performance ball screw working signal. The vibration signal of the ball screw is collected for 5 minutes at about 18 hours, 36 hours, 54 hours, 72 hours and 90 hours of the accelerated degradation experiment for evaluation.
[0055] After evaluation by the method, the health values of the ball screw at about 18 hours, 36 hours, 54 hours, 72 hours and 90 hours of the accelerated degradation experiment are 0.862, 0.685, 0.671, 0.548 and 0.468 respectively. It can be seen that the health value of the ball screw decreases continuously with the increase of the degradation time, indicating that the health degree gradually decreases and the performance continuously declines. It also shows the effectiveness of the ball screw health evaluation method provided by the present application in actual use.
[0056] The present application directly inputs the original vibration signal of the ball screw into the model for automatic feature extraction, and only uses the signal collected under normal performance to establish the evaluation model. The original vibration signal is extracted by using the denoising convolution autoencoder, and the baseline feature set is established for the signal under normal performance. The maximum mean discrepancy is used to measure the distribution difference between the features obtained under the state to be evaluated and the baseline feature set, and the health value is mapped.
[0057] The above specific embodiments can be partially adjusted in different ways by those skilled in the art without departing from the principles and purposes of the present application, the protection scope of the present application is subject to the claims and is not limited by the above specific embodiments, each implementation within the scope is subject to the present application.
Claims
1. A ball screw health assessment method based on denoising convolution autoencoder and maximum mean discrepancy, characterized in that, The denoising convolutional autoencoder is trained by collecting signals of the ball screw under normal performance to generate training samples, and the training samples are input into the trained denoising convolutional autoencoder to obtain baseline features; the working signals of the ball screw to be evaluated are collected to generate evaluation samples, and the trained denoising convolutional autoencoder is used to extract sample features; and then the maximum mean difference value between the baseline features and the sample features is mapped to the health value of the ball screw to be evaluated. The collected signals under normal performance are vibration signals of the ball screw under stable and continuous working conditions collected by a vibration sensor installed on the ball screw, which are further divided into multiple signal samples for model training. The vibration sensor is arranged at the screw nut and the bearing seat at both ends of the ball screw, and the collected vibration signals include at least one vibration direction. The denoising convolutional autoencoder includes an encoder and a decoder for feature learning of input information, wherein the encoder maps the input space to the feature space, and the decoder maps back to the input space to minimize the reconstruction error and learn the most effective information contained in the input. The encoder sequentially includes a noise layer and multiple one-dimensional convolution layers and batch normalization layers connected in alternation. The decoder includes multiple one-dimensional deconvolution layers and batch normalization layers connected in alternation. The training is performed by using a mean square error loss function, and the loss function value is iteratively optimized until the preset iteration step is reached, and the training of the denoising convolutional autoencoder is completed, and the training process specifically includes layer-by-layer training and overall fine-tuning training of the denoising convolutional autoencoder, wherein the layer-by-layer training process first connects the first one-dimensional convolution layer and the batch normalization layer of the encoder and the last one-dimensional convolution layer and the batch normalization layer of the decoder in the original order, and iteratively trains the new denoising convolutional autoencoder; then the first two one-dimensional convolution layers and the batch normalization layers of the encoder and the last two one-dimensional convolution layers and the batch normalization layers of the decoder are connected in the original order, and the new denoising convolutional autoencoder is iteratively trained; and the process is repeated until all layers are included; and the overall fine-tuning training process is performed after the layer-by-layer training, and the complete denoising convolutional autoencoder is trained.
2. The ball screw health assessment method based on the denoising convolution autoencoder and the maximum mean difference according to claim 1, characterized in that, The division refers to combining the collected vibration signals of multiple channels into a multi-channel signal for subsequent processing, and dividing the vibration signals into multiple signal samples according to a specified sample length N, wherein the sample length N contains at least 0.1 seconds of signal to ensure sufficient information, but not more than 1 second of signal to avoid repeated information reducing the calculation efficiency. 3.The ball screw health evaluation method based on the de-noising convolution auto-encoder and the maximum mean difference according to claim 1, characterized in that, The one-dimensional convolution specifically refers to: when the time step of a sliding convolution kernel is... The input sequence has If there are 1 channel, then a convolution kernel of a one-dimensional convolutional layer is a convolution kernel with 1 channel. A one-dimensional vector with elements In each step, the input sequence at the corresponding position of the convolution kernel is... First, the parts are assembled, specifically as follows: This is the one-dimensional vector formed after concatenation. This indicates a concatenation operation; the result of the subsequent one-dimensional convolution operation can be obtained through a dot product operation, specifically: This is the bias parameter vector.
4. The ball screw health assessment method based on the denoising convolution autoencoder and the maximum mean difference according to claim 1, characterized in that, The maximum mean difference value mapping is a nonlinear mapping, specifically is the maximum mean difference, is the mapped health value, and is a selected parameter, used to adjust the health value distribution in the [0, 1] interval according to the actual situation, so as to more intuitively understand the current health condition of the ball screw.
5. A system for implementing the ball screw health assessment method based on the denoising convolution autoencoder and the maximum mean difference according to any one of claims 1-4, characterized in that, The method comprises the following steps: The vibration signal acquisition unit, the denoising convolutional autoencoder training unit, the baseline feature unit and the health value calculation unit are connected with the vibration sensor arranged on the ball screw and acquire the vibration signal during stable and continuous operation, the denoising convolutional autoencoder training unit uses the sample obtained by dividing the collected ball screw vibration signal under normal performance as input for training, the baseline feature unit inputs the sample obtained by dividing the collected ball screw vibration signal under normal performance into the trained denoising convolutional autoencoder and constitutes the baseline feature set with all outputs of the encoder, and the health value calculation unit inputs the sample obtained by dividing the collected ball screw vibration signal under the state to be evaluated into the trained denoising convolutional autoencoder and calculates the maximum mean difference between all outputs of the encoder and the baseline feature and maps the maximum mean difference as the health value.