Search

CN-121981988-A - Bobbin state identification method and system based on deep learning

CN121981988ACN 121981988 ACN121981988 ACN 121981988ACN-121981988-A

Abstract

The invention provides a bobbin state identification method and system based on deep learning, and relates to the technical field of image processing, wherein the method comprises the steps of collecting a bobbin state image; the method comprises the steps of constructing an image data set for spool state detection, using a YOLO11 model as a base line model, improving the base line model through a lightweight design, constructing a DML-YOLO model, training the DML-YOLO model, compressing the trained DML-YOLO model through a channel pruning technology based on L1 regularization to obtain a DMLP-YOLO model, performing fine tuning training on the DMLP-YOLO model, deploying the DMLP-YOLO model after the fine tuning training to edge equipment, collecting real-time spool images, inputting the real-time spool images into the DMLP-YOLO model after the fine tuning training deployed in the edge equipment, and outputting spool state identification results.

Inventors

  • DAI NING
  • DU ZIHENG

Assignees

  • 浙江理工大学

Dates

Publication Date
20260505
Application Date
20260122

Claims (10)

  1. 1. The spool state identification method based on deep learning is characterized by comprising the following steps of: s1, acquiring a bobbin status image; s2, constructing an image data set for detecting the state of the bobbin according to the state image of the bobbin; S3, taking a YOLO11 model as a baseline model, improving the baseline model through a lightweight design, and constructing a DML-YOLO model; s4, training the DML-YOLO model based on the image data set; S5, compressing the trained DML-YOLO model by a channel pruning technology based on L1 regularization to obtain a DMLP-YOLO model; S6, performing fine tuning training on the DMLP-YOLO model based on the image dataset, and deploying the DMLP-YOLO model after the fine tuning training to edge equipment; s7, acquiring a real-time bobbin image; and S8, inputting the real-time bobbin image into a DMLP-YOLO model which is deployed in the edge equipment and subjected to fine adjustment training, and outputting a bobbin state identification result.
  2. 2. The method for identifying a bobbin state based on deep learning according to claim 1, wherein the step S2 specifically comprises: s201, defining the state of a yarn tube as a plurality of category labels based on spinning process requirements; S202, marking the data of the bobbin status image according to the class label; S203, performing quality check on the marked result after the data marking by a cross verification technology; S204, dividing the image data subjected to quality verification into a training set, a verification set and a test set according to a preset proportion to form the image data set.
  3. 3. The method for identifying a bobbin state based on deep learning according to claim 1, wherein the step S3 specifically comprises: S301, introducing a C3k 2-dynamic serial mixer module into a core feature extraction unit of the YOLO11 model to replace an original standard convolution structure; S302, based on an improved core feature extraction unit, introducing a double-channel gating network into a backbone network of the YOLO11 model to replace an original single-path backbone network; S303, combining the target dimensional characteristics of the yarn tube, removing redundant detection heads in the YOLO11 model, and performing structure reduction treatment on the improved backbone network to complete construction of the DML-YOLO model.
  4. 4. A method for identifying a bobbin status based on deep learning according to claim 3, wherein the step S301 specifically comprises: S3011, constructing a serial double residual block variant; S3012, performing multipath dynamic convolution operation in the multi-branch dynamic convolution block of the serial double residual block variant; S3013, constructing the C3k 2-dynamic serial mixer module based on the serial double residual block variant; s3014, replacing the original standard convolution structure in the YOLO11 model by the C3k 2-dynamic serial mixer module.
  5. 5. The method for identifying the state of the bobbin based on deep learning according to claim 4, wherein the processing step of the tandem double residual block variant specifically comprises: S30111, carrying out batch normalization processing on input characteristic tensors; S30112, mixing the characteristics of the space dimension and the channel dimension of the characteristics after normalization processing through the multi-branch dynamic convolution block; S30113, regularizing DropPath through a learnable scaling parameter, and processing the mixed features; s30114, carrying out first residual connection on the first regularized characteristics and the input characteristic tensor to obtain first-stage output characteristics; s30115, carrying out batch normalization processing on the output characteristics of the first stage; S30116, performing depth channel feature interaction on the first stage output features after batch normalization through a gating linear unit realized based on convolution; s30117, regularizing the DropPath by the learnable scaling parameters, and processing the interacted characteristics; And S30118, performing secondary residual connection on the features subjected to secondary regularization and the first-stage output features to obtain final output of the serial double residual block variant.
  6. 6. The method for identifying the state of the bobbin based on deep learning according to claim 4, wherein the multi-path dynamic convolution operation specifically comprises: S30121, carrying out global information coding on the input tensor through self-adaptive average pooling; s30122, generating a weight vector by utilizing convolution of a preset size based on the input tensor after global coding; s30123, carrying out splitting rearrangement operation on the weight vectors to form a plurality of independent groups; S30124, carrying out normalization processing on the independent groups to obtain corresponding weight coefficients; and S30125, carrying out weighted summation on the output of each branch according to the weight coefficient to obtain a final output.
  7. 7. The method for identifying a state of a spool based on deep learning according to claim 3, wherein the step of constructing the two-channel gating network specifically comprises: S3021, shallow feature extraction is carried out on an input image by introducing HGStem modules into the backbone network of the YOLO11 model; S3022, splitting the shallow features into heterogeneous branches parallel to dynamic feature branches and feature reuse branches on a plurality of scales respectively; s3023, in the dynamic feature branch, after downsampling the shallow features through standard convolution, connecting the C3k 2-dynamic serial mixer module to perform dynamic multi-scale feature extraction; S3024, in the feature reuse branch, after downsampling the shallow features through depth separable convolution, connecting the HGStem modules to perform feature reuse processing; S3025, when each scale is processed, sending the dynamic multi-scale features outputted by the dynamic feature branch and the features outputted by the feature reuse branch into a multi-scale self-adaptive space attention gating module for depth fusion; And S3026, integrating the features after the scale fusion to obtain a highly fused feature so as to complete the construction of the dual-channel gating network and replace the original single-path backbone network in the YOLO11 model.
  8. 8. The method for identifying a bobbin state based on deep learning according to claim 1, wherein the step S5 specifically comprises: s501, carrying out batch normalization processing on the output of each channel in the DML-YOLO model, and evaluating the importance of each channel; s502, based on the importance, performing sparse training on the DML-YOLO model through the L1 regularization; S503, removing a channel with a scaling factor lower than the global pruning threshold after sparse training; And S504, performing fine tuning training on the model with the channels removed to obtain the DMLP-YOLO model.
  9. 9. The bobbin state identification system based on deep learning is characterized by comprising a processor and a memory; the memory stores a program or instructions executable on the processor, which when executed by the processor, implement the steps of the deep learning based bobbin state identification method of any one of claims 1 to 8.
  10. 10. A readable storage medium, characterized in that it has stored thereon a program or instructions which, when executed by a processor, implement the steps of the deep learning based bobbin status identification method according to any one of claims 1 to 8.

Description

Bobbin state identification method and system based on deep learning Technical Field The invention relates to the technical field of image processing, in particular to a bobbin state identification method and system based on deep learning. Background Along with the deep advancement of intelligent transformation in the textile industry, efficient and accurate identification of the state of the yarn tube becomes a key link for improving the production automation level and the quality control capability. The automatic detection and identification of the state of the yarn tube are realized by utilizing a machine vision technology, and the method has important significance for reducing the manual dependency, improving the production efficiency and reducing the operation and maintenance cost. Currently, for bobbin state identification based on deep learning, a visual detection model based on deep learning is mainly adopted to perform bobbin state identification, and high detection precision can be achieved under certain conditions by training a large-scale image data set. And part of researches optimize detection performance by introducing data enhancement, model fine adjustment and other modes, so that a relatively mature bobbin state analysis scheme based on a general target detection framework is formed. However, the detection model in the prior art is high in computational complexity and large in parameter quantity, is difficult to realize efficient deployment and real-time reasoning in the edge equipment with limited resources, limits the practical application of the detection model in a production field, and meanwhile, the prior art is focused on precision optimization and lacks systematic design aiming at light weight and reasoning speed, so that the situations of high delay and high energy consumption can occur in a real industrial scene. Disclosure of Invention In view of the shortcomings of the prior art, the embodiment of the invention aims to provide a bobbin state identification method based on deep learning, which can solve the technical problems that a detection model in the prior art is high in computational complexity and large in parameter quantity, is difficult to realize efficient deployment and real-time reasoning in edge equipment with limited resources, and limits the practical application of the detection model in a production field, and meanwhile, the prior art focuses on precision optimization and lacks systematic design aiming at light weight and reasoning speed, so that the situations of high delay and high energy consumption can occur in a real industrial scene. In a first aspect of the embodiment of the present invention, a method for identifying a state of a bobbin based on deep learning is provided, including: s1, acquiring a bobbin status image; s2, constructing an image data set for detecting the state of the bobbin according to the state image of the bobbin; S3, taking a YOLO11 model as a baseline model, improving the baseline model through a lightweight design, and constructing a DML-YOLO model; s4, training a DML-YOLO model based on the image data set; S5, compressing the trained DML-YOLO model by a channel pruning technology based on L1 regularization to obtain a DMLP-YOLO model; S6, performing fine tuning training on the DMLP-YOLO model based on the image data set, and deploying the DMLP-YOLO model after the fine tuning training to edge equipment; s7, acquiring a real-time bobbin image; And S8, inputting the real-time bobbin image into a DMLP-YOLO model which is deployed in edge equipment after fine adjustment training, and outputting a bobbin state identification result. In a second aspect of the embodiment of the invention, a bobbin state identification system based on deep learning is provided, which comprises a processor and a memory; the memory stores a program or instructions executable on the processor which when executed by the processor implement the steps of the deep learning based bobbin status identification method according to the first aspect. In a third aspect of the embodiments of the present invention, a readable storage medium is provided, on which a program or an instruction is stored, which when executed by a processor, implements the steps of the method for identifying a bobbin status based on deep learning according to the first aspect. The technical scheme provided by the embodiment of the invention has the beneficial effects that at least: According to the embodiment of the invention, through the design of the lightweight model and the pruning compression of the channel, the parameter number and the computation complexity of the model are remarkably reduced, the model can be efficiently deployed on edge equipment with limited computation power and memory resources, meanwhile, the training process is optimized by combining an online data enhancement technology, the detection speed is greatly improved on the premise of ensuring the state recognition pre