A Multi-Source Information Fusion Method for Tool Wear Prediction

Abstract

The present invention discloses a multi-sensor signal fusion method for tool wear prediction. By using the method of average segmented aggregation, the data volume can be reduced while preserving the original data trend, thereby shortening the operation time. The present invention proposes a method using Markov transition fields to encode the time series data generated during the machining process into two-dimensional images, and realizes the fusion of multi-channel pictures through the method of Laplacian image pyramid, achieving the fusion of multi-sensor information at the data level and avoiding the loss of the time relationship between sensing data. The present invention encodes the multi-sensor signals during the machining process into two-dimensional images through the method of Markov transition fields, and uses the method of Laplacian image pyramid to realize the fusion of multi-channel pictures, and then realizes the encoding of multi-sensor signals and the fusion into multi-channel two-dimensional images, which can maximize the preservation of the time relationship inside the original signals.

Description

Technical Field

[0001] The present invention belongs to the field of tool wear prediction, and relates to a multi-sensor signal fusion method, and particularly to a multi-source information fusion method for tool wear prediction. Background Art

[0002] During the machining process, as a key component for removing materials, the tool state directly affects the machining quality and efficiency. A severely worn tool may even cause the machine tool to stop, resulting in reduced machine tool accuracy and even harm to the operator. Therefore, real-time prediction of tool wear during the machining process has great practicality and economy.

[0003] In recent years, research has shown that compared with using a single sensor signal to predict tool wear, the data obtained by using a multi-sensor signal fusion method has richer semantics and higher resolution, and the established model has higher accuracy and system reliability. According to the different levels of processing information sources, the fusion methods of multi-source information include data layer fusion, feature layer fusion, and decision layer fusion. Although feature layer fusion and decision layer fusion can achieve data compression and reduce the computational complexity to a certain extent, they have the disadvantage of large information loss.

[0004] Data fusion based on the data layer is excellent for information retention and perception ability. However, at the same time, it often has high requirements for computational performance. In addition, data layer fusion often requires complex signal preprocessing methods and a large amount of expert experience knowledge, which not only consumes a lot of manpower and material resources, but also lacks universality.

[0005] Chinese invention patent CN 112767296 B discloses a multi-sensor data fusion method and system based on time series imaging, and proposes a method of using angles and triangular matrices to encode time series data into images and fuse them. However, the proposed image fusion method is limited to fusing an even number of images, and no fusion strategy is given when the number of images is odd, which has certain limitations. Chinese patent CN 113369993 B discloses a tool wear state monitoring method under small samples, and proposes to encode time series data into images by generating a gray distance map, and use the generated images to train a neural network, and then perform tool wear prediction. However, the method described is limited to processing single sensor data and cannot be applied to the scenario of multi-sensor prediction. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a multi-sensor signal fusion method for tool wear prediction, which can encode and fuse multi-sensor signals into multi-channel two-dimensional images, and maximize the retention of the time relationship inside the original signal.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A multi-sensor signal fusion method for tool wear prediction, comprising the following steps:

[0009] Step S1: Conduct a machining experiment, and obtain multi-source time series data X = {x1, x 2, x3, …, x N} during the experiment. Where x i represents the data value corresponding to the i-th sampling point.

[0010] Step S2: Perform data preprocessing, and use the average segmented aggregation method to reduce the size of the time series data and normalize the data.

[0011] Step S21: Divide the time series data X = {x1, x 2, x3, …, x N} containing N sampling points into M segments. When determining the number of segments M, it satisfies where a is the average value of the number of sampling points between two adjacent wave troughs of the time series data X.

[0012] Step S22: Calculate the mean value of each segment of data Construct a new sequence data

[0013] Step S23: Normalize the aggregated data using the following formula:

[0014]

[0015] where, is the value after normalizing the i-th point to [0, 1], and respectively represent the maximum and minimum values in the time series Find the normalized value of each data to obtain a new sequence data

[0016] After that, use the clustering algorithm to cluster the normalized data to obtain the clustering result Q of the data.

[0017] Step S3: Encode the preprocessed data into a single-channel two-dimensional image, construct a Markov transition matrix using a one-dimensional Markov chain, and construct a Markov transition based on the Markov transition matrix.

[0018] Step S31: Divide the data normalized in Step S2 into Q quantile intervals, and assign each data point to the corresponding quantile interval q j, j ∈ [1, Q].

[0019] Step S32: Construct a Markov state transition matrix:

[0020]

[0021] where w ij represents the frequency of the sampling point in the time series data X transferring from the quantile interval q i to the quantile interval q j .

[0022] Step S33: Construct a Markov transition field:

[0023]

[0024] where w ij |x i ∈ q i , x j ∈ q j represents the probability that the sampling point x t transfers from the quantile interval q i to the quantile interval q j , that is, w ij |x i ∈ q i , x j ∈ q j = P(x t+1 = j|x t = i).

[0025] Step S34: Use a mean fuzzy kernel to perform an average aggregation operation on the constructed Markov transition field:

[0026] M' = M * K

[0027] where K is the fuzzy kernel and M' is the matrix after the aggregation operation.

[0028] After that, convert the aggregated matrix M' into an image to obtain the final single-channel image.

[0029] Step S4: Based on the Laplacian image pyramid decomposition, fuse multiple single-channel images into one multi-channel image.

[0030] Step S41: Assume that a total of 7 images are obtained after Step S3. Use Image 1 to represent the x-direction force, Image 2 to represent the y-direction force, Image 3 to represent the z-direction force, Image 4 to represent the x-direction vibration, Image 5 to represent the y-direction vibration, Image 6 to represent the z-direction vibration, and Image 7 to represent the acoustic emission. Construct the Gaussian pyramids of Image 1 and Image 2 respectively.

[0031] Step S42: Construct a Laplacian pyramid for each layer of the constructed Gaussian pyramid respectively.

[0032] Step S43: Stitch the Laplacian pyramid images of each layer to construct a Laplacian pyramid composed of stitched images.

[0033] Step S44: Reconstruct the final fused image through the Laplacian pyramid of the stitched images.

[0034] Step S45: After the fusion of Image 1 and Image 2, fuse the fused image with Image 3, and so on, until the images of multiple channels are fused. Finally, obtain the final image after encoding and fusing the multi-sensor signals, realizing the fusion of multi-sensor signals at the data level.

[0035] Further, the multi-sensors in Step S1 include a force sensor, a vibration sensor, and an acoustic emission sensor.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. The present invention proposes a method for reducing the size of time series data. By using the method of average piecewise aggregation, the data volume can be reduced while retaining the original data trend, thereby shortening the operation time.

[0038] 2. The present invention proposes a method using a Markov switching field to encode the time series data generated during the processing into a two-dimensional image, and realizes the fusion of multi-channel pictures through the method of Laplacian image pyramid. This method realizes the fusion of multi-sensor information at the data level, avoiding the loss of the time relationship between sensing data. This method also provides a new method for processing time series data in the processing process using mature machine vision technology.

[0039] 3. The present invention realizes the fusion of multi-sensor information at the data level, avoiding the loss of the time relationship between sensing data. It encodes the multi-sensor signals in the processing process into a two-dimensional image through the method of Markov switching field, and uses the method of Laplacian image pyramid to realize the fusion of multi-channel pictures, and then realizes the encoding and fusion of multi-sensor signals into a multi-channel two-dimensional image. Compared with the prior multi-sensor signal fusion methods, the method proposed by the present invention can retain the time relationship inside the original signal to the greatest extent. It provides a new method for processing time series data in the processing process using machine vision technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flowchart of the present invention.

[0041] Figure 2This is the flowchart for data preprocessing in the embodiments of the present invention.

[0042] Figure 3 This is the flowchart for the present invention to encode time series data using a Markov transition matrix to obtain a single-channel two-dimensional image.

[0043] Figure 4 This is the flowchart for image fusion proposed in this application. Detailed implementation manners

[0044] The following further elaborates on this application in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely used to explain the present invention and are not used to limit the present invention.

[0045] As Figure 1 shown, a multi-sensor signal fusion method for tool wear prediction includes the following steps:

[0046] Step S1: Conduct a machining experiment, and obtain multi-source time series data through multi-sensors during the experiment.

[0047] In the embodiment, the sensors used include a force sensor, an acceleration sensor, and an acoustic emission sensor. The sampling frequency is 5000 Hz, and seven time series signals, namely force signals in the x, y, and z directions, acceleration signals in the x, y, and z directions, and acoustic emission signals, can be obtained in each experiment.

[0048] Step S2: Conduct data preprocessing, including reducing the size of time series data and normalizing the data using the average segmentation aggregation method

[0049] In the embodiment, through step S1, a total of 7 time series data X = {x1, x 2, x3, …, x 6000} containing 6000 sampling points are obtained. The average value a of the number of sampling points between two adjacent wave valleys in the original data is 170. At this time To ensure that the aggregated data can retain the trend of the original data, the number of segments M = 15 is taken, and the original data is divided into 15 segments. Calculate the mean value of each segment of data Construct a new sequence data using the obtained 400 mean values

[0050] In the embodiment, the data scaled by the average segmentation aggregation method is obtained After that, the following formula is used to normalize the aggregated data:

[0051]

[0052] Where is the value after the i-th point is normalized to [0, 1], and respectively represent the maximum and minimum values in the time series . The data preprocessing process is as Figure 2 shown.

[0053] Find the normalized value of each data to obtain a new sequence of data . After that, the normalized data is clustered using a clustering algorithm to obtain the clustering result Q of the data. In the embodiment, the clustering algorithm used is the k-means clustering algorithm.

[0054] Step S3: Encode the preprocessed data into a single-channel two-dimensional image, including constructing a Markov transition matrix using a one-dimensional Markov chain and constructing a Markov transition based on the Markov transition matrix.

[0055] In the embodiment, the data preprocessed in step S2 is divided into lengths of 400, the data is divided into four quantile intervals, and each data point is assigned to the corresponding quantile intervals Q1, Q2, Q3, Q4. Then

[0056] Construct a Markov state transition matrix:

[0057]

[0058] where w ij represents the frequency of the sampling point in the time series data X transitioning from the quantile interval q i to the quantile interval q j .

[0059] Step S33: Construct a Markov transition field:

[0060] Then, construct a Markov transition field:

[0061]

[0062] where w ij |x i ∈q i ,x j ∈q j represents the probability that the sampling point x t transitions from the quantile interval q i to the quantile interval a j , that is, w ij |x i ∈q i ,x j ∈q j =P(x t+1 =j|x t =i).

[0063] Step S34: Use a mean blur kernel to perform an average aggregation operation on the constructed Markov transition field:

[0064] M′ = M * K

[0065] where K is the blur kernel and M' is the matrix after the aggregation operation.

[0066] After that, convert the aggregated matrix M' into an image to obtain the final single-channel image. The construction process of the image is as Figure 3 shown. Step S4: Based on the Laplacian image pyramid decomposition, fuse multiple single-channel pictures into one multi-channel picture.

[0067] In the embodiment, a total of 7 images are obtained after step S3. Image 1 represents the x-direction force, Image 2 represents the y-direction force, Image 3 represents the z-direction force, Image 4 represents the x-direction vibration, Image 5 represents the y-direction vibration, Image 6 represents the z-direction vibration, and Image 7 represents the acoustic emission. Construct a Laplacian pyramid for each layer of the constructed Gaussian pyramid., splice the Laplacian pyramid images of each layer to construct a Laplacian pyramid composed of spliced images., reconstruct the final fused image through the Laplacian pyramid of the spliced images.

[0068] After Images 1 and 2 are fused, fuse the fused image with Image 3, and so on, until the images of multiple channels are fused. Finally, obtain the final image after encoding and fusing the multi-sensor signals, realizing the fusion of multi-sensor signals at the data level. The process of image fusion in the embodiment is as Figure 4 shown.

[0069] The present invention is not limited to this embodiment. Any equivalent conceptions or changes within the technical scope disclosed in the present invention are included in the protection scope of the present invention.

Claims

1. A multi-sensor signal fusion method for tool wear prediction, characterized in that: Including the following steps: Step S1: Conduct a machining experiment. During the experiment, multi-source time series data X = {x1, x2, x3, …, x N} is obtained through multiple sensors; where x i represents the data value corresponding to the i-th sampling point; Step S2: Perform data preprocessing, using the average segmentation aggregation method to reduce the size of time series data and normalize the data; Step S21: Divide the time series data X = {x1, x2, x3, …, x N} into M segments; when determining the number of segments M, satisfy where a is the average number of sampling points between two adjacent wave troughs of the time series data X; Step S22: Calculate the mean value of each data segment Construct a new sequence of data Step S23: Normalize the aggregated data using the following formula: Among them, is the value after the i-th point is normalized to [0, 1], and respectively represent the maximum and minimum values in the time series ; Find the normalized value of each data to obtain a new sequence of data After that, use the clustering algorithm to cluster the normalized data to obtain the clustering result Q of the data; Step S3: Encode the preprocessed data into a single-channel two-dimensional image, construct a Markov transition matrix using a one-dimensional Markov chain, and construct a Markov transition field based on the Markov transition matrix; Step S31: Divide the data normalized in Step S2 into Q quantile intervals, and assign each data point to the corresponding quantile interval q j , j ∈ [1, Q]; Step S32: Construct a Markov state transition matrix: where, w ij represents the frequency of the sampling points in the time series data X transferring from the quantile interval q i to the quantile interval q j ; Step S33: Construct a Markov transition field: where w ij |x i ∈q i ,x j ∈q j represents the probability that the sampling point x t transfers from the quantile interval q i to the quantile interval q j , that is, w ij |x i ∈q i ,x j ∈q j =P(x t+1 =j|x t =i); Step S34: Use a mean blur kernel to perform an average aggregation operation on the constructed Markov transition field: M′ = M * K where K is the blur kernel and M′ is the matrix after the aggregation operation; After that, convert the aggregated matrix M′ into an image to obtain the final single-channel image; Step S4: Based on the Laplacian image pyramid decomposition, fuse multiple single-channel pictures into a multi-channel picture; Step S41: Assume that a total of 7 images are obtained after Step S3. Use Image 1 to represent the x-direction force, Image 2 to represent the y-direction force, Image 3 to represent the z-direction force, Image 4 to represent the x-direction vibration, Image 5 to represent the y-direction vibration, Image 6 to represent the z-direction vibration, and Image 7 to represent the acoustic emission. Respectively construct Gaussian pyramids for Image 1 and Image 2; Step S42: For each layer of the constructed Gaussian pyramid, construct a Laplacian pyramid respectively; Step S43: Stitch the Laplacian pyramid images of each layer to construct a Laplacian pyramid composed of stitched images; Step S44: Reconstruct the final fused image through the Laplacian pyramid of the stitched images; Step S45: After Image 1 and Image 2 are fused, fuse the fused image with Image 3, and so on, until the images of multiple channels are fused. Finally, obtain the final image after encoding and fusing multi-sensor signals, realizing the fusion of multi-sensor signals at the data level.

2. The multi-sensor signal fusion method for tool wear prediction according to claim 1, wherein: The multi-sensors in Step S1 include force sensors, vibration sensors, and acoustic emission sensors.

Images