CN116228641A - Micro fatigue crack length calculation method based on U-net network - Google Patents

Abstract

The invention discloses a micro fatigue crack length calculation method based on a U-net network, which relates to the technical field of crack detection and comprises the following steps: training the optimized U-net network by using a certain type of initial images of the fretting fatigue cracks to obtain a trained U-net network model; acquiring the micro fatigue crack image similar to the step S1 as an experimental image; identifying micro fatigue cracks in the experimental image by adopting the trained U-net network model to obtain a segmentation detection result graph; and measuring the length of the fretting fatigue crack in the segmentation detection result graph. The first convolution pooling layer group and the fusion layer which are sequentially connected are arranged between the up-sampling module and the down-sampling module in the U-net network structure, so that the calculated amount can be reduced, the network speed can be improved, and meanwhile, the characteristic diagram is expanded to capture local information and detail information, so that the characteristic information of the micro fatigue crack can be accurately extracted.

Description

Calculation method of fretting fatigue crack length based on U-net network

technical field

The invention relates to the technical field of crack detection, in particular to a method for calculating the length of a fretting fatigue crack based on a U-net network.

Background technique

Under long-term high-frequency fretting conditions, mechanical parts are prone to fretting fatigue cracks locally, and with the increase of load and cycle times, the crack growth rate will continue to increase, and the early tiny fatigue cracks will continue to develop and form significant cracks. Weaken the local stiffness of components, and even induce local damage, threatening the safety of the overall mechanical structure. Therefore, monitoring the length of fretting fatigue cracks is beneficial to quantify the specific impact of fretting cracks on equipment performance, which is of great significance for the safety life prediction of the overall mechanical structure.

In the prior art, the patent document with the publication number CN111445446B has developed a method for detecting cracks on the concrete surface based on improved U-net. The neural network template used in this method has a relatively large amount of calculation and can only obtain long cracks; The patent document No. CN113284107A discloses a real-time detection method for concrete cracks by introducing an attention mechanism improved U-net, but this detection model can only obtain fine cracks, but cannot obtain the length information of fine cracks.

Generally speaking, deep learning methods have been used to identify cracks. However, the size of fretting fatigue cracks is usually small, generally between one hundred microns and several hundred microns, and their length and width are extremely small. Therefore, it is necessary to develop an identification method suitable for small fatigue cracks such as engine blade tenon samples to identify small cracks, so as to realize fretting. Early identification of fatigue cracks.

Contents of the invention

In view of the above problems, the object of the present invention is to provide a method for calculating the length of fretting fatigue cracks based on the U-net network, which is suitable for monitoring the fretting fatigue cracks produced in the fretting of parts such as engine blade tenon samples with different cycle times length.

Technical scheme of the present invention is as follows:

A method for calculating the length of a fretting fatigue crack based on a U-net network, comprising the following steps:

S1. Use the original image of a certain type of fretting fatigue crack to train the optimized U-net network to obtain the trained U-net network model, which specifically includes the following steps:

S11. Obtain the original image of a certain type of fretting fatigue crack. Since the fretting cracks produced by different equipment and different movement methods are not the same, the original image required to be acquired is the fretting fatigue crack obtained by the same kind of equipment using the same movement method , such as the fretting fatigue cracks of the engine blade tenon sample subjected to cyclic stress, such as the fretting fatigue cracks of steel wires subjected to cyclic stress, such as the fatigue cracks of high-speed train axles subjected to cyclic stress.

S12. Label the crack area in the original image pixel by pixel so that the crack area forms a closed figure. There are many specific standard software, such as LabelMe, labelimg and rolabelimg.

S13, use the image segmentation suite to batch-organize the marked original images into a complete data set format and use it as the training set, test set, and verification set of the optimized U-net model, and train the optimized U-net model after training The U-net network model. The image segmentation suite for this step is not limited to the PaddleSeg and vox datasets.

S2. Obtain an image of fretting fatigue cracks similar to step S1 as an experimental image; use the trained U-net network model to segment and detect the experimental image to identify fretting fatigue cracks in the experimental image to obtain a segmentation detection result map.

S3, measuring the length of the fretting fatigue crack in the segmented detection result graph, there are multiple specific measurement methods, and the present invention provides a specific measurement method, including the following steps:

S31. Generate the central axis of the fretting fatigue crack in the segmented detection result graph by the central axis algorithm;

S32. Calculate the length of the central axis to obtain the length of the fretting fatigue crack: count the number of pixels on the central axis of the fretting fatigue crack, and according to the resolution size of the segmentation test result map, obtain the total number of pixels after unit conversion The total length of the central axis of the fretting fatigue crack.

In the U-net network structure, downsampling will reduce the probability of pixels being correctly marked, and the upsampling of the U-Net network has limited ability to recover feature information, making the crack width information not obvious, and the phenomenon of small cracks cannot be detected. The patent of the present invention proposes an optimized U-net network, which adds and fuses the maximum pooling layer, upsampling layer, and convolution kernel into small-sized depth-separable convolutions to construct a brand-new module expansion resolution rate, expand the depth of the network to increase the receptive field, and maintain the balance between the receptive field and resolution through multi-scale connections. On the premise of the maximum receptive field, the maximum resolution is maintained, so as to realize the detection of small cracks. Specifically, the optimized U-net network includes.

The optimized U-net network includes an upsampling module, a Concatenate operation, a downsampling module, and a layer fusion module between the bottom of the upsampling module and the bottom of the downsampling module;

The layer fusion module includes the first convolution pooling layer group and the fusion layer connected in sequence; the first convolution pooling layer group includes a parallel depth separable convolution module and 4 first convolution pooling layers, each The first convolutional pooling layer includes a maximum pooling layer with a stride of 2, a 1×1 depth-separable convolution module, and a 2×2 upsampling layer; the feature map output by the downsampling module is input into the first convolution In the pooling layer group, the fusion layer adds and fuses the input image features and the feature map output by the fusion layer is input into the upsampling module; the U-net network uses the nonlinear activation function ReLU to process the output of each layer , which can improve the nonlinear expression of the network and reduce the gradient disappearance phenomenon in the convolution process.

The depth of the feature vector of the U-net network is 2048, and the up-sampling module and the down-sampling module include five second convolutional layer groups connected in sequence, and each second convolutional layer group includes sequentially connected residual blocks, Attention mechanism module and 2 transposed convolutional layers;

The residual block includes depthwise separable convolution, transposed convolutional layer and BN layer. The input of the residual block is fused with the output of the convolution of the residual block through skipconnection in the addition layer;

The attention mechanism module includes an encoding block and a decoding block, the encoding block includes a maximum pooling layer and a convolutional layer, and the decoding block includes an upsampling layer and a convolutional layer; the feature information output by the residual block is input into the attention mechanism, and the soft The attention mechanism will be over-parameterized during training. The input of the attention mechanism module is fused with the output of the Sigmoid function classification in the attention mechanism module through a skipconnection, and the two are fused as one output, which solves the problem in the soft attention mechanism. Redundant parameter computation problem, optimized for convolutional neural networks.

The Concatenate operation splices the feature map generated by the upsampling module and the downsampled feature map with the np.concatennate function. In the Concatenate operation, a BN layer is added before the nonlinear activation function ReLU to ease the input distribution of the previous layer to the nonlinear function. The two ends are slowly approaching, and the BN layer normalizes the input data with a normal distribution of N (0, 1), and finally inputs it to the activation function ReLU, which can generate more obvious gradients in backpropagation, effectively helping the network Convergence is performed to improve the phenomenon of gradient dispersion.

Beneficial effect:

(1) The method of the present invention is an identification method capable of accurately identifying tiny fatigue cracks and measuring the length of the cracks.

(2) In the U-net network structure of the present invention, the first convolution pooling layer group and the fusion layer connected in sequence are arranged between the up-sampling module and the down-sampling module, which can reduce the amount of calculation, improve the network speed, and expand the features at the same time The map captures local information and detailed information, so that tiny fretting fatigue cracks can be accurately identified. In addition, splicing the channel dimension through layer fusion can reduce the training time of the network model, expand the feature map, and increase the resolution. Setting the depth separable convolution module in the first convolution pooling layer group can effectively reduce the amount of parameter calculation, and setting the convolution kernel size of the separable convolution module to 1×1 can reduce the number of channels.

Description of drawings

Fig. 1 is the block flow diagram of the optimized U-net network of 1 example of the embodiment of the present invention;

Fig. 2 is a block flow diagram of a layer 1 fusion module according to an embodiment of the present invention;

FIG. 3 is a flow diagram of a residual block in Embodiment 1 of the present invention;

Fig. 4 is a flow diagram of the attention mechanism module of Embodiment 1 of the present invention;

Fig. 5 is a flow chart of calculating the crack length of the experimental image of Example 1 of the present invention.

Detailed ways

In order to have a clearer understanding of the technical features, purpose and beneficial effects of the present invention, an embodiment of the present invention will be further described in conjunction with the accompanying drawings. The embodiments are only used to further illustrate the present invention, and cannot be interpreted as limiting the protection scope of the present invention. Some non-essential improvements and adjustments made by those skilled in the art according to the content of the present invention also belong to the protection scope of the present invention.

Example 1

In this embodiment, the fretting fatigue crack generated by the tenon of the engine blade through cyclic vibration is taken as an example to illustrate the method for calculating the length of the fretting fatigue crack of the present invention, which specifically includes the following steps:

S1. Train the optimized U-net network with the original image of the fretting fatigue crack generated by the cyclic vibration of the engine blade tenon to obtain the trained U-net network model, which specifically includes the following steps:

S11. Obtain the original image of the fretting fatigue crack generated by the cyclic vibration of the tenon of the engine blade: In order to obtain a clear original image of the fretting fatigue crack, this embodiment adopts the Zhongji GPS100 high-frequency fatigue testing machine, and in the device Basically, a digital microscope was placed on the side of the tenon sample to monitor the crack growth on both sides of the tenon sample, and image data was collected.

S12. Use the LabelMe software to mark the cracks pixel by pixel in the crack area in the original image, so that the crack area forms a closed figure.

S13. Use the image segmentation suite PaddleSeg to batch-organize the marked original images into a complete data set format and use them as the training set, test set, and verification set of the optimized U-net model, and train the optimized U-net model to get training The following U-net network model.

In the U-net network structure, downsampling will reduce the probability of pixels being correctly marked, and the upsampling of the U-Net network has limited ability to recover feature information, making the crack width information not obvious, and the phenomenon of small cracks cannot be detected. The patent of the present invention proposes an optimized U-net network, which adds and fuses the maximum pooling layer, upsampling layer, and convolution kernel into small-sized depth-separable convolutions to construct a brand-new module expansion resolution rate, expand the depth of the network to increase the receptive field, and maintain the balance between the receptive field and resolution through multi-scale connections. On the premise of the maximum receptive field, the maximum resolution is maintained, so as to realize the detection of small cracks. Specifically, the optimized U-net network includes.

Fig. 1 is the flowchart of the optimized U-net network of the present embodiment, the optimized U-net network includes an upsampling module, a Concatenate operation, a downsampling module and a layer between the bottom of the upsampling module and the bottom of the downsampling module Fusion module.

Figure 2 is a flow diagram of the layer fusion module, the layer fusion module includes the first convolution pooling layer group and the fusion layer connected in sequence; the first convolution pooling layer group includes a parallel depth separable convolution module, 4 A first convolutional pooling layer, each first convolutional pooling layer includes a maximum pooling layer with a stride of 2, a depthwise separable convolution module of 1×1, and an upsampling layer of 2×2; downsampling The feature map output by the module is input into the first convolution pooling layer group, the fusion layer adds and fuses the input image features and the feature map output by the fusion layer is input into the upsampling module; this U-net network uses a nonlinear activation function ReLU processes the output of each layer, which can improve the nonlinear expression of the network and reduce the gradient disappearance in the convolution process.

The depth of the feature vector of this U-net network is 2048. Please continue to refer to Figure 2. Both the up-sampling module and the down-sampling module include five second convolutional layer groups connected in sequence, and each second convolutional layer group includes A sequentially connected residual block, attention mechanism module, and 2 transposed convolutional layers.

Please refer to Figure 3. Figure 3 is a flowchart of the residual block. The residual block includes a depthwise separable convolution, a transposed convolution layer, and a BN layer. The input of the residual block is convolved with the output of the residual block through skipconnection. Fusion in additive layer.

Please refer to Figure 4, Figure 4 is a flow diagram of the attention mechanism module, the attention mechanism module includes an encoding block and a decoding block, the encoding block includes a maximum pooling layer and a convolutional layer, and the decoding block includes an upsampling layer and a convolutional layer; The feature information output by the residual block is input into the attention mechanism. The attention mechanism will be over-parameterized during training. The input of the attention mechanism module is fused with the output of the Sigmoid function classification in the attention mechanism module through a skipconnection. The two are fused to form an output, which solves the redundant parameter calculation problem in the soft attention mechanism and optimizes the convolutional neural network.

The Concatenate operation splices the feature map generated by the upsampling module and the downsampled feature map with the np.concatennate function. In the Concatenate operation, a BN layer is added before the nonlinear activation function ReLU to ease the input distribution of the previous layer to the nonlinear function. The two ends approach slowly, and the BN layer normalizes the input data with a normal distribution of N (0, 1), and finally inputs the value of the activation function ReLU, which can produce a more obvious gradient in backpropagation, effectively Help the network to converge to improve the phenomenon of gradient dispersion.

In this embodiment, the first convolutional pooling layer group and the fusion layer connected sequentially between the up-sampling module and the down-sampling module can reduce the amount of calculation, improve the network speed, and expand the feature map to capture local information and detailed information. In addition, splicing the channel dimension through layer fusion can reduce the training time of the network model, expand the feature map, and increase the resolution. Setting the depth separable convolution module in the first convolution pooling layer group can effectively reduce the amount of parameter calculation, and setting the convolution kernel size of the separable convolution module to 1×1 can reduce the number of channels.

S2. Obtain the fretting fatigue crack image similar to step S1 as the experimental image; use the trained U-net network model to segment and detect the experimental image to identify the fretting fatigue crack in the experimental image to obtain the segmentation detection result map; use the central axis The algorithm obtains the central axis in the crack segmentation result graph; the non-zero element statistical function is used to count the number of pixels contained in the central axis, and the specific detection and identification process is shown in Figure 5.

S3. Measuring the length of the fretting fatigue crack in the segmented test result graph, which specifically includes the following steps:

S31. Generate the central axis of the fretting fatigue crack in the segmented detection result graph by using the central axis algorithm.

S32. Calculate the length of the central axis to obtain the length of the fretting fatigue crack: count the number of pixels on the central axis of the fretting fatigue crack, and according to the resolution size of the segmentation test result map, obtain the total number of pixels after unit conversion The total length of the central axis of the fretting fatigue crack. For example, the number of pixels on the central axis of a crack in an image is 80 according to the statistics, and the number of pixels on the diagonal of the image is 1000, and the length of the diagonal is 43 cm. That is to say, in this In this case, 1000px=43cm, that is, 1px=0.43mm, the length of the central axis is: 80×0.43mm=34.4mm.

The relevant content of the present invention has been described above, and those skilled in the art will be able to realize the present invention based on these descriptions. Based on the above content of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

Claims (4)

1. A fretting fatigue crack length calculation method based on U-net network, is characterized in that, comprises the following steps: S1. Use the original image of a certain type of fretting fatigue crack to train the optimized U-net network to obtain the trained U-net network model; S2. Obtain an image of a fretting fatigue crack similar to step S1 as an experimental image; use the trained U-net network model to segment and detect the experimental image to identify the fretting fatigue crack in the experimental image to obtain a segmentation detection result map; S3, measuring the length of the fretting fatigue crack in the segmentation test result graph; The optimized U-net network includes an upsampling module, a Concatenate operation, a downsampling module and a layer fusion module between the bottom of the upsampling module and the bottom of the downsampling module; Among them, the layer fusion module includes the first convolution pooling layer group and the fusion layer connected in sequence; the first convolution pooling layer group includes 1 depth separable convolution module and 4 first convolution pooling layers connected in parallel , each first convolutional pooling layer includes a maximum pooling layer with a stride of 2, a depthwise separable convolutional module of 1×1, and an upsampling layer of 2×2; the feature map output by the downsampling module is input into the first In the convolution pooling layer group, the fusion layer adds and fuses the input image features and the feature map output by the fusion layer is input into the upsampling module; the nonlinear activation function used in this U-net network is ReLU; The depth of the feature vector of the U-net network is 2048, and the up-sampling module and the down-sampling module include five second convolutional layer groups connected in sequence, and each second convolutional layer group includes sequentially connected residual blocks, Attention mechanism module and 2 transposed convolutional layers; The residual block includes depthwise separable convolution, transposed convolutional layer and BN layer. The input of the residual block is fused with the output of the convolution of the residual block through skipconnection in the addition layer; The attention mechanism module includes an encoding block and a decoding block, the encoding block includes a maximum pooling layer and a convolutional layer, and the decoding block includes an upsampling layer and a convolutional layer; the feature information output by the residual block is input into the attention mechanism module, The input of the attention mechanism module is fused with the output of the Sigmoid function classification in the attention mechanism module through a skipconnection; The Concatenate operation splices the feature map generated by the upsampling module and the downsampled feature map with the np.concatennate function. In the Concatenate operation, a BN layer is added before the nonlinear activation function ReLU to perform N(0,1) on the input data. ) normalized normal distribution, and the normalized data is input to the activation function ReLU. 2. the fretting fatigue crack length calculation method based on U-net network according to claim 1, is characterized in that, step S1 comprises the steps: S11. Acquiring the original image of a certain type of fretting fatigue crack; S12. Marking the crack area in the original image pixel by pixel so that the crack area forms a closed figure; S13, use the image segmentation suite to batch-organize the marked original images into a complete data set format and use it as the training set, test set, and verification set of the optimized U-net model, and train the optimized U-net model after training The U-net network model. 3. the fretting fatigue crack length calculation method based on U-net network according to claim 1, is characterized in that, step S3 comprises the steps: S31. Generate the central axis of the fretting fatigue crack in the segmented detection result graph by the central axis algorithm; S32. Calculate the length of the central axis to obtain the length of the fretting fatigue crack. 4. the fretting fatigue crack length calculation method based on U-net network according to claim 3, is characterized in that, step S32 comprises the steps: Count the number of pixels on the central axis of fretting fatigue cracks, and according to the resolution size of the segmentation test result map, the total length of the central axis of fretting fatigue cracks is obtained after unit conversion from the total number of pixels.

Images