Method for predicting clamping force using convolutional neural network
The cepspectral image is generated and feature extraction is performed through the convolutional neural network method, combined with Adam optimization and loss function optimization, the complex and low-precision bolt clamping force prediction problem in the existing technology is solved, and the rapid and accurate real-time prediction effect is achieved. .
Patent Information
- Application Number
- CN201911157381.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2018-12-13
- Filing Date
- 2019-11-22
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2039-11-22
AI Technical Summary
The method of predicting bolt clamping forces using artificial neural networks in the prior art is complex, resulting in difficult real-time clamping force prediction and low accuracy, and a separate measuring device is required to measure the clamping forces of tightened bolts.
Convolutional neural network method is used to generate cepspectral images through signal processing and analysis equipment, extract feature images and synthesize them, optimize weight values using Adam optimization and loss functions, quickly predict clamping force, reduce image size to improve data processing speed and prediction accuracy.
It realizes fast and accurate real-time clamping force prediction, reduces dependence on measuring equipment, improves prediction accuracy, and is suitable for multi-step jointing processes at different speeds.
Smart Images

Figure CN111325314B_ABST
Abstract
Description
Technical Field
[0001] An exemplary embodiment of the present disclosure relates to a method for predicting clamping force using a convolutional neural network method, and more particularly, to a method for predicting clamping force using a convolutional neural network method for predicting bolt clamping force. Background Art
[0002] Bolts are widely used as an important element to join two or more parts. In particular, bolts that meet regulations must be used at the joints of steel structures, and the clamping force of the bolts must be strictly controlled.
[0003] For example, the Korean Industrial Standard (KS) classifies the types of bolts into Class A and Class B according to the torque coefficient and specifies the appropriate torque coefficient. The KS standard stipulates that for Class A bolts, the average value of the torque coefficient is suitable for 0.11 to 0.15, and for Class B bolts, the average value of the torque coefficient is suitable for 0.15 to 0.19.
[0004] The joint clamping force of the bolt can be calculated from the torque value, the torque coefficient and the diameter of the bolt, and the desired joint clamping force can be obtained by adjusting the torque value.
[0005] Tightening methods are commonly used to manage the clamping force of the bolt, which can be divided into torque control method and nut rotation method.
[0006] Since the nut rotation method uses the rotation angle of the nut, it has an advantage that the distribution of the clamping force is relatively small.
[0007] In contrast, the torque control method has the advantage of excellent workability because this method determines the upper and lower limits of the torque value and tightens the bolts using a torque wrench.
[0008] However, the torque coefficient of the bolt cannot remain constant but may change due to various factors such as the length of the bolt (which is a physical factor) and temperature (which is an environmental factor).
[0009] Therefore, the torque control method has a disadvantage that the distribution of the clamping force is large due to the change of the torque coefficient.
[0010] Furthermore, there is a disadvantage in that, after tightening the bolt using the torque wrench, it is necessary to measure the clamping force of the tightened bolt by a separate measuring device.
[0011] Therefore, it is important to predict the clamping force of the bolt so that the bolt maintains the proper clamping force.
[0012] Various prediction technologies using artificial intelligence have recently emerged. Among them, a prediction technology for predicting bolt clamping force using an artificial neural network has been introduced.
[0013] In the prior art method of calculating the clamping force using an artificial neural network, cepstral coefficients and short-time Fourier transform (STFT) are respectively derived from signals acquired by a sensor unit.
[0014] In the case of the cepstral coefficients, a high-pass filter is used to extract only the high frequency band of the acquired signal to reduce the influence of noise.
[0015] Then, the value is obtained by selecting a representative value in each frame, multiplying the weight value of each isosceles triangle distribution by hundreds of divided frequency domains, and then summing them.
[0016] The calculated values are multiplied by the logarithm and discrete cosine transformed to obtain the cepstral coefficients.
[0017] In the case of STFT, the acquired signal is transformed into frequency domain data having hundreds of frames by fast Fourier transform based on a short time with 50% overlap cycle.
[0018] The characteristics of the cepstral coefficients and the STFT data are input as supervisory signals, and a multi-layer weight matrix layer is formed by using the slope descent method of the cost function, and the clamping force prediction value is indicated using the multi-layer weight matrix layer.
[0019] However, since this conventional prediction technology is complicated, there are problems in that real-time clamping force prediction is difficult and has low accuracy. Summary of the invention
[0020] Embodiments of the present disclosure are directed to providing a method for predicting clamping force using a convolutional neural network method, which is simplified and capable of performing real-time prediction by improving accuracy.
[0021] According to an exemplary embodiment of the present disclosure, a method for predicting clamping force using a convolutional neural network method includes: generating a cepstrum image through a signal processing and analysis device; extracting a feature image by multiplying a predetermined weight value by the pixels of the generated cepstrum image through artificial intelligence learning; extracting the maximum pixel from the extracted feature image as a representative image; synthesizing an image by synthesizing the extracted representative image information; and predicting the clamping force by comparing the synthesized image with a predetermined value.
[0022] Adam optimization for optimizing predetermined weight values may be applied between the synthesis of the image and the prediction of the clamping force.
[0023] The loss function can be applied to Adam optimization.
[0024] In the extraction of feature images, convolution filters can be applied.
[0025] In the extraction of the representative image, a merging filter may be applied.
[0026] At least two convolution filters may be applied consecutively.
[0027] The extracted representative image may be subjected to at least one feature image extraction.
[0028] According to another exemplary embodiment of the present disclosure, a method for predicting clamping force using a convolutional neural network method includes: generating a cepstrum image from a signal processing analysis device; and predicting the clamping force by comparing the generated cepstrum image with a predetermined value.
[0029] According to the present disclosure, the following effects are obtained.
[0030] First, since the data processing speed is fast while reducing the image size, it has the advantage of easy real-time measurement.
[0031] Second, since Adam optimization is applied, it is advantageous to improve the accuracy of prediction by reducing the loss value because the weight value is sufficiently updated when the path is continuously accessed through iterative learning.
[0032] Third, it is advantageous to use this method in a real-time measurement system because it is faster than conventional slope-down methods.
[0033] Fourthly, since the magnitude of the signal during engagement is large, there is an advantage in that a clear distinction can be made when multi-step engagement is performed at different speeds.
[0034] Fifth, the advantage is that it is easy to use images of convolutional neural networks to derive feature maps because there are clear distinctions in multi-step connections. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 is a flow chart according to an exemplary embodiment of the present disclosure.
[0036] Figure 2 is a configuration diagram of a convolutional neural network method according to an exemplary embodiment of the present disclosure.
[0037] Figure 3 is a configuration diagram of a filter according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION
[0038] The present disclosure may have various modifications, and various embodiments and specific embodiments will be shown in the drawings and described in detail in the detailed description. However, this does not limit the present disclosure to specific embodiments, and it should be understood that the present disclosure covers all modifications, equivalents and replacements included in the thought and technical scope of the present disclosure.
[0039] Reference numerals refer to similar elements as each figure is described.
[0040] The terms including first, second, etc. are used to describe various constituent elements, but the constituent elements are not limited by these terms. The terms are used only to distinguish one constituent element from another element.
[0041] The term is used only to distinguish one constituent element from another element. The term "and / or" includes a plurality of related disclosed items or a combination of any items among the plurality of related disclosed items.
[0042] If there is no contrary definition, all terms used herein, including technical or scientific terms, have the same meanings as commonly understood by one of ordinary skill in the art.
[0043] Terms defined in general dictionaries should be interpreted as having the same meaning as in the context of the relevant technology, and should not be interpreted as ideal meanings or overly formal meanings unless explicitly defined in this application.
[0044] The flow of a method for predicting clamping force using a convolutional neural network method according to an exemplary embodiment of the present disclosure will be described.
[0045] Figure 1 is a flowchart according to an exemplary embodiment of the present disclosure, Figure 2 is a configuration diagram of a convolutional neural network method according to an exemplary embodiment of the present disclosure, and Figure 3 is a configuration diagram of a filter according to an exemplary embodiment of the present disclosure.
[0046] First, in a cepstrum image generating step (S1), a cepstrum image is generated by a signal processing and analyzing device (S20).
[0047] As the clamping force of the acquired signal increases, the generated cepstrum image contains frequency variation information as pixel values.
[0048] The cepstrum image may be obtained from a signal processing and analysis device (S10).
[0049] A learning cepstrum image may be acquired from the generated cepstrum image ( S30 ).
[0050] Next, in the feature image extraction step (S2), a feature image is extracted by multiplying a predetermined weight value by the pixels of the cepstrum image generated by artificial intelligence learning (S30).
[0051] At this time, in the feature image extraction step (S2), a convolution filter (S40) may be applied.
[0052] Convolution filters are small filters with weight values that are multiplied by the image pixels.
[0053] More specifically, a convolution filter is one of the layers of a convolutional neural network used to extract the input image.
[0054] In every deep learning, the values of the matrix of convolutional filters are called masks.
[0055] By using a convolution filter, the value of the mask learned through repetition is changed to an appropriate value for distinction, thereby improving the accuracy of learning.
[0056] When the cepstrum image is input to the convolution filter, the filter moves on the image and extracts the feature image by multiplying a predetermined appropriate weight value.
[0057] The extracted feature image is input to a merging filter (S50).
[0058] The merging filter (S50) extracts only representative values from the pixel information.
[0059] More specifically, the merging filter (S50) is used to derive the highest significant value of the feature map derived by the convolution, and is used to reduce the size of the image.
[0060] When the extracted feature image passes through the merging filter (S50), only representative information is stored and the size of the image is reduced.
[0061] At least two convolution filters may be applied consecutively.
[0062] Next, in the representative image extraction step (S3), the maximum pixel is extracted from the extracted feature image.
[0063] At this time, in the representative image extraction step (S3), a merging filter may be applied.
[0064] The extracted representative image may at least undergo the feature image extraction step again.
[0065] Next, in the image synthesis step (S4), a synthesized image is generated by synthesizing the extracted representative image information.
[0066] Finally, in the clamping force prediction step (S5), the predicted clamping force is indicated by comparing the synthesized image with a predetermined value (S80).
[0067] At this time, the information composed of representative values is finally synthesized by the dense layer, and the predicted value is represented as the clamping force with the highest probability.
[0068] The predicted value is indicated as a probability vector specific to the clamping force ( S70 ).
[0069] On the other hand, Adam optimization (S62) may be applied to optimize the weight value between the image synthesis step (S4) and the clamping force prediction step (S5).
[0070] Since the minimum is found using the Adam optimizer, it is suitable for use in real-time measurement systems due to its faster speed than the conventional slope descent method.
[0071] At this time, the loss function (S61) may be applied to Adam optimization (S62).
[0072] In other words, the Adam optimization technique helps in the optimization of the convolution filters.
[0073] A filter with sufficiently optimized weight values has to effectively extract features of the image.
[0074] The following will refer to Figure 3 The configuration of a filter according to an exemplary embodiment of the present disclosure is described in detail.
[0075] The cepstrum image 100 is acquired as data measured from the component, and at this time, the size of the image is a size of 599×32 pixels.
[0076] The convolution filter 200 is applied to the entire image, while the filter of size 3 x 3 is shifted one space to the right or down to extract the features of the cepstrum variation.
[0077] More specifically, the first convolution filter 210 and the second convolution filter 220 may be continuously used as the convolution filter 200 .
[0078] The size of the first feature image 310 extracted from the convolution filter 200 is maintained at a size of 599 x 32 pixels.
[0079] The first feature image 310 is input to the first merging filter 410 .
[0080] The merge filter 400 may include a first merge filter 410 and a second merge filter 420 in sequence.
[0081] When the first feature image 310 subjected to convolution enters the merging filter 400, all regions of the image are divided into 2×2 regions, and the maximum pixel value in each region is extracted.
[0082] This value is the first representative image 510 .
[0083] That is, since only one representative image value is obtained from four values as shown in the figure, the horizontal and vertical sizes of the image are reduced by half.
[0084] The first representative image 510 is derived again as the second feature image 320 through the third convolution filter 230 and the fourth convolution filter 240 .
[0085] The second feature image 320 passes through the second merging filter 420 again to obtain a second representative image 520 .
[0086] The synthesized image 600 becomes 1 by synthesizing the image information of the dense layer ( S60 ) to measure the similarity with each of the 31 categories of 50 kN to 81 kN and adding the similarities.
[0087] In the clamping force prediction step (S5), the class having the highest similarity among the measured similarities is confirmed as the determined clamping force.
[0088] In this way, the first convolution filter 210 and the second convolution filter 220 may be dually used as the convolution filter 200 .
[0089] In this way, a multi-filter structure can be applied, in which the feature image obtained by the first convolution filter 210 and the second convolution filter 220 is provided to the first pooling filter 410, and the feature image obtained by the third convolution filter 230 and the fourth convolution filter 240 are provided to the second filter 420.
[0090] In other words, a multi-filter structure may be applied using a triple filter structure that is a combination of a double convolution filter and a pooling filter.
[0091] In this way, deep features can be extracted by stacking convolutional filters in multiple layers.
[0092] Additionally, important information can be stored in smaller images at least once by using pooled filters, thus improving the speed and accuracy of real-time prediction.
[0093] According to an exemplary embodiment of the present disclosure, a method for predicting clamping force using a convolutional neural network method may be composed of a cepstrum image generating step (S31) generating a cepstrum image from a signal processing and analysis device, and a prediction step (S80) indicating a predicted clamping force by comparing the generated cepstrum image with a predetermined value.
Claims
1. A method for predicting clamping force using a convolutional neural network method, include: generating a cepstrum image containing frequency change information as pixel values by a signal processing and analysis device; Extracting a feature image by multiplying a predetermined weight value by pixels of the generated cepstrum image through artificial intelligence learning; Extracting the maximum pixel from the extracted feature image as a representative image; synthesizing an image by synthesizing the extracted representative image information through dense layers; and predicting the clamping force by comparing the synthesized image with a predetermined value, indicates the predicted clamping force with the highest similarity, and Select bolts based on the predicted clamping force.
2. The method according to claim 1, in, An Adam optimization for optimizing the predetermined weight values is applied between the synthesis of the image and the prediction of the clamping force.
3. The method according to claim 2, in, Apply the loss function to the Adam optimization.
4. The method according to claim 1, in, In the extraction of the feature image, a convolution filter is applied.
5. The method according to claim 1, in, In the extraction of the representative image, a merging filter is applied.
6. The method according to claim 4, in, At least two of said convolution filters are applied consecutively.
7. The method according to claim 6, in, The characteristic image is extracted at least once from the extracted representative image.
Citation Information
Patent Citations
Ceramic carrier embrittlement detection method in ternary catalytic converter assembly
CN101126744A
Image pixel classifying method based convolutional neural network
CN106651887A
Intelligent spectrogram analysis method of deep learning based on convolutional neural network
CN108446631A
Device and method for classifying open and close eyes based on deep residual learning
KR101903437B1
Method for estimating axial force of a bolt
KR1020090015359A
