A method for identifying long-sequence current signals in the operating state of industrial robots

Abstract

The present invention relates to the field of industrial robots and discloses a TYX‑net-based method for recognizing long-sequence current signals of multi-joint industrial robots. By using a deep learning approach and robot-coupled current signals to perform a health assessment on the robot, efficient recognition of the robot's long signals is achieved through the pyramid-structured hierarchical multi-scale feature extraction network TYX‑net.

Description

Technical Field

[0001] The present invention relates to the field of industrial robots, and in particular to a method for identifying long-sequence current signals of an industrial robot's operating state. Background Art

[0002] When an industrial robot continues to work, the degradation of its various components will have a coupling effect on the industrial robot, resulting in reduced positioning accuracy of the industrial robot, vibration and noise, and affecting the quality of the final product.

[0003] Current performance degradation detection methods for industrial robots primarily focus on extracting signals from individual robot components. These methods are relatively mature and widely used in industrial production. However, the overall performance degradation of an industrial robot is the result of the coupled degradation of its individual components. Existing methods still have shortcomings in processing these coupled signals and are unable to effectively identify their characteristics.

[0004] Existing deep learning-based methods for assessing the health of industrial robots primarily focus on short-term and vibration signals, with limited application to long-sequence signals and current signals common in industry. The mainstream approach for short-term signals is RNN-based models. However, RNN models are computationally expensive and time-consuming when processing long sequences, making them poorly suited for industrial deployment. Furthermore, vibration signals also present challenges such as high acquisition costs and susceptibility to environmental interference. Summary of the Invention

[0005] The present invention aims to provide a method for identifying long-sequence current signals of an industrial robot's operating status, so as to evaluate the health status of the industrial robot with lower cost and higher accuracy.

[0006] To achieve the above object, the present invention adopts the following technical solution: a method for identifying long-sequence current signals of an industrial robot in operation state, comprising the following steps:

[0007] A. Collect the current signals of the industrial robot joints. After collection, divide the current signals into 120 stages according to the number of operating days. The first 57 stages are divided into the running-in period.

[0008] B. Perform data preprocessing to remove the reset and pause signals from the robot's original signals and retain only the main program signals. Each signal is 122 seconds long, the sampling frequency is 250 Hz, and the single sample length is 31,000. The collected signals are divided into training and test sets in an 8:2 ratio.

[0009] C. Use the training set to train the model TYX-3 to process long sequence signals, and use the training set to train the model TYX-4 to process ultra-long sequence signals;

[0010] D. Use the network output in step C to predict the operating cycle of the industrial robot arm. Starting from the start of the robot arm operation, the predicted value is calculated in days, with each day of operation as an operating cycle. The operating cycle of the main program signal is used as the sample label. The predicted value is compared with the sample label. If it is the same as the label value, it is determined that the prediction is accurate.

[0011] Preferably, as an improvement, TYX-3 adopts a 38-layer network structure, including a PatchEmbedding layer as the input layer of the network, two PatchMerging layers for downsampling, three TYX modules including convolutional attention blocks, BatchNorm layers and forward propagation layers, and a fully connected layer for output.

[0012] Preferably, as an improvement, TYX-4 adopts a 53-layer network structure, including a PatchEmbedding layer as the input layer of the network, three PatchMerging layers for downsampling, four TYX modules including convolutional attention blocks, BatchNorm layers and forward propagation layers, and a fully connected layer for output.

[0013] Preferably, as an improvement, the PatchEmbedding layer is a convolutional layer, and a convolution with a convolution kernel size of 1×10 and a step size of 10 is used to perform preliminary information encoding.

[0014] Preferably, as an improvement, the PatchMerging layer is an element-by-element sampling layer. Under each channel, the first four elements of the vector are sampled every four elements, and the taken elements are spliced to obtain four new vectors, each of which is 1 / 4 of the original vector in length. LayerNorm operation and fully connected layer operation are then performed, so that the final number of output channels is halved to twice the number of input channels, with a length of 1 / 4.

[0015] Preferably, as an improvement, the convolutional attention block includes a multi-scale convolutional attention block MSCA, two channel mixing layers, and an activation function layer. The channel mixing layer uses 1×1 convolution for channel mixing operation, and the activation function layer is a GELU activation function.

[0016] Preferably, as an improvement, the multi-scale convolutional attention block MSCA includes an A model branch, a B model branch, a C model branch, a D model branch and a channel mixing layer; the A model branch is used to extract the small-scale features of the signal, the B model branch is used to extract the medium-scale features of the signal, the C model branch is used to extract the large-scale features of the signal, and the D model branch is used to extract the ultra-large-scale features of the signal.

[0017] Preferably, as an improvement, a 1×3 convolution kernel is used as the A model branch; a 1×5 convolution kernel is used as the B model branch; a 1×11 convolution kernel is used as the C model branch; and a 1×21 convolution kernel is used as the D model branch.

[0018] Preferably, as an improvement, the channel mixing layer in the multi-scale convolutional attention block adds the tensors of the A, B, C, and D branches at the channel level, uses 1×1 convolution as the channel fusion operation, and then multiplies the output by the input matrix of MSCA to perform a multi-scale attention operation.

[0019] Preferably, as an improvement, an ETA (Error Tolerability Accuracy) threshold is introduced in step D to further estimate the model results. The prediction of the robot's operating cycle often has errors within a few days, and these errors are sometimes acceptable. Therefore, this threshold is introduced. If the prediction error is within the range, the prediction is considered accurate, thereby evaluating the overall prediction effect.

[0020] Existing technologies often target rolling elements with constant rotational speeds, resulting in short sampling times and stable motion. However, when the target's motion trajectory is complex and the operating cycle is long, existing technologies are unable to effectively assess health status. Consequently, a health assessment network for long sequences of complex signals is needed. This present invention proposes a multi-joint robot operating cycle prediction method based on TYX-net. This method is an end-to-end approach that does not require manual feature extraction and complex modeling. It demonstrates excellent performance in processing long sequences of signals, and its computational complexity is lower than that of most existing methods. This method achieves high accuracy in predicting the operating cycle of industrial robots, demonstrating the present invention's strong feature extraction capabilities for robot arm signals, enabling accurate judgment of the robot's current operating status and high reliability. A convolutional attention mechanism is proposed for long signal diagnosis. This attention mechanism utilizes multi-scale feature extraction and channel fusion methods, employing a newly proposed convolutional attention module to effectively extract multi-scale complex features from long sequence signals within a neural network. The newly proposed dimensionality reduction method reduces computational complexity while ensuring that sequence information is not lost, thereby improving recognition accuracy. The proposed method employs four multi-scale branches to extract complex signal features and reflect positional relationships, helping to reduce complexity and computational overhead. A stackable module TYX-block is proposed to adapt to long signal inputs of different lengths by changing the number of modules in TYX-net.

[0021] This paper utilizes deep learning methods to assess robot health using coupled current signals, achieving excellent results. Using the pyramid-structured hierarchical multi-scale feature extraction network TYX-net, the proposed method achieves efficient recognition of long robot signals. Experimental analysis of data of varying lengths demonstrates the effectiveness of the proposed method, providing valuable insights for future long signal processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is an overall flow chart of an embodiment of the present invention.

[0023] Figure 2 2 is a network structure diagram in an embodiment of the present invention.

[0024] Figure 3 Detailed diagram of the network structure in an embodiment of the present invention.

[0025] Figure 4 This is a flow chart of current signal acquisition in an embodiment of the present invention.

[0026] Figure 5 Schematic diagram of PatchMerging operation in an embodiment of the present invention.

[0027] Figure 6 TYX structural parameter diagram in an embodiment of the present invention.

[0028] Figure 7 This is a comparison test diagram of 7550 unit length signals in an embodiment of the present invention.

[0029] Figure 8 This is a data representation diagram of a 7550 unit length signal in an embodiment of the present invention.

[0030] Figure 9 This is a comparison test diagram of 15500 unit length signals in an embodiment of the present invention.

[0031] Figure 10 This is a data representation diagram of a 15500 unit length signal in an embodiment of the present invention.

[0032] Figure 11 This is a comparison test diagram of 31000 unit length signals in an embodiment of the present invention.

[0033] Figure 12 This is a data representation diagram of a 31000 unit length signal in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The following is further described in detail through specific implementation methods:

[0035] The embodiment is basically as shown in the attached Figure 1A method for identifying long-sequence current signals of an industrial robot in operation state includes the following steps:

[0036] A. Figure 4 As shown in the figure, the current signal of the industrial robot joint is collected. After collection, the current signal is divided into 120 stages according to the number of operating days. The first 57 stages are divided into the running-in period.

[0037] B. Perform data preprocessing to remove the reset and pause signals from the robot's original signals and retain only the main program signals. Each signal is 122 seconds long, the sampling frequency is 250 Hz, and the single sample length is 31,000. The collected signals are divided into training and test sets in an 8:2 ratio.

[0038] C. Figure 2 、 Figure 3 As shown in the figure, the training set training model TYX-3 is used to process long sequence signals. TYX-3 adopts a 38-layer network structure, including a PatchEmbedding layer as the input layer of the network, two PatchMerging layers for downsampling, three TYX modules including convolutional attention blocks, BatchNorm layers and forward propagation layers, and a fully connected layer for output; the training set training model TYX-4 is used to process ultra-long sequence signals. TYX-4 adopts a 53-layer network structure, including a PatchEmbedding layer as the input layer of the network, three PatchMerging layers for downsampling, four TYX modules including convolutional attention blocks, BatchNorm layers and forward propagation layers, and a fully connected layer for output. The TYX structure parameters are shown in the figure. Figure 6 As shown;

[0039] D. Use the network output in step C to predict the operating cycle of the industrial robot arm. Starting from the start of the robot arm operation, the predicted value is calculated in days, with each day of operation as an operating cycle. The operating cycle of the main program signal is used as the sample label. The predicted value is compared with the sample label. If the predicted value is the same as the label value, it is determined that the prediction is accurate. The ETA (Error Tolerability Accuracy) threshold is introduced to further estimate the model results.

[0040] Among them, the PatchEmbedding layer is a convolutional layer, which uses convolution with a kernel size of 1×10 and a step size of 10 to perform preliminary information encoding, such as Figure 5As shown in the figure, the PatchMerging layer is an element-by-element sampling layer. For each channel, the first four elements of the vector are sampled every four elements. These elements are concatenated to obtain four new vectors, each 1 / 4 the length of the original vector. LayerNorm and fully connected layer operations are then performed, halving the number of output channels to twice the number of input channels, with a length of 1 / 4. The convolutional attention block consists of a multi-scale convolutional attention block (MSCA), two channel mixing layers, and an activation function layer. The channel mixing layer uses 1×1 convolution for channel mixing, and the activation function layer uses the GELU activation function. The multi-scale convolutional attention block (MSCA) consists of an A-model branch, a B-model branch, a C-model branch, a D-model branch, and a channel mixing layer. The A-model branch extracts small-scale features of the signal and uses a 1×3 convolution kernel. The B-model branch extracts medium-scale features and uses a 1×5 convolution kernel. The C-model branch extracts large-scale features and uses a 1×11 convolution kernel. The D-model branch extracts very large-scale features and uses a 1×21 convolution kernel. The channel mixing layer in the multi-scale convolutional attention block performs element-wise addition of the tensors from the A, B, C, and D branches, performs a 1×1 convolution as a channel fusion operation, and then performs a multi-scale attention operation by matrix multiplying the output with the MSCA input.

[0041] The technical solution of this embodiment is used to predict the robot operation cycle on the 7550 unit length signal, the 15500 unit length signal, and the 31000 unit length signal, as shown in FIG. Figure 7 、 Figure 8 As shown in Figure 2, the accuracy of predicting the robot's operating cycle on a 7550-unit length signal is 86.68%, ETA1 is 95.48%, and ETA2 is 98.11%. Figure 9 、 Figure 10 As shown in Figure 2, the accuracy of predicting the robot's operating cycle on a 15,500 unit length signal is 88.32%, ETA1 is 96.71%, and ETA2 is 98.35%. Figure 11 、 Figure 12 As shown, the accuracy of predicting the robot's operating cycle based on a 31,000-unit length signal is 94.49%, with ETA1 at 98.19% and ETA2 at 99.17%. This demonstrates that this method can accurately determine the robot's operating status. Compared to other models, this method is highly reliable and has a strong ability to extract features from long sequence signals. It can effectively identify and predict the robot's operating cycle, thereby improving the factory's ability to accurately determine the operating cycle, prevent failures, and reduce management costs.

[0042] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the description can be used to interpret the content of the claims.

Claims

1. A method for identifying long-sequence current signals of an industrial robot's operating status, characterized by: The following steps are involved: A. Collect the current signals of the industrial robot joints. After collection, divide the current signals into 120 stages according to the number of operating days. The first 57 stages are divided into the running-in period. B. Perform data preprocessing to remove the reset and pause signals from the robot's original signals and retain only the main program signals. Each signal is 122 seconds long, the sampling frequency is 250 Hz, and the single sample length is 31,000. The collected signals are divided into training and test sets in an 8:2 ratio. C. Use the training set to train the model TYX-3 to process long sequence signals, and use the training set to train the model TYX-4 to process ultra-long sequence signals; D. Use the network output in step C to predict the operating cycle of the industrial robot arm. Starting from the start of the robot arm operation, the predicted value is calculated in days, with each day of operation as one operating cycle. The operating cycle of the main program signal is used as the sample label. The predicted value is compared with the sample label. If the predicted value is the same as the label value, it is determined that the prediction is accurate. The TYX-3 adopts a 38-layer network structure, including a PatchEmbedding layer as the network input layer, two PatchMerging layers for downsampling, three TYX modules including convolutional attention blocks, BatchNorm layers, and forward propagation layers, and a fully connected layer for output; the TYX-4 adopts a 53-layer network structure, including a PatchEmbedding layer as the network input layer, three PatchMerging layers for downsampling, four TYX modules including convolutional attention blocks, BatchNorm layers, and forward propagation layers, and a fully connected layer for output; The convolutional attention block includes a multi-scale convolutional attention block MSCA, two channel mixing layers, and an activation function layer. The channel mixing layer uses 1×1 convolution for channel mixing operation, and the activation function layer is a GELU activation function. The multi-scale convolutional attention block MSCA includes an A model branch, a B model branch, a C model branch, a D model branch, and a channel mixing layer. The A model branch is used to extract small-scale features of the signal, the B model branch is used to extract medium-scale features of the signal, the C model branch is used to extract large-scale features of the signal, and the D model branch is used to extract ultra-large-scale features of the signal. The channel mixing layer in the multi-scale convolutional attention block adds the tensors of the A, B, C, and D branches at the channel level, uses 1×1 convolution as the channel fusion operation, and then multiplies the output by the input matrix of MSCA to perform a multi-scale attention operation.

2. The method for identifying long-sequence current signals of an industrial robot's operating status according to claim 1, characterized in that: The PatchEmbedding layer is a convolutional layer, which uses a convolution with a kernel size of 1×10 and a step size of 10 to perform preliminary information encoding.

3. The method for identifying long-sequence current signals of an industrial robot's operating status according to claim 2, characterized in that: The PatchMerging layer is an element-by-element sampling layer. Under each channel, the first four elements of the vector are sampled every four elements. The sampled elements are spliced to obtain four new vectors, each of which is 1 / 4 the length of the original vector. LayerNorm operation and fully connected layer operation are then performed, so that the number of output channels is halved to twice the number of input channels, with a length of 1 / 4.

4. The method for identifying long-sequence current signals of an industrial robot's operating status according to claim 3, characterized in that: A 1×3 convolution kernel is used as the A model branch; a 1×5 convolution kernel is used as the B model branch; a 1×11 convolution kernel is used as the C model branch; and a 1×21 convolution kernel is used as the D model branch.

5. The method for identifying long-sequence current signals of an industrial robot's operating status according to claim 4, characterized in that: In step D, an ETA threshold is introduced to further estimate the model results.

Images