Automated welding system, learning device, neural network system, and estimation device

A learning device comprising:

• a memory configured to store a computer-readable program which, when executed, causes the learning device to generate a trained model for controlling a welding robot to perform arc welding at a groove;
• and
• a processor which, when executing the computer-readable program, is configured to control the learning device to perform a method of generating the trained model for controlling the welding robot to perform the arc welding at the groove,
• wherein the method of generating the trained model includes:
• acquiring a data set including a learning image and region data indicating a region that includes a feature point in the learning image and has a probability distribution in which a probability increases as the region comes close to the feature point;
• and
• generating a learned model by supervised learning so as to output a probability distribution image that indicates, when an image including the feature point is input, a probability of a feature point in each pixel and that includes a feature region in which a probability increases toward a position corresponding to the feature point, by using the learning image as input data and using the region data as teacher data,
• wherein the learning image is an image obtained by capturing an image of a molten pool and an arc generated at a groove formed between two members to be welded by an arc welding, and
• wherein the feature point includes a feature point of the arc and a feature point of the molten pool.

A neural network system for causing a computer to output a probability distribution image that indicates a probability of a feature point in each pixel and includes a feature region in which a probability increases toward a position corresponding to the feature point, based on an input image including the feature point, the neural network system comprising:

• a first convolution network including a convolution layer;
• a first deconvolution network including a convolution layer;
• a second convolution network including a pooling layer and a convolution layer;
• and
• a second deconvolution network including a convolution layer and an upsampling layer,
• wherein the first convolution network performs a convolution processing on an input image and outputs a first feature image that is generated,
• wherein the second convolution network performs a pooling processing on the first feature image, further performs a convolution processing, and outputs a second feature image that is generated,
• wherein the second deconvolution network performs a deconvolution processing on the second feature image, further performs an upsampling processing, and outputs a third feature image that is generated,
• wherein the first deconvolution network performs a deconvolution processing on a composite image obtained by combining the first feature image and the third feature image, and outputs an output image that is generated,
• wherein the first convolution network, the first deconvolution network, the second convolution network, and the second deconvolution network perform learning so as to reduce a difference between a region that includes a feature point included in the input image and has a probability distribution in which a probability increases toward the feature point and a feature region included in the output image, and
• wherein the computer is configured to perform a calculation on the input image including the feature point based on the learned first convolution network, the first deconvolution network, the second convolution network, and the second deconvolution network and output a probability distribution image including a feature region in which a probability increases toward a position corresponding to the feature point.

A welding support device comprising:

• a memory configured to store a computer-readable program which, when executed, causes the welding support device to perform a method of controlling a welding robot to perform arc welding at a groove;
• and
• a processor which, when executing the computer-readable program, is configured to control the welding support device to perform the method of controlling the welding robot to perform the arc welding at the groove,
• wherein the method of controlling the welding robot to perform the arc welding at the groove includes:
• acquiring a camera image generated by a camera;
• determining a probability distribution image based on the camera image by using a learned model generated in advance by supervised learning so as to output the probability distribution image that indicates, when an image including a feature point is input, a probability of a feature point in each pixel and that includes a feature region in which a probability increases toward a position corresponding to the feature point by using a learning image as input data and using region data including a feature point in the learning image and indicating a region that has a probability distribution in which a probability increases toward the feature point as teacher data;
• extracting a region having a probability equal to or higher than a predetermined probability from the probability distribution image;
• and
• selecting a representative point in the region having a probability equal to or higher than the predetermined probability.